High School Sociology

Fifteen point nine million people, five weeks

On 8 May 2020 the Bureau of Labor Statistics released the April jobs figures. The unemployment rate had risen 10.3 percentage points in a single month, from 4.4 percent in March to 14.7 percent in April. The number of people counted as unemployed rose by 15.9 million, to 23.1 million. In the monthly records that begin in January 1948, nothing like that had happened before.

Hold that number and ask a blunt question. Did 15.9 million Americans become lazy, unqualified or careless in five weeks? Obviously not. Their skills in April were the skills they had in March. What changed was outside them: state closure orders, restaurants and hotels and cinemas told to shut, a whole set of jobs that stopped existing at once. If you want to explain why one specific person lost one specific job, you might talk about that person. If you want to explain a 10.3 point move in five weeks, you cannot. You have to talk about the arrangement everyone was living inside.

That switch, from the person to the arrangement, is the move this whole course is built on.

What sociology is, and what it hands to other subjects

Sociology is the systematic study of groups, group interactions, societies and social interaction. Two words there are load-bearing. Systematic means the claims have to be checked against evidence that someone else could go and collect again, not against how obvious they feel. Groups means the unit of attention is usually more than one person: a family, a team, a school, a company, a religion, a country.

It helps to see what the neighbouring subjects would do with the same fact. Take a 17-year-old working 20 hours a week at a supermarket checkout.

None of those three is a rival to the others. They answer different questions. Sociology's question is the one about the group and the structure.

Key idea: Sociology does not study people one at a time. It studies the patterns many people make together and the arrangements that make those patterns likely.

Social facts: the things that stay put while the people change

A social fact is a feature of social life that exists outside any one individual and presses on all of them: laws, customs, rituals, morals, fashions, the rules about who speaks first. You did not invent the rule that you face forward in a lift, and you cannot repeal it by deciding not to.

Here is the property that makes social facts worth studying. Imagine every driver in your state is replaced next year by a different driver. The individuals are all new. The number of traffic deaths per 100,000 people will land close to last year's number anyway, because the roads, the speed limits, the enforcement and the vehicles are the same. The rate belongs to the arrangement, not to the personalities.

That is also a warning. A rate describes a population, not a person. Knowing that a group has a high rate of something tells you nothing certain about any individual in it. Treating a group average as a fact about individuals is a mistake with a name, the ecological fallacy, and you will meet it again in Module 2 and in the lesson on Durkheim.

Mills and the switch between trouble and issue

C. Wright Mills published The Sociological Imagination in 1959. His first chapter contains a section called The Personal Troubles of Milieu and the Public Issues of Social Structure, and that title is the whole idea in one line. A trouble is a problem located in one person's life and immediate surroundings. An issue is a problem located in the way society is organised.

Mills used unemployment as his example, and it is still the clearest one. When one man in a city of 100,000 has no work, you can reasonably look at that man: his skills, his health, his choices. When millions in the same country have no work at the same time, looking at each man's character explains nothing, because the character of millions did not change simultaneously. The problem has moved from the man to the structure of the economy. Mills wrote that the sociological imagination lets us "grasp history and biography and the relations between the two within society".

The skill is not to decide that everything is society's fault. The skill is to notice which of the two you are looking at, and to check. Here are three sentences a person might say about their own life, with the number that turns each one into an issue.

Read the middle column again. Each of those numbers is made out of millions of private decisions, and none of them can be explained by any single decision. That is what a public issue looks like on paper.

The point: A trouble and an issue can be the same event. Which one you are studying depends on whether you are looking at one biography or at a rate.

Micro and macro, and why the level matters in an argument

Micro-level analysis looks at small groups and face-to-face interaction: who talks in a classroom, how two friends repair an argument, what happens in the first ten seconds of a job interview. Macro-level analysis looks at large patterns and whole institutions: graduation rates by state, the share of jobs in manufacturing since 1980, whether a country's religious attendance is falling.

Mixing the levels up is the single most common way a sociological argument goes wrong. Watch it happen:

• Claim: Students at this school do not care about class. Evidence offered: three students in one lesson were on their phones. That is micro evidence for a macro claim. It could be true of the school, but three students cannot show it.
• Claim: My cousin was rude to me because Americans are individualistic. Evidence offered: a national cultural pattern, applied to one conversation. That is macro evidence for a micro claim, and it explains nothing about your cousin.

The repair is the same in both directions: say which level your claim is about, then bring evidence from that level. Micro claims need observed interaction. Macro claims need counts across many cases.

What matters here: Name the level before you argue. Evidence from one level rarely settles a claim about the other.

Structure constrains, it does not decide

Social structure is the stable pattern of positions, relationships and institutions that people move inside: schools with grades, workplaces with managers, families with roles, laws with penalties. Structure changes the odds. It does not write the outcome.

Go back to April 2020. Some of those 23.1 million people were back in work within weeks, some were out for a year, some changed occupation entirely. Their individual efforts and contacts mattered to which of those happened. But no amount of individual effort could have kept the national rate at 4.4 percent that April, because the jobs themselves were gone. Both things are true at once, and a sociologist who forgets either one is doing it badly. Structure sets the distribution; people act inside it.

Common misconceptions

• "Sociology is just opinions about society." The claims in this course are attached to counts you can go and check: a federal survey, a census table, a published crime series. When a claim cannot be checked, the course says so.
• "Sociology says people have no choice." It says choices are made under conditions that are not chosen, and that the conditions show up in the rates. Individual action still explains who ends up where.
• "A high rate in a group tells you about a person from that group." It does not. That is the ecological fallacy, and it survives being repeated confidently.
• "Sociology is the same thing as social work." Social work is a profession that helps people; sociology is a research discipline that studies patterns. Some sociologists do work that informs policy, but the subject itself is an investigation, not a service.

Where this leaves us

• The April 2020 unemployment rate rose 10.3 points in one month to 14.7 percent, and 23.1 million people were unemployed: a change no account of individual character can explain.
• Sociology is the systematic study of groups, societies and social interaction, and its unit of attention is the pattern rather than the person.
• Social facts exist outside individuals and press on them; rates built from social facts stay stable while the individuals behind them change.
• Mills separated the personal trouble of one biography from the public issue visible in a rate. Both can describe the same event.
• Micro-level analysis studies interaction; macro-level analysis studies large patterns. Evidence has to match the level of the claim.
• Structure changes the odds a person faces without deciding what that person does.

Sources

1. U.S. Bureau of Labor Statistics. (2020, May 13). Unemployment rate rises to record high 14.7 percent in April 2020. The Economics Daily. bls.gov
2. OpenStax. (2021). What is sociology? In Introduction to Sociology 3e. Rice University. openstax.org
3. Anderson, M., Faverio, M., and Gottfried, J. (2024). Teens, social media and technology 2024. Pew Research Center. pewresearch.org
4. Mills, C. W. (1959). The Sociological Imagination. Oxford University Press. Chapter 1, The Promise.
5. Hamilton, B. E., Martin, J. A., and Osterman, M. J. K. (2024). Births: Provisional data for 2023 (Vital Statistics Rapid Release Report No. 35). National Center for Health Statistics.

A table from 1897

In Le suicide, published in 1897, Emile Durkheim printed a table he numbered XVIII: suicides per million people of each religious confession, collected by several different statisticians across five German-speaking states. Six of its twelve rows look like this.

Read the Protestant column against the Catholic column. In every one of the twelve rows, including the six not shown, the Protestant rate is higher. Durkheim said the gap ranged from about 20 or 30 percent at the low end to 300 percent at the high end. Now read the Jewish column, and notice that the popular summary of this table is wrong: Jews are lowest in Austria and Prussia in the 1850s, but higher than Catholics in Prussia after 1869, in Bavaria and in Wurttemberg. A pattern can be real in one comparison and absent in another, and the only way to know is to look at the column rather than the slogan.

Durkheim then pushed the comparison inside one country. Sorting the fourteen provinces of Prussia between 1883 and 1890 by how Protestant they were gives this:

The rate falls as the Protestant share falls, step by step. In Switzerland the same ordering appeared across cantons: 86.7 per million in Catholic cantons, 212.0 in mixed cantons, 326.3 in Protestant ones, and it held whether the canton was French-speaking or German-speaking.

What Durkheim did with the pattern

Durkheim was not interested in religion as belief. He was interested in what a church did to the group around it. His explanation was social integration: Catholic parishes of that period bound members into a dense web of shared practice and shared authority, while Protestantism placed the individual alone before conscience and scripture. More integration, fewer suicides. The act looks like the most private thing a person can do, and yet its rate is set by the structure of the group.

He sorted suicide into four types by two dials, how integrated a person is and how regulated their desires are.

Anomie, that third condition, is the one you will meet again in the lessons on deviance and on social change. It means normlessness: the rules that told people what to want have loosened, and nothing has replaced them.

Worth holding on to: Durkheim's argument runs from a rate to a structure. He never claimed to know why any one person died.

Where the table is weak

Two criticisms matter, and a good high school answer names both. First, every figure is an aggregate: a rate for a province, not a record of individuals. Concluding that Protestant individuals were more likely to kill themselves is the ecological fallacy, because the province rate could in principle be driven by its Catholic minority. Later researchers with individual-level records have found the Protestant-Catholic difference smaller than Durkheim's provinces suggest, and largely confined to German-speaking Europe. Second, the underlying counts came from officials who recorded cause of death, and coroners in Catholic districts, where suicide was a grave sin, had reason to record an ambiguous death as an accident. If registration itself varies by religion, part of the gap is a measurement artefact.

Neither criticism makes the work worthless. It survives as the first serious demonstration that a deeply personal act has a social rate, and as the reason sociologists still start with published statistics rather than intuition.

Marx: the question is who owns the machinery

Karl Marx wrote before Durkheim and asked a different question: not what holds a society together, but what tears it apart. His unit was class, and he defined class by a relationship to the means of production, the factories, land, tools and capital used to make things. The bourgeoisie owns them. The proletariat owns nothing to sell but labour time. The relationship is not a difference in income; it is a difference in position, and Marx thought the interests of the two positions could not be reconciled.

The Manifesto of the Communist Party, in Samuel Moore's 1888 English translation, opens its first chapter with the line that carries the whole theory: "The history of all hitherto existing society is the history of class struggles." Engels added a footnote conceding that this meant written history, since he and Marx had come to think early societies were communal.

Marx also gave sociology the idea of alienation: work that belongs to someone else turns the worker into a stranger to what she makes, to the making of it, to other workers and to herself. You can test that idea on a summer job. If the task is fixed by a screen, the pace by a timer, and the product sold before you see it, that is what Marx meant, and you do not have to accept his politics to notice the experience.

The upshot: For Marx, the engine of history is conflict over who owns productive property, and every institution, including law, religion and schooling, has to be examined for whose interests it serves.

Weber: rates are not enough, meaning matters

Max Weber agreed that class matters, then insisted that status and prestige work separately from wealth, and that no explanation is finished until it captures what the action meant to the person doing it. His term for that interpretive step is verstehen, understanding from the inside. A sociologist studying a fast at Ramadan can count how many people fast; Weber says the count is not an explanation until you can state the meaning of the fast for the faster.

Two more Weberian tools recur in this course. Rationalisation is the long historical drift from custom and faith toward calculation, efficiency and written rules, whose purest institutional form is bureaucracy. An ideal type is a deliberately simplified model of an institution, built so that real cases can be measured against it, not a claim that any real case is perfect. In The Protestant Ethic and the Spirit of Capitalism he argued that a religious ethic of disciplined work and reinvestment helped make industrial capitalism thinkable, and then warned that the resulting order had hardened into a cage of rules that no one intended and no one can easily leave. Lesson 19 tests that thesis against modern data.

Three questions, one school corridor

Sociology did not arrive fully formed from these three. Auguste Comte named it in the 1830s. Harriet Martineau translated Comte into English and wrote her own comparative study of American institutions in the 1830s, including a chapter on the position of women. W. E. B. Du Bois published The Philadelphia Negro in 1899, based on house-to-house interviews across a whole city ward: American empirical sociology arguably starts there, and it was left out of the textbooks for decades.

So what?: The three founders are not three opinions about society. They are three questions, and each one tells you what to go and measure.

Common misconceptions

• "Durkheim proved that being Protestant makes a person more likely to take their own life." He showed a difference between group rates. Individual-level conclusions do not follow, and later individual data narrowed the gap.
• "Marx was describing the Soviet Union." Marx died in 1883, 34 years before the Russian revolution. His analysis was of industrial capitalism in Britain and Germany, and judging it by later states is a category error in both directions.
• "Weber admired bureaucracy because he called it efficient." He called it technically superior to older forms and said it produced a cage of rules that outlives anyone's purposes. Description is not endorsement.
• "Conflict theory means Marx, and functionalism means Durkheim, so they contradict each other flatly." They ask different questions. Durkheim's integration and Marx's class conflict can both be operating in the same school, and often are.

Putting it together

• Durkheim's Table XVIII shows Protestant suicide rates above Catholic rates in all twelve of its comparisons, with Jewish rates sometimes lowest and sometimes above Catholic rates.
• Sorting Prussian provinces by Protestant share gives means of 264.6, 220.0, 163.6 and 95.6 per million: the rate tracks the composition of the province.
• Durkheim explained the pattern by social integration, and classified suicide as egoistic, altruistic, anomic or fatalistic.
• The table's weaknesses are the ecological fallacy and possible differences in how deaths were registered by religion.
• Marx defined class by the relation to the means of production and read institutions for whose interests they serve.
• Weber added verstehen, ideal types, rationalisation and bureaucracy, and insisted that meaning is part of the explanation.
• Comte named the field, Martineau brought it into English with her own comparative study, and Du Bois built American empirical sociology from door-to-door data in 1899.

Sources

1. Durkheim, E. (1897). Le suicide: Etude de sociologie, Livre II, chapitre 2, Tableau XVIII. Felix Alcan. classiques.uqam.ca
2. OpenStax. (2021). The history of sociology. In Introduction to Sociology 3e. Rice University. openstax.org
3. Marx, K., and Engels, F. (1888). Manifesto of the Communist Party (S. Moore, Trans.; original work published 1848). wikisource.org
4. Morselli, H. (1881). Catholicism, Protestantism, and suicide. Popular Science Monthly, 20. wikisource.org
5. Wikipedia. (2025). Suicide (Durkheim book). wikipedia.org

29,873 dollars and 9,552 dollars

In fiscal year 2022 the state of New York spent 29,873 dollars per public school student. Utah spent 9,552. The national figure was 15,633, up 8.9 percent in a year, the biggest one-year jump in more than two decades. Those are Census Bureau numbers from the annual survey of school system finances, and every sociologist in this lesson accepts them. What they disagree about is what the numbers mean.

That disagreement is the point of this lesson. Sociology has three classic perspectives, and a perspective is not a belief about whether schools are good. It is a decision about where to look. Give the same high school to three sociologists and you will get three different research projects, three different pieces of evidence, and, on some questions, three different answers.

The functionalist look: what jobs does this building do?

Functionalism treats society as a set of interrelated parts, each doing work that keeps the whole running, and asks of any institution: what is it for? Robert Merton sharpened the question by splitting the answers into three kinds.

• Manifest functions are the intended, stated purposes. A high school teaches literacy and mathematics, certifies completion, and prepares students for work or college.
• Latent functions are real consequences nobody planned. School supervises teenagers while adults work. It is where most people meet their friends and many meet future partners. It feeds children: federal school meal programmes reach tens of millions of students, which is a function of schooling that appears in no curriculum document.
• Dysfunctions are consequences that damage the system. A tracking system that parks a capable student in a low set at 14 wastes talent the economy wanted.

A functionalist reading of the spending gap would say a school system must reproduce the skills a society needs, and would ask whether 9,552 dollars per student is enough to do that job. Notice the shape of the claim: it is about the requirements of the whole, and its evidence is outcomes, graduation, literacy, employment.

Why this matters: Latent functions are where functionalism earns its keep. If you only list what an institution says it does, you will misunderstand what happens when it closes.

The conflict look: who benefits from this arrangement?

Conflict theory starts from competition over scarce resources and asks who gets what, and why the arrangement persists. Put the same spending numbers in front of a conflict theorist and the first question is structural: where does school money come from? In most states a large share comes from local property taxes, so districts with expensive houses can raise more per student than districts without, and the gap is not an accident but a predictable output of the funding rule.

The second question is about content. Samuel Bowles and Herbert Gintis argued in 1976 that schooling teaches punctuality, obedience to rules and acceptance of hierarchy, and that these lessons fit students to the workplaces waiting for them: the children of managers get schools that reward initiative, and the children of workers get schools that reward compliance. Sociologists call the unwritten part of this the hidden curriculum: the lessons about authority, time and self-presentation that no one writes into a syllabus.

A conflict analysis is testable, and this matters, because a perspective that cannot be wrong is not doing science. It predicts that measurable advantages track family resources rather than measured ability alone, that changes to funding formulas change outcomes, and that the rules governing who takes the advanced courses will favour families who can buy preparation.

The interactionist look: what happens in the first week?

Symbolic interactionism works at the micro level, on meanings built up in face-to-face contact. It descends from George Herbert Mead and was named by Herbert Blumer, whose rule is simple: people act toward things on the basis of what those things mean to them, and meanings come out of interaction.

Here is the documented case. In 1970 Ray Rist published a study of one kindergarten classroom in the Harvard Educational Review. By the eighth day of school, before any test had been given, the teacher had assigned the children to three tables. Rist found the assignments tracked family background and appearance, and that the children at the first table then received more instruction and more attention, while the third table was addressed more in terms of control. He followed the class and found the groupings persisted into later grades. The teacher did not announce a theory of ability; she formed expectations in the first week, acted on them, and the children grew into them. That is a self-fulfilling prophecy, and it is invisible to any study that only counts budgets.

Erving Goffman added the stage metaphor, dramaturgy: the corridor is a front stage where a student performs a self for an audience, and the bathroom or the group chat is the back stage where the performance is dropped and managed. Anyone who has behaved differently in front of a teacher than in front of friends has done fieldwork on this already.

In short: Interactionism explains how a category becomes a fate: a label is applied, behaviour follows the label, and the label looks confirmed.

The three, side by side

Putting them in real conflict: who ends up in the advanced class?

Take one fact you can check in your own school: students from higher-income families are over-represented in the most advanced courses. All three perspectives accept the fact. Their explanations differ, and so does the evidence each would bring.

• The functionalist explanation. Advanced courses sort students by demonstrated readiness, which is what a system needs if it is going to fill demanding roles. Income correlates with readiness because of everything that happened before age 14. Evidence for it: placement rules based on grades and tests do predict who passes the advanced exam.
• The conflict explanation. Readiness is partly purchased: tutoring, quieter homes, summer programmes, parents who know how to appeal a placement. Evidence for it: when districts change placement rules, the composition of advanced classes changes without any change in underlying ability. Studies of automatic-enrolment policies, which place every student above a test threshold into the advanced course unless they opt out, find more students from low-income families taking and passing those courses.
• The interactionist explanation. Placement follows teacher expectations formed early, in the way Rist observed, and students internalise the category they were put in. Evidence for it: the timing and basis of the original grouping, observable only by watching classrooms and reading the paperwork.

What would settle it? Not argument. A design. Take students whose test scores sit within a point or two of a placement cut-off, which makes them comparable, and compare what happens to those just above and just below. If the two groups end up in very different places, the placement rule itself is doing work rather than merely reading off ability. Add classroom observation to see whether teachers treat the two groups differently. Add family income to see whether appeals and transfers cluster in richer families. Each perspective has told you what to measure, which is what a perspective is for.

The core of it: Perspectives are not opinions to pick between. They are instructions about where to point the instrument, and they can be checked against each other on the same case.

Common misconceptions

• "Functionalism means everything in society is good." Merton built dysfunction into the framework precisely because parts of a system can damage it. A functionalist can condemn tracking on functionalist grounds.
• "Conflict theory is just Marxism with a new name." Marx is one source. Weber's work on status and authority, feminist analyses of gender and racial-conflict analyses all sit in the same family, and several reject Marx's economics.
• "Interactionism is too small to matter." The eight-day kindergarten sorting Rist documented shaped years of schooling. Micro processes accumulate into macro outcomes.
• "You have to pick one perspective and stick to it." Most working sociologists use whichever fits the question, and many studies combine a macro pattern with micro observation of how it is produced.

The short version

• Per-pupil spending in fiscal 2022 ran from 29,873 dollars in New York to 9,552 in Utah, against a national figure of 15,633.
• Functionalism asks what work an institution does, and distinguishes manifest functions, latent functions and dysfunctions.
• Conflict theory asks who benefits, and traces school inequality to funding rules and to the hidden curriculum.
• Symbolic interactionism asks what things mean to participants; Rist's 1970 study shows expectations formed in eight days shaping years.
• The three perspectives generate different predictions about the same fact, and a research design can test between them.
• Blind spots are real: necessity assumed, consent ignored, structure forgotten. Naming them is part of using the perspective well.

Sources

1. U.S. Census Bureau. (2024, May 9). Public school spending per pupil increased by 8.9 percent in fiscal year 2022. census.gov
2. OpenStax. (2021). Theoretical perspectives in sociology. In Introduction to Sociology 3e. Rice University. openstax.org
3. Rist, R. C. (1970). Student social class and teacher expectations: The self-fulfilling prophecy in ghetto education. Harvard Educational Review, 40(3), 411-451.
4. Bowles, S., and Gintis, H. (1976). Schooling in Capitalist America: Educational Reform and the Contradictions of Economic Life. Basic Books.
5. Wikipedia. (2025). Manifest and latent functions and dysfunctions. wikipedia.org

Ten million ballots, and the wrong answer

In 1936 The Literary Digest ran the largest opinion poll anyone had attempted. It mailed ten million straw ballots and got 2.38 million back, and on that mountain of paper it forecast that Alf Landon would take 57.08 percent of the popular vote. Franklin Roosevelt won every state except Maine and Vermont with 60.8 percent. The poll was wrong by about 39 points.

That same year George Gallup used roughly 50,000 respondents, a fiftieth as many, and came within 1.4 points of the result. Hold those two facts next to each other, because together they overturn the most natural assumption anyone has about surveys: that size is what makes a poll good.

Why the small poll won

The Digest drew its names from lists of magazine subscribers, telephone subscribers and registered car owners. In the middle of the Great Depression those lists selected people with money. That is a broken sampling frame: the list you actually draw from does not match the population you want to describe. Worse, only 24 percent of the ballots came back, and people who disliked Roosevelt were more motivated to send theirs in. That is nonresponse bias, and researchers now consider it the larger of the two errors.

No sample size fixes either problem. Ten million responses from car owners still describe car owners. This is the single most important idea in survey research, so it is worth stating flatly: how people get into the sample matters more than how many of them there are.

Key idea: A large biased sample is not better than a small fair one. It is more confidently wrong.

What a probability sample is

A probability sample is one in which every member of the target population has a known, non-zero chance of being selected, and chance rather than convenience decides who is in. That property is what licenses the arithmetic that follows. Four common designs:

The last one is how the General Social Survey works. Run by NORC at the University of Chicago since 1972, the GSS samples addresses across the country in stages, targets adults aged 18 and over living in households who can answer in English or Spanish, and has usually ended up with between about 2,765 and 2,992 completed interviews, each taking 70 to 100 minutes. The 2024 round collected data from April to December 2024 using more than one mode, with some respondents online and some interviewed in person.

Contrast that with a convenience sample: the students who happen to be in your class, the people who click a link in a video description, the customers who choose to fill in a receipt survey. None of them has a known chance of selection, so there is no way to say what population the result describes. You can still learn things from such data, but you cannot generalise from it, and the honest write-up says so.

Margin of error: what the plus or minus actually covers

Because selection was random, statisticians can calculate how much a sample estimate would bounce around if you drew the sample again and again. A rough rule for a percentage near 50 in a simple random sample is that the 95 percent margin of error is close to one divided by the square root of the sample size.

Two consequences that students regularly miss. First, the gain from extra interviews shrinks fast: going from 400 to 1,000 buys you two points, and going from 1,000 to 2,800 buys you one. That is why national surveys settle around a few thousand rather than a few million. Second, and more important, the margin of error only describes error from random selection. It says nothing about a broken frame, nonresponse, a leading question or a respondent who lies. Those errors can be far larger, and no formula reports them. The Literary Digest could have printed a margin of error of a tenth of a point and still been wrong by 39.

When response rates are low, researchers weight the data: if the sample has too few young men compared with census figures, each young man's answers count for more. Weighting repairs known imbalances; it cannot repair an unknown one.

What matters here: The margin of error is the smallest of a survey's errors, not the sum of them.

The question is part of the instrument

Two surveys can sample perfectly and still disagree, because the words are not neutral. Pew Research Center has run experiments where half the respondents get one wording and half get another. The halves are equivalent by random assignment, so any difference is caused by the words.

Look at the third row. The same respondents flipped by nearly 20 points because one abstract phrase was replaced by a concrete one. Neither answer is a lie. The question simply was not the question anyone thought it was.

Open and closed questions also part company. Asked in 2008 what mattered most in their vote with the options listed, 58 percent of respondents picked the economy; asked the same thing with no list, 35 percent said the economy, and 43 percent named something not on the closed list at all. A closed question tells you how people distribute themselves among your categories. An open question tells you what categories they use.

Then there is social desirability bias: people understate drinking, drug use and racial hostility and overstate voting, exercising, donating and attending religious services. This is why surveys of embarrassing behaviour are read with caution, and why a self-report of church attendance is not the same measurement as a count of people in pews.

Four broken questions and their repairs

The repaired versions share a habit worth copying: they ask about counted behaviour in a fixed time window rather than about a self-image.

Common misconceptions

• "A poll of two thousand people cannot possibly represent a country." It can, if selection is random, because the error depends on sample size and randomness rather than on the fraction of the population sampled. A properly drawn 2,800 describes 260 million adults better than a self-selected two million.
• "Margin of error tells you how wrong the poll might be." It tells you only how much random selection alone would move the number. Frame, nonresponse and wording errors sit outside it.
• "Online polls with huge numbers of votes are more reliable." Self-selected respondents have no known probability of selection, which is the Literary Digest error with a faster delivery mechanism.
• "If two surveys disagree, one of them must be dishonest." They may have asked different questions. Check the wording before the motives.

What to carry forward

• The Literary Digest's 2.38 million responses were wrong by 39 points because of a broken sampling frame and nonresponse; Gallup's 50,000 came within 1.4 points.
• A probability sample gives every member of the population a known chance of selection, which is what makes generalisation legitimate.
• The GSS uses multistage area probability sampling of US households, has run since 1972, and typically completes under 3,000 interviews of 70 to 100 minutes.
• Margin of error shrinks with the square root of sample size and covers random sampling error only.
• Question wording changes answers by measurable amounts: 68 to 43 percent on Iraq, 52 to 33 percent on domestic priorities.
• Closed questions constrain answers to your categories; open questions reveal the respondent's own.
• Ask about counted behaviour in a defined window, one idea per item, with no answer implied.

Sources

1. NORC at the University of Chicago. (2025). General Social Survey: Frequently asked questions. gss.norc.org
2. Pew Research Center. (n.d.). Writing survey questions. Methods 101. pewresearch.org
3. Wikipedia. (2025). The Literary Digest. wikipedia.org
4. OpenStax. (2021). Research methods. In Introduction to Sociology 3e. Rice University. openstax.org

39 percent, down from 48

Between 5 May and 20 December 2022, interviewers for the General Social Survey put a long-running question to 3,544 American adults. The survey reads out a list of institutions and asks, for each one, whether you have a great deal of confidence in it, only some confidence, or hardly any confidence at all. For the scientific community, 39 percent chose a great deal. In 2018 and again in 2021 the figure was 48 percent. The margin of error on the 2022 sample is about 3 percentage points.

That paragraph contains everything you need to read a survey result honestly: the exact words of the question, who was asked, when, how many, and how much slack the number carries. Strip any of those out and the number turns into a slogan. This lesson works through two real surveys line by line.

Survey one: the GSS confidence battery

The question wording matters more than usual here, because there are three options and most people pick the middle one. A fall in the share saying a great deal does not mean people moved to hardly any; they may have moved to only some. Keep that in mind while reading the table.

Three things to notice before drawing any conclusion. First, the starting levels are not the same: the press was already at 13 percent in 2018, so its fall of 6 points removed nearly half of what was left, while science fell 9 points from a much higher base. Percentage points and percentage change are different quantities, and headlines routinely confuse them.

Second, apply the error. Each estimate carries about 3 points of sampling error, and the difference between two estimates carries more than either one alone. A 9 point fall is comfortably outside that range. A 5 point fall, as with medicine, is close enough to the boundary that the careful sentence is that confidence in medicine fell modestly, not that it collapsed.

Third, the survey covers adults aged 18 and over living in US households who can be interviewed in English or Spanish. It does not cover people in prisons, in nursing homes or on military bases. That is the denominator, and it is not the same as all Americans.

The point: A survey number is a claim about a specific population, asked in specific words, in a specific season, with a known amount of slack. Report all five or you have not reported the finding.

The cross-tabulation that changes the story

A cross-tabulation splits one answer by another variable. Split the science item by party identification and the national average stops being the story.

Read the rows, not the last line. Among Democrats the figure did not move at all. Among Republicans it fell 21 points. The 9 point national fall is almost entirely the second group moving, and a 10 point party gap became a 31 point gap. Anyone who reported only the national average would describe a general loss of faith in science that, on these numbers, did not happen generally.

This is the everyday use of a cross-tab, and the reason sociologists distrust a single headline percentage. It is also a warning in the other direction: a cross-tab with small subgroups has large errors inside each cell, so a 4 point difference between two small groups may be nothing at all.

Survey two: Pew on trust in government

The Pew Research Center maintains a trend running back to 1958 on a different question: do you trust the government in Washington to do what is right just about always, most of the time, only some of the time, or never? In a September 2025 reading, 17 percent chose one of the first two answers, 2 percent saying just about always and 15 percent most of the time. In October 1964 the figure was 77 percent.

