STATISTICS FOR BEHAVIORAL SCIENCES

If you will ever read a research paper, run a study, or argue about a poll, this book gives you the grammar. It is written for psychology and behavioral science students who need to compute, interpret, and critique statistics, not just recognize them. Each chapter builds directly on the last, so by the end you can trace a single logic from a sample mean all the way to a confidence interval or an F-ratio.

1. 1Chapter 1

   Population, sample, and the gap that never closes

   A psychologist wants the average anxiety level of all college students in the United States. Measuring every single one is impossible, so the sample becomes 200 students from a single university. The psychologist computes a statistic from those 200 and hopes it describes a parameter of the whole population. The chapter distinguishes descriptive statistics ('statistical procedures used to summarize, organize, and simplify data') from inferential statistics ('techniques that allow us to study samples and then make generalizations about the populations from which they were selected'), the two general categories of statistical techniques. Between the sample statistic and the population parameter sits sampling error, a naturally occurring discrepancy that no clever design can eliminate.

   The idea

   Every research conclusion is a translation from a small group to a big one, and translation always loses something. Sampling error is not a mistake you made; it is the price of looking at only part of a population. Once you accept that the gap is permanent, the rest of the book makes sense: each technique is a different way of measuring, bounding, or living with that gap.

   Try this: Take any headline statistic you saw this week and write down what population it claims to describe, who was actually sampled, and what could differ between the two.

2. 2Chapter 2

   Organize the pile: frequency distributions

   Before any test or theory, you have a heap of scores. The book shows the first move: an organized tabulation of the number of individuals located in each category on the scale of measurement. From that table, proportion falls out immediately as f divided by N, and percentage is just that proportion times 100.

   The idea

   A frequency distribution is the bridge between raw data and every statistic that follows. Proportions and percentages are not decoration; they are the first honest summary of who landed where. Getting comfortable with f/N early means the later ideas about probabilities and expected frequencies feel familiar rather than foreign.

   Try this: Tally one category of your own data this week, emails by sender or steps by day, and compute the proportion for your top category.

3. 3Chapter 3

   The mean is a balance point

   Three contenders fight to represent a distribution. The mean is the sum of the scores divided by the number of scores, written M = ΣX/n for a sample and μ = ΣX/N for a population. The book states: 'The mean is the balance point of the distribution,' because the total distance below the mean equals the total distance above it. The median holds the midpoint, and the mode simply claims the most frequent score. Each measure answers a slightly different question about where the center sits.

   The idea

   Thinking of the mean as a seesaw's fulcrum explains why one extreme score drags it while leaving the median unmoved. The balance-point picture also previews the logic of deviations, which return later as the raw material for variance. Choosing among mean, median, and mode is choosing what kind of typical you want to claim.

   Try this: Compute the mean and median of your last ten purchases and notice which one the single biggest purchase moves.

4. 4Chapter 4

   Variability: how spread out is the crowd?

   Two classes can share the same average and still be wildly different places to sit. The range, the difference between the largest score and the smallest score, gets flagged as a crude measure because it listens only to the extremes. The real work is variance, the mean of the squared deviations: σ² = Σ(X−μ)²/N for a population, and for a sample s² = SS/(n − 1). The computational formula for SS is ΣX² − (ΣX)²/n. Standard deviation is the square root of variance, bringing the answer back to score units.

   The idea

   Dividing by n − 1 instead of n corrects for the bias in using sample data to estimate population variability. Squared deviations feel abstract until you take the square root and land back in units you can picture.

   Try this: Compute SS with the computational formula for five small numbers and verify it matches the sum of squared deviations done the long way.

5. 5Chapter 5

   z-scores put every score on one map

   One student scores 76 on a psychology exam; another scores 82 on a biology exam. Which performance is better? Raw numbers cannot say, because the exams have different means and spreads. The book converts each score with z = (X − μ)/σ, or z = (X − M)/s for a sample: the sign tells you direction from the mean, the number tells distance in standard deviations. Transforming X values into z-scores creates a standardized distribution with a mean of 0 and a standard deviation of 1, and the shape of the distribution is unchanged. The chapter uses the example of comparing two students' scores on different exams—one student scoring 76 on a psychology exam and another scoring 82 on a biology exam—to show how z-scores allow comparison of scores from different distributions.

   The idea

   A z-score is an address: same shape, new coordinates. That is what makes cross-distribution comparisons legitimate, because both scores now speak the language of standard deviations.

   Try this: Find the mean and standard deviation of two different scales you care about, like your sleep hours and your screen time, and convert one recent value of each to a z-score.

