Statistical Methods for the Social Sciences

If you read polls, run surveys, or evaluate programs, this book shows you how to tell signal from luck. It suits social science students, policy analysts, and self-taught data readers who want rigor without measure-theory proofs. Agresti's examples come from real research questions, so every method arrives attached to a problem you might actually face.

1. 1Chapters 1–3

   Start with the data you can see

   A student opens the book expecting formulas and instead gets a question: what do these numbers mean? Agresti begins with description—histograms, means, medians, standard deviations—using real social science data on things like income and education. Before any inference happens, you learn to summarize a messy sample into a few honest numbers. The payoff comes when those summaries stop being arithmetic exercises and become the evidence base for every later claim.

   The idea

   Description is not a warm-up act; it is the foundation of inference. A mean without a sense of spread tells you almost nothing, and a graph can reveal structure no test will flag. Agresti's ordering matters: you cannot judge whether a sample is unusual until you know what typical looks like. Skipping straight to p-values is like grading an exam without reading the answers.

   Try this: Take any dataset you have and compute the mean and median side by side; if they differ much, plot a histogram before believing either.

2. 2Chapter 4–5

   Probability as the language of uncertainty

   The book pivots from describing samples to reasoning about populations, and probability becomes the bridge. Agresti works through sampling distributions—what happens when you draw repeated samples from the same population—so the normal curve stops being decoration and starts doing work. The central limit theorem arrives as the reason means behave predictably even when individual observations do not. By the end of this stretch, randomness has a grammar.

   The idea

   Sampling distributions are the single most important idea in introductory statistics, and the hardest to grasp. Once you see a sample statistic as one draw from its own distribution, confidence intervals and significance tests stop being rituals. The normal distribution matters less because data is normal and more because averages of many observations converge to it. That convergence is what lets social science make claims from modest samples.

   Try this: Simulate 1,000 sample means from any dataset, even by spreadsheet, and watch the histogram of means tighten around the truth.

3. 3Chapter 5–6

   Confidence intervals: a range, not a verdict

   Agresti builds the confidence interval as the workhorse of estimation: a point estimate plus a margin of error, with 95% confidence as the social science default. He shows the interval for a mean, then for a proportion, and stresses what the confidence level actually promises across repeated samples. The stakes are practical—a survey report that gives only a percentage, with no interval, tells you less than it appears to.

   The idea

   A confidence interval is a statement about the method, not about one interval: use the procedure many times and it captures the parameter 95% of the time. This is subtler than the popular reading that a given interval has a 95% chance of containing the truth. Reporting intervals alongside estimates forces honesty about precision, and wide intervals are a warning, not a failure. Estimation usually answers the scientific question better than a bare accept-or-reject decision.

   Try this: Next time you see a poll number, demand the margin of error and compute the interval yourself before drawing a conclusion.

4. 4Chapter 6–7

   Significance testing and its temptations

   Here Agresti walks through the five-step logic of a significance test—assumptions, hypotheses, test statistic, p-value, conclusion—using examples like testing whether a new teaching method changes scores. The null hypothesis gets framed as the skeptic's position: no effect until evidence says otherwise. He shows how the p-value measures surprise under that null, and how a small p-value is a signal, not a proof.

   The idea

   A p-value is the probability of data at least this extreme if the null were true; it is not the probability the null is true. Statistical significance answers whether an effect is distinguishable from noise, never whether it matters. Agresti pushes readers to report effect sizes alongside p-values, because a trivially small effect can be highly significant with a big enough sample. Misreading this is the most common statistical error in the social sciences.

   Try this: For any significant result you encounter this week, ask what the estimated effect size is—and whether anyone would care at that size.

5. 5Chapter 7–8

   Comparing two groups

   The book turns to the comparison that drives most social research: is group A different from group B? Agresti covers the two-sample t test for means and the comparison of two proportions, with confidence intervals for the difference. Examples run through treatment-versus-control setups and demographic contrasts. The outcome is a toolkit for the question every study asks: did something actually change, or does it just look that way?

   The idea

   Comparing groups requires accounting for two sources of randomness at once, so the standard error of a difference combines both samples' variability. Independent samples and matched pairs demand different analyses, and mixing them up distorts your uncertainty. A difference of proportions, reported with an interval, is often more informative than either proportion alone. Most policy debates are secretly about differences, not levels.

   Try this: When you read a claim that one group outperforms another, check whether the reported difference comes with a standard error or interval.

6. 6Chapter 8–9

   Categorical data and the chi-square test

   Agresti's home turf arrives: contingency tables, where rows and columns are categories and the cells hold counts. He builds the chi-square test of independence step by step—expected counts under independence, the statistic that sums squared deviations, degrees of freedom from table shape. An example like party affiliation by gender shows how the test flags association without saying what the association is.