The party split is instructive in a way the science item was not. In September 2025, 26 percent of Republicans and 9 percent of Democrats expressed trust. Across the full trend, the party holding the presidency reports more trust, and the lines cross over after each change of administration. So the partisan gap in this series is largely about who is in office, while the partisan gap on confidence in science in 2022 was not.

Why this matters: Two good surveys can give different pictures of public opinion because they asked different questions of different samples in different ways. Comparing them requires matching all three.

Five checks before you quote a percentage

1. Exact wording. Confidence in medicine is not confidence in doctors, and approval of a policy is not approval of a politician.
2. Denominator. All adults, registered voters, parents of school-age children and teenagers are four different populations. A share of a subgroup can move while the share of everyone does not.
3. Date and field period. The 2022 GSS was collected across eight months; an event in July falls inside it, which blurs any before-and-after story.
4. Error. Compare the size of the change with the margin of error, and remember that comparing two estimates compounds the error.
5. Trend integrity. Only compare identically worded items. If the mode changed from in-person to online, part of any movement may be the mode rather than the public.

Common misconceptions

• "Trust in institutions has collapsed across the board." Confidence in the press was already at 13 percent in 2018, and the sharpest fall in the science item came from one political group rather than from everyone.
• "A three point move in a poll is news." With a margin of error of about 3 points on each estimate, a three point move is inside the noise.
• "If a survey's margin of error is 3 points, the result cannot be wrong by more than 3 points." Only random sampling error is inside that figure. Wording, nonresponse and mode sit outside it.
• "A fall in the share saying a great deal means people now say hardly any." With three options, people can move to the middle. Check the other categories before describing a collapse.

What you now know

• The 2022 GSS interviewed 3,544 adults from May to December 2022, with a margin of error of about 3 points.
• Great-deal confidence fell between 2018 and 2022 for science (48 to 39), medicine (39 to 34), organized religion (22 to 15), major companies (21 to 15) and the press (13 to 7).
• Cross-tabulating the science item by party shows Democrats unchanged at 55 percent and Republicans down 21 points to 24 percent.
• Pew's trust-in-government series stood at 17 percent in September 2025 against 77 percent in October 1964, and its party gap tracks who holds the presidency.
• Percentage points and percentage change differ, and both depend on the starting level.
• Wording, denominator, date, error and trend integrity are the five checks to run before quoting any percentage.

Sources

1. AP-NORC Center for Public Affairs Research. (2023). Major declines in the public's confidence in science in the wake of the pandemic. Analysis of the 2022 General Social Survey. apnorc.org
2. Pew Research Center. (2025, December 4). Public trust in government: 1958-2025. pewresearch.org
3. NORC at the University of Chicago. (2025). General Social Survey: Frequently asked questions. gss.norc.org
4. OpenStax. (2021). Research methods. In Introduction to Sociology 3e. Rice University. openstax.org

A conclusion that looks airtight

A student surveys 200 juniors and finds this: students in three or more school activities have a mean grade average of 3.4, and students in none have 2.8. She writes her conclusion. Joining activities raises grades by about 0.6 of a grade point, so the school should require every student to join a club.

The number is real. The recommendation does not follow, and this lesson is about the exact place where the argument breaks. There are three breaks, and they are the three you will make yourself if nobody names them.

Break one: the arrow might point the other way

Call joining activities the independent variable, the thing you think does the causing, and grade average the dependent variable, the thing you think is caused. The data show the two vary together, which is a correlation. Correlation is symmetric: it does not know which of the two came first. Students with strong grades have free evenings, parental permission and eligibility rules on their side, so good grades can produce club membership just as easily as the reverse. That is reverse causation, and the survey cannot tell the two stories apart, because both predict exactly the pattern she found.

Break two: the students selected themselves

Nobody assigned these students to clubs. They chose, which means the two groups differ in every way that goes into choosing: how organised they already were, whether they have a job after school, whether someone drives them home at six. Any of those differences could produce the grade gap on its own. When the groups being compared were formed by their own choices, the comparison carries all their prior differences with it.

A difference that could produce the pattern without the supposed cause is a confounding variable. Here is one that would do it alone: students with a paid evening job cannot join clubs and have less study time, so employment lowers club membership and lowers grades, manufacturing a correlation between the two with no causal link between them at all. A relationship produced entirely by a third factor is called spurious.

Break three: the recommendation assumes the cause transfers

Even if clubs did raise grades for students who chose them, it does not follow that they would raise grades for students compelled into them. A treatment's effect on volunteers is not its effect on conscripts. This is the most common failure in policy arguments built on survey data, and it survives even when the causal claim is correct.

The point: Three questions repair almost every bad causal claim. Could the arrow run backwards? Could a third factor produce both? Would the effect hold for people who did not choose it?

What random assignment buys you

An experiment fixes break one and break two at a stroke, because the researcher rather than the participant decides who gets the treatment. Random assignment means using chance to sort participants into a treatment group and a control group, so that on average the two groups are alike in every respect, including respects nobody thought to measure. Any later difference between them has only one place to have come from.

Two real examples, both famous, both in this course for a reason.

A field experiment on hiring. Marianne Bertrand and Sendhil Mullainathan answered help-wanted advertisements in Boston and Chicago with fictitious resumes. The resumes were randomly assigned names that sounded either White or African American, holding the qualifications on the page constant. Applicants with White-sounding names received 50 percent more callbacks. Because the names were randomly attached to the same resumes, no difference in skill, no difference in effort and no difference in neighbourhood can explain the gap. This design, sending matched applications and varying one attribute, is called an audit study.

A social experiment on neighbourhoods. In the 1990s the federal housing department ran Moving to Opportunity in Baltimore, Boston, Chicago, Los Angeles and New York. About 4,600 low-income families volunteered and were randomly assigned to three groups: a voucher usable only in a low-poverty area plus counselling, an unrestricted voucher, or no voucher. Early results found better neighbourhoods and no gain in test scores, which was widely read as showing that neighbourhoods do not matter much. Years later, Raj Chetty, Nathaniel Hendren and Lawrence Katz linked the families to tax records and found that children who moved before age 13 had adult incomes nearly a third higher than the control group, while children who moved as teenagers did no better and in some respects worse. The experiment had been measuring the wrong outcome at the wrong time.

Worth holding on to: Random assignment is the only method that licenses a plain causal claim, and even then the claim is about the people studied, the treatment delivered and the outcome measured.

Why sociologists cannot always experiment

You cannot randomly assign a person's race, gender, religion, parents' income or country of birth, and many of the most important sociological questions are about exactly those. You also cannot ethically assign people to harm. So the field uses two other families of method.

Ethnography means studying a group by spending extended time in its world, usually through participant observation: you are present, you take part, and you write field notes. Matthew Desmond moved into a trailer park and then a rooming house in Milwaukee in 2008 and 2009 and followed tenants through eviction court for Evicted. No survey would have produced what he got: the sequence of events by which a rent shortfall becomes a court date becomes a lost job. Annette Lareau sat in homes, classrooms and doctors' offices with a small number of Black and White families and came away with two distinct styles of raising children, which she named concerted cultivation and the accomplishment of natural growth. Ethnography cannot tell you how common anything is. It tells you what the mechanism looks like from inside, which is what you need before you can even write a good survey question.

Existing data means analysing records someone else already collected: census tables, crime reports, school district finances, court filings, anonymised tax records. It is cheap, it can cover the whole population instead of a sample, and it cannot ask a question the record does not contain. A relative of it is content analysis, the systematic coding of texts or images: count how many of 200 children's books published in 1960 and in 2020 show a father cooking, and you have measured something real about culture without interviewing anyone.

The half-measures in between

When you cannot randomise, you can sometimes find a natural experiment: something outside the researcher's control that assigned people to conditions in a way close to random. A lottery for school places, a policy that begins on one side of a state line, a cut-off date that puts children born in August in a different year group from children born in September. These are the workhorses of modern sociology, and their weakness is always the same question: was the assignment really as good as random?

One more trap belongs here. The Hawthorne effect is the change in behaviour that comes from being studied. It is named after factory experiments at the Hawthorne Works in the 1920s and 1930s, whose original data have since been reanalysed and found much weaker than the legend, so treat the phrase as a real hazard with an unreliable origin story. It matters for your own capstone: the students who know you are watching them are not behaving quite as they would otherwise.

Common misconceptions

• "A strong correlation with a large sample proves causation." Sample size shrinks random error. It does nothing about direction or confounding, and a big sample makes a spurious relationship look more certain rather than less.
• "Controlling for other variables statistically is as good as an experiment." It handles the confounders you measured. Randomisation handles the ones you never thought of, which is why it remains the standard.
• "Qualitative work is just anecdotes." Systematic fieldwork with recorded observation over months answers questions about process that no questionnaire reaches. Its claim is mechanism, not frequency.
• "Correlation means nothing." It is usually the first evidence there is, and it is enough to justify the harder study. The error is stopping there and writing a recommendation.

Looking back

• Correlation is symmetric and cannot establish direction; reverse causation fits the same data.
• Self-selection carries every prior difference between groups into the comparison, and a confounder such as paid employment can manufacture a correlation with no causal link.
• An effect on volunteers is not an effect on people compelled to take part.
• Random assignment makes groups alike in measured and unmeasured ways, which is what licenses a causal claim.
• Randomised names on identical resumes produced 50 percent more callbacks for White-sounding names; Moving to Opportunity found large income gains for children who moved before age 13.
• Ethnography establishes mechanism, existing data establishes scale, surveys establish distribution, experiments establish cause.
• Natural experiments approximate randomisation, and being observed changes behaviour.

Sources

1. Bertrand, M., and Mullainathan, S. (2004). Are Emily and Greg more employable than Lakisha and Jamal? A field experiment on labor market discrimination. American Economic Review, 94(4), 991-1013. nber.org
2. Wikipedia. (2025). Moving to Opportunity. wikipedia.org
3. OpenStax. (2021). Research methods. In Introduction to Sociology 3e. Rice University. openstax.org
4. Desmond, M. (2016). Evicted: Poverty and Profit in the American City. Crown.
5. Lareau, A. (2011). Unequal Childhoods: Class, Race, and Family Life (2nd ed.). University of California Press.

Macon County, Alabama, 1932

Doctors from the United States Public Health Service arrived in Macon County, Alabama, in the autumn of 1932 and offered something scarce: free medical examinations, a hot meal on examination day, free treatment for minor complaints and burial insurance worth fifty dollars. Six hundred Black men signed up, most of them sharecroppers, many of whom had never seen a doctor. Three hundred and ninety-nine of them had latent syphilis. Two hundred and one did not, and they served as the comparison group.

The men were told they were being treated for bad blood. They were not told the name of the disease, they were not told that the point of the exercise was to watch untreated syphilis run its course, and the painful spinal taps performed on them were described as free treatment. The study was presented to them as care. It was observation.

Forty years

The Tuskegee Study of Untreated Syphilis in the Negro Male was planned to last six months. It ran from 1932 to 1972. By 1947 penicillin was the standard treatment for syphilis and was being distributed through public clinics across the country, and the men in the study were still not given it; researchers went as far as asking local draft boards to keep participants off the list of men requiring treatment before military service.

Nothing about this was secret from the profession. Findings were published in medical journals for decades under the study's own name. What ended it was one employee. Peter Buxtun, a venereal disease investigator hired by the Public Health Service in 1965, argued inside the agency for years, got nowhere, and took the documents to the press. The Associated Press story ran on 25 July 1972. The study was shut down on 16 November 1972, and a federal advisory panel concluded it had been ethically unjustified from the start.

The accounting of harm, as recorded in the later litigation and reviews: 28 men had died of syphilis, about 100 more of related complications, 40 wives had been infected, and 19 children had been born with congenital syphilis. A 1974 out-of-court settlement paid roughly 10 million dollars to survivors and families. President Clinton apologised on behalf of the government on 16 May 1997, to five surviving participants who came to the White House.

Why this matters: Almost every rule you will follow in your own study was written by people looking at this case. The rules are not bureaucratic decoration. They are the residue of somebody being harmed.

What the exposure produced

Congress passed the National Research Act in 1974, which created a national commission and required institutions receiving federal research money to run review boards. The commission's report, published in 1979 and named after the conference centre where the drafting began, is the Belmont Report. It is eight pages long and it holds three principles that you can apply to any study, including yours.

The regulations that put those principles into force are known as the Common Rule, published in 1991 as Part 46 of Title 45 of the Code of Federal Regulations and substantially revised with effect from 2018. Part 46 defines who counts as a human subject, what has to be told to them, which low-risk categories are exempt from full review, and, in its Subpart D, the extra protections owed to children.

The sociologist's version of the same problem

Medicine is not the only field with a case it teaches as a warning. Laud Humphreys wanted to know who was having anonymous sex with other men in public restrooms in the 1960s, at a time when doing so was a criminal offence in most states. He took the role of watchqueen, the lookout who warns of approaching police, which let him observe without being questioned. Then he wrote down the licence plate numbers of the cars outside, used motor vehicle records to get names and addresses, and a year later appeared at those homes as an interviewer for what he described as a health survey, without saying what he actually knew.

The findings, published as Tearoom Trade in 1970, cut against the stereotype of the time: by his classification 38 percent of the men he traced were neither homosexual nor bisexual in their own lives, and 24 percent were married men whose encounters were hidden. That result mattered. It also could have destroyed the men it described, because a single leaked notebook meant arrest, dismissal and public exposure for people who had agreed to none of it. The dispute helped split his department, and the American Sociological Association's code of ethics now sets out the standards, informed consent, confidentiality, protection from harm, and disclosure of funding, that Humphreys did not meet.

Notice what the case does not settle. Some sociologists argue that covert observation is the only way to study people who would never agree to be studied, including police misconduct and racist organisations, and that ruling it out protects the powerful. Others answer that a finding obtained by deceiving people cannot be owed to them and that the profession's licence to work at all depends on subjects trusting it. Both positions are still argued in print. What is settled is the arithmetic of risk: Humphreys held data that could have imprisoned his subjects.

What a review board actually does

An institutional review board, or IRB, is a committee at a university, hospital or school district that reads a research plan before data collection begins. It is not a panel of the researcher's friends: federal rules require at least five members, at least one whose training is not scientific, and at least one with no other connection to the institution. It asks a short list of questions.

• Are risks minimised by the design itself, not just promised away?
• Is the remaining risk reasonable in relation to what will be learned?
• Is selection of subjects equitable?
• Will consent be sought, from whom, recorded how, and in language the person can read?
• Are the data kept in a way that protects privacy, and for how long?

Boards can approve, require changes, or refuse. Most classroom research falls into a lighter category: Part 46 exempts some research in established educational settings and some anonymous survey work from full review, which is why your teacher may only need a short approval rather than a full application. Exempt does not mean unsupervised. Someone other than you still decides that it is exempt.

Consent when the subject is sixteen

A 16-year-old cannot give legally effective informed consent in most research settings. The structure has two parts, and mixing them up is the commonest error in a school project. Permission comes from a parent or guardian. Assent comes from the student, and it is a real decision: a student whose parent said yes can still decline, and that refusal ends the matter for that student.

Three practical consequences for the capstone in Module 8. First, your consent script has to be written, short and honest about the topic, because a parent who signs a form saying general school survey has not been told anything. Second, the right to skip a question has to be stated on the page, not merely felt. Third, a study that collects nothing identifying, no names, no class period, no student numbers, is far easier to justify than one that does, and it is usually just as informative.

The upshot: Voluntary means a student can say no without cost. If your respondents are your classmates and you are standing over the desk while they answer, saying no costs something, and your design has to remove that cost rather than deny it exists.

Anonymous and confidential are different words

An anonymous questionnaire is one where nobody, including you, can link an answer to a person, because the link was never created. A confidential study creates the link and promises to protect it. Confidentiality is much harder than it sounds, and the way it usually fails is not theft but arithmetic.

Suppose you publish a table of your 60 respondents broken down by year group, gender and whether they have a part-time job. One cell contains two people: senior girls with a job. Anyone in that school who knows both of them can now read their answers to every other question in your results. This is deductive disclosure, and professional statistical agencies spend enormous effort on it. The Census Bureau suppresses or noises small cells for exactly this reason. Your defence is the same as theirs: report a cell only when it is large enough to hide an individual, and merge categories when it is not.

One more idea belongs here, because it is an ethical commitment as much as a methodological one. Max Weber argued for value neutrality, meaning that a researcher's personal opinions must not shape how the evidence is read or reported. You do not have to pretend to have no views about school discipline or inequality. You do have to design the study so that your view could lose, and report the result that comes out. A questionnaire built to produce the answer you already hold is not research, and a reader who spots the loaded wording will discount the whole thing.

Common misconceptions

• "Tuskegee was a case of the government infecting people with syphilis." It did not infect anyone. The men arrived already infected, and the wrong was in withholding a known treatment for decades while telling them they were being treated. Holding the actual facts matters, because the myth is easier to dismiss.
• "A signature on a form means you have informed consent." The form is evidence of a conversation, not a substitute for one. Consent requires that the person understood the purpose, the risks and the right to stop, in language they can read.
• "Ethics rules only apply to medical research." The Common Rule applies to surveys, interviews and observation as well. A questionnaire about drug use or family income can do real damage if the answers are traceable.
• "If the data are anonymous there is nothing to worry about." Anonymity removes names. It does not stop a table with a cell of two people from identifying them.

What to remember

• 600 men enrolled at Tuskegee in 1932, 399 with latent syphilis, told they were treated for bad blood; the study ran until 1972 and penicillin was withheld after 1947.
• Exposure by Peter Buxtun and the Associated Press in July 1972 led to the National Research Act of 1974, review boards, and the Belmont Report of 1979.
• Belmont's three principles are respect for persons, beneficence and justice, and each one maps onto a design decision in a school survey.
• Humphreys' Tearoom Trade showed that sociological methods alone can put subjects at risk of arrest, and the argument about covert observation is not closed.
• A review board checks risk, benefit, equitable selection, consent and data protection; exempt categories still need someone else to declare them exempt.
• For a minor, permission comes from a parent or guardian and assent comes from the student; either refusal ends participation.
• Anonymity prevents the link from existing; confidentiality protects a link that does exist, and small cells in a published table can undo it.

Sources

1. Wikipedia. (2025). Tuskegee Syphilis Study. wikipedia.org
2. National Commission for the Protection of Human Subjects of Biomedical and Behavioral Research. (1979). The Belmont Report. Department of Health, Education, and Welfare. hhs.gov
3. Office for Human Research Protections. (2018). Protection of human subjects, 45 CFR Part 46. ecfr.gov
4. OpenStax. (2021). Ethical concerns. In Introduction to Sociology 3e. Rice University. openstax.org
5. Humphreys, L. (1970). Tearoom Trade: Impersonal Sex in Public Places. Aldine.

Two numbers, fifty-seven years apart

In 1965, 42.4 percent of American adults smoked cigarettes. In 2022, 11.6 percent did, which is about 28.8 million people. Cigarettes were legal throughout that period, sold in every state, and are still stocked behind the counter of most pharmacies. No prohibition was ever passed.

Something moved roughly thirty percentage points of adult behaviour without banning the product. Explaining that is the work of this lesson, and it needs a small set of tools: the parts a culture is made of, the grades of norm, the sanctions that enforce them, and the difference between what a group says it values and what it does. Keep the two numbers in view. At the end you will explain the change three ways, and the three explanations point at different causes.

The two halves of a culture

Culture is the whole of what a group has learned and passes on. Sociologists split it in two. Material culture is the physical stuff: cars, phones, church buildings, cigarettes, school desks bolted in rows. Non-material culture is everything not physical: language, values, beliefs, norms, the rule that you do not read over a stranger's shoulder on a bus.

Both halves run on symbols, things that stand for something else by agreement. A red octagon means stop only because English-speaking drivers learned it does. The most powerful symbol system is language, and it carries a famous overclaim. The strong version of the Sapir-Whorf hypothesis, that your language determines what you can think, has not survived testing. A weak version has: the categories your language marks routinely, such as whether it distinguishes light blue from dark blue with separate basic words, measurably shift how fast speakers sort colours. The distinction matters because it is the difference between a claim that is false and a claim that is small and real.

Values are not behaviour

Values are a group's abstract standards of what is good and desirable. Beliefs are what it holds to be true. Ideal culture is what a group says it stands for; real culture is what members actually do. The gap between them is not hypocrisy to be scolded, it is data to be measured, and it is where a lot of sociology lives.

Values also differ by country in ways you can count. The Pew Research Center asked 18,850 adults in 17 advanced economies, between February and May 2021, an open question: what makes your life meaningful? Because the question was open, people supplied their own categories rather than ticking a box.

Read the religion row carefully, because it is the kind of finding that gets misreported in both directions. Fifteen percent is a minority of Americans. It is also three times the highest figure anywhere else in the survey. Both statements come from the same cell.

Key idea: A value is invisible until you find the behaviour or the answer that reveals it. Sociology measures values through what people do, say and choose, never by asserting what a culture believes.

Norms come in grades

A norm is a rule for behaviour in a situation. William Graham Sumner separated the weak from the strong in 1906, and the grades still organise the subject.

The grades are not fixed for all time, and watching one move is watching culture change. Smoking indoors moved from folkway to law in about twenty-five years. Wearing a hat indoors moved the other way, from a more to a matter nobody notices.

Sanctions are how a norm bites

A sanction is the response a norm provokes. Sanctions are formal when an institution administers them, a fine, a suspension, a licence revoked, and informal when people do it themselves: a stare, a joke at your expense, being left out of a group chat. They are positive as well as negative; a certificate and an approving nod are both sanctions.

Informal sanctions do most of the day-to-day work, for a simple reason. There is one police force and there are thousands of onlookers. A school can suspend a student for fighting, but the reason most students do not fight at lunch on most days is that everyone would see it.

Now explain the thirty points

Take the three perspectives from Module 1 and set them on the smoking decline. They use the same data and they disagree about what did the work.

A functionalist account. New information arrived, in the form of the Surgeon General's 1964 report linking smoking to lung cancer, and the culture's norms adjusted to keep the system working. Institutions realigned one after another: warning labels on packs, the Public Health Cigarette Smoking Act signed on 1 April 1970 that took cigarette advertising off television and radio from 2 January 1971, California's statewide workplace ban in 1994 extended to bars in 1998. On this reading the norm changed because the belief changed, and the institutions followed.

A conflict account. Look at who won and who lost, and at whose behaviour changed least. Quitting is not distributed evenly. Federal survey data for 2022 found recent successful quitting ranged from 16.8 percent among adults with a graduate degree down to 4.0 percent among adults with no high school diploma, and from 11.9 percent among higher-income adults to 7.5 percent among lower-income adults. The decline was real and it was uneven, which a conflict analyst reads as a story about resources, workplace rules that protect some employees and not others, and an industry that kept selling to the groups with the least protection.

A symbolic-interactionist account. What changed was what a cigarette means. In 1965 a cigarette in a film signified adulthood, poise, sometimes glamour. By 2005 the same object in the same shot signified something closer to a character flaw. Removing broadcast advertising in 1971 removed one large machine for making the first meaning, and indoor bans physically relocated smokers to doorways, where the act became visible, separate and marked. On this reading the behaviour fell because the meaning of the act was renegotiated in a million ordinary interactions.

The upshot: The three accounts are not three opinions about the same sentence. They predict different things. The functionalist expects change to follow information; the conflict analyst expects the gap between groups to persist; the interactionist expects behaviour to track the meaning attached to it. Each prediction can be checked against data, which is why the perspectives are tools rather than beliefs.

Inside a culture there are smaller ones

A subculture is a group with its own norms and symbols that still sits inside the larger culture. A counterculture rejects central values of the larger culture and builds alternatives.

The clearest measured case in the United States is the Old Order Amish. The Young Center at Elizabethtown College puts the North American Old Order population at about 421,000 in 2026, growing near 3 percent a year, which doubles the population roughly every twenty years; total fertility ran about 5.3 children per woman in the 2010s. Their rules are not vague sentiment: each settlement's Ordnung, reviewed twice a year by the members, specifies what is permitted, from power from the grid to the style of a coat. Note what that implies about norms in general. The Amish are unusual not because they have detailed norms but because theirs are written down and reviewed, while yours are enforced constantly and never published anywhere.

Ethnocentrism, relativism, and what relativism does not ask of you

Sumner also gave us ethnocentrism: judging another culture by the standards of your own and finding it deficient because it is not yours. Its opposite as a working method is cultural relativism, associated with Franz Boas and his students: describe a practice in terms of the system it belongs to before evaluating it. Ask what the practice does, who performs it, what would happen if it stopped.

Two clarifications, because this is where the idea usually gets mangled. Relativism as a method does not require you to approve of anything; it requires you to understand before you judge, which is a sequence, not a surrender. And relativism does not treat a culture as a single mind: inside every culture some people benefit from a practice and others bear its cost, and asking which is which is ordinary sociology rather than ethnocentrism. The uncomfortable feeling you get when the rules around you stop being the ones you know is culture shock, and it is evidence of how much of your own behaviour is culturally supplied rather than chosen.

When the gadget arrives before the rule

William F. Ogburn named the pattern in 1922: material culture changes faster than non-material culture, so there is a period in which the object exists and the norms for it do not. He called it culture lag.

You have lived through several. Smartphones reached most American teenagers years before any settled norm existed about photographing a classmate without asking, and school policies were written afterwards, unevenly, district by district. Group chats arrived before any rule about who may screenshot them. Generative software that writes essays arrived in classrooms before the rules on what counts as your own work were rewritten. Culture lag predicts exactly this shape: the tool, then the argument, then the rule.

Common misconceptions

• "Culture means art, museums and classical music." Sociologists use the word for the whole learned way of life, including the rule about which side of an escalator you stand on. Highbrow taste is one small part of it.
• "Your language limits what you can think." The strong Sapir-Whorf claim failed. The surviving version is narrower: the distinctions your language marks routinely can speed up or slow down how you sort things, which is a real but modest effect.
• "Cultural relativism means all practices are equally good." It is a rule about the order of operations in analysis, describe the practice in context first, not a rule forbidding judgement. It also does not require pretending every member of a culture benefits equally.
• "Norms are what the law says." Law is one grade of norm, and the least of the three in daily volume. Most of what regulates your behaviour today was never written down, and its sanctions are faces rather than fines.

Pulling it together

• Adult smoking fell from 42.4 percent in 1965 to 11.6 percent in 2022 with no prohibition, which is what a change in norms, sanctions and meaning looks like at national scale.
• Culture divides into material and non-material, and both run on symbols; the strong language-determines-thought claim is false, the weak one holds.
• Values are abstract standards and are measured through behaviour and answers: in 2021, family was the top source of meaning in 14 of 17 advanced economies, and religion was named by 15 percent of Americans against under 5 percent elsewhere.
• Norms grade from folkways through mores and taboos to laws, and sanctions can be formal or informal, positive or negative, with informal sanctions doing most of the enforcing.
• Functionalist, conflict and interactionist readings of the same decline point at information, resources and meaning respectively, and make different checkable predictions.
• Subcultures sit inside a culture, countercultures reject its central values, and the Old Order Amish are the best documented American case at about 421,000 people in 2026.
• Ethnocentrism judges by your own standards; relativism describes before judging and still allows you to ask who inside the culture pays.
• Culture lag names the gap between a new object and the norms for it, which is why the rules about phones and AI in school arrived years late.

Sources

1. Pew Research Center. (2021). What makes life meaningful? Views from 17 advanced economies. pewresearch.org
2. OpenStax. (2021). Elements of culture. In Introduction to Sociology 3e. Rice University. openstax.org
3. Wikipedia. (2026). Amish. wikipedia.org
4. Centers for Disease Control and Prevention. (2024). Current cigarette smoking among adults in the United States. National Center for Chronic Disease Prevention and Health Promotion. Cited without a link because the host refuses automated requests.
5. Centers for Disease Control and Prevention. (2024). Adult smoking cessation, United States, 2022. Morbidity and Mortality Weekly Report, 73(29). Cited without a link for the same reason.

A locked room in Temple City

In November 1970 a woman walked into a social services office in California looking for services for the blind, and brought her 13-year-old daughter with her. The child weighed 59 pounds, could not straighten her arms or legs, and did not speak. From roughly the age of 20 months she had been kept in a locked room by her father, strapped to a child's toilet during the day or bound in a crib at night, with almost no language addressed to her. Researchers called her Genie to protect her identity.

By January 1971 her receptive vocabulary was measured at her own name, a few familiar names, and about 15 to 20 words. She could produce two phrases: stop it, and no more. Over the next four years she learned rapidly. By mid-1975 she could name most objects she came across. What she never acquired was grammar. She stayed at the level of short strings of words, and researchers still disagree about how much of that was the isolation and how much was other damage.

Hold on to the shape of that outcome, because it is the evidence for a claim this whole lesson rests on. A human being is not a member of a society by birth. Membership is produced, by other people, in interaction, and the process has a name: socialisation.

What one case can and cannot establish

Genie's case is quoted constantly and misused often, so be precise about its limits. It is a sample of one. The deprivation came bundled with malnutrition, beatings and darkness, so no one can say which of those produced which deficit. And nobody chose the conditions, which means there is no comparison group of children deprived of language but fed and unharmed.

Kingsley Davis published two earlier American cases, in 1940 and 1947, that show why a single case cannot settle the question. A girl he called Anna, found at about six after years confined in a storage room, gained limited speech and died young. A girl he called Isabelle, found at a similar age in similar conditions, recovered much further and reached something close to ordinary speech. Same broad deprivation, different endings. That is what a set of case studies does: it establishes that the process matters enormously, and it cannot tell you how much, for whom, or at what point recovery stops being possible.

What matters here: The isolation cases prove that social contact is necessary for normal development. They do not measure how much of anything, because the conditions were never assigned and nothing was controlled.