6. 6Chapter 6

   Probability as proportion of the curve

   Probability gets a plain definition: a fraction or a proportion of all the possible outcomes, computed as the number of outcomes classified as A over the total number of possible outcomes. For a normal distribution, that proportion is literally area under the curve. The unit normal table becomes the tool of the trade, listing proportions that correspond to each z-score location, so you can find the probability of a score above, below, or between given z values.

   The idea

   This chapter is where location becomes likelihood. Because the z-scores from Chapter 5 standardize any normal distribution, one table serves them all.

   Try this: Use a z-table to find the proportion of scores above one standard deviation above the mean, and check it against the 68-95 rule you may already know.

7. 7Chapter 7

   The distribution of sample means and the central limit theorem

   Take every possible random sample of size n from a population and collect all their means: that set is the distribution of sample means, and its center sits exactly at the population mean, μM = μ. The central limit theorem then delivers the payoff: for any population with mean μ and standard deviation σ, the distribution of sample means has mean μ, standard deviation σ/√n, and approaches normal as n grows; with n ≥ 30 it is almost perfectly normal. That standard deviation gets its own name, the standard error of the mean, and it measures the standard distance between a sample mean and the population mean. Bigger samples shrink it, which is the law of large numbers in action.

   The idea

   This is the hinge of the whole book: it explains why a single sample mean can say anything trustworthy about a population. The standard error turns an abstract worry, how far off might my sample be, into a computable number.

   Try this: Simulate 30 samples of size 5 and size 30 from any dataset, plot the two sets of means, and watch the spread collapse as n grows.

8. 8Chapter 8

   Hypothesis testing: four steps, two ways to be wrong

   The chapter introduces the four steps of hypothesis testing: (1) state the null and alternative hypotheses, (2) set the criteria for a decision (alpha level, typically .05 or .01), (3) compute the test statistic, and (4) make a decision to reject or fail to reject the null hypothesis (reject H0 if the test statistic is in the critical region). The null hypothesis (H₀) is defined as 'a statement about the population that predicts no effect or no difference'—'the hypothesis that states there is no change, no effect, or no difference'—and the alternative hypothesis (H₁) predicts that there is an effect or difference. To make the stakes human, the book stages a court trial where the null hypothesis is that the defendant is innocent: convicting an innocent person is a Type I error, acquitting a guilty one is a Type II error. The alpha level (level of significance), typically α = .05, is defined as 'a probability value that is used to define the concept of “very unlikely” in a hypothesis test.' It sets the risk of a Type I error and defines the boundaries of the critical region for rejecting the null hypothesis. Beta is the risk of the second, and power, the probability of correctly rejecting a false null, equals 1 − β and grows with bigger samples and bigger effects.

   The idea

   The trial analogy makes the asymmetry unforgettable: the system presumes innocence, so a false alarm and a miss are different crimes with different costs.

   Try this: For your next decision, write the null hypothesis in plain words and name what a Type I and a Type II error would each cost you.

9. 9Chapter 9

   When σ is unknown: the t statistic

   Real research rarely hands you the population standard deviation, so the z-test's key ingredient is missing. The fix is to estimate σ from the sample's own s, giving an estimated standard error sM = s/√n and the statistic t = (M − μ)/sM. Because you estimated, you pay in uncertainty: the t distribution is a flatter, more variable approximation of the normal, and its exact shape is set by degrees of freedom, df = n − 1.

   The idea

   The t statistic is honesty built into a formula: substituting an estimate for a known value widens the distribution to admit the extra uncertainty.

   Try this: Look up the critical t for df = 5 and df = 30 at α = .05 and compare each to 1.96 to see the price of small samples.

10. 10Chapter 10

    Two independent samples and pooled variance

    Now the question changes shape: two separate groups, two means, one comparison. The independent-measures t statistic divides the difference M₁ − M₂ by an estimated standard error that combines the error from both samples, and the null hypothesis is written H₀: μ₁ − μ₂ = 0. When the two samples differ in size, the book introduces pooled variance, a weighted average of the two sample variances: s²p = (SS1 + SS2)/(df1 + df2).

    The idea

    Combining error from both samples acknowledges that each mean carries its own uncertainty, and the null hypothesis of a zero difference gives the test a precise target. This design also sets up the contrast with repeated measures that comes next.

    Try this: Take two small lists of numbers with different lengths, compute each SS, pool the variances, and confirm the longer list dominates the estimate.