   The idea

   Chi-square tests whether two categorical variables are independent, but a significant result is only the beginning: you must inspect standardized residuals to find which cells drive the association. The test needs adequate expected counts in the cells, a condition students routinely ignore. Association in a table is not causation, and lurking variables can manufacture patterns. This chapter is the seed of Agresti's later career-defining work on categorical data analysis.

   Try this: Cross-tabulate two categorical variables you care about, compute expected counts, and check whether any cell falls below the usual threshold before trusting chi-square.

7. 7Chapter 9

   Regression: one predictor, many uses

   Simple linear regression enters as the method for a quantitative response and a quantitative predictor: fit a line, interpret the slope, test whether the slope is zero. Agresti uses scatterplots first, so you see the relationship before you model it. The correlation coefficient gets reinterpreted as a standardized slope, tying earlier descriptive tools to the inferential machine. The chapter ends with prediction and its honest limits.

   The idea

   Regression's slope answers a precise question: how much does the response change, on average, per one-unit increase in the predictor, holding everything else out of the picture. A strong correlation never establishes causation, and extrapolating beyond the observed range of x is where predictions go to die. The line is a summary, not a law. Residual plots are the diagnostic that tells you whether the model deserves belief.

   Try this: Fit a line to any two quantitative variables you have, then plot the residuals; curvature there means the straight line is lying to you.

8. 8Chapter 11

   Multiple regression: holding other variables constant

   The book's central method arrives: multiple regression, where several predictors share one model. Agresti shows how a slope changes meaning in the presence of other variables—it becomes the effect of one predictor holding the others fixed. Examples of income predicted by education and work experience reveal how a bivariate relationship can shrink, flip, or vanish once a third variable enters. This is the machinery behind 'controlling for' in every social science paper.

   The idea

   The multiple regression coefficient is the workhorse of observational social science because it approximates what an experiment would show—adjusting for measured confounders. But adjustment only works for variables you measured and included; omitted variables keep their bias. Multicollinearity can make individual coefficients unstable even when the model predicts well. Reading a regression table is a skill of interpretation, not arithmetic.

   Try this: Find a regression table in a news article or report and identify which coefficient is the headline claim and what variables it was adjusted for.

9. 9Chapter 12

   Comparing several groups: analysis of variance

   When more than two groups need comparing, Agresti introduces analysis of variance: partition total variability into between-group and within-group pieces, then test whether the between-group part is larger than chance allows. An example comparing mean responses across several treatment categories shows the F statistic at work. The test's significance is only the opening move; follow-up comparisons identify which groups differ.

   The idea

   ANOVA is regression with categorical predictors in disguise, and seeing that equivalence is the book's quiet unification. A significant F test says the group means are not all equal, not which ones differ—so planned comparisons or confidence intervals for pairs must follow. The within-group variability is the yardstick that decides whether between-group differences are impressive. Running many pairwise tests without adjustment inflates the chance of false positives.

   Try this: If you compare more than two groups, plan a follow-up procedure for pairwise differences before you run the overall F test.

10. 10Chapter 14–15

    Combining methods: model building and diagnostics

    Toward the end, Agresti assembles the toolkit: model selection, interaction terms, and diagnostics that check whether a regression model fits. Examples show interaction—when the effect of one predictor depends on the level of another—changing the interpretation entirely. Residual analysis and checks of assumptions get treated as routine hygiene, not optional extras. The outcome is a researcher who can build a model and then interrogate it.

    The idea

    A model is a claim about how the world generates the data, and diagnostics are how you test your own claim. Interaction terms matter because averaged effects can mislead when effects differ across subgroups. Automatic variable selection cannot substitute for thinking about which variables belong in the model. The best analysts spend more time checking models than fitting them.

    Try this: In your next regression, add one interaction term you suspect matters and compare whether the main effects' interpretation changes.

11. 11Chapter 15

    Logistic regression for binary outcomes

    The book closes its modeling arc with logistic regression, Agresti's signature subject: modeling a binary response—success or failure, voted or not—using predictors. He introduces odds and odds ratios, then shows how the model links predictors to the probability of success. An example predicting a binary outcome from continuous and categorical predictors shows the method's reach. The chapter connects back to the contingency tables from earlier chapters, now with the power to adjust for other variables.

    The idea

    Logistic regression extends the chi-square era's two-way tables into models with many predictors, which is why it dominates social science research with binary outcomes. Odds ratios are the natural effect measure, but they are not risk ratios, and confusing them exaggerates effects when baseline risk is high. The model works on the log-odds scale, so coefficients need translating before they mean anything to a reader. This is the gateway to Agresti's advanced work on categorical data.

    Try this: Take one binary outcome you care about and compute the odds ratio between two groups by hand before fitting any model.

Social science students taking their first serious statistics course, and practitioners who want a rigorous but readable reference for regression, ANOVA, and categorical data analysis. Anyone who has memorized formulas without understanding why they work will find the connected explanations clarifying.

FAQ