The experiment nobody may run on children

You cannot deprive a child of contact to see what happens. Harry Harlow did it with rhesus monkeys, reporting the results in an address titled The Nature of Love in 1958. Infant monkeys were raised with two artificial mothers: one of wire that dispensed milk, one covered in cloth that dispensed nothing. If attachment were simply about feeding, the infants should have preferred the wire mother. They overwhelmingly clung to the cloth one, going to the wire frame only to feed. In the fear test, an infant with no cloth mother present cowered away from a noise-making toy; an infant with the cloth mother reachable touched it, then explored and attacked the toy.

The finding reframed attachment as a need in its own right rather than a by-product of feeding. The studies are also now regarded as unethical, and they helped provoke the modern rules on laboratory animals. Both statements are true at once, and a course that reports the finding without the second one is not teaching you the state of the field.

The experiment that actually was run

Romania in the 1990s had thousands of children in state institutions, and the alternative on offer was not obviously better, because the country had almost no foster care system. Between 2000 and 2005 a team led by Charles Nelson, Nathan Fox and Charles Zeanah assessed 136 institutionalised children in Bucharest, average age 21 months, and then randomly assigned them either to a newly built foster care programme or to continued institutional care. A group of children from the same city who had never been institutionalised served as a comparison.

Random assignment is what makes this study unusual and what makes it hard to read comfortably. The results were consistent: the children moved into families did better on cognitive measures at four and a half years than those who stayed, and both institutionalised groups scored below the never-institutionalised children. On attachment security, roughly half of the children placed with families were classified as securely attached against fewer than one in five of those who remained in institutions, and children placed before 24 months did better than those placed later.

Notice the ethical structure, because it connects to the previous lesson. Nobody could ethically assign children to institutional care. The researchers assigned some children to something better than the existing default, and the control group received what the state was already providing. That design decision is what made the evidence possible and it is still argued about.

Cooley: you see yourself in other people

Charles Horton Cooley proposed in 1902 that the self is built out of other people's reactions, in a three-step loop he called the looking-glass self. You imagine how you appear to another person, you imagine their judgement of that appearance, and you develop a feeling about yourself from what you imagine.

Work it through on something concrete. You answer a question in class and get it wrong. Step one: you picture how you looked answering. Step two: you guess what the other students concluded, that you had not read the chapter. Step three: the feeling arrives, and it attaches to a self-description, that you are not good at this subject. The crucial and easily missed detail is that step two is a guess. Cooley's mechanism runs on your interpretation of others' judgements, not on their actual judgements, which is why a student can build a durable belief about their own ability out of a reaction that never happened.

Mead: the self arrives in stages

George Herbert Mead pushed the argument further: the self develops through learning to take the role of another person, and it comes in a sequence you can watch in any family.

The generalised other is the pay-off concept. Once you carry it, society does not need a supervisor in the room; you supervise yourself using an internalised audience. Mead split the self into the I, the spontaneous actor that does something before it is checked, and the me, the part that observes and evaluates using that audience. The argument in class you almost started and did not is the me overruling the I.

Who does the socialising

The groups and institutions that carry out socialisation are called agents of socialisation. They do not teach the same things, and where they conflict is where a lot of adolescence happens.

It does not stop when you grow up

Preparing for a role you do not yet hold is anticipatory socialisation: the senior reading a college's orientation pages in March, the apprentice copying how the qualified electrician talks to a client. Sociologists studying the life course treat later transitions the same way, with new norms to learn at each one.

Resocialisation is the harder version: dismantling one set of norms and installing another. It happens most visibly in what Erving Goffman called a total institution, described in his 1961 book Asylums: a place where people are cut off from wider society, live and work in one location under a single authority, and pass through a standard entry procedure. The procedure strips the old identity, with uniforms, a number, the removal of personal possessions, a new name or rank. Military basic training, prisons, some hospitals and boarding schools share the structure. Recognising it in a place you know is the point; the machinery is identical whether the outcome is a soldier or a patient.

Remember: Socialisation is not indoctrination, and it is not one-way. A person selects, resists, mocks and reinterprets what the agents offer, which is why two children of the same parents with the same school and the same neighbourhood do not come out the same.

Nature and nurture, stated accurately

Twin and adoption studies, including the long-running Minnesota Twin Family Study, find substantial heritability for traits from height to measured personality. Sociology does not dispute that, and you should not pretend it does. What sociology insists on is what a heritability estimate actually is: a statement about how much of the variation in a trait within a particular population under particular conditions tracks genetic variation. It is not a proportion of any individual's trait, and it changes when the environment changes. If every child in a country gets adequate food, almost all remaining variation in height will look genetic; introduce severe malnutrition in half the country and the same trait's heritability drops, with no gene having changed.

That is why the isolation cases and the Bucharest experiment matter. Whatever a child's inheritance, its expression runs through a social environment, and remove the environment and the capacity does not appear on schedule by itself.

Common misconceptions

• "Genie proves there is a critical period for language that closes at puberty." Her case is consistent with that hypothesis and cannot establish it. A single child, with malnutrition and abuse alongside the isolation, cannot separate the causes or fix the age at which recovery becomes impossible.
• "Socialisation is what happens to children." It runs the whole life course. Starting a job, becoming a parent, entering the military and moving country all involve learning norms you did not have.
• "The looking-glass self means you become what others think of you." It means you become what you think others think of you. The step is your interpretation, which can be wrong and still shape a self-image for years.
• "If a trait is heritable, social conditions cannot matter." Heritability describes variation in one population in one setting. Change the setting and the number changes, which is exactly why studies of deprivation find what they find.

The takeaway

• Genie, found in November 1970 at 13, had a receptive vocabulary of about 15 to 20 words and two spoken phrases; she later named most objects but never acquired grammar.
• Case studies of isolation establish that social contact is necessary and cannot measure how much, because nothing was assigned or controlled.
• Harlow's infant monkeys preferred the cloth mother that gave no food, which reframed attachment as a need in itself; the studies are also now considered unethical.
• The Bucharest project randomly assigned 136 institutionalised children, average age 21 months, to foster care or continued institutional care, and those placed with families did better on cognition and attachment, especially if placed before 24 months.
• Cooley's looking-glass self works through your imagination of others' judgements, not their actual judgements.
• Mead's stages run from imitation through play and game to the generalised other, and the self divides into the spontaneous I and the evaluating me.
• Family, school, peers, media and workplace teach different and sometimes conflicting things.
• Anticipatory socialisation prepares you for a role you do not hold; resocialisation replaces one set of norms with another, most systematically inside Goffman's total institutions.
• Heritability is a property of variation in a population and setting, not a share of any individual's trait.

Sources

1. Wikipedia. (2025). Genie (feral child). wikipedia.org
2. OpenStax. (2021). Theories of self development. In Introduction to Sociology 3e. Rice University. openstax.org
3. Wikipedia. (2025). Harry Harlow. wikipedia.org
4. Nelson, C. A., Zeanah, C. H., Fox, N. A., Marshall, P. J., Smyke, A. T., and Guthrie, D. (2007). Cognitive recovery in socially deprived young children: The Bucharest Early Intervention Project. Science, 318(5858), 1937-1940.
5. Davis, K. (1940). Extreme social isolation of a child. American Journal of Sociology, 45(4), 554-565.

Eight men and a card of lines

In 1951 Solomon Asch sat groups of eight young men in a row at Swarthmore College and showed them a card with a single reference line and three comparison lines. The task was to say aloud which comparison line matched. It was easy: in a control condition where people answered alone, the error rate was under 0.7 percent. In the experimental groups, seven of the eight were working for Asch and, on 12 of the 18 trials, all gave the same wrong answer before the real participant spoke.

Across the 50 experimental subjects, 35.7 percent of answers on the critical trials followed the incorrect majority. Roughly three-quarters of participants went along at least once. Twenty-six percent never did. A later series with 123 students at three universities produced the same headline: conformity on about one third of critical trials.

Then Asch changed one thing. He put a single confederate in the group who gave the correct answer. Conformity to the majority collapsed to about 5 percent of responses. That comparison, and not the headline number, is the finding worth carrying: the pressure is a property of the group's structure, and one visible ally dismantles most of it.

Obedience is a different mechanism

Conformity is pressure from equals. Obedience is compliance with an instruction from someone defined as having the right to give it, and Stanley Milgram set out to measure it at Yale from August 1961. A participant was told to deliver increasing electric shocks to a learner in the next room, who was an actor receiving nothing. Forty psychiatrists asked to predict the results estimated that about one tenth of one percent of people would go to the maximum 450 volts. In the first official experiment every participant went to at least 300 volts, and 65 percent went to 450.

The number is famous, and the caveats are part of the finding. Gina Perry's 2012 work in Milgram's archives found that only about half of the participants fully believed the shocks were real, and of those who did, 66 percent disobeyed. The experimenter also improvised prods beyond the written script. So the accurate statement is narrower than the legend: under specific conditions, a large minority of ordinary people continued to a point they had said they would not, and the exact percentage is not a constant of human nature.

The third study everyone has heard of has not survived as well. Compare all three.

The core of it: The three studies are usually recited as one lesson about human nature. They are three different claims with three different evidential strengths, and the honest ranking puts Asch first, Milgram second with conditions attached, and the prison study third as a cautionary tale about research design.

The positions a person occupies

A status in sociology is not prestige; it is a position in a social structure, and everyone holds many at once. That collection is a status set: student, daughter, employee, goalkeeper, neighbour.

An ascribed status is assigned without effort or choice: age, the family you were born into, in most contexts race and sex. An achieved status is one you reach through what you do: graduate, captain, shift supervisor, convicted felon. The line blurs and that is where the interesting work is, because achieving a status usually requires resources that were themselves ascribed.

A master status is one that overrides the others in how people treat you. It is rarely chosen. A student known for one fight in ninth grade can find every teacher's expectation running through that fact three years later, which is a master status doing the work that a whole status set should be doing.

Two ways a role goes wrong

A role is the set of expectations attached to a status. Status is the position; role is the script. Two distinct failures get confused constantly, so fix them with the definition and one example each.

• Role conflict is between two statuses. Your shift starts at 5 pm and practice runs until 5.30. Employee and team member cannot both be satisfied.
• Role strain is inside one status. As team captain you are expected to encourage the weakest player and to win the match, and in the final minute those two expectations pull against each other. One status, two incompatible demands.

Leaving a role that has become part of your identity is role exit, and it takes work: an ex-athlete, an ex-member of a religious community and a retiree all have to build a new answer to what they do.

Groups, compared

Georg Simmel made the observation that group size changes group behaviour by itself. A dyad of two is the most intense and the most fragile, because either member's departure ends it. Add one person and the structure changes qualitatively: a triad allows coalitions, mediation and being outvoted.

Read the last row against the third. Your reference group need not be a group you belong to, which is why a student with good grades in a selective school can feel like a failure, and a student with identical grades elsewhere can feel successful. Nothing about the grades changed. The comparison set did.

One group pathology is worth naming because it has a body count in aviation and policy: groupthink, Irving Janis's term for a cohesive group suppressing dissent to preserve agreement. Notice that Asch already gave the remedy. One person willing to say the obvious thing out loud changes what everyone else does.

Networks: the acquaintance finds the job

A social network is the web of ties a person has, and the shape of the web does work that the individuals do not. Mark Granovetter surveyed 282 professional, technical and managerial workers in Newton, Massachusetts, for his 1970 dissertation, published as Getting a Job in 1974. Among those who found their job through a personal contact, he asked how often they saw that contact:

Most jobs came through people the worker barely saw. Granovetter's explanation is the strength of weak ties: your close friends know what you know, because you all move in the same circles, so the genuinely new information about an opening arrives from the acquaintance who is connected to a different pool. When workers were asked whether a friend told them about the job, the commonest answer was that it was not a friend but an acquaintance.

Networks also turn out to be short. Milgram's small-world studies from 1967 asked people to forward a packet toward a named stranger through personal contacts. In one run, 160 letters were sent and 24 arrived; completed chains averaged about five and a half to six intermediaries, which is where six degrees of separation comes from. Judith Kleinfeld's later critique matters: in one study 232 of 296 letters never arrived at all, and because longer chains are more likely to break, the famous average probably understates the true distance.

Bureaucracy: what the rules buy and what they cost

Max Weber described the bureaucratic form as the most technically efficient way yet found to administer large numbers of people, and as a cage. Both halves are in the table.

Weber's phrase for the result is usually rendered as the iron cage: rational systems, each sensible in itself, combining into a structure no individual chose and none can easily leave. George Ritzer updated the argument as McDonaldization, arguing that four principles, efficiency, calculability, predictability and control, have spread from fast food into schooling, medicine and work, and that they produce their own irrationalities. Robert Michels added the iron law of oligarchy: even organisations founded to be democratic tend to concentrate power in a small leadership, because running things requires expertise, continuity and time that ordinary members do not have.

Your school is the nearest example of the whole table. Bells, a timetable, a discipline code applied to cases it did not anticipate, records that follow you for years, and staff selected on credentials. Complaining about it is easy. Reading it as the price of treating two thousand people consistently is the sociological move.

Common misconceptions

• "Asch showed that most people are conformists." He showed that 35.7 percent of critical responses conformed, that 26 percent of participants never did, and that one ally cut conformity to about 5 percent. The result is about situations, not personalities.
• "Milgram proved that 65 percent of people would torture someone on command." Under his conditions 65 percent reached the maximum shock. Archival work found that only about half fully believed the setup, and the figure moves substantially when the conditions change.
• "The Stanford prison experiment shows that roles turn ordinary people cruel." Later archival work found guards were coached toward harshness and several participants said they were performing, so the study cannot carry that claim.
• "Bureaucracy means inefficiency and red tape." In Weber's analysis it is the form adopted because it is efficient at scale. The pathologies are the price of that efficiency, not evidence that the form fails.
• "Your closest friends are your most useful contacts." In Granovetter's data most jobs came through people seen occasionally or rarely, because close friends share your information rather than adding to it.

Summing up

• Asch: error under 0.7 percent alone, 35.7 percent of critical responses conforming in groups, about 5 percent with one dissenting ally present.
• Milgram: 65 percent to 450 volts against a predicted 0.1 percent, qualified by evidence that only about half of participants fully believed the shocks.
• The Stanford prison study is now the weakest of the three, because guards were coached and participants reported acting.
• Status is a position and role is its script; ascribed statuses arrive without choice and a master status overrides the rest.
• Role conflict is between two statuses; role strain is between demands inside one.
• Group size changes behaviour by itself, and a reference group supplies the standard by which you judge your own results.
• Among workers who found jobs through contacts, 55.6 percent saw that contact only occasionally and 27.8 percent rarely, which is the strength of weak ties.
• Chains between strangers averaged about six intermediaries among completed chains, with most letters never arriving, so the estimate is probably low.
• Weber's six bureaucratic features each buy something and cost something; Ritzer and Michels extend the argument to fast food and to democratic organisations respectively.

Sources

1. Wikipedia. (2025). Asch conformity experiments. wikipedia.org
2. Wikipedia. (2025). Milgram experiment. wikipedia.org
3. OpenStax. (2021). Formal organizations. In Introduction to Sociology 3e. Rice University. openstax.org
4. Granovetter, M. S. (1974). Getting a Job: A Study of Contacts and Careers. Harvard University Press.
5. Le Texier, T. (2019). Debunking the Stanford Prison Experiment. American Psychologist, 74(7), 823-839.

A nine-page article from March 1982

The Atlantic Monthly published an article by James Q. Wilson and George Kelling in March 1982 built around one image. If a window in a building is broken and nobody fixes it, the rest of the windows will soon be broken. Their argument was that visible disorder, graffiti, litter, public drinking, broken glass, signals that nobody is watching, which invites more disorder and eventually serious crime. The policy conclusion followed: enforce the small things and the big things will fall.

Broken windows became one of the most influential ideas in American policing, and whether it is true is still argued. Deciding what you think requires the five explanations of deviance that sociologists work with, because each one predicts something different about what fixing a window accomplishes. Build the five first. The dispute is waiting at the end of the lesson.

Durkheim: a society of saints would still punish

Deviance is behaviour that violates a norm, which is not the same as crime; crime is the subset that breaks a law. Emile Durkheim made the argument that sounds wrong at first and turns out to be hard to escape: deviance is normal, in the statistical sense of being present in every society ever studied, and it does work for the group.

His illustration is a monastery of saints. Nothing we would call crime happens there, and yet the community will still find offences to punish, because behaviour that is merely slightly worse than the rest becomes the boundary case. Punishment does three things at once. It tells everyone where the line is. It brings the group together against the offender, which strengthens solidarity. And where the offender turns out to have public sympathy, it can move the line, which is how norms change. Durkheim also argued a society can have too little regulation, a condition of anomie, which you met in Module 1.

Merton: the gap between what you are told to want and what you can reach

Robert K. Merton took Durkheim's anomie and located it somewhere specific: a society that presses everyone to pursue the same goal, in the American case material success, while distributing the legitimate means to reach it unequally. The result is strain, and Merton set out five ways people respond.

Strain theory predicts that deviance concentrates where the gap between goals and means is widest, which is a testable claim about neighbourhoods and schools rather than about individuals.

Sutherland: deviance is learned, in groups, like everything else

Edwin Sutherland argued in the 1930s and 1940s that criminal behaviour is learned in interaction with other people, and that what is learned includes technique, motives and the justifications that make the act acceptable to yourself. Differential association says the balance matters: a person becomes deviant when definitions favourable to breaking a rule outweigh definitions unfavourable to it, weighted by how early, how often, how long and how intensely the associations run.

Sutherland also coined the term white-collar crime, in 1939, and the reason it belongs here is that it embarrasses the theories that explain crime by poverty. An accountant who moves client money does not lack legitimate means. He has learned, in an office, a set of definitions under which the act is a technical adjustment rather than theft.

Becker and Lemert: the label does work of its own

Labelling theory moves the question from why someone breaks a rule to what happens after they are caught. Edwin Lemert separated primary deviance, the act itself, which almost everyone commits at some point, from secondary deviance, the behaviour that follows once a person is publicly labelled and begins to be treated, and to see themselves, as that kind of person. Howard Becker put the strong version in Outsiders in 1963: deviance is not a quality of the act but the consequence of a group applying a rule and a sanction to it.

William Chambliss supplied the case that makes the mechanism visible. In a study published in 1973 he followed two groups of boys in the same high school. The Saints came from wealthier families, had cars, and committed vandalism, petty theft and drunk driving away from the neighbourhood; they were seen as good students headed for college and almost never punished. The Roughnecks came from poorer families, had no cars, and fought and stole in public view where the same conduct was seen and recorded. The Saints all finished college. Several Roughnecks were convicted. The acts overlapped substantially; the labels did not, and the labels predicted the outcomes.

Why this matters: If who gets labelled depends partly on class, visibility and transport, then official crime statistics are records of official reactions as well as of behaviour. You will spend the whole of the next lesson on that problem.

Hirschi: the question is why most people do not

Travis Hirschi inverted the puzzle in 1969. Rule-breaking is often easy and rewarding, so the thing to explain is restraint. His social control theory says four bonds hold a person to conventional behaviour, and deviance rises as they weaken.

• Attachment: ties to people whose opinion of you matters.
• Commitment: an investment you would lose, a place at a college, a job, a reputation.
• Involvement: hours occupied by conventional activity.
• Belief: acceptance that the rules are legitimate.

Control theory is why after-school programmes are defended on the argument about involvement, and why the same programmes are criticised for treating the symptom rather than the strain.

Now the dispute: what does each theory predict about the broken window?

The evidence offered for the claim

New York City is the exhibit. From 1993 the police department under William Bratton pursued minor offences systematically: fare evasion, public drinking, graffiti, unlicensed venues. Serious crime then fell for two decades, and it kept falling. By 2023 the city's homicide rate was 4.1 per 100,000 people, below the national rate of 5.6, which would have been a startling claim to make about New York in 1990. Kelling and a co-author published an analysis in 2001 arguing that both petty and serious crime dropped after the change in policing.

There is also careful experimental support for the psychological mechanism, from a place unconnected to American policing. Kees Keizer and colleagues at the University of Groningen ran six field experiments in the Netherlands, published in Science in 2008. In one, they measured how many people littered a flyer in an alley with a clean wall, then repeated it with graffiti on the wall. Graffiti roughly doubled the share who littered. In another, an envelope with a visible banknote protruded from a postbox: disorder in the surroundings more than doubled the share of passers-by who took it. Seeing one norm broken raised the rate at which people broke a different one.

The evidence offered against it

The problem with the New York exhibit is that everything else changed too. Crime fell in the 1990s in cities that did nothing like New York's policing, and the candidate causes are numerous: the end of the crack epidemic, a sharp fall in unemployment, a rise in incarceration from earlier drug laws, an ageing population, gentrification, and falling childhood lead exposure. Steven Levitt's work attributed much of the decline to increased incarceration, more police and the waning of crack, along with a much-argued claim about abortion legalisation that other economists have challenged.

Two more direct tests cut against the theory. Bernard Harcourt and Jens Ludwig reanalysed the New York data and concluded that the pattern was explained by mean reversion, the tendency of the places that spiked highest to fall the furthest, with or without a policing change. They then used the Moving to Opportunity experiment from Lesson 6: families were randomly assigned vouchers that moved them to more orderly neighbourhoods, and the tenants who moved continued to be arrested at about the same rate. If disorder itself caused crime, moving people out of disorder should have reduced their offending, and it did not.

Robert Sampson and Stephen Raudenbush offered a different mechanism: what predicts crime in a neighbourhood is collective efficacy, the willingness of residents to intervene for the common good, and both disorder and crime are consequences of its absence. On that reading, painting over graffiti treats a symptom. And when the New York police sharply reduced minor-offence enforcement during a slowdown in late 2014 and early 2015, major crime complaints did not rise; the measured effect on murder, rape, robbery and car theft was not statistically significant.

Bench Ansfield added a historical point about the original article: Wilson and Kelling supported their central claim by citing a single vandalism demonstration by Philip Zimbardo, and used it for more than its author had concluded.

What would settle it

The two bodies of evidence are not actually about the same claim, and seeing that is the skill this lesson is for. The Dutch experiments randomised disorder and measured small norm violations minutes later. The New York argument is about serious crime, over decades, in one city, with no control group. A finding that graffiti doubles littering does not establish that painting walls reduces robbery.

What would help: randomise at the level of the intervention and measure the outcome that is in dispute. Assign disorder repair or minor-offence enforcement to some city blocks and not to closely matched others, by lottery, and track both serious crime and the number of people acquiring records. Policing experiments of this kind have been run in several American cities with mixed results, and the honest summary today is that the micro mechanism has support, the city-scale causal claim does not, and the cost side, arrests and records for minor offences, is a measurable harm that the original article did not weigh.

The point: The dispute is not settled by choosing the theory you find congenial. It is narrowed by asking of each study: what was assigned, what was measured, over what period, and compared with what.

Common misconceptions

• "Deviance and crime are the same thing." Crime is the part of deviance that breaks a law. Standing too close to people in a lift is deviant and not criminal; some laws are broken so routinely that violation is not treated as deviant at all.
• "Durkheim thought crime was good." He argued deviance is universal and performs functions for the group, including marking boundaries and enabling change. That is an analytic claim about societies, not approval of any offence.
• "Labelling theory says deviant acts do not really happen." It says the act alone does not make a person deviant in the social sense; the reaction, and who is available to be labelled, do part of the work. The Saints really did drive drunk.
• "New York's crime drop proves broken windows works." Crime fell in cities with very different policing during the same period, and a reanalysis attributes much of the pattern to mean reversion. One city with no control group cannot settle a causal claim.

Recap

• Wilson and Kelling's 1982 article claimed visible disorder invites serious crime, and it reshaped policing.
• Durkheim: deviance is universal and functional, marking boundaries, building solidarity and allowing norms to move.
• Merton: strain between prescribed goals and available means yields conformity, innovation, ritualism, retreatism or rebellion.
• Sutherland: deviance is learned in groups, and white-collar crime shows that poverty is not the mechanism.
• Lemert and Becker: the label converts primary deviance into secondary deviance, and Chambliss's Saints and Roughnecks committed similar acts to very different official effect.
• Hirschi: attachment, commitment, involvement and belief restrain people, so weakening bonds rather than motivation explains deviance.
• For the claim: New York's sustained decline, a 2023 city homicide rate of 4.1 against 5.6 nationally, and Dutch field experiments in which graffiti roughly doubled littering.
• Against it: mean reversion, randomised moves out of disorder that did not reduce offending, collective efficacy as the common cause, and a 2014 to 2015 enforcement slowdown with no significant effect on major crime.
• The two evidence bases measure different outcomes at different scales, which is why the micro mechanism can hold while the policy claim remains unproven.

Sources

1. Wikipedia. (2025). Broken windows theory. wikipedia.org
2. OpenStax. (2021). Theoretical perspectives on deviance. In Introduction to Sociology 3e. Rice University. openstax.org
3. Keizer, K., Lindenberg, S., and Steg, L. (2008). The spreading of disorder. Science, 322(5908), 1681-1685.
4. Chambliss, W. J. (1973). The Saints and the Roughnecks. Society, 11(1), 24-31.
5. Harcourt, B. E., and Ludwig, J. (2006). Broken windows: New evidence from New York City and a five-city social experiment. University of Chicago Law Review, 73(1), 271-320.

A claim that looks like arithmetic

Here is a paragraph from a student paper. Everything in it is a real number, correctly quoted.

The FBI reports a violent crime rate of 380.7 per 100,000 people. The Bureau of Justice Statistics reports 22.5 violent victimisations per 1,000 persons, which is 2,250 per 100,000. The survey therefore finds nearly six times as much violent crime as the FBI admits to, which shows that police records conceal most of the violence in the country.

The conclusion is wrong, and it is wrong in four separate places. Finding all four is the work of this lesson, and by the end you will be able to say exactly what each of the two national crime numbers measures and which questions each one can answer.

Error one: the two numbers do not count the same offences

The FBI figure comes from the Uniform Crime Reporting programme, in which roughly 18,000 law enforcement agencies submit counts to the FBI voluntarily. Its violent crime category has four offences: murder and non-negligent manslaughter, rape, robbery, and aggravated assault.

The survey figure comes from the National Crime Victimization Survey, which has run since 1973. Its violent crime category has four offences too, and they are not the same four: rape or sexual assault, robbery, aggravated assault, and simple assault. Simple assault, an attack without a weapon and without serious injury, is the most common of the four and it is not in the FBI's violent crime index at all. Meanwhile murder cannot be in the survey, because the survey works by interviewing victims.

Error two: the denominators are different populations

The FBI rate is per 100,000 residents of all ages. The survey rate is per 1,000 persons age 12 or older, which is a smaller population, so the same number of crimes produces a higher rate. The survey also excludes people living on military bases, in prisons, in hospitals and other institutions, and people who are homeless, which are populations with distinctive victimisation risks.

Converting 22.5 per 1,000 into 2,250 per 100,000 is arithmetically correct and substantively meaningless, because the 100,000 in the FBI figure and the 100,000 the student constructed are not the same 100,000 people.

What matters here: Before you compare two rates, write down the numerator definition, the denominator population, the unit and the year for each. If any of the four differ, the comparison is about the definitions and not about crime.

Error three: one measures police records, the other measures reported experience

This is the error with the most content in it, because the gap is real and it is measurable. In 2023 the survey found 22.5 violent victimisations per 1,000 persons age 12 or older, and 10.1 per 1,000 that were reported to police. Divide one by the other and you have the reporting rate: 44.7 percent of violent victimisations in 2023 were reported to police. The remainder is what criminologists call the dark figure of crime.

The reporting rate is not a constant. It differs sharply by offence, and it moves from year to year.

So the student's sentence about concealment has the mechanism backwards. Police records do not contain unreported crimes because nobody told the police about them. That is a property of victims' decisions, and of what those decisions respond to: whether the offender is known, whether insurance requires a report, whether the victim expects to be believed.

Error four: the FBI's own count changed shape in 2021

For ninety years the FBI collected monthly summary counts under the hierarchy rule: if several offences happened in one incident, only the most serious was recorded. A burglary that became an assault was counted once, as the assault. The Summary Reporting System was retired on 1 January 2021 in favour of the National Incident-Based Reporting System, which records every offence in an incident, uses 24 detailed offence categories instead of eight, and adds information about victims, offenders and circumstances.

The gain in detail came with a gap in coverage, because agencies had to rebuild their reporting systems. At the start of 2022 the FBI was receiving NIBRS data from 11,794 agencies covering 69 percent of the population, and two of the largest departments in the country, New York City and Los Angeles, were not yet submitting. By the last quarter of 2023 participation had reached 15,199 agencies and 82 percent of the population. During the gap the FBI filled in estimates, and it said plainly that confident statements about national trends were not available.

That has a consequence you should carry into any argument about recent crime. A change in a police-record series across 2021 and 2022 may be a change in crime, a change in which agencies reported, a change in how incidents are counted, or all three.

Doing the comparison properly

The two sources are not rivals. They answer different questions, and the correct move is to pick the one that matches your question.

On the long trend they do agree, which is why the decline since the early 1990s is treated as solid. The survey's violent victimisation rate fell from 79.8 per 1,000 in 1993 to 22.5 in 2023, and the reported-to-police component fell from 33.8 to 10.1 per 1,000 over the same period. The FBI series, collected in a completely different way from completely different informants, shows the same shape, with the violent crime rate near a fifty-year low in 2022 at 380.7 per 100,000 alongside a property crime rate of 1,954.4 and a murder rate of 6.3.

How a reporting change manufactures a crime wave

Look again at one row of the reporting table. Between 2022 and 2023 the share of rape and sexual assault victimisations reported to police rose from 21.4 percent to 46.0 percent. Now imagine that the true number of such offences had not changed at all. A measure built only on police records would show reports of that offence roughly doubling, and a newspaper reading only that series would report a doubling of rape.