11. 11Chapter 11

    Same people, twice: the repeated-measures t

    Instead of two crowds, imagine one crowd measured under both treatments. The repeated-measures t works on difference scores, D = X2 − X1, testing H0: μD = 0 with t = (MD − μD)/sMD and df = n − 1, where n counts pairs. The design's hidden advantage is that individual differences cancel: because the same subjects appear in both treatments, the analysis eliminates the problems created by individual differences. Less noise from person-to-person variation means a real treatment effect is easier to detect.

    The idea

    Every participant is their own control, so the quirks that make one person high on both measures subtract out of the difference scores. That variance reduction is free power, which is why repeated-measures designs tempt researchers with small samples.

    Try this: Track one personal metric before and after a small change this week, compute your difference scores, and see how much steadier they are than raw values.

12. 12Chapter 12

    ANOVA and the F-ratio: comparing many means at once

    Three or more treatments make pairwise t tests clumsy, so analysis of variance steps in, using sample variances as the basis of its analysis to test mean differences among groups simultaneously. The machinery is sums of squares: SStotal = ΣX² − (G²/N) with G the grand total, SSwithin as the sum of SS inside each treatment, SSbetween = SStotal − SSwithin, and degrees of freedom dfbetween = k − 1, dfwithin = N − k, with MS = SS/df. The F-ratio is the ratio of the variance between treatments to the variance within treatments, F = MSbetween. An F-ratio near 1.00 indicates no treatment effect.

    The idea

    ANOVA reframes the question: instead of asking whether means differ directly, it asks whether variance between groups exceeds what variance within groups would predict by chance. The F-ratio's baseline of 1.00 gives you an instant sanity check before any table lookup.

    Try this: Split any dataset of yours into three groups, compute the variance of group means versus the average within-group variance, and eyeball your own F-ratio.

13. 13Chapters 13-14

    Repeated-measures ANOVA and two factors at once

    The same-participants design scales up: in repeated-measures ANOVA, individual differences are measured and removed from the denominator of the F-ratio, sharpening the test. Then the design grows a second independent variable, and two-factor ANOVA produces three separate F-ratios: a main effect of factor A, a main effect of factor B, and the interaction between them. An interaction occurs when the effects of one factor depend on the different levels of a second factor, and the book recommends graphing the cell means to see it.

    The idea

    Removing individual differences from the denominator is the ANOVA version of the repeated-measures t's advantage, applied to many conditions. The interaction is often the real finding, because main effects can hide stories where a treatment helps at one level and hurts at another.

    Try this: Sketch two lines on a graph, one for each level of factor B across levels of factor A, and decide for yourself whether they tell an interaction story.

14. 14Chapters 15-16

    Correlation, regression, and the line that fits

    Two variables, one scatterplot: the Pearson correlation measures the direction and degree of the linear relationship, running from −1.00 for a perfect negative to +1.00 for a perfect positive, with r = SP/√(SSX·SSY). Squaring it gives the coefficient of determination, r², the proportion of variance in one variable accounted for by its relationship with the other. For ordinal data, or monotonic but nonlinear patterns, Spearman's rank-order correlation takes over. The book is blunt about limits: correlation does not demonstrate causation, and a third-variable problem or a reverse cause-and-effect relationship may explain what you see. Regression then builds the prediction line Ŷ = bX + a by least squares, with slope b = SP/SSX and intercept a = MY − bMX, and the standard error of estimate measures the standard distance between the line and the actual data points, falling to zero only when r = ±1.00.

    The idea

    Correlation and regression are two views of one relationship: r describes its strength and direction, while the regression equation turns it into predictions.

    Try this: Pick two variables you can log daily for a week, compute a rough correlation, then list one third variable that could explain the link before you believe it.

15. 15Chapters 17-18

    Chi-square and the binomial: testing categories

    Not every datum is a number on a scale; some are just categories. The chi-square statistic, χ² = Σ(fO − fE)²/fE, compares observed frequencies with expected frequencies for nominal data. The goodness-of-fit test asks whether sample proportions match a specified population distribution, with df = k − 1 across k categories. The test for independence asks whether two categorical variables are related, working from a matrix of observed frequencies with df = (R − 1)(C − 1) and expected cell frequencies fe = (row total × column total)/N. For the special case of exactly two categories, the binomial test applies, and the chapter's example is a researcher flipping a coin many times to see whether the observed heads match the expectation under P = Q = 0.50.

    The idea

    These tests free you from means and variances entirely: the data are counts, and the question is whether the counts fit a claim.

    Try this: Flip a coin 20 times, compute the expected count of heads under P = 0.50, and calculate how far your observed count strays.

Psychology and behavioral science students who need to pass a statistics course without losing their minds, especially those who fear formulas more than failure. Also ideal for researchers and clinicians who run studies but want to finally understand what their software is actually doing under the hood.

FAQ