This is not hypothetical reasoning about statistics. It is the standard trap in any argument about sexual assault, domestic violence or hate crime, where campaigns to encourage reporting, changes in police recording practice and legal redefinitions all move the recorded number without moving the underlying rate, and can also move the underlying rate in the opposite direction at the same time. Two independent measures are the only defence, which is exactly why the United States pays for both of these systems.

Bottom line: An administrative count measures the behaviour of institutions as well as the behaviour of offenders. A victimisation survey measures what people say happened to them, with sampling error and recall problems of its own. Neither is the truth, and knowing which distortions each one carries is what lets you use both.

The survey's own limits

Being fair to the FBI series means being equally hard on the survey. The NCVS interviews about 240,000 persons in about 150,000 households a year, with each household staying in the sample for three and a half years and being interviewed every six months. That design is expensive and good, and it still has four known weaknesses. It relies on memory over a six-month reference period. It depends on respondents being willing to tell an interviewer about an assault by a family member, sometimes with that person in the house. It cannot produce reliable estimates for small subgroups or small areas, because the sample in any one cell gets thin fast. And it excludes the institutionalised population entirely, which is a large exclusion when you are studying violence.

Common misconceptions

• "The FBI hides crime." The FBI compiles what agencies submit. Unreported crime is absent because victims did not report it, and 44.7 percent reporting for violent crime in 2023 is a measured fact rather than an accusation.
• "The two sources contradict each other, so crime statistics are useless." They measure different offences over different populations in different units. Where the questions align, as on the long trend, they agree closely.
• "A rise in recorded offences always means more crime." It can also mean more reporting, better recording, a legal redefinition, or more agencies submitting data. Rape reporting rising from 21.4 to 46.0 percent in one year is the clean illustration.
• "Survey figures are estimates, so record counts are more accurate." Record counts are also estimates of crime, and in 2021 and 2022 the FBI's national figures included estimation for agencies that had not yet moved to NIBRS.

Where this leaves us

• The FBI index counts murder, rape, robbery and aggravated assault per 100,000 residents; the survey counts rape or sexual assault, robbery, aggravated assault and simple assault per 1,000 persons age 12 or older.
• Simple assault is the largest component of the survey's violent crime and is not in the FBI index; murder is in the FBI index and cannot be in the survey.
• In 2023 the survey found 22.5 violent victimisations per 1,000 and 10.1 reported to police, which is a 44.7 percent reporting rate; property crime ran at 102.2 per 1,000 households, 13.6 million victimisations, with 30 percent reported.
• Reporting rates move: rape and sexual assault reporting went from 21.4 percent in 2022 to 46.0 percent in 2023, while robbery reporting fell from 64 to 42 percent.
• The FBI retired summary reporting on 1 January 2021; NIBRS drops the hierarchy rule, and coverage went from 69 percent of the population at the start of 2022 to 82 percent by late 2023.
• Both series show the same long decline: 79.8 per 1,000 in 1993 to 22.5 in 2023 in the survey, and a near fifty-year low in the FBI violent crime rate in 2022.
• Pick the source that matches the question, and treat agreement between an administrative count and an independent survey as the strongest available evidence.

Sources

1. Tapp, S. N., and Coen, E. J. (2024). Criminal victimization, 2023 (NCJ 309335). Bureau of Justice Statistics. bjs.ojp.gov
2. Bureau of Justice Statistics. (2025). National Crime Victimization Survey. bjs.ojp.gov
3. Federal Bureau of Investigation. (2025). Crime Data Explorer. cde.ucr.cjis.gov
4. Wikipedia. (2025). National Incident-Based Reporting System. wikipedia.org
5. OpenStax. (2021). Crime and the law. In Introduction to Sociology 3e. Rice University. openstax.org

$87,460 and the household in the middle

On 15 September 2026 the Census Bureau reported that the median American household had an income of $87,460 in 2025, the highest figure in a series that starts in 1967. The same release put the mean household income at $126,700. Both numbers are correct, both describe the same 137.1 million households, and they are $39,240 apart. By the end of this lesson you will be able to rebuild each of them from the published tables, say which question each one answers, and then do the same for wealth, where the mean is five and a half times the median.

The income figures come from the Current Population Survey, the monthly survey of about 60,000 households that also produces the unemployment rate. In February, March and April each year it adds the Annual Social and Economic Supplement, which asks the people in each sampled household about every source of income they had in the previous calendar year. Two definitions decide what the resulting numbers mean.

• A household is everyone living in one housing unit, related or not. A widow living alone is a household; so is a couple with three children.
• Money income is cash income before taxes: wages and salaries, self-employment earnings, Social Security, pensions, interest, dividends and cash assistance. It leaves out taxes paid, tax credits received, and help given in kind, such as food benefits.

Definitions change answers. In 2025 family households had a median income of $112,900, while people living alone or with non-relatives had $52,200. A country in which more people live alone will report a lower median household income even if nobody's pay changes, which is why a careful reader always asks: median of what?

Step one: line the households up and find the middle

The Bureau publishes the share of households in each income bracket. Add the shares as you go down, and the running total tells you how far along the line you have walked.

The median is the income of the household standing at the 50 percent mark. The running total reaches 43.3 at the top of the $50,000 bracket and 55.3 at the top of the next one, so the median lies somewhere between $75,000 and $99,999. To place it, assume the households are spread evenly across that bracket. You need 50.0 minus 43.3, which is 6.7 points, out of the bracket's 12.0 points. 6.7 divided by 12.0 is 0.56, so you go 56 percent of the way across a $25,000 range: $75,000 plus 0.56 times $25,000 is about $89,000. The Bureau runs the same interpolation inside much narrower $2,500 brackets and gets $87,460. Your rough answer lands within 2 percent of it, which is how you know the step worked.

The same walk finds any percentile. The Bureau's 2025 values: the 10th percentile was $20,010, the 20th $35,800, the 80th $182,400, the 90th $261,300 and the 95th $354,000. Divide the 90th by the 10th and you get 13.06: a household just inside the top tenth had about thirteen times the income of a household just inside the bottom tenth. In 1967, with incomes converted to 2025 dollars, the same ratio was 9.23.

Step two: set the mean beside the median

The mean is total income divided by the number of households, $126,700 in 2025. When the mean sits well above the median, a long tail of high values is pulling it upward, the shape statisticians call right skew.

You can watch the mechanism with five households, one standing in for each fifth of the country. The Bureau publishes the mean income inside each fifth: $19,100, $51,700, $88,250, $142,500 and $331,800. The median of the five is the middle one, $88,250. The mean is their total, $633,350, divided by five: $126,670, within a few dollars of the national figure. Now change one input. Give the top household $10 million instead of $331,800. The median does not move, because it only asks which household stands in the middle. The mean jumps to $2,060,310, sixteen times what it was. A statistic that one household can multiply by sixteen is a poor description of the typical case, though it remains an exact description of the total.

Key idea: The median answers the question "what does the household in the middle have?" The mean answers "what would each household have if the total were shared out equally?" When the two diverge, the size of the gap is itself a measurement of inequality.

Step three: cut the line into fifths and add up the shares

Divide the ranked households into five equal groups, called quintiles, and ask what share of all household income each group received. If incomes were equal, every fifth would receive 20 percent. Here are the Bureau's figures for the first and the latest year of the series.

Read the table in two directions. Down the 2025 column, the top fifth received more than twice as much as the bottom three fifths together: 52.4 percent against 3.0 plus 8.2 plus 13.9, which is 25.1 percent. Along the rows, every fifth had a higher real income in 2025 than in 1967, the bottom fifth by 52 percent and the top fifth by 136 percent, yet the bottom four fifths all hold smaller shares than they did. A falling share and a rising income are not a contradiction. The income columns tell you whether a group is better off than its predecessors were; the share columns tell you how the growth was divided.

One caution belongs with any long series. The survey has changed its methods several times, and the Bureau flags every change in its tables. The largest came with the income figures for 1993, when interviewers moved from paper questionnaires to computers and the highest earnings the survey would record rose to $999,999. The Gini index, which you are about to compute, jumped from 0.433 to 0.454 in that single year, and part of that jump is the new method rather than a change in American incomes.

Step four: draw the Lorenz curve and estimate the Gini index

Turn the shares into running totals. In 2025 the bottom 20 percent of households received 3.0 percent of income; the bottom 40 percent received 3.0 plus 8.2, which is 11.2; the bottom 60 percent, 25.1; the bottom 80 percent, 47.6; and everyone, 100. Plot those points with the share of households along the bottom and the share of income up the side, and you have a Lorenz curve. If incomes were equal, the curve would be the straight diagonal. The further it sags below the diagonal, the more unequal the distribution.

The Gini index measures the sag. It is the area between the diagonal and the curve, divided by the whole triangle under the diagonal, so it runs from 0, perfect equality, to 1, where one household has everything. Because that triangle has an area of one half, the Gini is 1 minus twice the area under the curve. You can estimate the area with five trapezoids, one for each fifth. Each trapezoid is 0.2 wide, and its area is 0.2 times the average of the running totals at its two edges.

So the estimate is 1 minus 2 times 0.2738, which is 0.45. Run the same steps on the 1967 shares and you get 0.37. The Bureau's published values, computed from every household in the sample rather than from five points, are 0.490 for 2025 and 0.397 for 1967. Your estimates are too low in both years, and the reason is worth knowing: a five-point curve treats every household inside a fifth as having the same income, so it misses the inequality within each fifth, above all within the top one, where the top 5 percent alone took 23.5 percent of all income. The direction of change comes out the same either way. The distribution was considerably more unequal in 2025 than in 1967.

The upshot: Every inequality statistic in the news is one of three objects: a percentile ratio, a share, or a summary of the Lorenz curve such as the Gini index. You can now say what each is made of, and a Lorenz curve is nothing more than a running total of shares.

Step five: wealth is a different ladder

Income is what comes in during a year. Net worth, which is what sociologists mean by wealth, is what a family owns on the day it is asked, minus what it owes: houses, savings, retirement accounts, shares and businesses, less mortgages, car loans, student debt and credit card balances. The best American source is the Federal Reserve's Survey of Consumer Finances, run every three years. Its 2022 round interviewed 4,602 families, and its design shows how hard wealth is to measure. An ordinary random sample of addresses would contain almost none of the families who own most of the country's businesses and shares, so the survey adds a second sample, drawn from tax records, that deliberately over-represents the wealthy; 1,304 of the 4,602 interviews came from that list. Weights then scale each group back to its true size in the population. The 400 people on the Forbes list of the richest Americans are excluded from sampling altogether.

Two things jump out. First, for all families the mean is five and a half times the median, $1,063,700 divided by $192,900, against less than one and a half for income, $126,700 divided by $87,460. Second, the bottom quarter's mean is negative. On average, the families in it owe more than they own.

The Federal Reserve also publishes the Distributional Financial Accounts, which take the national totals of household assets and debts every quarter and divide them among wealth groups, using the survey for the shape of the distribution. They report the share of all household wealth held by each group.

Set this beside the income table. The top fifth of households received about half of all income; the top tenth held more than two thirds of all wealth, 68.9 percent, and the bottom half of the country held 2.3 percent. The kinds of wealth differ as well. In 2026 the bottom half of households held 9.7 percent of the nation's residential real estate but 0.6 percent of its corporate shares and mutual funds, while the top 1 percent held 50.9 percent of those shares. For most families in the middle, wealth means a house. At the top, it means owning companies.

The second worked example: change the measure and the ranking flips

Here is a question with an apparently obvious answer: which age group is worst off? Rank households by the age of the person who owns or rents the home, and start with the Census income table. Then put the wealth figures beside it.

By income, households headed by someone 65 or older have the lowest median of any age group, lower even than households headed by people aged 15 to 24. Switch the input from income to wealth and the ranking reverses: families headed by someone aged 65 to 74 have the highest median net worth of any age group, more than ten times the median for families headed by someone under 35. Nothing about the people changed between the two columns. Only the measure did.

The reason is the difference between a flow and a stock. Income is a flow, measured over the last twelve months, and most people over 65 are no longer in paid work. Net worth is a stock, the accumulated result of decades of saving, paying down a mortgage and holding assets that gained value. Before you trust the reversal, check what else differs between the columns: different surveys, households against families, 2025 against 2022 dollars. None of those could produce a tenfold gap, so the reversal is real. The general procedure is the one you just ran. Before accepting a claim that some group is the poorest or the richest, ask which measure was used, and rerun the ranking with the other one.

What the three perspectives make of the same tables

Now the numbers can be argued over, and each perspective argues from a different row.

• Functionalist. In 1945 Kingsley Davis and Wilbert Moore argued that unequal rewards are how a society persuades its most able people to undertake the long training its most important positions require, the Davis-Moore thesis. The education rows of the Census table are its best exhibit: in 2025 households headed by someone with a bachelor's degree had a median income of $138,300, against $60,790 with a high school diploma and $37,090 with no diploma. Melvin Tumin replied in 1953 that stratification itself stops qualified people from reaching that training, because access to education depends on the family you are born into, so the system wastes talent lower down.
• Conflict. Marx asked who owns the means of production, and the Distributional Financial Accounts answer him directly: the top 1 percent hold about half of all corporate shares and the bottom half hold 0.6 percent. On this reading, returns to capital flow to a small group by construction, and the rising top shares since 1967 and 1989 are what the theory predicts.
• A Weberian addition. Weber argued that inequality has at least three dimensions, class, status and party, which need not line up; this is the three-component theory. The age table is his point in miniature: the same retired household stands low on one ladder and high on another.
• Symbolic interactionist. Class is also displayed and read. Thorstein Veblen named conspicuous consumption in 1899: spending that exists to be seen. Even "middle class" is a category you can watch being built. The Pew Research Center defines it as living in a household with two thirds to double the national median income, adjusted for household size. By that rule 61 percent of Americans were middle class in 1971 and 51 percent in 2023, while the share in upper-income households rose from 11 to 19 percent. Change the rule and the size of the middle class changes with it.

The perspectives do not simply describe the tables differently; they predict different things. If Davis and Moore are right, rewards should track training and the scarcity of skills. If the conflict theorists are right, the largest gains should go to owners of capital whatever their training. The wealth table is the conflict theorist's strongest exhibit and the education gradient is the functionalist's, which is why neither side has been able to close the argument with one table.

Common misconceptions

• "The average American household earns $126,700." That is the mean, pulled up by a long tail of high incomes. The household in the middle had $87,460, and nearly two thirds of households had less than the mean.
• "If the bottom fifth's share fell, the poor must have become poorer." Between 1967 and 2025 the bottom fifth's share fell from 4.0 to 3.0 percent while its mean real income rose from $12,590 to $19,100. Shares measure division, not level.
• "Income and wealth are two ways of measuring the same thing." Income is a yearly flow and wealth an accumulated stock. Households headed by people over 65 are at the bottom of one ranking and near the top of the other.
• "A country has one true Gini index." It depends on the income concept and the unit. For 2025 the Bureau reports 0.490 for pretax money income, 0.448 after taxes and credits, and 0.468 after adjusting for household size, all from the same survey.

Pulling it together

• The Census Bureau's median household income for 2025 was $87,460 and the mean $126,700, from the Current Population Survey's annual income supplement, covering 137.1 million households.
• A median is found by walking a running total to 50 percent and interpolating within the bracket; percentiles work the same way, and the ratio of the 90th to the 10th percentile was 13.06 in 2025 against 9.23 in 1967.
• A mean far above the median signals right skew, and one extreme household can multiply a mean while leaving the median untouched.
• Quintile shares in 2025 ran from 3.0 percent for the bottom fifth to 52.4 percent for the top fifth, against 4.0 and 43.6 in 1967, while real mean income rose in every fifth.
• Running totals of shares make a Lorenz curve; the Gini index is 1 minus twice the area under it, estimated from quintiles at 0.45 for 2025 and published from the full data at 0.490.
• Wealth is far more concentrated: median net worth $192,900 and mean $1,063,700 in 2022, and in 2026 the top 1 percent held 32.5 percent of household wealth and the bottom half 2.3 percent.
• Switching from income to wealth reverses the ranking of age groups, because income is a flow and wealth a stock.
• Functionalists point to the education gradient, conflict theorists to the ownership of capital, Weber to the several ladders, and interactionists to how class is displayed and defined.

Sources

1. Kollar, M., and Scherer, Z. (2026). Income in the United States: 2025 (Current Population Reports P60-289). U.S. Census Bureau. census.gov
2. Aladangady, A., Bricker, J., Chang, A. C., Goodman, S., Krimmel, J., Moore, K. B., Reber, S., Volz, A. H., and Windle, R. A. (2023). Changes in U.S. family finances from 2019 to 2022: Evidence from the Survey of Consumer Finances. Board of Governors of the Federal Reserve System. federalreserve.gov
3. Board of Governors of the Federal Reserve System. (2026). Distributional Financial Accounts. federalreserve.gov
4. Kochhar, R. (2024). The state of the American middle class. Pew Research Center. pewresearch.org
5. OpenStax. (2021). Theoretical perspectives on social stratification. In Introduction to Sociology 3e. Rice University. openstax.org

Three times the price of a cheap diet

In 1963 Mollie Orshansky, an economist at the Social Security Administration, made a calculation that still decides who is officially poor in the United States. She took the cheapest of the Department of Agriculture's food plans, the economy plan, priced it for families of different sizes, and multiplied the cost by three. The three came from the Agriculture Department's 1955 Household Food Consumption Survey, which had found that families of three or more people spent about a third of their after-tax money income on food. That finding described families at every income level, not only poor ones. Orshansky herself called the resulting thresholds "arbitrary, but not unreasonable".

The Office of Economic Opportunity, set up to run the War on Poverty, adopted the lower of her two sets of thresholds as its working definition in May 1965, and in August 1969 the Bureau of the Budget made a revised version the government's official statistical definition of poverty. Since then the thresholds have been raised every year by the rate of price inflation and by nothing else. In 2025 the official line for two adults and two children was $32,649, and the official poverty rate was 10.2 percent: 34.5 million people.

Since 2011 the Census Bureau and the Bureau of Labor Statistics have also published a second number, the Supplemental Poverty Measure, or SPM. For 2025 it gave 13.1 percent: 44.4 million people. Two federal agencies, one survey, one year, and almost ten million people apart. The way to understand the gap is to set the two measures side by side and read the comparison row by row, because every row records a decision about what poverty is.

The two measures side by side

Two rows drive most of the difference in the overall rate. The SPM line is higher, because modern spending on housing and other necessities has outrun a 1963 diet times three, and the SPM subtracts expenses the official measure never sees. Pulling the other way, the SPM counts benefits the official measure ignores. The Bureau is explicit that the SPM does not replace the official measure and is not designed to decide who qualifies for programs; it is a second instrument, built to answer a slightly different question.

What matters here: A poverty rate is the output of a set of definitions. Before you compare two rates, or the same rate over time, check that every row of this table is held constant, or you will be comparing decisions rather than people.

Reading the table: who gets poorer and who gets less poor

The overall gap of 2.9 points hides groups that move in opposite directions. The official figures below are the Bureau's official+ version, which adds unrelated children under 15 so that both columns cover exactly the same people.

Now match each movement to a row of the comparison table.

• People 65 and older have the largest gap, 5.6 points. The Bureau's explanation is medical spending: it is subtracted from resources in the SPM and ignored in the official measure. In 2025, medical expenses alone raised the SPM rate for people 65 and older by 3.8 points.
• People in cohabiting couples go the other way, from 21.7 to 13.1 percent. The official measure recognises only relatives, so if one unmarried partner earns $50,000 and the other earns nothing, the second is counted as having no income at all. The SPM treats partners as sharing.
• The West and the Midwest have identical official rates and very different SPM rates. The official line is the same in Los Angeles and in rural Iowa; the SPM line rises where housing costs more. The same row explains why people outside metropolitan areas look slightly less poor under the SPM, and why renters look poorer.
• Children come out at 13.4 percent on both measures in 2025. The Bureau notes that the official rate for children has usually run higher than the SPM rate, and the next table shows the likely reason: tax credits aimed at families with children count only in the SPM. A sharp fall in the official child rate in 2025 brought the two together.

What each program does to the count

Because the SPM counts benefits and expenses item by item, the Bureau can remove one item at a time and recalculate. The table shows how many percentage points each element moved the SPM rate in 2025. A negative number means the element lifts people out of poverty; a positive one means it pushes people in.

Read the Social Security row slowly. Social Security lifted 28.8 million people out of SPM poverty in 2025. Take it away and the SPM rate for people 65 and older would be 15.37 plus 32.13, or 47.5 percent, instead of 15.4. Set that beside the long official series, in which the poverty rate for people 65 and older fell from 35.2 percent in 1959 to 9.8 percent in 2025, and you can see why the program sits at the centre of every argument about old age and poverty. For children, the biggest single element is the tax system: refundable credits cut the child rate by 4.5 points.

These are mechanical calculations, and the Bureau says so: each one assumes nobody would behave differently if the program vanished. Critics of large transfers argue that some people would work or save more without them, so the true effect is smaller; defenders reply that some benefits, such as help with childcare or food, make work possible, so the true effect could be larger. The table cannot settle that argument. What it does settle is the size of the first-round effect, which is the number most policy arguments leave out.

The child series also contains a natural before-and-after. Children's SPM poverty rate was 12.5 percent in 2019 and 9.6 percent in 2020. In 2021, when the Child Tax Credit was temporarily enlarged, made fully refundable and partly paid in monthly instalments, it fell to 5.1 percent. In 2022, after the expansion lapsed and pandemic stimulus payments ended, it was 12.0 percent, and in 2025, 13.4. The Census Bureau attributed the 2022 increase to the expiry of the temporary Child Tax Credit and Earned Income Tax Credit expansions and the end of stimulus payments. Other things changed in those years too, which is why the attribution rests on the item-by-item accounting and not on the timing alone.

Even the yardstick gets corrected

In July 2026 the Bureau of Labor Statistics, which calculates the SPM thresholds, announced that it had found coding errors in the thresholds for 2019 to 2024. It corrected them and reset the line to 82 percent of median spending, and the Census Bureau revised every SPM poverty rate for those six years before publishing the 2025 report. The rates you have just read for 2019 to 2024 are the revised ones. Nothing about American families changed; the instrument did. This is ordinary, and it is exactly why a careful reader records which release a number came from.

Bottom line: The official measure asks whether a family's pretax cash clears a price-indexed 1960s line. The SPM asks whether a household's resources after taxes, benefits and unavoidable costs clear a line based on current spending where it lives. Neither is the true rate. Each is the right answer to its own question.

A different question: where do the children end up?

Poverty rates describe one year. Mobility describes movement: within one person's working life, which is intragenerational mobility, or between parents and their children, which is intergenerational mobility. Intergenerational mobility itself comes in two kinds that are constantly confused.

Absolute mobility: from 90 percent to 50

In 2017 Raj Chetty and five colleagues compared children's household incomes at age 30 with their parents' household incomes at the same age, adjusted for inflation. About 90 percent of children born in 1940 earned more than their parents had. For children born in the 1980s the figure was about 50 percent. Comparing sons only with fathers, it fell from 95 percent for sons born in 1940 to 41 percent for sons born in 1984. The decline happened in all 50 states and was steepest in the industrial Midwest, in states such as Michigan and Illinois.

The authors then asked why. Two things changed at once: the economy grew more slowly after the 1970s, and the growth that did occur went disproportionately to the top. When they recalculated with each factor held at its mid-century value, most of the decline came from the more unequal distribution of growth, not from the slowdown. On their numbers, restoring the 1940s rate with today's distribution would require real growth above 6 percent a year.

Relative mobility: reading a transition table

Relative mobility is about rank, and the clearest way to see it is a transition table. In 2014 Chetty, Nathaniel Hendren, Patrick Kline and Emmanuel Saez linked tax records for millions of children born in 1980 to 1982 to their parents' records. Each row below takes the children whose parents were in one fifth of the income distribution and shows where those children were as adults. Every row adds to about 100.

Start with the benchmark. If parents' income made no difference, every cell would be 20 percent. Now read the corners. A child born into the bottom fifth had a 7.5 percent chance of reaching the top fifth; a child born into the top fifth had a 36.5 percent chance of staying there, almost five times as high. At the same time the table is not a caste system: 38.2 percent of bottom-fifth children reached the middle fifth or above, and 39.8 percent of top-fifth children fell to the middle or below. The authors set the 7.5 percent beside comparable figures of 11.7 percent for Denmark and 13.4 percent for Canada.

The same national figure hides very different places. Measured by where children grew up, the chance of rising from the bottom fifth to the top was 4.4 percent in Charlotte, 10.8 percent in Salt Lake City and 12.9 percent in San Jose, nearly three times Charlotte's rate and comparable to Denmark's and Canada's. Areas with high upward mobility shared five features: less residential segregation, less income inequality, better primary schools, more social capital and more stable families. The authors state plainly that this descriptive analysis does not identify the causal mechanisms.

What each perspective makes of the five correlates

The list of five is where the theoretical argument starts, because each perspective reads it differently. A functionalist sees family stability and social capital as Durkheim's integration: places whose institutions hold together pass advantages on to every child. A conflict theorist sees segregation and inequality as the story: where the advantaged can separate their children's schools and neighbourhoods from everyone else's, they keep the top places for their own. A symbolic interactionist looks at what happens inside the schools and neighbourhoods themselves: whether teachers and employers read a child from a particular street as a future graduate or as trouble. The cross-country version of the same argument is the Great Gatsby curve, the pattern popularised in 2012 by the economist Alan Krueger in which more unequal countries tend to show less relative mobility; critics reply that the usual mobility statistic is partly built out of inequality, so some of the correlation is mechanical. Correlations of this kind cannot pick between the stories. A design that moves children between places at random, like the Moving to Opportunity experiment in Lesson 6, can.

Common misconceptions

• "The poverty line reflects what a family needs today." The official line is a 1963 diet times three, raised only for price inflation since. That is why the SPM line for renting families of four, $41,701, sits so far above the official $32,649.
• "Orshansky assumed poor families spend a third of their income on food." The one-third came from families at all income levels in 1955. Getting this wrong is, according to the historian of the thresholds Gordon Fisher, the commonest error in describing them.
• "The higher SPM rate is the real poverty rate and the official one is a cover-up." They answer different questions with different units, resources and lines. Some groups, such as people in cohabiting couples and people outside metropolitan areas, are less poor on the SPM.
• "If mobility were higher, every child could move up." That is true of absolute mobility, which depends on growth, and false of relative mobility, which is about rank: someone's rise is someone else's fall.

The short version

• Mollie Orshansky's 1963 thresholds price a minimal diet and multiply by three; they became official in 1969 and have been updated only for inflation since. The 2025 line for two adults and two children was $32,649 and the rate 10.2 percent.
• The SPM, published since 2011, bases its line on current spending, adjusts for housing costs and tenure, counts benefits and tax credits, subtracts taxes and necessary expenses, and pools unmarried partners. Its 2025 rate was 13.1 percent.
• People 65 and older are poorer on the SPM (15.4 against 9.8 percent) mainly because of medical spending; people in cohabiting couples are less poor (13.1 against 21.7) because partners' incomes are pooled.
• Social Security lowered the 2025 SPM rate by 8.5 points overall and 32.1 points for people 65 and older; refundable tax credits lowered the child rate by 4.5 points. These are static calculations that assume no change in behaviour.
• Measures are revised: in 2026 corrected thresholds changed every SPM rate for 2019 to 2024.
• Absolute mobility fell from about 90 percent for children born in 1940 to about 50 percent for those born in the 1980s, mostly because growth became less evenly shared.
• In relative terms, 7.5 percent of children from the bottom fifth reached the top fifth, against 36.5 percent of top-fifth children who stayed there, with large differences between cities and five correlates whose causes are still argued over.

Sources

1. Bijou, C., and Shrider, E. A. (2026). Poverty in the United States: 2025 (Current Population Reports P60-290). U.S. Census Bureau. census.gov
2. U.S. Census Bureau. (2023, September 12). Income, poverty and health insurance coverage in the United States: 2022 (Press release CB23-150). census.gov
3. Fisher, G. M. (1997). The development of the Orshansky poverty thresholds and their subsequent history as the official U.S. poverty measure. U.S. Census Bureau. census.gov
4. Chetty, R., Hendren, N., Kline, P., and Saez, E. (2014). Where is the land of opportunity? The geography of intergenerational mobility in the United States. Quarterly Journal of Economics, 129(4), 1553-1623. opportunityinsights.org
5. Chetty, R., Grusky, D., Hell, M., Hendren, N., Manduca, R., and Narang, J. (2017). The fading American dream: Trends in absolute income mobility since 1940. Science, 356(6336), 398-406. opportunityinsights.org

Eight words for a person in 1890

For the census of 1890, enumerators going door to door were told to write one of eight words beside each person's name: White, Black, Mulatto, Quadroon, Octoroon, Chinese, Japanese or Indian. Quadroon and Octoroon, meant to record a quarter and an eighth of what the instructions called "Black blood", were there because Congress required them. The enumerator decided which word applied, by looking and by asking. In 1930 the census added "Mexican" as a race; in 1940, after Mexican Americans lobbied against it with the support of the Mexican government, the category was dropped and people of Mexican descent were counted as White again. Only in 1960 did most Americans begin reporting their own race, and only in 2000 were they allowed to mark more than one.

If race were a fixed fact of biology, the list of races would not change by act of Congress, and a whole population would not change race between two censuses. That is what sociologists mean when they say race is socially constructed: the categories, their boundaries and the rules for placing people in them are made by societies, and they differ from one society and one decade to the next. Race refers to physical differences that a society treats as significant. Ethnicity refers to shared culture: language, religion, ancestry and custom. The census treats Hispanic or Latino origin as an ethnicity with its own question, so a Hispanic person can be of any race.

Constructed does not mean imaginary. In 1928 W. I. Thomas and Dorothy Swaine Thomas put the point in one sentence now called the Thomas theorem: situations that people define as real are real in their consequences. The 1930 census instructions applied the one-drop rule, recording anyone with any known Black ancestry as Negro, and rules of that kind decided where people could live, which schools would admit them and whom they could marry. Those decisions left traces that are still visible in the data, which is why this lesson has two halves: first, how the measuring moves the numbers; second, four competing explanations for the gaps the measuring reveals.

2020: the count that moved because the question did

For the 2020 census the Bureau kept its two separate questions, one on Hispanic origin and one on race, but redesigned them and improved the way it processed and coded what people wrote. The results changed sharply.

A reader who took the White alone row literally would conclude that nineteen million White Americans vanished in a decade. The Bureau's own analysts said the differences were largely due to the improved design of the questions and the coding of answers, along with some real demographic change. The White in combination population, people reporting White plus at least one other race, rose 316 percent, and Some Other Race alone or in combination grew to 49.9 million, mostly people of Hispanic origin. The same people were, in large part, being counted differently.

The questions are about to change again. In March 2024 the Office of Management and Budget revised the federal standards for race and ethnicity data to combine race and Hispanic origin into a single question and to add a Middle Eastern or North African category, and the Bureau had already reported that its research found a combined question would give a more accurate picture of how people identify. Every trend line that crosses one of these changes has a break in it, which is the first thing to check before believing a headline about a group shrinking or growing.

The point: A racial category is a decision about where to draw lines, and the count in each category depends on the question, the answer boxes and the coding rules. Treat any change across a redesign as partly a change in the instrument.

The measured gaps

The categories are constructed, and the gaps measured with them are large. These figures come from the Census Bureau's 2025 income and poverty reports and the Federal Reserve's 2022 Survey of Consumer Finances, which you worked with in Lessons 13 and 14. The survey groups are not defined identically: the Census columns classify households by the race of the householder, and the Federal Reserve classifies families by the race of the respondent.

Work two ratios. Median Black household income was 62 percent of the White non-Hispanic median: $59,980 divided by $96,710. Median Black family wealth was 16 percent of the White median: $44,900 divided by $285,000. The wealth gap is roughly two and a half times as wide as the income gap, and any explanation has to account for both numbers. Note also what the Asian column conceals: a single median for a category that includes families from dozens of countries, with very different histories of migration, hides large differences inside it.

Sociologists offer four main explanations for the Black and White gaps in particular. They are not mutually exclusive, and the real dispute is over how much each contributes. Each makes different predictions, and each has been tested.

Explanation one: discrimination now

The claim is that employers, landlords and lenders treat otherwise identical people differently today. The strongest evidence comes from field experiments, the audit design you met in Lesson 6, where Bertrand and Mullainathan's CVs with White-sounding names received 50 percent more callbacks than identical CVs with Black-sounding names. In 2017 Lincoln Quillian, Devah Pager, Ole Hexel and Arnfinn Midtbøen gathered every available American hiring field experiment on African American or Latino applicants: 28 studies, 55,842 applications, 26,326 positions. In the 24 studies conducted since 1989, White applicants received on average 36 percent more callbacks than African American applicants and 24 percent more than Latino applicants. They found no change in the level of discrimination against African Americans over 25 years and modest evidence of a decline for Latinos, and accounting for applicants' education and gender, the method, the occupation and local labour markets did little to alter the result.

The design is the explanation's great strength: race is assigned on paper by the researcher, so a gap in callbacks can only be discrimination. Its limit is scope. It measures one gate, the callback, at one moment. It cannot tell you how much of a $36,730 gap in median household income that gate produces, because pay, promotion, lending and housing are different gates.

Explanation two: history stored in wealth and in place

The claim is that discrimination that was legal and official for most of American history built up advantages and disadvantages that would persist even if all current discrimination stopped, mainly through wealth, which is inherited, and neighbourhoods, which shape schools and property values. The wealth ratio is this explanation's first exhibit: a gap in a stock passed between generations should be wider than a gap in yearly income, and at 16 cents against 62 it is.

The second exhibit is a test of a specific policy. In the 1930s the federal Home Owners' Loan Corporation drew maps grading urban neighbourhoods for mortgage risk, with the lowest grade outlined in red, the origin of the term redlining. Most residents of the red-graded areas were White, but those areas contained most of the country's urban Black households, and historians have argued over how much the maps themselves did. In 2021 Daniel Aaronson, Daniel Hartley and Bhashkar Mazumder compared neighbourhoods on either side of the boundary lines, which were otherwise similar, and cities just above and below the population cutoff that decided whether a city was mapped at all. Both comparisons pointed the same way: the maps led to lower rates of home ownership, lower house values and rents, and more racial segregation in later decades.

The limit is that history in wealth and place cannot explain everything. If it did, Black and White children growing up with the same parental income in the same neighbourhood should end up alike, and as the next explanation shows, the boys do not.

Explanation three: class, not race

In 1978 William Julius Wilson argued in The Declining Significance of Race that after the civil rights era, class position had become more important than race in shaping the life chances of Black Americans. The explanation makes a sharp prediction: compare Black and White children raised by parents with the same income, and the gap should shrink toward zero.

In 2020 Raj Chetty, Nathaniel Hendren, Maggie Jones and Sonya Porter tested it with tax and census records covering nearly the whole population from 1989 to 2015. The result split by sex. For women the prediction held: conditional on parental income, there was no gap between Black and White women's earnings. For men it failed: Black men raised at the same parental income as White men had lower wages and lower employment, and Black boys had lower adult incomes than White boys from families with the same income in 99 percent of Census tracts. Hispanic Americans, by contrast, were moving up the income distribution quickly across generations. So class accounts for much of the gap in some groups and very little of it for Black men.

Explanation four: culture and family

The strong cultural claim has two famous versions. In 1965 the Moynihan Report, written by the Assistant Secretary of Labor Daniel Patrick Moynihan, argued that the rise of single-mother families among poor Black Americans, which he traced back to slavery and discrimination, would keep the gap open by itself. In 1986 Signithia Fordham and John Ogbu proposed in The Urban Review that some Black students treat academic effort as acting white, an oppositional culture that depresses achievement.

The evidence against the strong forms is substantial. Chetty and colleagues found that differences in parental marital status, education and wealth explain very little of the gap between Black and White men once parental income is held equal. James Ainsworth-Darnell and Douglas Downey, analysing national survey data in the American Sociological Review in 1998, found that Black students reported more pro-school attitudes and more optimistic expectations about education than White students did, the opposite of what oppositional culture predicts.

The evidence for weaker forms is also real. The same Chetty study found that the places where the Black and White gap for boys was smallest were low-poverty neighbourhoods with low measured racial bias among White residents and high rates of fathers present among Black families, a pattern at the level of the neighbourhood rather than the household; fewer than 5 percent of Black children grew up in such places. And Roland Fryer and Paul Torelli, using the Add Health survey in the Journal of Public Economics in 2010, found a racial difference in how popularity relates to grades, concentrated among students with a grade point average of 3.5 or higher and strongest in schools with more interracial contact, though they called their estimates imprecise.

Worth holding on to: Each explanation survives in some form and fails in its strongest form somewhere. What decides the argument is not which story sounds right but which predictions hold when a design isolates them.

What would settle it

The table does not name a winner, and it does not need to. It shows that the strongest version of the class explanation fails for men, that the strongest version of the family explanation fails on the measures tested, that discrimination at the point of hire is measured and has not declined, and that one historical policy left effects that are still visible. What remains open is the share of the gap each mechanism accounts for, and that question needs designs that follow the same people through several gates at once.

The three perspectives line up behind different rows. Conflict theorists read the first two as a system that one group maintains to its own advantage. The functionalist tradition's closest relative is Robert E. Park's race relations cycle of contact, competition, accommodation and assimilation, which fits the rapid upward mobility of Hispanic Americans and does not fit the persistent gap for Black men. Symbolic interactionists supply the first half of this lesson: a category applied by an enumerator's glance, and later chosen on a form, is an identity made and remade in interaction.

Common misconceptions

• "If race is socially constructed, it has no real effects." The Thomas theorem runs the other way. Categories applied by law decided where families could buy houses, and the 1930s maps still predict home ownership and segregation decades later.
• "The White population shrank by 8.6 percent between 2010 and 2020." The Bureau attributed the change largely to redesigned questions and coding. The White in combination population rose 316 percent over the same decade.
• "Hispanic is a race on the census." It is an ethnicity asked in a separate question, and Hispanic respondents report every race. The 2024 standards will combine the two questions, which will create a new break in the series.
• "The income gap between groups is really a class gap." For women, conditional on parental income, it largely is; for men it is not. Black boys earned less than White boys from families with the same income in 99 percent of Census tracts.

What to remember

• Census race categories have been created and abolished by Congress and the Bureau: Quadroon and Octoroon in 1890, Mexican in 1930 only, self-reporting from 1960, more than one race from 2000, a combined question and a Middle Eastern or North African category under the 2024 standards.
• Redesigned questions and coding in 2020 moved the Two or More Races count from 9.0 to 33.8 million and the White alone count down 8.6 percent, largely through measurement.
• Measured gaps are large: median Black household income was 62 percent of the White non-Hispanic median in 2025, and median Black family wealth 16 percent of the White median in 2022.
• Discrimination now: 28 field experiments show 36 percent more callbacks for White than African American applicants since 1989, with no decline.
• History in wealth and place: boundary comparisons show the 1930s redlining maps lowered home ownership and values and raised segregation for decades.
• Class, not race: at equal parental income there is no gap between Black and White women, but a large one between Black and White men.
• Culture and family: family characteristics explain little of the gap for men and Black students report pro-school attitudes, while neighbourhood father presence and a peer penalty in integrated schools have some support.

Sources

1. Jones, N., Marks, R., Ramirez, R., and Ríos-Vargas, M. (2021, August 12). 2020 Census illuminates racial and ethnic composition of the country. U.S. Census Bureau. census.gov
2. U.S. Census Bureau. (n.d.). Measuring race and ethnicity across the decades: 1790-2010. census.gov
3. Quillian, L., Pager, D., Hexel, O., and Midtbøen, A. H. (2017). Meta-analysis of field experiments shows no change in racial discrimination in hiring over time. Proceedings of the National Academy of Sciences, 114(41), 10870-10875. pmc.ncbi.nlm.nih.gov
4. Chetty, R., Hendren, N., Jones, M. R., and Porter, S. R. (2020). Race and economic opportunity in the United States: An intergenerational perspective. Quarterly Journal of Economics, 135(2), 711-783. opportunityinsights.org
5. Aaronson, D., Hartley, D., and Mazumder, B. (2021). The effects of the 1930s HOLC "redlining" maps. American Economic Journal: Economic Policy, 13(4), 355-392. aeaweb.org

A paragraph with the right number and the wrong meaning

In January 2026 the Bureau of Labor Statistics published its annual report on women's earnings. Women who worked full time in wage and salary jobs had median weekly earnings of $1,043 in 2024; men had $1,261. Divide one by the other and you get 0.83: 83 cents on the dollar. Here is a paragraph built on that figure, of the kind you will meet in speeches, on posters and in student essays.

Women earn 83 cents for every dollar a man earns for doing the same job. The gap has not moved since the 1970s. Since women and men are doing the same work, the only possible explanation is that employers deliberately pay women less.

The one number in it is real, and nearly every inference built on it is wrong. This lesson traces the errors one at a time, then puts the opposite mistake, the claim that the gap disappears once you compare like with like, through the same checks. On the way you need the concepts sociologists use for gender and sexuality, because the pay gap is where most people first meet them.

First, the words: sex, gender and orientation

Sociologists separate terms that everyday speech runs together, following the distinctions set out in OpenStax's chapter on sex and gender. Sex refers to biological characteristics such as chromosomes, hormones and anatomy. Gender refers to the roles, behaviours and expectations a society attaches to being a woman or a man, which differ between societies and change over time. Gender identity is a person's own sense of their gender, and sexual orientation is a person's pattern of attraction.

The pay statistics sort workers by the sex recorded in the survey. Gender enters as the explanation: the expectations that steer people toward some occupations and away from others, the division of unpaid care at home, and the way employers read workers. In 1987 Candace West and Don Zimmerman gave the interactionist version of the idea its name, doing gender: gender is not simply something a person has but a routine accomplishment, produced in everyday interaction, in who speaks in a meeting, who is asked to take notes and who is expected to leave when a child is sick.

Error one: "for doing the same job"

The 83 percent compares the median of all women working full time with the median of all men working full time, whatever their jobs, hours, ages or experience. It is the raw, or unadjusted, gap, and the BLS says in the report itself that its comparisons do not control for skills, experience or the other factors that affect pay. The occupations are not the same:

• 1 percent of women working full time were in natural resources, construction and maintenance occupations, against 16 percent of men; 16 percent of women were in office and administrative support, against 5 percent of men.
• Among professionals, 12 percent of women worked in the well-paid computer, mathematical, architecture and engineering occupations, against 50 percent of men; 65 percent of professional women worked in education and healthcare, against 28 percent of men.
• The three commonest jobs for women were elementary and middle school teacher ($1,226 a week), registered nurse ($1,476) and customer service representative ($836). The commonest job for men, by far, was truck driver ($1,062).

Hours differ too, even among full-time workers: 21 percent of men and 12 percent of women usually worked more than 40 hours a week. Change what you hold constant and the ratio changes with it.

There is no single gender pay gap. There is a family of ratios, each answering a different question, and the last two rows show that even the choice between a weekly and an annual measure moves the figure: for 2024 the BLS weekly ratio was 83 percent and the Census annual ratio 80.6.

Why this matters: Before quoting any pay gap, name its numerator, its denominator, the time unit and what is held equal. A raw gap and an adjusted gap answer different questions, and the error in "for doing the same job" is to present the first as if it were the second.

Error two: "the gap has not moved since the 1970s"

In 1979, the first year of comparable BLS data, women working full time earned 62 percent of men's median weekly earnings. The ratio rose from 64 to 70 percent during the 1980s and from 72 to 81 percent between 1990 and 2005. Since 2010 it has stayed between 81 and 84 percent. So the true history has two chapters: a large convergence, then a plateau. On the Census annual measure, the 2025 ratio of 83.9 percent was the first statistically significant yearly increase since 2016, after two years of decline. Whether that is the start of a third chapter is a question one year of data cannot answer.

Error three: "the only possible explanation is deliberate underpayment"

To separate explanations, economists and sociologists use a decomposition. The question it asks is: how much of the gap would disappear if women had the same measured characteristics as men, rewarded at the rates men receive? Francine Blau and Lawrence Kahn ran one on the Panel Study of Income Dynamics, for full-time wage and salary workers aged 25 to 64. In 2010 women's hourly wages were 79.3 percent of men's before any adjustment. Here is how the gap divided.

Measured differences, above all in occupation and industry, account for 62 percent of the gap, so "the only possible explanation" is wrong. Once those characteristics are held equal, the ratio rises from 79.3 to 91.6 percent. But the remaining 38 percent does not measure deliberate underpayment either, and here the paragraph makes a subtler mistake. The unexplained share contains whatever the model did not measure: discrimination, certainly, but also differences in the flexibility of jobs, in hours within a job title, and anything else left out. And discrimination can hide inside the explained part, if women are steered away from engineering or if occupations pay less because women fill them. Blau and Kahn's reading of the experimental research is that discrimination cannot be discounted; their reading of the decomposition is that it cannot, on its own, tell you how large discrimination is.

The opposite error: "compare like with like and the gap disappears"

Now run the rival paragraph through the same checks: once you compare women and men with the same occupation, hours and experience, it says, the gap vanishes and is simply the sum of choices. It fails on two counts.

First, it does not vanish. After everything Blau and Kahn could measure was held equal, women still earned 91.6 percent of what men earned, an 8.4-point gap. Second, holding occupation equal assumes that which occupation a person enters has nothing to do with gender, and that is exactly what is in question. If the expectations girls meet at school steer them away from engineering, then controlling for occupation removes part of the effect of gender and calls it choice.

The strongest evidence about what the choices are comes from what happens when children arrive. Henrik Kleven and colleagues followed women's and men's earnings year by year around the birth of a first child in six countries. Before the birth, women's and men's earnings paths moved together. After it, women's fell and men's did not. The long-run child penalty, the average shortfall of women relative to men five to ten years after the first birth, was:

The penalty exists everywhere and its size varies threefold, which is itself evidence that it is not biology alone: the authors point to differences in parental leave, childcare and gender norms. The BLS age pattern is consistent with the same story, since the ratio is 88 to 91 percent below age 35 and 77 to 81 percent above it, although a cross-section like that also mixes in differences between generations. Claudia Goldin, awarded the 2023 Nobel prize in economics for her research on women's work, puts the timing in one phrase: the remaining gap largely arises with the birth of the first child. Fathers, by contrast, earned more than men without children, $1,469 a week against $1,179 in the 2024 BLS data.

In short: The raw gap is not a same-job gap, the measured factors explain much but not all of it, the unexplained remainder is not a pure measure of discrimination, and the child penalty shows that "choice" happens under constraints that differ sharply between countries.

What the three perspectives say about the same numbers

The three perspectives read the child penalty differently. A functionalist treats a division of labour between paid work and care as a household arrangement that once fitted the demands of pregnancy and nursing and has persisted by being passed on. A conflict theorist, in a line running back to Friedrich Engels, sees the household reproducing the relation between owner and worker, with women's unpaid labour and earnings dependence as the mechanism. An interactionist looks at doing gender: the daily negotiations over who leaves work early, and the way a manager reads a mother's request for flexible hours. The six-country table is a test between them. If the penalty were a fixed functional necessity, it should not vary threefold between rich democracies with similar technologies.

Sexuality: what the surveys can and cannot see

Sexual orientation can be measured three ways: by the identity people claim, by their behaviour, or by whom they are attracted to, and the three do not give the same numbers. The best-known American series measures identity. In 2025 Gallup's telephone surveys of 13,454 adults found 9.0 percent identifying as lesbian, gay, bisexual, transgender or something other than heterosexual, up from 3.5 percent in 2012, the first year Gallup asked. Bisexual was the most common identity, 5.3 percent of all adults. Identification fell steeply with age: 23 percent of adults aged 18 to 29, 10.4 percent of those 30 to 49, 3.1 percent of those 50 to 64 and 2.3 percent of those 65 and older. It was 10.5 percent among women and 5.6 percent among men, and 5 percent of respondents declined to answer.

Read the trend the way this course reads every trend. The survey records the labels people choose to report to an interviewer. A rise from 3.5 to 9.0 percent is consistent with more people having these orientations, with more people willing to adopt and report the labels as legal and social acceptance grew, including after the Supreme Court's 2015 decision in Obergefell v. Hodges, or with both, and a survey that asks only about identity cannot tell these apart. The steep age gradient is equally ambiguous in a single year: it could reflect generations that will keep their rates as they age, which is Gallup's interpretation, or differences in how people report at different ages. The interactionist point is that a label such as bisexual is adopted and given meaning in interaction, which is why its share can change faster than any underlying biology.

Common misconceptions

• "83 cents means a woman is paid 17 percent less for the same job." It compares the medians of all full-time women and all full-time men across different occupations, hours and ages. The adjusted gap in Blau and Kahn's data was about 8 points, not 17 to 21.
• "The unexplained part of a decomposition is the amount of discrimination." It is whatever the model did not measure, which includes discrimination and much else, while discrimination can also sit inside the explained part.
• "The gap is the same at every age." In 2024 women earned 91 percent of men's earnings at 16 to 24 and 77 to 81 percent at 35 and older, and the event-study evidence locates the widening at the first birth.
• "Rising LGBTQ+ identification proves more people are gay or bisexual than before." An identity survey cannot separate a change in orientation from a change in willingness to report it.

Looking back

• In 2024 women working full time had median weekly earnings of $1,043 and men $1,261, a ratio of 83 percent; it was 62 percent in 1979 and has stayed between 81 and 84 percent since 2010.
• The raw ratio compares all full-time women with all full-time men; holding hours at 40 gives 87 percent, hourly workers 91 percent, and the Census annual measure 80.6 percent for 2024 and 83.9 for 2025.
• Occupations differ sharply: 1 percent of women against 16 percent of men in construction and maintenance, and 12 percent of professional women against 50 percent of men in computing and engineering.
• Blau and Kahn's 2010 decomposition attributes 32.9 percent of the gap to occupation, 17.6 to industry and 14.1 to experience, with 38 percent unexplained; the adjusted ratio is 91.6 percent.
• The unexplained share is not a direct measure of discrimination, and controlling for occupation can hide the effect of gender on which occupation a person enters.
• Long-run child penalties range from 21 percent in Denmark to 61 percent in Germany, with 31 percent in the United States, and the gap largely opens with the first child.
• Gallup found 9.0 percent of adults identifying as LGBTQ+ in 2025, up from 3.5 percent in 2012, with 23 percent among adults under 30; an identity survey cannot tell a change in orientation from a change in reporting.

Sources

1. U.S. Bureau of Labor Statistics. (2026). Highlights of women's earnings in 2024 (BLS Reports, Report 1117). bls.gov
2. Kollar, M., and Scherer, Z. (2026). Income in the United States: 2025 (Current Population Reports P60-289). U.S. Census Bureau. census.gov
3. Blau, F. D., and Kahn, L. M. (2017). The gender wage gap: Extent, trends, and explanations. Journal of Economic Literature, 55(3), 789-865. nber.org
4. Kleven, H., Landais, C., Posch, J., Steinhauer, A., and Zweimüller, J. (2019). Child penalties across countries: Evidence and explanations. AEA Papers and Proceedings, 109, 122-126. aeaweb.org
5. Jones, J. M. (2026, February 16). LGBTQ+ identification holds at 9% in U.S. Gallup. news.gallup.com

Early on 11 July 1958

In the early hours of 11 July 1958, local police entered the house in Central Point, Caroline County, Virginia, where Richard and Mildred Loving were asleep. They had married five weeks earlier in Washington, D.C. When the officers found them in bed, Mildred pointed to the marriage certificate on the bedroom wall. The officers told her it was not valid in Virginia.

Richard was classified as white and Mildred as colored, and Virginia law, built on its Racial Integrity Act of 1924, made it a crime for such a couple to marry out of state and come home, and treated the marriage itself as a felony. On 6 January 1959 the Lovings pleaded guilty. They were sentenced to a year in prison, suspended on condition that they leave Virginia and not return together for 25 years. They moved to Washington, and their case took eight years to reach the Supreme Court.

Follow this one family and most of the concepts sociologists use for the family arrive on their own: what an institution is, who decides what counts as a family, how the shape of family life changed in the half-century after their case, and why the three perspectives read those changes so differently.

What the state was enforcing

A social institution is an established pattern of norms, roles and organisations built around a basic need of a society, and marriage is one of the oldest. The Lovings' case shows its two faces at once. To the couple, the marriage was a personal commitment made in a Washington courthouse. To Virginia, it was a legal status that the state granted or refused, and the rules about who could marry whom were enforced by the police at four in the morning.

Every society has such rules. Endogamy is marriage within one's own group, and exogamy is marriage outside it; societies differ in which boundaries they enforce and how. Virginia enforced a racial boundary with criminal law. Sociologists also distinguish, as OpenStax's chapter on marriage and the family sets out, between a nuclear family of parents and their children and an extended family that includes grandparents, aunts, uncles or cousins.

The Census Bureau, whose tables supply most of the numbers in this lesson, uses two definitions that you met in Lesson 13. A household is everyone living in one housing unit. A family is two or more people related by birth, marriage or adoption who live together. A grandmother raising her grandson is a family; two unrelated flatmates are a household but not a family. The Lovings, once married and living together, were a family in the Census's eyes and a crime in Virginia's.

The family pattern of the Lovings' generation

What did family life look like in the years when the Lovings married? The Census Bureau's historical tables give the picture. In 1958 the median age at first marriage was 22.6 for men and 20.2 for women. Married couples headed 75 percent of all households. In the 1960 census 87.7 percent of children lived with two parents, the average household had 3.33 people, and only 13.1 percent of households were a person living alone. In 1967, 82.7 percent of women aged 25 to 34 lived with a husband.

It is tempting to treat this as the traditional family against which everything since is a departure. The long series says otherwise. In 1890 the median age at first marriage was 26.1 for men and 22.0 for women, and it fell to its lowest point, 22.5 and 20.1, in 1956. The early marriages of the Lovings' generation were the unusual chapter in the record, not its baseline.

12 June 1967

On 12 June 1967 the Supreme Court decided Loving v. Virginia unanimously. Chief Justice Earl Warren's opinion struck down Virginia's law on two grounds: it classified people by race in violation of the Equal Protection Clause, and it deprived them of the freedom to marry, which the Court called a fundamental right. Marriage across racial lines became legal throughout the country.

In 1967, 3 percent of American newlyweds married someone of a different race or ethnicity. By 2015 it was 17 percent, according to a Pew Research Center analysis of Census data, and the increase among Black newlyweds, from 5 percent in 1980 to 18 percent, and white newlyweds, from 4 to 11 percent, was a large part of it. Attitudes moved faster still. In the General Social Survey, 63 percent of non-Black adults said in 1990 that they would be very or somewhat opposed to a close relative marrying a Black person; in the survey Pew analysed for its 2017 report, 14 percent did. A law, a behaviour and an attitude changed together, and it is not possible to say from these numbers alone which moved which.

The half-century after the decision

Richard Loving was killed on 29 June 1975, at 41, when a drunk driver ran an intersection and struck the couple's car; Mildred lost an eye in the crash. She lived until 2 May 2008. The table uses those years, with the start and the present, to show what happened to the American family around their story.

Three stories are in those rows. The first is later marriage: the typical woman now marries eight years later than her 1960 counterpart. The second is the rise of living alone: in 2025, 39.7 million households, 29 percent, were a single person, and married couples, 47 percent of households, were no longer a majority. The third is the story most often told wrongly. The share of children living with two parents fell steeply from the 1960s to about 2000, when it reached 69.1 percent, and has barely moved since: 69.9 percent in 2008, 70.4 percent in 2025. A claim that the two-parent family is still in freefall does not match the table.

Marriage has also stopped being the only form a couple takes. Among adults aged 25 to 34, living with an unmarried partner rose from 0.2 percent of women in 1967 to 16.4 percent in 2023, and 18.8 percent of men that age were living in a parent's home as the householder's child, against 9.1 percent in 1967. The number of opposite-sex unmarried couples counted by the survey rose from about 439,000 in the 1960 census to 9.5 million in 2023, although the early figures come from an indirect method, later replaced by direct questions, so the size of the rise is less certain than its direction.

Remember: The phrase "the American family" describes several separate trends. Some have continued for sixty years, such as later marriage and more people living alone; one, the fall in two-parent living for children, stopped around 2000. Check which trend a claim is about before you accept its direction.

Same-sex marriage, and a census that followed

On 12 June 2007, the fortieth anniversary of her case, Mildred Loving issued a rare public statement supporting the right of same-sex couples to marry. In 2015 the Supreme Court's decision in Obergefell v. Hodges made same-sex marriage legal nationwide, relying in part on Loving. The statistics then had to follow the law. The Census Bureau's marriage estimates include same-sex married couples from 2019, and in the same year the survey switched to gender-neutral codes for identifying a child's parents, so that a child with two mothers or two fathers is counted as living with two parents. Every one of those changes is a footnote on the tables above, and every one reflects a decision about what a family is.

Three readings of the same trends

The perspectives do not disagree about the numbers. They disagree about what the numbers mean.

• Functionalist. The family performs functions a society needs: reproduction, the socialisation of children, economic cooperation and emotional support. Later marriage and more solitary living suggest that some of those functions have moved to other institutions, such as schools, employers and paid services, and a functionalist asks whether the functions that remain, above all raising children, are still being done well. The stability of two-parent living since 2000 matters to that question.
• Conflict and feminist. The family is also a site of unequal work. In the Bureau of Labor Statistics' 2025 American Time Use Survey, 87 percent of women and 75 percent of men did some household activity on an average day, and on those days women spent 2.8 hours on it and men 2.1; men had 5.6 hours of leisure a day and women 4.8. On this reading, part of what the tables record is women leaving arrangements that cost them more than they cost men.
• Symbolic interactionist. A family is also a shared definition. The Lovings defined themselves as married and hung the certificate where they would see it every day; the state defined them as criminals. Interactionists study how people build and defend such definitions, whether in a couple deciding that living together is a family or a census deciding how to code a child's two mothers.

The perspectives can be set against evidence. If the functionalist worry is right, children's outcomes should track the living arrangements in the table; if the conflict reading is right, the division of unpaid work should predict who leaves marriages and when. Both are testable with the kinds of data you have now seen, which is why the argument continues in journals rather than being settled by opinion.

Common misconceptions

• "The 1950s family was the historical norm." Median ages at first marriage in 1956, 22.5 and 20.1, were the lowest in a series that starts at 26.1 and 22.0 in 1890. The Lovings' generation married unusually young.
• "The two-parent family is still collapsing." The share of children living with two parents fell from 85.4 percent in 1968 to 69.1 percent in 2000 and was 70.4 percent in 2025.
• "A household and a family are the same thing." A person living alone is a household but not a family; 36 percent of households in 2025 were not families.
• "Marriage law is a private matter the state has always left alone." Until 1967 many states banned marriage across racial lines, and until 2015 many states refused marriage to same-sex couples.

Summing up

• Marriage is a social institution that the state defines: Virginia enforced racial endogamy by criminal law until Loving v. Virginia struck it down on 12 June 1967.
• In Census usage, a household is everyone in a housing unit and a family is two or more related people living together.
• Newlyweds marrying someone of a different race or ethnicity rose from 3 percent in 1967 to 17 percent in 2015, and General Social Survey opposition to a relative marrying a Black person fell from 63 to 14 percent.
• Median age at first marriage rose from 22.8 and 20.3 in 1960 to 30.8 and 28.4 in 2025; married couples fell from 74 to 47 percent of households, and one-person households rose from 13 to 29 percent.
• Children living with two parents fell from 87.7 percent in 1960 to about 69 percent in 2000 and has hardly changed since.
• Measurement follows law: same-sex married couples enter the marriage estimates from 2019, with gender-neutral parent codes.
• Functionalists ask whether the family's remaining functions are done well, conflict theorists point to unequal unpaid work, 2.8 hours against 2.1, and interactionists study how families define themselves.

Sources

1. U.S. Census Bureau. (2025, December 2). Census Bureau releases new estimates on America's families and living arrangements (Press release CB25-TPS.78). census.gov
2. U.S. Census Bureau. (2025). Historical households tables (HH-1, HH-4), with the historical marital status table MS-2 and the living arrangements tables CH-1 and AD-3. census.gov
3. Livingston, G., and Brown, A. (2017). Intermarriage in the U.S. 50 years after Loving v. Virginia. Pew Research Center. pewresearch.org
4. Loving v. Virginia, 388 U.S. 1 (1967). courtlistener.com
5. U.S. Bureau of Labor Statistics. (2026, June 25). American Time Use Survey: 2025 results (USDL-26-1022). bls.gov

$31,918 and $11,060

On 7 May 2026 the Census Bureau reported what American public schools spent per pupil in fiscal year 2024. The national figure was $17,619, up 6.6 percent in a year. New York spent the most, $31,918 a pupil, and within it New York City spent $35,796. Idaho spent the least, $11,060, with Utah close behind at $11,347. The figures are not adjusted for differences in prices between places, but even a large adjustment for cost of living would not close a gap of nearly three to one.

The gap exists because of where the money comes from. In 2024 state governments supplied 45.2 percent of school revenue, the federal government 11.6 percent, and local sources 43.2 percent. Of the local money, 63.3 percent came from property taxes, which means that part of what a school can spend depends on the value of the houses and businesses in its district.

In 1968 parents in Edgewood, a poor district in San Antonio, sued Texas over exactly this. In 1973 the Supreme Court decided San Antonio v. Rodriguez against them by five votes to four: relying on local property taxes did not violate the Equal Protection Clause, and education was not a fundamental right under the federal Constitution. The fight moved to state courts, which over the following decades ordered many states to change how they funded schools. Those orders matter to this lesson, because they became the best evidence on the question underneath the whole argument: does the money change anything for the students?

The first answer: Coleman, 1966

The Civil Rights Act of 1964 required a survey of educational opportunity, and the sociologist James Coleman led it. In September and October 1965 his team gathered questionnaires and test results from 4,000 public schools and more than 645,000 pupils in grades 1, 3, 6, 9 and 12. The report, Equality of Educational Opportunity, is still one of the largest social surveys ever run, and its central finding surprised almost everyone. Variations in schools' facilities and curriculums accounted for relatively little of the variation in pupils' test scores. Family background mattered more, and so did the social composition of the student body: children from similar homes did better in schools with more advantaged classmates.

Two details of the report are often forgotten. School differences mattered more for minority pupils than for white ones: about 20 percent of the achievement of Black pupils in the South was associated with the particular school they attended, against 10 percent for white pupils there. And the report found that the schools attended by minority pupils were less well equipped in several respects, such as science laboratories, which showed a small but consistent relationship with achievement.

The headline became a slogan, money does not matter, and the economist Eric Hanushek's reviews of the research in the 1980s reinforced it: he reported no consistent relationship between school resources and student outcomes and argued that the problem was how money was spent, not how much. Critics reanalysed the same studies and reached the opposite conclusion. The argument continued for decades because the studies on both sides shared a weakness.

The second answer: follow the court orders

In 2016 C. Kirabo Jackson, Rucker Johnson and Claudia Persico took a different route. Court-ordered finance reforms changed how much was spent on a child for reasons that had nothing to do with that child's family: two children born in the same district a few years apart could receive quite different spending because a court ruled in between. That is the logic of the natural experiment you met in Lesson 6. The authors linked the timing and type of reforms to data on children born between 1955 and 1985 who were followed into adulthood through 2011.

Their estimate: a 10 percent increase in per-pupil spending in every one of a child's twelve years of school led to 0.27 more completed years of education, 7.25 percent higher adult wages, and a 3.67 percentage-point fall in the yearly chance of being in poverty as an adult, with much larger effects for children from low-income families. The extra money had bought smaller classes, higher teacher salaries and longer school years.

Why the two answers differ

The main threat in the Coleman column is the confounding problem from Lesson 6 in a new form. If districts with more disadvantaged pupils receive extra compensatory money, a snapshot shows higher spending alongside lower scores, and the true effect of money is hidden. The court-order design breaks that link. It also measures outcomes that tests taken in the same year cannot see. Read carefully, the two studies are less opposed than the slogans: Coleman's own finding that school quality mattered most for the most disadvantaged pupils is what Jackson and his colleagues found forty years later.

So what?: The money question was not settled by more data of the same kind. It moved when researchers found a design in which spending changed for reasons unrelated to the students. When you meet a claim that money does or does not matter, ask which design produced it.

The tracking dispute

Tracking means grouping students into separate classes or course sequences by measured achievement: honors and standard sections, or separate paths through mathematics. It is common in American high schools, and it is argued over in terms that map neatly onto the three perspectives.

• The functionalist case for it. Schools sort students for the positions they will later fill, and a teacher facing a class of similar achievement can pitch the lesson at the right level for everyone in it.
• The conflict case against it. In Keeping Track: How Schools Structure Inequality (Yale University Press, 1985), Jeannie Oakes reported that high-track teachers taught material demanding critical thinking while low-track classes drew heavily on workbooks, and that poor and minority students were placed in low tracks out of proportion to their measured ability. On this reading tracking does not sort by ability so much as convert class background into different educations.
• The interactionist mechanism. A track is also a label. The self-fulfilling prophecy you met in Lesson 3 predicts that a student placed in a low track will be taught, and will come to see themselves, as a low achiever.

The strongest single piece of evidence comes from outside the United States. In an experiment published in 2011, Esther Duflo, Pascaline Dupas and Michael Kremer worked with 121 primary schools in Kenya that each received money for an extra teacher, allowing the first grade to be split into two classes. In 61 schools pupils were assigned to the two classes at random; in 60 they were assigned by initial achievement. After 18 months, pupils in the tracking schools scored 0.14 standard deviations higher, the gain persisted a year after the program ended, and pupils at every level of initial achievement benefited, including those in the lower section. Pupils near the middle did about equally well whether they were put in the upper or the lower class, which is not what a strong labelling effect would predict.

That is a randomised result in favour of tracking, and it deserves weight. It also has limits that matter for an American reader. It concerns first graders following the same curriculum in both sections, whereas American high school tracks often decide which courses a student may take at all, which is precisely the mechanism Oakes described. Oakes's evidence, in turn, is observational: it shows what low tracks looked like and who was in them, but not what the same students would have learned without tracking. A study that could settle the American question would randomly assign schools to tracked and untracked systems, keep the content of the low track visible, and follow students into course-taking and college, and no study of that scale has been done.

What the outcome data show now

Both arguments matter because of where American results stand. The 2024 Nation's Report Card, the National Assessment of Educational Progress, found national scores below their 2019 levels in every grade and subject tested. A third of 8th graders, 33 percent, scored below the Basic level in reading, the highest share ever recorded. And the gap between higher and lower performers kept widening, a trend the National Assessment Governing Board says has run for more than a decade: in 4th grade mathematics the national gain of 2 points came from middle and higher performers while the lowest performers stayed flat, and in 8th grade mathematics a flat national average concealed gains at the top and declines at the bottom. Any policy about money or tracking now has to answer for what it does to the students at the bottom of that distribution.

Key idea: Both school disputes turn on the same methodological point. A comparison of existing schools cannot separate what a school does from who attends it; a design that changes the school, by court order or by lottery, can.

Common misconceptions

• "The Coleman Report proved that schools don't matter." It found that facilities and curriculums explained little variation in test scores in a cross-section, that schools mattered more for minority pupils, and that classmates mattered. It did not show that changing a school's resources would change nothing.
• "More money always improves results." The best evidence shows large long-run effects when money was spent on things such as smaller classes and teacher pay; how money is spent was Hanushek's point, and the finance-reform studies do not refute it.
• "Research has shown that tracking hurts low achievers." The best randomised evidence, from Kenya, found that tracking helped pupils at every level. The American concern is about tracks that change course content, which that experiment did not test.
• "Spending per pupil is a measure of school quality." It varies with prices, salaries, student needs and what the money buys; New York's $31,918 and Idaho's $11,060 are not directly comparable measures of quality.

The takeaway

• Public schools spent $17,619 per pupil in fiscal 2024, from $11,060 in Idaho to $31,918 in New York; states supplied 45.2 percent of revenue, local sources 43.2 percent, mostly property tax, and the federal government 11.6 percent.
• San Antonio v. Rodriguez (1973) upheld property-tax funding by five to four, sending finance cases to state courts.
• The Coleman Report surveyed more than 645,000 pupils in 4,000 schools in 1965 and found that facilities and curriculums explained relatively little variation in test scores, while family background and classmates mattered more.
• Using court-ordered reforms, Jackson, Johnson and Persico found that 10 percent more spending for twelve years raised schooling by 0.27 years and wages by 7.25 percent, with larger effects for low-income children.
• The two studies differ in design and outcome; the cross-section is confounded by compensatory spending, which the natural experiment avoids.
• Oakes documented thinner teaching in low tracks and unequal placement; a Kenyan randomised trial found that tracking raised scores by 0.14 standard deviations at every level. Neither settles tracks that decide course content in American high schools.
• The 2024 NAEP found scores below 2019 everywhere, 33 percent of 8th graders below Basic in reading, and a widening gap between higher and lower performers.

Sources

1. U.S. Census Bureau. (2026, May 7). Public school spending per pupil reaches historic high in 2024. census.gov
2. Coleman, J. S., Campbell, E. Q., Hobson, C. J., McPartland, J., Mood, A. M., Weinfeld, F. D., and York, R. L. (1966). Equality of educational opportunity. U.S. Department of Health, Education, and Welfare, Office of Education. files.eric.ed.gov
3. Jackson, C. K., Johnson, R. C., and Persico, C. (2016). The effects of school spending on educational and economic outcomes: Evidence from school finance reforms. Quarterly Journal of Economics, 131(1), 157-218. nber.org
4. Duflo, E., Dupas, P., and Kremer, M. (2011). Peer effects, teacher incentives, and the impact of tracking: Evidence from a randomized evaluation in Kenya. American Economic Review, 101(5), 1739-1774. nber.org
5. National Assessment Governing Board. (2025). 10 takeaways from the 2024 NAEP results. nagb.gov

36,908 interviews and one number

Between 17 July 2023 and 4 March 2024, the Pew Research Center interviewed 36,908 American adults, in English and Spanish, for its third Religious Landscape Study. The headline was a single number: 62 percent of adults identified as Christian. In the first study, in 2007, the figure had been 78 percent, and in 2014, 71 percent. The religiously unaffiliated, people who describe themselves as atheist, agnostic or nothing in particular, were 29 percent.

Read that way, it is a story of decline. But the same survey found that the Christian share had hovered between 60 and 64 percent since 2019, that the share of nones had stopped growing, and that 44 percent of adults prayed at least daily and 33 percent attended services at least monthly, both figures stable for several years. Large majorities held spiritual beliefs: 86 percent said people have a soul or spirit, 83 percent believed in God or a universal spirit, and 70 percent believed in an afterlife. Whether American religion is fading, levelling off or changing shape is a live argument, and sociology has been arguing about it since its founders. This lesson sets their explanations side by side, then puts the data against them.

Three founders, three questions

Durkheim, Marx and Weber, whom you met in Lesson 2, each asked a different question about religion, and OpenStax's chapter on religion traces the modern perspectives back to them.

A fourth, interactionist, line runs through the sociologist Peter Berger. In The Sacred Canopy (1967) he argued that beliefs stay certain only while they are held up by what he called plausibility structures: the everyday conversations with people who share them. When a society becomes plural, and your neighbours, classmates and colleagues believe different things, each faith has to be chosen rather than taken for granted. Pew's data contain a trace of that mechanism: among adults aged 18 to 24 who were raised in highly religious homes, only 28 percent were highly religious themselves.

The upshot: The three founders are not rival answers to one question. Durkheim asks what religion does for a group, Marx whose interests it serves, and Weber how its ideas change what people do. A good analysis uses the one that fits the question being asked.

Weber's thesis, tested a century later

Weber's argument was historical and careful. He observed that in the Germany of his day Protestants were more prominent in business and skilled work than Catholics, and he traced this to ideas: Calvinist teaching about a calling in worldly work and about predestination encouraged disciplined work, saving and reinvestment, a spirit that later outlived its religious origin. For a century the thesis was argued mainly with examples. Two recent studies tested it with data from the German lands where the Reformation began.

Read the table carefully. Neither study finds a Protestant work ethic doing the work Weber gave it. Becker and Woessmann find a real Protestant advantage but a different mechanism, human capital rather than a spirit of capitalism; Cantoni finds no advantage in city growth at all. Neither directly measures the attitudes Weber wrote about, and Weber's claim was about the origins of a mentality in a particular century, so a defender can argue that the tests measure something else. What they establish is narrower and useful: in the places where the thesis should most clearly apply, the economic gap either disappears or is explained by literacy.

The secularisation argument

The broader question is whether modern societies inevitably become less religious. Three positions have the best evidence behind them.

Berger himself became a witness in the argument. Having predicted worldwide secularisation, he wrote in 1999 of the desecularisation of the world, pointing to religious revivals from Iran to post-Soviet Russia, and concluded that Western Europe, not the rest of the world, was the exception that needed explaining.

Putting the data against the theories

Pew's 2025 study of the world between 2010 and 2020 supplies the global test. Christians remained the largest group, 2.3 billion people, but their share of the world's population fell 1.8 points to 28.8 percent. Muslims grew fastest, by 347 million, to 25.6 percent, which Pew attributes largely to a young age structure and high fertility. People with no religious affiliation rose to 1.9 billion, 24.2 percent, and they did so despite being older on average and having fewer children, because many people raised in a religion, mainly Christians, stopped identifying with it. The centre of Christianity moved too: 30.7 percent of the world's Christians now live in sub-Saharan Africa, against 22.3 percent in Europe.

Each theory can claim part of that picture. Existential security fits it best as a whole: switching out of religion concentrated in richer countries, and growth through births in poorer ones. Classic secularisation fits the switching but not the absence of any global decline in religious share beyond a single point. The religious economies view fits the long American exception but has more trouble with the American decline since 2007, which happened in the most competitive religious market in the world.

The American figures add a warning about time. The gap between age groups is large: 46 percent of adults aged 18 to 24 identified as Christian, against 80 percent of those 74 and older, and 43 percent of the youngest were unaffiliated, against 13 percent of the oldest. Yet the youngest cohort, born from 2000 to 2006, was no less religious than those born in the 1990s. Pew states the condition plainly: because the older, more religious cohorts are shrinking, stability can last only if today's young adults become more religious as they age or later generations turn out more religious than their parents. A single survey cannot tell you which will happen. Following the same cohorts over time can.

What matters here: The theories predict different patterns in different places, not just more or less religion. Compare them on the pattern, which is why a global dataset that separates births from switching tells you more than any single national trend.

Common misconceptions

• "America is rapidly losing its religion, year after year." Christian identification fell from 78 to 62 percent between 2007 and 2023-24 but has been roughly stable since 2019, and daily prayer and monthly attendance have held steady for several years.
• "Weber showed that Protestants work harder." Weber traced a historical mentality to Calvinist ideas. Modern tests find either that literacy explains the Protestant prosperity gap or that Protestantism had no effect on city growth.
• "Secularisation means religion is disappearing across the world." Between 2010 and 2020 the religiously affiliated share of the world fell by less than one point, to 75.8 percent, while Muslims grew to 25.6 percent.
• "Christianity is a Western religion." In 2020, 30.7 percent of the world's Christians lived in sub-Saharan Africa and 22.3 percent in Europe.

What you now know

• Pew's 2023-24 Religious Landscape Study of 36,908 adults found 62 percent Christian, down from 78 percent in 2007 and 71 in 2014 but stable since 2019, and 29 percent unaffiliated.
• Durkheim explains religion by the cohesion it produces, Marx by the inequality it consoles and protects, and Weber by the action its ideas inspire; Berger's plausibility structures give the interactionist mechanism.
• Becker and Woessmann found that Protestants' higher literacy accounts for their prosperity gap in Prussia; Cantoni found no effect of Protestantism on the growth of 272 German cities.
• Classic secularisation predicts continuing decline, religious economies predict vigour where groups compete, and existential security predicts decline where life is secure and persistence where it is not.
• From 2010 to 2020 the world's unaffiliated share rose to 24.2 percent through switching, while Muslims grew to 25.6 percent through young populations and high fertility.
• American religious identity differs sharply by age, 46 percent Christian at 18 to 24 against 80 percent at 74 and older, and whether the recent stability lasts depends on how today's young adults change as they age.

Sources

1. Smith, G. A., Cooperman, A., Alper, B. A., Mohamed, B., Rotolo, C., Tevington, P., Nortey, J., Kallo, A., Diamant, J., and Fahmy, D. (2025). Decline of Christianity in the U.S. has slowed, may have leveled off. Pew Research Center. pewresearch.org
2. Hackett, C., Stonawski, M., Tong, Y., Kramer, S., Shi, A., and Fahmy, D. (2025). How the global religious landscape changed from 2010 to 2020. Pew Research Center. pewresearch.org
3. Becker, S. O., and Woessmann, L. (2009). Was Weber wrong? A human capital theory of Protestant economic history. Quarterly Journal of Economics, 124(2), 531-596. iza.org
4. Cantoni, D. (2015). The economic effects of the Protestant Reformation: Testing the Weber hypothesis in the German lands. Journal of the European Economic Association, 13(4), 561-598. ideas.repec.org
5. OpenStax. (2021). The sociological approach to religion. In Introduction to Sociology 3e. Rice University. openstax.org

From 30 percent to 42

Every two years the Centers for Disease Control and Prevention asks a national sample of American high school students the same question: during the past 12 months, did you ever feel so sad or hopeless almost every day for two weeks or more in a row that you stopped doing some usual activities? In 2013, 30 percent said yes. In 2019 it was 37 percent, in 2021 42 percent, and in 2023 40 percent. Among girls the figure went from 39 percent in 2013 to 57 percent in 2021, then eased to 53; among boys from 21 to 29, then 28. Among LGBTQ+ students in 2023 it was 65 percent.

Something happened, and the candidate cause most people name is in their pocket. In The Anxious Generation (2024), the social psychologist Jonathan Haidt argued that the spread of smartphones and social media around 2010 caused a rewiring of childhood, pointing out that measures of adolescent distress turned upward at about that time while changing far less for adults. The candidate is certainly everywhere. In Pew Research Center's autumn 2025 survey of 1,458 teens aged 13 to 17, roughly nine in ten used YouTube and about three quarters visited it daily, 61 percent visited TikTok daily, and 36 percent said they were on at least one of five major platforms almost constantly.

This lesson treats the question as a problem to be worked, not a verdict to be announced. The steps are the ones you would apply to any claim that a medium changes people: what exactly is measured, what the timing can and cannot show, how large the association is, and which designs can isolate a cause.

Step one: what exactly is measured

The survey measures a self-report of sadness serious enough to interrupt ordinary life. That is a meaningful measure, and it is also one that can rise if young people become more willing to describe their feelings, or learn a new vocabulary for them, as well as if more of them are actually suffering. The same survey offers a partial check. Behaviour-based measures moved less than the feelings question: the share who seriously considered suicide went from 17 percent in 2013 to 22 percent in 2021 and 20 percent in 2023, and the share who attempted suicide stayed between 7 and 10 percent throughout. That pattern does not settle anything, but it tells you that the measure matters, the same lesson you drew from the LGBTQ+ identification data in Lesson 16.

Step two: timing is evidence, not proof

Haidt's strongest argument is timing: the turn upward came as smartphones spread, and it came first and hardest for girls, who, in Pew's data, are also the heaviest users of Instagram and Snapchat. Timing is real evidence, because a cause must come before its effect. But it cannot rule out other changes in the same years, which is the confounding problem from Lesson 6. Anything else that shifted for teenagers after about 2010, in schools, families, the economy after the 2008 recession, or how mental health is discussed, is a rival explanation until something excludes it. A national trend line cannot do that by itself.

Step three: how big is the association?

If social media drives distress, heavier users should be more distressed. In 2019 Amy Orben and Andrew Przybylski tested this across three large surveys covering 355,358 adolescents, running every reasonable way of analysing the data, a method called specification curve analysis, so that no single analytic choice could manufacture a result. The association between digital technology use and well-being was negative but small: it explained at most 0.4 percent of the variation in well-being. The authors judged it too small to warrant policy change.

Two replies are possible, and both are fair. First, a small average can hide a larger effect for some users, such as heavy users or girls, and Pew's 2024 survey of teens and parents found girls more likely than boys to say social media had hurt their mental health (25 against 14 percent), their confidence (20 against 10) and their sleep (50 against 40). Second, a correlation cannot tell you which way the arrow points. The developmental psychologist Candice Odgers, reviewing Haidt's book in Nature, argued that the evidence did not show a large or consistently negative effect and that distressed teenagers may turn to social media more, which would produce the same correlation in reverse.

The point: A trend line, a correlation and a teenager's own impression are three different kinds of evidence. None of them alone can show that a medium causes harm, and none of them alone can show that it does not.

Step four: designs that can isolate a cause

Two studies come closest to an experiment, and both point toward some harm.

• A natural experiment. Facebook did not arrive everywhere at once: it opened college by college from 2004. Luca Braghieri, Ro'ee Levy and Alexey Makarin compared student mental health surveys from colleges that had just received Facebook with those from colleges that had not yet got it. The arrival worsened student mental health and raised the share of students reporting that poor mental health had impaired their academic work, and their evidence on mechanisms pointed to unfavourable social comparison.
• A randomised experiment. Hunt Allcott, Luca Braghieri, Sarah Eichmeyer and Matthew Gentzkow, writing in the American Economic Review in 2020, randomly assigned adult Facebook users to deactivate their accounts for the four weeks before the 2018 midterm elections. Deactivation increased subjective well-being and time spent on offline activities, including socialising with family and friends, reduced both factual news knowledge and political polarisation, and led to a lasting drop in Facebook use afterwards.

The limits are as important as the findings. The college study concerns Facebook in 2004 to 2006, a very different product from TikTok today, used by students rather than younger teenagers. The deactivation study concerns adults over four weeks. Both show that the platforms can affect well-being; neither can tell you how much of the rise from 30 to 42 percent they explain. That is where the dispute honestly stands.

What teenagers themselves say

Pew's survey of 1,391 teens and their parents in September and October 2024 adds the teenagers' own view. Forty-eight percent said social media had a mostly negative effect on people their age, up from 32 percent in 2022, but only 14 percent said it had a negative effect on them. Forty-five percent said they spent too much time on it, up from 36 percent. At the same time, 74 percent said it made them feel more connected to their friends. Pew itself warns that such answers show perceptions, not causes. The gap between 48 and 14 percent has a name in media research, the third-person effect: people tend to believe that media influence others more than themselves.

The general lesson: how sociologists judge a media effect

Worry about new media is old. In the 1950s comic books were blamed for juvenile delinquency, and in 1972 the British sociologist Stanley Cohen gave the pattern its name in Folk Devils and Moral Panics, a study of the reaction to the rival mod and rocker youth gangs of the 1960s. A moral panic, in Cohen's account, occurs when a condition or group is defined as a threat to a society's values, in claims that exaggerate its seriousness; he called those who lead the alarm moral entrepreneurs and the blamed group folk devils. The concept is a warning, not a verdict. Some of Haidt's critics have called the concern about phones a moral panic; Haidt rejects the label. A panic can be about a real problem, which is why the evidence steps above matter more than the label.

Media effects are also sometimes real and well measured. During the 1968 presidential election Maxwell McCombs and Donald Shaw surveyed 100 residents of Chapel Hill, North Carolina, and compared the issues they ranked most important with the issues local and national media had emphasised, finding a strong match. Their idea, agenda-setting, is that media influence what people think about more than what they think, and it has been tested in hundreds of studies since.

The three perspectives, as OpenStax's chapter on media sets them out, predict different things about the platforms:

• Functionalist. Media perform functions, keeping people informed, connected and entertained, and can have dysfunctions. The 74 percent who feel more connected to friends are the functional side; displaced sleep and time with friends would be dysfunctions.
• Conflict. Platforms are businesses that earn money from attention, so the question is who benefits from designs that keep a user scrolling. The conflict prediction is that design follows revenue, whatever it costs users.
• Symbolic interactionist. Online life is a stage for the self-presentation that Goffman described in Lesson 3, with an audience that counts its approval in public. The social-comparison mechanism in the Facebook college study is an interactionist mechanism.

Bottom line: The case against phones is stronger than a moral panic and weaker than its most confident advocates claim. What would settle it is the same thing that settled the school-money question in Lesson 18: more designs in which exposure changes for reasons unrelated to how teenagers already feel.

Common misconceptions

• "The rise in teen sadness proves that phones caused it." Timing is consistent with a cause but cannot exclude other changes in the same years, and a correlation can run in reverse.
• "Studies show social media has no effect, so there is nothing to worry about." The average association is small, but a natural experiment and a randomised experiment both found effects on well-being, and averages can hide larger effects for particular groups.
• "If most teens say social media hurts people their age, it must hurt them." Only 14 percent said it hurt them personally, and perceptions of this kind are subject to the third-person effect.
• "Calling something a moral panic means the problem is imaginary." Cohen's concept concerns an exaggerated reaction; the underlying issue may still be real.

Recap

• The CDC's survey found persistent sadness or hopelessness among high school students rose from 30 percent in 2013 to 42 percent in 2021 and 40 percent in 2023, from 39 to 57 then 53 percent among girls.
• Social media use is near universal: about three quarters of teens visit YouTube daily and 36 percent are almost constantly on at least one major platform.
• The survey measure is a self-report, and behaviour-based measures such as suicide attempts moved much less than the feelings measure.
• Timing supports the phone explanation but cannot exclude confounders; across 355,358 adolescents the association explained at most 0.4 percent of the variation in well-being; and distress may drive use as well as the reverse.
• Facebook's staggered arrival at colleges worsened student mental health, and randomly assigned deactivation raised adults' well-being, so the platforms can cause harm; how much of the national rise they explain is still unknown.
• Agenda-setting, the third-person effect and moral panic are the concepts that let you place a new media scare in the long history of media research.

Sources

1. Centers for Disease Control and Prevention. (2024). Youth Risk Behavior Survey data summary and trends report: 2013-2023. U.S. Department of Health and Human Services, Division of Adolescent and School Health.
2. Faverio, M., and Sidoti, O. (2025). Teens, social media and AI chatbots 2025. Pew Research Center. pewresearch.org
3. Faverio, M., Anderson, M., and Park, E. (2025). Teens, social media and mental health. Pew Research Center. pewresearch.org
4. Orben, A., and Przybylski, A. K. (2019). The association between adolescent well-being and digital technology use. Nature Human Behaviour, 3(2), 173-182. pubmed.ncbi.nlm.nih.gov
5. Braghieri, L., Levy, R., and Makarin, A. (2022). Social media and mental health. American Economic Review, 112(11), 3660-3693. aeaweb.org

341,784,857 people with a median age of 39.4

On 1 July 2025, according to the Census Bureau's estimates, the United States had 341,784,857 residents. Half were older than 39.4 years, up from 35.6 in 2001. Since the 2020 census the number of people aged 65 and older had grown by 16.2 percent, to 64.6 million, while the number of children under 18 had fallen by 2.4 percent, to 72.0 million. The country is ageing, and so is most of the world.

Demography is the study of population size, composition and change, and most of its questions reduce to arithmetic on three flows: births, deaths and migration. This lesson is a procedure. You will compute the standard rates for real countries from the World Bank's World Development Indicators, which are built on United Nations population estimates, place those countries on a model of how populations change, and then change one input and watch a ranking turn upside down. The last part turns from arithmetic to sociology: who in the United States lives longest, and how the three perspectives explain what ageing means.

Step one: the basic rates

The crude birth rate is births per 1,000 people in a year, and the crude death rate is deaths per 1,000. The difference, divided by 10, is the rate of natural increase in percent. The total fertility rate is the number of children a woman would have over her life at current rates; about 2.1 is replacement level, the rate at which each generation replaces itself. Life expectancy at birth is the average number of years a newborn would live if current death rates at every age held throughout its life.

Work the first row. Niger's rate of natural increase is 41.4 minus 8.7, which is 32.7 per 1,000, or 3.27 percent a year. To turn a growth rate into a doubling time, divide 70 by the percentage, the rule of 70: 70 divided by 3.27 is about 21 years. A child born in Niger today can expect the country's population to double before she turns 22. India: 16.0 minus 6.6 is 9.4 per 1,000, 0.94 percent, a doubling time of about 74 years. Japan: 5.7 minus 13.3 is minus 7.6 per 1,000, a natural decrease of 0.76 percent a year, so without migration its population shrinks.

Now the United States: 10.6 minus 9.0 is 1.6 per 1,000, just 0.16 percent a year. Yet the Census Bureau estimates the population grew from 331.5 million in April 2020 to 341.8 million in July 2025, an increase of 10.3 million or 3.1 percent. Natural increase at 0.16 percent a year for five and a quarter years would have added only about 2.8 million: 331.5 million times 0.0016 times 5.25. Most of the growth, therefore, came from the third flow, net international migration. Population change is always natural increase plus net migration, and in the United States the second term is now the larger one.

National statistical offices publish their own figures, which differ slightly from these internationally harmonised ones; the World Bank series is used here so that every country is measured in the same way. The world as a whole, by the United Nations' count, reached 8.2 billion people in 2024 with a fertility rate of 2.25, down from 3.31 in 1990.

Step two: place each country on the demographic transition

In 1930 the American demographer Warren Thompson noticed that countries seemed to pass through the same sequence, and in the 1940s Frank Notestein and others developed it into the model of the demographic transition. The model has four classic stages, and some demographers add a fifth.

The model describes a sequence; it does not guarantee it. It was built from European history, and the timing and causes of falling fertility, whether economic development, education for girls, contraception or urban living, are still argued over. What the table does show is how much momentum the young stages carry. The United Nations projects that the world's population will peak at about 10.3 billion in the mid-2080s, and that 79 percent of the growth to 2054 will come from the youthful age structure already in place, because a large generation of young people will have children even at low fertility rates.

Step three: age structure and the dependency ratio

The dependency ratio compares the people of dependent ages, under 15 and 65 or older, with the people of working age, 15 to 64, per 100. It is the demographer's quick measure of how many people each worker's output must support. Compute it from the table.

• Niger: 46.6 plus 2.6 is 49.2 percent dependent, leaving 50.8 percent aged 15 to 64. 49.2 divided by 50.8, times 100, is 96.8 dependents per 100 people of working age.
• Japan: 11.4 plus 29.8 is 41.2 percent, leaving 58.8. The ratio is 70.1.
• The United States: 17.3 plus 17.9 is 35.2 percent, leaving 64.8. The ratio is 54.4.

Niger and Japan both carry heavy dependency loads, and the loads are opposites. In Niger the dependents are overwhelmingly children, who will become workers; in Japan they are mostly older people, who will not. A population pyramid makes the difference visible at a glance: a wide base narrowing sharply for Niger, a narrow base with a bulge in the upper ages for Japan.

In short: Natural increase is births minus deaths; population change adds migration; doubling time is 70 divided by the growth rate; and the dependency ratio tells you how heavy the load is but not what kind it is. Always look at the age structure behind the ratio.

The second worked example: change the measure and the ranking flips

Which country in the table has the worst mortality? Rank by crude death rate and the answer is Japan, at 13.3 deaths per 1,000, far above Niger's 8.7. Now switch the input to life expectancy. Japan's is 84.0 years and Niger's 61.4. The ranking has reversed completely, and the second answer is the right one.

The crude death rate is crude because it ignores age. In an old population many people are at the ages where death is common, so the rate is high even though each person, at each age, is less likely to die than in a young country. Japan has 29.8 percent of its people aged 65 or older; Niger has 2.6 percent. That is why careful comparisons use either life expectancy, which is calculated from death rates at every age, or an age-adjusted death rate, which applies each population's age-specific rates to one standard age structure. The National Center for Health Statistics reports American mortality exactly this way: 722.1 deaths per 100,000 on the age-adjusted measure in 2024, down 3.8 percent from 750.5 in 2023.

The same trap waits in the dependency ratio. The World Bank's figures for the United States and for the world as a whole are almost identical, about 54 dependents per 100 of working age, yet the United States has 17.9 percent of its people aged 65 and older against 10.2 percent for the world, and 17.3 percent under 15 against 24.7 percent. Same number, different societies.

Health: who lives longest

American life expectancy was 79.0 years in 2024, up from 78.4 in 2023 and 77.5 in 2022 as the country recovered from the pandemic years; it was 81.4 for women and 76.5 for men, and 3,072,666 Americans died that year. Globally, the United Nations reports life expectancy at birth of 73.3 years in 2024, 8.4 years more than in 1995, after a fall from 72.6 in 2019 to 70.9 during 2020 and 2021.

Averages like these hide a steep social gradient. Raj Chetty and colleagues linked 1.4 billion tax records to Social Security death records to estimate life expectancy at age 40 by income. The gap between the richest 1 percent and the poorest 1 percent was 14.6 years for men and 10.1 years for women. Between 2001 and 2014, life expectancy rose by 2.34 years for men and 2.91 years for women in the top 5 percent of incomes, and by only 0.32 and 0.04 years in the bottom 5 percent. Among low-income people, life expectancy varied by about 4.5 years between the best and worst places to live, and those local differences were strongly correlated with health behaviours such as smoking (r = minus 0.69), but not significantly with access to medical care. Health is socially patterned, and the pattern runs through class, place and behaviour more than through clinics.

Some of the pattern is changing in unexpected directions. The Census Bureau reports that among Americans aged 65 and older there were 70.6 men for every 100 women in 2001 and 81.6 in 2025, because death rates for older men have fallen faster than for older women: the age-adjusted death rate for men 65 and older fell from 8,285.0 per 100,000 in 1970 to 5,205.7 in 2022, against a fall from 5,621.3 to 3,918.7 for women.

Three perspectives on growing old

Sociologists do not agree on what ageing means for a person's place in society, and OpenStax's chapter on aging sets out the three families of theory.

• Functionalist. Disengagement theory, from Elaine Cumming and William Henry in 1961, held that withdrawal from social roles is a natural part of ageing that lets a society hand positions to the young. Activity theory, developed in the same years, answered that staying active and involved is the key to ageing well.
• Conflict. Modernisation theory (Cowgill and Holmes, 1972) argued that older people lose status as industrial societies stop valuing the experience they hold, and age stratification theory (Riley, Johnson and Foner, 1972) treated age as a system of inequality like class, race and gender. On this reading the fight over Social Security, which you saw in Lesson 14 keeping about 32 points of older Americans out of poverty, is a contest between age groups over resources.
• Symbolic interactionist. Ageing is also a matter of meaning: how people interpret getting older and how others treat them. Lars Tornstam's idea of gerotranscendence holds that many people become less self-centred and more at peace as they age, the reverse of the decline story.

The theories make testable predictions. If disengagement were natural and beneficial, older people who keep working or volunteering should fare no better than those who withdraw; activity theory predicts they fare better. Conflict theory predicts that the share of public spending going to older people will track their political power as well as their needs. Each prediction can be checked against survey and budget data, which is where the argument now lives.

Worth holding on to: A population's future is written in its age structure. The United Nations projects that by the late 2070s people aged 65 and older will outnumber children under 18 worldwide, and almost every question in this course, from families to schools to poverty, looks different in a society where that has happened.

Common misconceptions

• "A high death rate means people die young." Japan's crude death rate is 13.3 against Niger's 8.7, but its life expectancy is 84.0 against 61.4. The crude rate reflects age structure.
• "A population starts shrinking as soon as fertility falls below 2.1." A young age structure carries momentum: the United Nations attributes 79 percent of world growth to 2054 to it.
• "The American population grows mainly through births." Natural increase of about 0.16 percent a year would have added roughly 2.8 million people from 2020 to 2025; the population grew by 10.3 million.
• "Life expectancy is mainly about access to doctors." Among low-income Americans, local life expectancy correlated strongly with smoking and not significantly with access to medical care.

Putting it together

• The United States had 341.8 million residents in July 2025, a median age of 39.4, 64.6 million people aged 65 and older (up 16.2 percent since 2020) and 72.0 million children (down 2.4 percent).
• Rate of natural increase is births minus deaths per 1,000, divided by 10; doubling time is 70 divided by that percentage: about 21 years for Niger, while Japan shrinks by 0.76 percent a year.
• Population change adds net migration to natural increase, which is why the United States grew by 10.3 million from 2020 to 2025 when births minus deaths explain about 2.8 million.
• The demographic transition runs from high births and deaths to low ones, with a debated fifth stage of natural decrease; young age structures carry momentum toward a world peak of about 10.3 billion in the mid-2080s.
• The dependency ratio was 96.8 in Niger, 70.1 in Japan and 54.4 in the United States, with opposite kinds of dependents in Niger and Japan.
• Crude death rates rank countries the wrong way round; life expectancy and age-adjusted rates correct for age structure.
• American life expectancy was 79.0 in 2024, and the gap between the richest and poorest 1 percent was 14.6 years for men and 10.1 for women.
• Disengagement and activity theories, modernisation and age stratification theories, and interactionist accounts such as gerotranscendence offer competing, testable readings of ageing.

Sources

1. World Bank. (2026). World Development Indicators: birth rate, death rate, fertility rate, life expectancy and population by age, 2024. worldbank.org
2. United Nations, Department of Economic and Social Affairs, Population Division. (2024). World population prospects 2024: Summary of results. un.org
3. U.S. Census Bureau. (2026, June). Populations in all age groups growing in the South, driven by outlying counties in metro areas (Vintage 2025 population estimates). census.gov
4. Xu, J., Murphy, S. L., Kochanek, K. D., and Arias, E. (2026). Mortality in the United States, 2024 (NCHS Data Brief No. 548). National Center for Health Statistics.
5. Chetty, R., Stepner, M., Abraham, S., Lin, S., Scuderi, B., Turner, N., Bergeron, A., and Cutler, D. (2016). The association between income and life expectancy in the United States, 2001-2014. JAMA, 315(16), 1750-1766. pmc.ncbi.nlm.nih.gov

A letter to the mayor, 21 May 1954

On 21 May 1954 Jo Ann Robinson, an English teacher at Alabama State College and president of Montgomery's Women's Political Council, wrote to Mayor W. A. Gayle. Two months earlier the council, a group of Black professional women founded in 1946, had asked the city for three changes on its buses: no one made to stand over empty seats; an end to the rule that Black passengers pay at the front and then enter at the rear; and stops at every corner in Black neighbourhoods, as in white ones. Nothing had changed. Robinson's letter warned that there had been talk from twenty-five or more local organisations of a city-wide boycott of the buses.

The boycott that made Martin Luther King Jr. famous began eighteen months later and lasted 381 days, from 5 December 1955 to 20 December 1956. It is usually told as a story about one woman's tired feet and one young minister's voice. Sociologists tell it as a story about organisation, and following it day by day is the clearest way to see how a movement is built.

Crowds and movements

Collective behaviour is activity outside the normal institutions of a society in which many people take part voluntarily: a crowd gathering after an accident, a rumour spreading through a school, a fashion sweeping a year group. A social movement is something more durable, a sustained and organised effort to bring about or resist social change. David Aberle's classification sorts movements by how much change they seek and for whom; the Montgomery boycott began as a reform movement, seeking to change specific rules rather than the whole social order.

The first theorists of collective behaviour distrusted crowds. In 1895 the French writer Gustave Le Bon described the crowd as a single irrational mind that swept individuals along. Later sociologists replaced that picture. Ralph Turner and Lewis Killian's emergent norm theory held that people in an unfamiliar situation work out new rules together in small groups, and Neil Smelser's value-added theory of 1962 listed conditions, from underlying strain to a precipitating event, that must accumulate before collective action occurs. As OpenStax's chapter on collective behaviour puts it, crowds came to be seen as the rational behaviour of people pursuing goals. Montgomery is a strong test case for the change of view.

Two arrests that did not start a boycott

In March 1955 a 15-year-old, Claudette Colvin, was arrested for refusing to give up her seat on a Montgomery bus. Seven months later an 18-year-old, Mary Louise Smith, was arrested for the same refusal. Both arrests were discussed as possible test cases. Neither mobilised the city.

Then, on Thursday 1 December 1955, Rosa Parks refused to give up her seat. Parks was 42, a seamstress, and no accidental protester: she had joined the Montgomery branch of the NAACP in 1943 and served as its secretary, and in the summer of 1955 she had attended the Highlander Folk School, a training centre for activists in Tennessee. Her refusal was not planned in advance, and she later rejected the story that she had simply been tired: the only tired she was, she wrote, was tired of giving in.

Why this matters: The grievance on the buses was the same in March, October and December 1955. What differed in December was who had been arrested, who was ready to act, and what they had prepared. A grievance alone does not produce a movement; this is the starting point of every modern theory of social movements.

Four days: 1 to 5 December

The response came from organisations that already existed. Robinson and the Women's Political Council called for a one-day boycott on Monday 5 December. That night Robinson, two students and John Cannon, who chaired the college's business department, ran off about 52,500 leaflets on the mimeograph at Alabama State College, according to the King Institute's biography of Robinson, and groups handed them out across Black Montgomery. E. D. Nixon, a past leader of the local NAACP, secured Parks's bail with Clifford and Virginia Durr and telephoned the city's Black ministers, who met at Dexter Avenue Baptist Church on 2 December and agreed to announce the boycott from their pulpits.

On 5 December about 90 percent of Montgomery's Black citizens stayed off the buses. That afternoon the city's ministers and leaders formed the Montgomery Improvement Association and elected as its president the 26-year-old pastor of Dexter Avenue, Martin Luther King Jr., whom Parks later said was an advantage precisely because he was new to the city and had made neither strong friends nor enemies. That evening several thousand people filled Holt Street Baptist Church and voted to continue.

The association's demands, issued on 8 December, were strikingly modest: courteous treatment by drivers; first-come, first-served seating, with Black passengers filling the bus from the back and white passengers from the front; and Black drivers on mainly Black routes. They did not yet ask for an end to segregated seating. That is framing at work. David Snow and Robert Benford distinguish a diagnostic frame, which names the problem; a prognostic frame, which proposes a solution; and a motivational frame, which gives people reasons to act now. A demand that sounded reasonable to moderate listeners was a prognostic frame chosen to widen support.

Keeping a city off its buses for a year

A one-day protest needs enthusiasm. A 381-day protest needs resources. When the city began penalising Black taxi drivers who carried boycotters at bus fares, the association built a carpool of about 300 cars, copying the system that T. J. Jemison had used in a 1953 bus boycott in Baton Rouge. The women who had started the protest ran its committees and volunteer networks: Robinson, Johnnie Carr and Irene West among them. Mary Fair Burks of the Women's Political Council later credited the victory to the nameless cooks and maids who walked endless miles for a year.

The costs were real. Early in 1956 the homes of King and Nixon were bombed. City officials obtained injunctions against the boycott and indicted more than 80 of its leaders under a 1921 law against conspiracies that interfered with lawful business; King was convicted and ordered to pay $500 or serve 386 days in jail. Support arrived from outside: Bayard Rustin and Glenn Smiley came to advise on nonviolent methods, and Rustin, Ella Baker and Stanley Levison founded In Friendship to raise money in the North.

Resource mobilisation theory, set out by John McCarthy and Mayer Zald in 1977, explains movements by exactly those things: their ability to gather time, money, people and organisation. The sociologist Aldon Morris, in The Origins of the Civil Rights Movement (Free Press, 1984), emphasised that the crucial resources were indigenous: Black churches, colleges and local associations that already had members, meeting places, communication channels and trusted leaders.

The courts, and the end

The boycott did not win alone. On 5 June 1956 a federal district court ruled in Browder v. Gayle, a case brought on behalf of four Black women including Colvin and Smith, that segregation on the city's buses was unconstitutional, and in November 1956 the Supreme Court affirmed. The boycotters stayed off the buses until the order actually reached Montgomery. On 20 December King called the boycott off, and the next morning he rode an integrated bus.

That combination points to a third theory. Doug McAdam's political process model, set out in 1982, holds that movements rise when three things coincide: political opportunities, such as federal courts newly willing to strike down segregation after Brown v. Board of Education in 1954; indigenous organisation of the kind Morris described; and a shift in consciousness in which people come to believe both that the situation is unjust and that it can be changed. Montgomery had all three.

Greensboro, 1 February 1960: how a tactic travels

Four years later the movement's next tactic came from students. On 1 February 1960 four students from North Carolina A and T College, Joseph McNeil, Ezell Blair, Franklin McCain and David Richmond, bought a few items at the Woolworth's in downtown Greensboro and sat down at the lunch counter reserved for white customers. They were refused service and stayed until closing. The next morning about two dozen students came. By the end of February there had been sit-ins at more than 30 locations in 7 states, and by the end of April more than 50,000 students had taken part.

A tactic that spreads that fast is travelling along networks: between colleges, through churches and student groups, and through the news. In April 1960 Ella Baker, then executive director of the Southern Christian Leadership Conference, called the sit-in leaders together, and 120 students from 12 southern states founded the Student Nonviolent Coordinating Committee. The sequence matches the stages that Herbert Blumer and Charles Tilly described, which OpenStax's chapter on social movements summarises: a preliminary stage in which people become aware of an issue, coalescence as they organise, institutionalisation as organisations form, and eventually decline.

Who takes the risk?

In the summer of 1964 more than a thousand volunteers from outside the South went to Mississippi for Freedom Summer, a voter registration campaign in the state with the lowest share of Black citizens registered to vote. Three civil rights workers, James Chaney, Andrew Goodman and Michael Schwerner, were murdered in its first days. Many applicants who had been accepted did not go.

Doug McAdam used the project's application forms to compare 720 people who went with 241 who withdrew. The ones who went were more likely to have a history of prior activism and to be tied into networks of other participants, and they identified strongly with the movement's values; people without those ties more often stayed home. McAdam called this high-risk activism, and his finding is the network version of the point Montgomery made about organisations: people join dangerous causes through people they already know.

What the movement changed

Outcomes can be measured, and the Voting Rights Act, signed on 6 August 1965, is the clearest case. When it was adopted, according to the Department of Justice, only a third of voting-age African Americans in the states it specially covered were registered to vote, against two thirds of eligible white citizens. The act ended literacy tests in six southern states and many counties of North Carolina, and allowed federal examiners to register voters. By the end of 1965, the National Archives records, a quarter of a million new Black voters had been registered, a third of them by federal examiners, and by the end of 1966 only four of the thirteen southern states had fewer than half of their African American citizens registered.

The core of it: Every theory in this lesson replaces the image of the irrational crowd with something that can be counted: organisations and their members, cars in a carpool, courts willing to rule, students connected across campuses, applicants with ties to other applicants, registration rolls before and after a law.

Common misconceptions

• "Rosa Parks was a tired seamstress who acted alone." She was a longtime NAACP officer and trained activist, and the boycott was organised by the Women's Political Council, ministers and the NAACP within days of her arrest.
• "The boycott started spontaneously." The council had been pressing the city since 1954 and had warned of a boycott; two earlier arrests had been discussed as test cases.
• "Protest movements are irrational crowds." The boycott ran for 381 days on committees, a 300-car carpool, mass meetings and legal strategy, which is why sociology moved away from Le Bon's view.
• "The boycott alone desegregated the buses." The decisive ruling came in Browder v. Gayle; the boycott sustained pressure and support while the case was decided.

What to carry forward

• Collective behaviour is non-institutional group activity; a social movement is sustained, organised action for or against change. Montgomery began as a reform movement.
• Early theories treated crowds as irrational; emergent norm and value-added theories, and then resource mobilisation, treated collective action as organised and goal-directed.
• The Women's Political Council warned of a boycott in May 1954; arrests in March and October 1955 did not mobilise the city, and Parks's arrest on 1 December did, because organisations were ready.
• About 90 percent stayed off the buses on 5 December 1955; modest demands framed the protest; a 300-car carpool, churches and Northern fund-raising sustained it for 381 days.
• Browder v. Gayle ended bus segregation in 1956, fitting McAdam's political process model of opportunity, organisation and changed consciousness.
• The Greensboro sit-in of 1 February 1960 spread to more than 30 locations in a month and involved more than 50,000 students by April, and SNCC was founded that spring.
• Freedom Summer volunteers who went, compared with those who withdrew, had more prior activism and ties to other participants.
• After the Voting Rights Act a quarter of a million new Black voters registered by the end of 1965.

Sources

1. Martin Luther King, Jr. Research and Education Institute. (n.d.). Montgomery bus boycott. Stanford University. kinginstitute.stanford.edu
2. Martin Luther King, Jr. Research and Education Institute. (n.d.). Sit-ins. Stanford University. kinginstitute.stanford.edu
3. McAdam, D. (1986). Recruitment to high-risk activism: The case of Freedom Summer. American Journal of Sociology, 92(1), 64-90.
4. National Archives. (n.d.). Voting Rights Act (1965). Milestone Documents. archives.gov
5. U.S. Department of Justice, Civil Rights Division. (2015). Introduction to federal voting rights laws: The effect of the Voting Rights Act. justice.gov

Forty-two million people in one city

In 2025, according to the United Nations' World Urbanization Prospects, the most populous city on Earth was Jakarta, with nearly 42 million inhabitants. Dhaka, with almost 37 million, was second and growing fast enough to become the largest by the middle of the century; Tokyo, with 33 million, was third and shrinking. There were 33 megacities of 10 million or more, 19 of them in Asia, up from 8 in 1975. Cities held 45 percent of the world's 8.2 billion people, against 20 percent of 2.5 billion in 1950, and the UN projects that two thirds of the world's population growth between now and 2050 will happen in cities.

This lesson compares theories at three scales: what cities do to the people who live in them, why some countries grew rich while others did not, and what drives social change in general. Each comparison ends by setting the theories against numbers.

First, what counts as a city

Every urban figure rests on a definition, and definitions differ. The 2025 UN report uses a harmonised method called the Degree of Urbanisation: a city is a population centre of at least 50,000 people with at least 1,500 people per square kilometre, and a town a cluster of at least 5,000 people with at least 300 per square kilometre. On those definitions 36 percent of the world lived in towns in 2025, down from 40 percent in 1950, and 19 percent in rural areas, half the 1950 share.

Countries use their own rules, and the rules move. After the 2020 census the Census Bureau raised the threshold for an urban area to 5,000 people or 2,000 housing units, from 2,500 people, and began to identify urban land mainly by housing density. The urban share of Americans fell from 80.7 percent in 2010 to 80.0 percent in 2020, and 1,140 areas with about 4.2 million people that had counted as urban became rural. The Bureau stated plainly that this was not a sign of people moving from cities to the countryside. The change was in the ruler.

The 1920 census was the first in which most Americans were urban. Its population, like everyone's since, had to adjust to living among strangers, and that adjustment is the first question urban sociology asked.

Comparison one: what does a city do to its people?

Nineteenth-century sociologists framed the question as a contrast between two kinds of social bond. Ferdinand Tönnies set Gemeinschaft, community built on kinship and place, against Gesellschaft, association built on contract and self-interest, and Durkheim distinguished the mechanical solidarity of similar people from the organic solidarity of people made interdependent by a division of labour. At the University of Chicago, Robert Park's human ecology and Ernest Burgess's concentric zone model of 1925 treated the city as an ecosystem in which groups compete for space and succeed one another outward from the centre, the functionalist strand that OpenStax's chapter on urbanisation describes. Three later answers to the question disagree sharply.

Read across the table and you can see that the three are not simply rival opinions; they predict different things you could measure. Wirth predicts isolation rising with city size. Gans found a community of relatives and peers in the middle of Boston that Wirth's theory said should not exist, and he criticised the planners who demolished it as a slum. Fischer turned Wirth upside down: a city of millions can sustain a club of chess players, a congregation of recent migrants or a scene of experimental musicians that a village of 500 never could. The UN's data add a correction to all three: most city dwellers do not live in megacities. Of the world's roughly 12,000 cities, 96 percent have fewer than a million inhabitants and 81 percent fewer than 250,000.

The conflict perspective asks a different question: who decides how cities grow? Its answer is the political and economic leaders who direct investment and regulate land use, which is why OpenStax sets it beside human ecology. The redlining maps in Lesson 15 and the neighbourhood effects in Lesson 14 are its evidence that the shape of a city is decided, not natural.

Remember: Whether a city isolates people is an empirical question, and the answer so far is that it depends on class, ethnicity and the groups a city is large enough to support, not on size alone.

Comparison two: why are some countries rich and others poor?

Globalisation, the growing integration of economies, cultures and populations across borders, is the setting for three theories of global inequality, summarised in OpenStax's chapter on global stratification.

Reading the evidence: extreme poverty, 1990 to 2024

The World Bank measures extreme poverty as living on less than $3.00 a day, adjusted for differences in prices between countries. The number of people below that line fell from about 2.3 billion in 1990 to about 831 million in 2025.

Each theory can claim a row, and none can claim the whole table. East Asia's fall from 67.0 to 2.0 percent, with China's own rate going from 83.1 percent in 1990 to effectively zero by 2019, is the strongest exhibit for modernisation theory: countries that industrialised and traded their way into world markets escaped mass poverty within a generation, which simple dependency theory said they could not. World-systems theory reads the same row as a movement of countries from the periphery toward the semi-periphery, exactly the mobility between tiers it allows. Sub-Saharan Africa's figure, still 45.1 percent in 2024, is the dependency theorist's exhibit: integration into world markets, largely as an exporter of raw materials, has not produced the same result there. The World Bank itself notes that extreme poverty has become increasingly concentrated in Sub-Saharan Africa and in places affected by conflict and fragility, and that global poverty reduction slowed in the last decade. Why the regions diverged, whether because of institutions, geography, colonial history, conflict or policy, is the live question, and it is argued with exactly these figures.

So what?: A theory of global inequality has to explain both the fastest escape from poverty in history and the regions it has not reached. Test any theory you hear against both rows.

Comparison three: what drives social change?

Social change is the transformation of a society's institutions, culture and relationships over time, and this course has already met most of its engines, which OpenStax's chapter on social change groups under technology, social institutions, population and the environment.

A symbolic interactionist adds the step both columns skip: change becomes real when people redefine a situation, as when the Lovings' certificate stopped being a crime and became a marriage, or when Montgomery's Black residents came to believe that walking was a form of power.

Common misconceptions

• "Most city dwellers live in megacities." Of the world's roughly 12,000 cities, 96 percent have fewer than a million people, and most urban residents live in small and medium-sized centres.
• "Americans are leaving cities for the countryside." The fall from 80.7 to 80.0 percent urban between 2010 and 2020 came from a new definition, as the Census Bureau itself explained.
• "Cities destroy community." Gans found close-knit urban villages, and Fischer showed that large cities support subcultures that small places cannot.
• "Globalisation has made the poor poorer everywhere", or its mirror, "globalisation has lifted everyone out of poverty." Extreme poverty fell from 43.4 to 10.4 percent of the world, but 45.1 percent of people in Sub-Saharan Africa remained below the line in 2024.

Where this leaves us

• In 2025 cities held 45 percent of the world's 8.2 billion people, against 20 percent in 1950; there were 33 megacities, led by Jakarta, Dhaka and Tokyo, and two thirds of growth to 2050 is projected to be urban.
• Urban figures depend on definitions: the American urban share went from 51.2 percent in 1920 to 80.7 percent in 2010, and a new definition moved it to 80.0 percent in 2020.
• Wirth predicted that city size, density and variety weaken personal ties; Gans documented close-knit urban villages; Fischer argued that size lets subcultures thrive.
• Modernisation theory explains poverty by lack of industrialisation, dependency theory by exploitation by the core, and world-systems theory by position in a core, semi-periphery and periphery hierarchy.
• Extreme poverty below $3.00 a day fell from 43.4 percent of the world in 1990 to 10.4 percent in 2024, from 67.0 to 2.0 percent in East Asia but only from 61.5 to 45.1 percent in Sub-Saharan Africa.
• Technology, population, collective action and law all drive social change, and functionalist, conflict and interactionist accounts read each of them differently.

Sources

1. United Nations, Department of Economic and Social Affairs, Population Division. (2025). World urbanization prospects 2025: Summary of results. un.org
2. U.S. Census Bureau. (2022, December 29). Nation's urban and rural populations shift following 2020 Census. census.gov
3. U.S. Census Bureau. (1995). Urban and rural population: 1900 to 1990 (Table 1). census.gov
4. World Bank. (2026). Poverty headcount ratio at $3.00 a day (2021 PPP), World Development Indicators, and Poverty overview. data.worldbank.org
5. OpenStax. (2021). Theoretical perspectives on global stratification. In Introduction to Sociology 3e. Rice University. openstax.org

Eighty-eight percent or seventy-eight

In December 2008 the Pew Research Center built an experiment into one of its national polls. The question was plain: all in all, are you satisfied or dissatisfied with the way things are going in this country today? Respondents who heard it straight after a question asking whether they approved of the way George W. Bush was handling his job as president said dissatisfied 88 percent of the time. Respondents who met it without that lead-in said dissatisfied 78 percent of the time. The words were identical. Ten points came from the question that went before. Approval of the president, for its part, barely moved whichever came first.

Your survey will be smaller than Pew's and simpler, but every choice in it, which words, in what order, handed to whom, with what promise, moves your numbers the same way. This lesson makes those choices on one worked example, in the order you will make them, from a hunch to a box of completed forms. Then it changes a single input and shows the answer turning upside down.

About the example. Ridgeway High is invented, and so is every number that comes from it in this lesson and the next; the numbers are there so that you can check each step yourself. The procedures, rules and research findings are real and come from the sources listed at the end.

Step 1: turn a hunch into a question with countable answers

Lesson 10 supplied the hunch. Role conflict is a clash between two statuses: when a shift starts at 5 pm and practice ends at 5.30, the employee and the team member cannot both be satisfied. Lesson 6 met the same after-school job from the other side, as the confounding variable that could lower club membership and grades at once. Put the two together and you have a question worth asking in your own school: do students who work for pay take part in school activities less often than students who do not?

As written, nobody can answer it, because work and taking part are not yet things anyone can count. The fix is an operational definition: each concept restated as an observable condition inside a time window, so that two students reading the question would count the same thing.

The independent variable is the suspected cause and the dependent variable the suspected effect. The finished research question reads: among Ridgeway's juniors and seniors, were students who worked for pay in the last seven days less likely than other students to take part in a school club, team or arts group in the same seven days?

Two design choices are hidden in that sentence. The seven-day window trades typicality for accuracy: a student can report last week's hours and practices, while a phrase like "this school year" invites a guess shaped by self-image. The price is that one week can be unusual, so the survey goes into an ordinary week, not one with a holiday, exams or the end of a sports season, and the write-up states the dates. The second choice is that the question can come back either way. Role conflict predicts that workers take part less. A rival idea, that energetic, organised students pile on both, predicts no gap or even a reversed one. A study that could only confirm your hunch would not be research, which is what Weber's value neutrality in Lesson 7 asked of you.

Bottom line: A research question is ready when every concept in it has become a countable answer in a stated time window, and when more than one result could come back.

Step 2: write a draft, then break it on purpose

Lesson 4 repaired four classic faults: the double-barrelled item, the leading question, vague frequency words, and a sensitive question with a name attached. The first draft of the Ridgeway questionnaire had six more, each common enough to hunt for in your own.

The repairs follow Pew's published guidance. Closed questions should include every reasonable answer, which makes them exhaustive, and no answer should fit two boxes, which makes them mutually exclusive. Four or five options are usually enough for an opinion question. Simple, concrete words beat jargon and double negatives; Pew's own cautionary example asks whether you favour or oppose not allowing something. And agree-disagree items carry a known distortion, acquiescence bias: some respondents, more often those with less schooling or less information about the topic, tend to agree with whatever statement is put to them, and more so when an interviewer is present. Pew's better practice is to offer a choice between alternative statements. For lists, Pew has also found that asking yes or no about each item gives more accurate answers than a single select-all-that-apply box, which is why question 6 below asks about each kind of group separately.

Step 3: put the questions in an order that does not steer

The December 2008 result is one case of a general pattern that Pew calls an order effect: earlier questions supply context for later ones. Asking a specific question before a general one about the same thing, happiness in one's marriage before happiness in general, is a known way to change the general answer. So the Ridgeway form follows four ordering rules.

• General before specific. The opinion question comes first. A student who has just counted her own work hours and missed practices would judge student jobs in general with those details fresh.
• Easy questions early, background last. Pew advises opening with simple, engaging questions and keeping demographic items away from the start, so grade is the final question.
• Printed routing. "If no, go to question 5" keeps students who did not work from inventing hours, and every skip is printed on the page.
• Balanced option order. On a form people fill in themselves, answers near the top of a list are chosen more often; on the telephone, the last one heard is. Question 6 lists four kinds of group with no natural order, so the form is printed in two versions, A and B, with that list reversed, and the two are shuffled together before each class. Hours and day counts keep their natural order, because there the order carries meaning.

The result fits on one side of one sheet. This is Form A, part of the invented example:

1. Which statement comes closer to your own view, even if neither is exactly right? During the school year, a part-time job mostly helps students / During the school year, a part-time job mostly gets in the way of school / Not sure
2. In the last seven days, did you do any work for pay? Count babysitting, yard work, a regular job, or work in a family business if you were paid. Yes / No (if no, go to question 5)
3. In the last seven days, about how many hours did you work for pay? 1-5 / 6-10 / 11-15 / 16-20 / More than 20
4. On how many of the last five school days did you work for pay after school? 0 / 1 / 2 / 3 / 4 / 5
5. In the last seven days, did you go to a practice, meeting, rehearsal, game or performance for any school club, team or arts group? Yes / No (if no, go to question 8)
6. In the last seven days, did you take part in each kind of group? Sports team: yes or no. Arts group, such as band, choir, drama or art: yes or no. Academic or interest club: yes or no. Service club or student government: yes or no.
7. On how many of the last five school days did you take part in any school group? 0 / 1 / 2 / 3 / 4 / 5
8. This school year, did you stop going to a school club, team or arts group that you had belonged to? Yes / No (if no, go to question 10)
9. In a few words, what was the main reason you stopped? (write in, or leave blank)
10. What grade are you in? 11 / 12

Across the top runs one line: do not write your name, and you may skip any question. Question 9 is the only open question. Lesson 4 showed that open questions reveal the categories respondents themselves use, at the cost of reading and coding every answer by hand, which Lesson 25 does.

Step 4: pretest on students who will not be in the sample

Pew calls pretesting an essential step in questionnaire design, and it tests new questions with focus groups, cognitive interviews and pilot tests before they reach a real survey. The school-sized version is the cognitive interview. Sit with three or four students like the ones you will survey, but from classes that were not drawn in step 5, and ask each to read every question aloud, say what they think it is asking, and answer it while thinking out loud. Note every hesitation. Time the whole form.

In the invented example, the pretest earned its fifteen minutes twice. An early draft asked, "Do you have a job?" Two of the four pretesters, who babysit for neighbours most weekends, said no, because to them a job meant a boss and a paycheck; that is why question 2 asks about work for pay and names babysitting. A third counted sitting in the stands at a basketball game as taking part in a school group, which is why question 5 names the activities it means. All four finished in about four minutes.

In short: You cannot read your own questions the way a stranger reads them. A few students reading aloud catch misreadings that rereading your own draft never will.

Step 5: choose a frame, then let chance choose the students

The population is Ridgeway's 480 juniors and seniors. Lesson 4 showed that how people get into a sample matters more than how many there are, so the first decision is the frame, the actual list you will draw from.

The English sections win. Number them 1 to 20 from the master schedule, then, in front of your teacher, draw six numbers with a random number generator or from folded slips in a container, and write down the numbers and the method. Everyone in a drawn section is invited. This is a single-stage cluster sample: chance picks the groups, and the groups bring their members.

State its limits now, before any results can tempt you to forget them. Students in one class resemble each other more than 150 names from a hat would; one section may be honours, another may meet at 7.30 in the morning. A cluster sample therefore carries more sampling error than a simple random sample of the same size, which statisticians express as the design effect, and with only six clusters that extra error cannot be estimated well. The frame misses any junior or senior not taking English at Ridgeway this term. And whatever the results, they describe Ridgeway's juniors and seniors in one week: not sophomores, not other schools, not teenagers in general.

Worth holding on to: The frame decides who can be in the study, and chance decides who is. Write down both, because no reader can judge a result without them.

Change one input: the gym lobby at five o'clock

Keep the questionnaire exactly as it is and change only the frame. Suppose the researcher, short of time, gives the form to everyone leaving the gym lobby at 5 pm on a Tuesday as practices end. In the invented example 48 students fill it in: 15 had worked for pay in the last seven days and 33 had not. Every one of the 48 had been to a practice that week, which is why they were in the lobby.

The lobby did not produce a slightly worse estimate. It produced the opposite conclusion, and a bigger crowd from the same lobby would only repeat it with more confidence. The flaw is structural: the frame chose students by the very outcome being measured, a form of selection bias. A student who went to work instead of practice could not have been standing there. The bus line at 3 pm would fail in the other direction, filling the sample with students who go straight home. To test any frame, ask who could not possibly be in it.

Step 6: ask an adult before you ask a single student

Everything so far is a plan on paper, and it stays on paper until an adult with authority has read it. Take your teacher five things: the research question, the questionnaire, the sampling plan, the consent note and permission slip, and a sentence on how the forms will be stored and destroyed. Your teacher knows your school's rules, which may call for a principal's signature, a school research committee or a letter to parents. Two sets of outside rules are worth knowing before that conversation.

• Science fairs. If the project will be entered in a fair affiliated with the Regeneron International Science and Engineering Fair, Society for Science's rules count any survey or questionnaire as research with human participants. A school institutional review board of at least three people, an educator who is not your sponsoring teacher, a school administrator, and a medical or mental health professional, must approve the plan before anyone is recruited. Participants under 18 give their own assent and bring written permission from a parent or guardian, with the questionnaire attached so that the parent sees exactly what will be asked. Consent may not involve coercion, and any change to the plan after approval must be approved again.
• Federal law on school surveys. The Protection of Pupil Rights Amendment covers the programs of schools and districts that receive U.S. Department of Education funds. It names eight protected areas: political affiliations or beliefs; mental or psychological problems; sexual behaviour or attitudes; illegal, anti-social, self-incriminating or demeaning behaviour; critical appraisals of close family members; legally privileged relationships, such as those with lawyers, doctors or ministers; religious practices, affiliations or beliefs; and income. No student in a federally funded program may be required to answer a survey that reveals such information without a parent's prior written consent, and districts must tell parents about surveys containing such items and let them opt their child out.

Whether or not the law reaches a particular class project, the practical rule is the same: keep every question outside all eight areas. The Ridgeway form asks how many hours students worked, not what they earned, and whether they went to club meetings, not what they believe. That is why the income question in step 2 was dropped rather than repaired.

Lesson 7 set out the structure of consent for a minor: permission from a parent or guardian, assent from the student, and either refusal ends the matter. The American Sociological Association's code of ethics adds a standard aimed squarely at you: when the people you study are students at your own institution, take special care to protect them from any negative consequence of declining. It also rules out inducements so large that they unduly push people into taking part. At Ridgeway, that means five design decisions.

• Students whose permission slips came back signed get a form. Students without one get a short reading, so nobody sits idle and conspicuous.
• A student with permission who decides not to take part hands in the form blank. Everyone folds their form and drops it into a sealed box at the end, so no one, including the teacher, can tell a blank form from a completed one.
• The teacher stays at the desk and never handles forms. There is no extra credit, because a grade reward would make saying no cost something.
• No names, no student numbers. The only background question is grade, with about 240 students in each. The box is emptied into a labelled envelope after each class, so the response rate can be counted by section without linking any form to a person.
• No table will report a cell of fewer than five students, the defence against deductive disclosure from Lesson 7. The forms stay sealed in the teacher's room and are shredded at the end of the semester.

The note read aloud in each class and printed at the top of the form:

My name is (your name), and I am a student in (teacher's name)'s sociology class. I am doing a class project about paid work and school activities. Your English class was chosen by a random draw. If you agree to take part, you will answer ten questions, which takes about five minutes. Do not write your name anywhere. Taking part is your choice. You can skip any question or hand in a blank form, and nothing happens either way: no grade or credit depends on it, and nobody can tell a blank form from a completed one. I will report only totals, never a group of fewer than five students, and the forms will be shredded at the end of the semester. If you have questions, ask me or (teacher's name) in room (number).

The permission slip sent home a week earlier, with both forms of the questionnaire attached:

Your child's English class was chosen at random for a five-minute anonymous survey, part of a sociology class project on paid work and school activities. The questions ask whether students worked for pay in the last week and for how many hours, and whether they took part in school clubs, teams or arts groups. They do not ask about pay, family income, beliefs or health. No names are collected, and results will be reported only as totals. The full questionnaire is attached. Taking part is voluntary, and your child may still decline on the day. Please tick one: I give permission for my child to take part / I do not give permission. Parent or guardian signature and date. Questions: (teacher's name and school email).

Step 7: run it the same way in every room

• Survey all six sections in one ordinary week, on the same day if the timetable allows.
• Read the same note in every room, word for word, so that no class hears a more persuasive version.
• Shuffle Forms A and B together before each class.
• Keep a field log: for each section, the number enrolled, permissions returned, absences, blank forms and completed forms, plus anything unusual, such as a fire drill or a test that period.

The response rate is completed questionnaires divided by everyone sampled: 112 of 150, or 75 percent. It is not 112 of 121, which quietly drops the students whose families never returned a slip, and it is not 112 of 480, which confuses the sample with the population. Then ask what the missing 38 might change. The six sections held 78 juniors and 72 seniors, 52 and 48 percent; the completed forms came from 58 juniors and 54 seniors, again 52 and 48 percent. On grade, at least, the respondents look like the sample. On things the log cannot see, such as students absent because they were travelling with a team, they may not.

Common misconceptions

• "Random means whoever happens to be around." In sampling, random means a chance procedure you can describe and someone else could repeat: numbered sections and a recorded draw. Whoever happens to be around is a convenience sample, and the gym lobby showed what that can do.
• "An anonymous survey can ask about anything." Anonymity protects who answered, not what was asked. The eight protected areas, your school's rules and the small-cell problem from Lesson 7 all still apply.
• "Pretesting is for long professional surveys." Four students reading ten questions aloud found that babysitting did not count as a job to them, a misreading no amount of rereading by the author would have caught.
• "Agree or disagree is the neutral way to ask an opinion." It invites agreement. A choice between two stated positions, with a not sure option, measures the view more cleanly.
• "Once the teacher says yes, the design is fixed and finished." Approval covers the plan you showed. Under science fair rules any change needs approval again, and the same courtesy is owed to your teacher.

The short version

• In Pew's December 2008 experiment, 88 percent said they were dissatisfied with the country's direction when the question followed one on presidential approval, against 78 percent without it: question order is part of the measurement.
• Operational definitions turn concepts into countable answers in a stated window, here any work for pay and any school group activity in the last seven days.
• Closed questions need exhaustive, mutually exclusive options in plain words; agree-disagree items invite acquiescence, so offer a choice between stated views; ask general questions before specific ones and background questions last.
• Pretest with a few students like your respondents who are not in the sample, and time the form.
• Randomly drawing sections of a required course gives every student the same known chance; its costs are a larger error than a simple random sample and a frame that misses anyone outside the course.
• A frame that selects on the outcome, like a gym lobby after practice, can reverse the answer.
• Get adult approval first, stay outside the eight protected areas, collect permission and assent, make declining invisible and free, and never report a cell under five.
• The response rate is completed forms over everyone sampled: 112 of 150, or 75 percent, in the example.

Sources

1. Pew Research Center. (n.d.). Writing survey questions. Methods 101. pewresearch.org
2. Society for Science. (n.d.). Human participants. In International rules for pre-college science research: Guidelines for science and engineering fairs. societyforscience.org
3. U.S. Department of Education, Student Privacy Policy Office. (n.d.). What is the Protection of Pupil Rights Amendment (PPRA)? studentprivacy.ed.gov
4. Protection of pupil rights, 20 U.S.C. 1232h. Legal Information Institute, Cornell Law School. law.cornell.edu
5. American Sociological Association. (2018). Code of ethics (Standards 11, 11.2 and 12.3). asanet.org

112 forms and one confident paragraph

The invented study from Lesson 24 ends with 112 completed questionnaires in a sealed box, collected from six randomly drawn English sections at Ridgeway High, a school that does not exist. Here is the first draft of the results paragraph, as a student might write it on the evening the tally is finished:

Our survey proves that jobs are pulling students out of school life. Of the students who do no activities, 51 percent have jobs. Students with jobs are 25 percent less likely to take part, and seniors are the worst: only 38 percent of working seniors take part in anything. With a margin of error of plus or minus 5 points, these results apply to teenagers across the country, and the school should limit how many hours students may work.

Every number in that paragraph comes straight from the tally, and the arithmetic behind each one is correct. The paragraph still contains seven errors, at least one in every sentence. This lesson rebuilds the analysis from the unopened box, and each time the rebuilt version reaches a claim from the draft, it marks exactly where the reasoning breaks. At the end is a complete write-up that says what the forms support and nothing more.

Before counting: a codebook and a data sheet

Open the box and write a number from 1 to 112 in the corner of each form as it is unfolded. The number is linked to no person; it only lets you find a form again when something looks wrong. Then write a codebook, a list of every variable with the question it comes from and a code for each possible answer.

Questions 4, 6 and 7 follow the same pattern. Two codes do most of the work. The American Association for Public Opinion Research, AAPOR, advises that a question with no response get a different value from answers such as none, don't know or prefer not to say, and the same logic separates 8 from 9 here. A student who did not work has no hours to report, which is an answer. A blank where hours belong is missing data.

Enter each form as one row of a spreadsheet or of squared paper, one column per variable, and then check the rows for answers that contradict each other. Decide the rule for a contradiction before you look at any results, and write it down, so that your hopes cannot choose the rule. At Ridgeway, two forms said no to question 2 but still ticked an hours box. Under the rule written in advance, both were coded 9 on WORK and HOURS: each form contradicts itself, and picking the answer you prefer would put your judgement into the data. Finally, have a classmate enter all 112 forms separately and compare the two sheets. In the example the sheets disagreed on 3 forms, all typing slips, each settled by going back to the numbered original. That comparison is a small version of inter-rater reliability.

Counting: the frequency table and its denominators

Now count. Tally marks on paper or a count function in a spreadsheet give the same product, a frequency table: how many students gave each answer, and what percentage that is of the students who answered.

Three habits are visible in that table, and each prevents a real mistake. First, every percentage has a stated denominator. Workers were 36.4 percent of the 110 students who answered question 2, not of the 112 who returned forms, which would give 35.7, and not of Ridgeway's 480 juniors and seniors, most of whom were never asked. The denominator for hours is the 40 workers alone. Second, the missing answers are shown rather than silently dropped, so a reader can judge how much they could matter. Third, the categories 16-20 and more than 20 were merged because the second held a single student. Lesson 7's rule, no cell below five, applies to every table you publish, and merging is how you keep it without throwing data away.

The rest of the tally, briefly: on question 1, 50 students said a job mostly helps, 37 that it mostly gets in the way, 23 were not sure and 2 left it blank; and 21 students said they had stopped going to a school group during the year.

Mistake 1: the percentage that runs the wrong way

The research question compares students who worked with students who did not, so the table that answers it is a cross-tabulation of work by taking part, for the 110 students who answered both questions. The rule is to compute percentages within each category of the independent variable, then compare across. With work in the columns, each column totals 100 percent, the layout AAPOR describes as typical.

Read across the Yes row: 45 percent of workers took part, against 70 percent of students who did not work. That comparison answers the research question. The draft's figure came from percentaging the same counts the other way, along each row: of the 43 students who did not take part, 22 had worked, which is 51.2 percent.

That number answers a different question, what non-participants are like, and it depends heavily on how common jobs are. Suppose only 20 of the 110 students had worked, and the two groups took part at exactly the same rates as before, 45 and 70 percent. Then 11 workers and 27 non-workers would not take part, and workers would make up 11 of the 38 non-participants, about 29 percent. The relationship would be identical, and the draft's statistic would have fallen from 51 to 29. A figure that moves when the relationship does not is measuring something else.

The core of it: Percentage within the groups you are comparing, then compare across. A percentage run the other way is not wrong arithmetic; it is the answer to a question nobody asked.

Mistake 2: percentage points are not percent

The gap between 70.0 and 45.0 is 25 percentage points, the quantity Lesson 5 taught you to keep separate from proportional change. In proportional terms, 45 is about 64 percent of 70, so workers took part at roughly two thirds of the non-workers' rate: about 36 percent less likely, not 25. Turned round, 55 percent of workers did not take part against 30 percent of non-workers, so workers were about 1.8 times as likely to have stayed away. All three statements are true. "25 percent less likely" is not, because it attaches the size of the point gap to the language of proportions. Say which measure you are using, every time.

Mistake 3: the wrong margin of error

The draft's plus or minus 5 points would need about 400 respondents, by the rough rule from Lesson 4 that a percentage from n cases carries a 95 percent margin of about 1 divided by the square root of n. For all 112 forms the rule gives about 9 points. But the claim being made is about two subgroups, and Pew's Andrew Mercer states the key fact plainly: estimates for subgroups rest on fewer cases, so their margins are larger, sometimes much larger.

The margin that matters is the one on the gap, and it is neither 16 nor 12. The Census Bureau's handbook for the American Community Survey tests a difference between two estimates by combining their errors: square each one, add the squares, and take the square root. Done with the two margins, that is the square root of 16 squared plus 12 squared, 256 plus 144, which is 400, so about 20 points. The observed gap of 25 points is larger than 20, so under simple random sampling a gap this size would rarely come from the luck of the draw alone.

Notice that the two plausible ranges overlap, from 58 to 61 percent. A common shortcut would call the difference inconclusive for that reason. The Census Bureau warns against exactly this: overlapping intervals are not a reliable test of statistical significance. Test the difference itself.

Then add the caveat the draft never considered. The rough rule assumes 110 separate random draws, and Ridgeway drew six whole classes. Classmates resemble each other, so six sections carry less independent information than 110 scattered names, and the true margin is wider than 20 points by an amount that six clusters are too few to estimate well. The same handbook lists errors that no margin covers at all: problems in the frame or the questionnaire, mistakes in coding, and nonresponse bias. The honest summary is that the gap is larger than the rough margin, which is encouraging, and that the real margin is larger than the rough one, which is why the result is suggestive rather than settled.

Mistakes 4 and 5: one dramatic cell, and a cause nobody tested

"Only 38 percent of working seniors take part" is a true count, 10 of 26, offered with nothing to compare it to. The comparison it needs is non-working seniors, and the confounding question from Lesson 6 needs the same table: seniors at Ridgeway both worked more and took part less than juniors, so part of the overall gap might be a gap between grades rather than between workers and non-workers. Splitting the cross-tabulation by grade holds grade constant.

Read it in three moves. Seniors did work more, 26 of 53 against 14 of 57 juniors, and did take part less, 49 percent against 72. Within each grade, workers still took part about 20 points less often than non-workers. So grade accounts for some of the overall gap, roughly 4 or 5 of its 25 points, and not the rest. Now the caution: the cells are small. The rough margin for the 14 working juniors is about 27 points, and neither grade's gap is large enough on its own to rule out chance. What carries weight is that the two grades, measured separately, point the same way by about the same amount. A third variable can do more than shrink a relationship: in Simpson's paradox the direction reverses once the groups are split, which is why the split is worth making even when you expect nothing.

The draft's first sentence holds the larger error. "Proves that jobs are pulling students out" is a causal claim, and this survey measured work and taking part at one moment. Lesson 6's alternatives all remain open. Reverse causation: a student cut from a team, or tired of the band, may take a job afterwards, and that sequence produces exactly the same table. Question 9 gives a glimpse. Of the 21 students who had stopped going to a school group during the year, 6 named a job as the main reason, 6 said they had lost interest, and 9 gave other reasons, grouped because none of those reasons was given by five or more students. So the job-first sequence happens; the survey cannot say how often, or whether it is the usual order. Confounding: family money could push a student toward work and also make activity fees, equipment and late buses harder to manage. It was not measured, deliberately, because income is a protected area. That is a trade-off to report, not a flaw to hide.

A design that could put the events in order would survey the same students twice, in September and again in March, and see whether those who started jobs then left activities, or the reverse. That is a panel or longitudinal design, and it is the next study, not a sentence to add to this one.

Mistakes 6 and 7: the whole country and the school board

"These results apply to teenagers across the country" confuses the sample's frame with the world. The frame was Ridgeway's juniors and seniors taking English, in one week in March. Whether the pattern holds for sophomores, for the fall sports season, for another school or for American teenagers is a question of external validity, and nothing in these forms answers it. National surveys can speak for the country because they sample the country; this one sampled six classrooms.

The last clause, that the school should limit students' working hours, is not a finding. It is a recommendation that depends on values the survey did not measure and consequences it did not study: why students work, what they earn and need, and what they would lose. The data could inform that argument, and people on opposite sides of it could cite the same table. The report's job is to state the pattern and its limits clearly enough that everyone arguing is arguing from the same facts.

The upshot: A one-time school survey can describe a pattern among the people it sampled, in the week it sampled them. Causes, other places and policies need other evidence.

What the numbers can and cannot say

Three perspectives read the same gap

A number does not interpret itself, and the three perspectives from Lesson 3 read this one differently. The useful part is that each points to a different next measurement.

A complete model write-up

Here is the whole report of the invented study, in the order a reader expects. It carries the items AAPOR lists for transparent reporting, the sample size, the margin of error, the full text of the questions, the mode, the population and how the sample was built and recruited, and it follows the reporting standard in the American Sociological Association's code of ethics: report findings fully, whether or not they support what you expected, state every relevant qualification, and disclose the methods and measures behind them.

Paid Work and School Activities Among Ridgeway High School Juniors and Seniors. A practice report for High School Sociology. The school and all data are invented.

Summary. In a one-week survey of 112 juniors and seniors in six randomly drawn English sections, 45 percent of students who had worked for pay in the previous seven days had taken part in a school club, team or arts group that week, compared with 70 percent of students who had not worked. The 25-point gap was about 20 points within each grade. Because work and participation were measured at the same time, the survey cannot show whether work reduces participation, whether leaving an activity leads students to work, or whether something else produces both.

Question. Role conflict, the clash between the expectations attached to two statuses, predicts that students who hold jobs will take part in school activities less often; the idea that energetic students take on both predicts no gap. The study asked: among Ridgeway's juniors and seniors, were students who worked for pay in the last seven days less likely than other students to take part in a school club, team or arts group in the same seven days?

Method. The population was Ridgeway's 480 juniors and seniors. Because every junior and senior takes one English class, 6 of the 20 English sections were drawn at random in front of the supervising teacher, and all 150 students in them were invited. The ten-item paper questionnaire, printed in two versions with one list in reverse order and pretested with four students from sections not drawn, took about four minutes. Work for pay included babysitting, yard work and paid work in a family business; taking part meant going to at least one practice, meeting, rehearsal, game or performance of a school group. The supervising teacher and the principal approved the plan; parents received a permission slip with the questionnaire attached a week in advance, and students gave assent on the day. No names or identifying numbers were collected. All six sections were surveyed on one day in an ordinary week in March. Of the 150 students, 121 returned signed permission, 6 of those were absent and 3 handed in blank forms, leaving 112 completed questionnaires, a response rate of 75 percent. The respondents' grade split, 52 percent juniors, matched that of the six sections. The data are unweighted. Two forms that contradicted themselves on the work questions were coded as missing under a rule written before data entry, and two students entered all forms independently and resolved 3 disagreements against the originals.

Results. Of the 110 students who answered the work question, 40, or 36.4 percent, had worked for pay in the previous seven days, 27 of them for 10 hours or fewer; 68 of all 112, or 60.7 percent, had taken part in a school group. Table 1 compares participation by work status, within each grade and overall.

Using the rough rule that a percentage from n cases has a margin of about 1 divided by the square root of n, the two overall percentages carry margins of about 16 and 12 points, and their difference a margin of about 20 points, so the 25-point gap is larger than sampling chance would usually produce under simple random sampling. Because whole classes were sampled, the true margin is wider, and the gap should be read as suggestive. Within each grade the groups are small, with a margin of about 27 points for the 14 working juniors, and neither within-grade gap is distinguishable from chance on its own; that the two grades show similar gaps is the stronger evidence. Of the 21 students who had stopped going to a school group during the year, 6 named a job as the main reason, 6 had lost interest, and 9 gave other reasons, grouped because none was given by five or more students.

What the results mean. The data fit the prediction from role conflict: in this week, students who worked were less likely to take part, and the difference was not simply a matter of seniors both working more and taking part less, since it appeared within each grade. The data cannot establish that work caused the difference. Students who left an activity for other reasons and then took a job would produce the same table; that 6 of the 21 who stopped named a job shows the job-first sequence occurs, but not how often. Family finances, which could affect both work and the costs of activities, were deliberately not measured, because family income is a protected area for school surveys.

Limitations. The results describe Ridgeway's juniors and seniors in one March week. Activity seasons and work hours change across the year, and nothing here describes sophomores, other schools or teenagers in general. The sample contained only six clusters, so its true sampling error is larger than the rough margins above. A quarter of the sampled students did not respond; if students absent for team travel, or whose families did not return the slip, differed from respondents, the percentages are biased in ways the survey cannot measure. All answers are self-reports about a single week, and hours were reported in bands.

Ethics. Participation was voluntary, with parental permission and student assent. Every participant handed in a folded form, so declining was invisible; no identifying information was collected; no reported cell holds fewer than five students; and the forms will be shredded at the end of the semester.

Next step. Surveying the same students in September and again in March would show whether starting a job usually comes before leaving an activity, which a one-time survey cannot.

Attached. The questionnaire in Forms A and B, the codebook, the field log and the permission slip.

Notice what the report does not do. It never says proves, it gives the denominator for every percentage, it states the margin and why the true one is wider, it reports the within-grade figures even though they are less dramatic than the headline, and it names the variable it chose not to measure. Had the gap come back at zero, the same report, with the same method section, would have been just as complete, because a result that contradicts the prediction still adds to what is known.

Common misconceptions

• "If the results do not support my hypothesis, the study failed." A clean result of no difference answers the question as fully as a gap does. The ASA code requires reporting results whether they support or contradict the expected outcome, and OpenStax makes the same point: contradicting results still add to understanding.
• "It does not matter which way you percentage a table, since the counts are the same." The two directions answer different questions, and in the example the wrong one fell from 51 to 29 percent while the relationship stayed fixed.
• "If the two ranges overlap, there is no real difference." The Census Bureau warns that overlapping intervals are not a reliable test. Compute the margin on the difference.
• "Beating the margin of error proves the finding." The margin covers only chance in drawing the sample. Frame, questionnaire, coding and nonresponse errors sit outside it, and so does every question about cause.
• "Leaving out the awkward numbers makes a report clearer." It makes it misleading. The within-grade figures are less dramatic than the headline, which is exactly why they belong in the report.

Looking back

• Number the forms, write a codebook that separates does not apply from no answer, fix the rule for contradictions before looking at results, and have a second person enter the data.
• Every percentage needs its denominator; show missing answers; merge any cell below five.
• Percentage within the independent variable and compare across: 45 against 70 percent answers the question, and 51 percent of non-participants does not.
• Say whether a gap is in percentage points or in percent: here 25 points, or about 36 percent less likely.
• Subgroup margins are large; the margin on a difference combines the two, about 20 points here; overlapping ranges are not a test; and a cluster sample's true margin is wider still.
• Hold a third variable constant by splitting the table: the gap survived within each grade at about 20 points.
• A one-time survey cannot order events or establish causes; a panel that surveys the same students twice could.
• A complete write-up gives the question, method, results, meaning, limitations and ethics, and reports what came back rather than what was hoped for. The same habits read every table in this course, from Durkheim's provinces to the Census Bureau's quintiles.

Sources

1. American Association for Public Opinion Research. (n.d.). Best practices for survey research. aapor.org
2. U.S. Census Bureau. (2020). Understanding and using American Community Survey data: What all data users need to know (Chapter 7, Understanding error and determining statistical significance). census.gov
3. Mercer, A. (2016, September 8). 5 key things to know about the margin of error in election polls. Pew Research Center. pewresearch.org
4. American Sociological Association. (2018). Code of ethics (Standard 12.4, Reporting on research). asanet.org
5. OpenStax. (2021). Approaches to sociological research. In Introduction to Sociology 3e. Rice University. openstax.org