STATISTICS FOR THE BEHAVIORAL SCIENCES

If you read research in psychology, education, or health sciences, this book decodes the numbers behind every claim you encounter. It is written for students who fear math: each formula arrives with a plain-language reason for existing, from the seesaw mean to the jury-trial logic of hypothesis testing. You finish able to read a journal article's results section and know what each statistic is vouching for.

1. 1Chapter 1

   Populations, samples, and the honest gap between them

   A researcher wants to know something about everyone of interest—say, all college students—but can only measure a handful of them. That full group is the population; the handful is the sample, chosen to represent it. From the sample come statistics, values describing the people actually measured; from the population come parameters, values describing everyone. Descriptive statistics summarize and simplify the data you have. Then comes the uncomfortable part the book puts at the heart of inference: the sample statistics are, by nature, imperfect and variable representations of the population parameters. That built-in discrepancy—sampling error—is why the rest of the book exists.

   The idea

   The entire discipline rests on a confession: your sample will never match the population exactly, and no amount of care removes that gap. What statistics offers instead is a way to measure the gap—to say how wrong a sample statistic probably is. Once you accept that every generalization carries this natural error, the machinery of later chapters stops looking like ritual and starts looking like damage control.

   Try this: Pick one statistic from a news headline this week and ask: was it measured on a sample or the whole population, and how big was the sample?

2. 2Chapter 3

   The mean is a seesaw

   Picture a seesaw with the scores as weights placed along a board. The mean is the balance point of the distribution—the spot where the board tips level, computed as the sum of the scores divided by the number of scores (μ = ΣX/N for a population, M = ΣX/n for a sample). Now slide one weight far out to the end: a single extreme score drags the balance point toward it. That is why the book recommends the median for skewed distributions, since the median is not affected by extreme scores—it is simply the midpoint of the ordered scores, the point below which 50% fall. For categories with no order at all, like eye color, only the mode survives: the score with the greatest frequency, and the only measure of central tendency usable with nominal data.

   The idea

   Choosing a measure of center is really choosing what kind of distortion you can live with. The mean answers to every score, which makes it precise and fragile; the median ignores magnitude, which makes it sturdy under skew; the mode ignores everything but popularity. The scale of measurement decides for you—nominal data leaves no choice at all.

   Try this: Take your last five purchases, compute the mean and the median, and see which one better represents a 'typical' amount.

3. 3Chapter 4

   Standard deviation: the average distance from the mean

   Two classes can share the same mean while one is full of near-identical scores and the other scatters everywhere. To capture that, the book measures distance: subtract the mean from each score, square the deviations, and average them—that is variance, the mean of the squared deviations, σ² = Σ(X−μ)²/N for a population. The sum of squared deviations, SS, can be found with either the definitional or the computational formula. Square-root the variance and you get standard deviation, σ = √[Σ(X−μ)²/N], a measure of the standard, or average, distance from the mean. For samples there is a correction: divide SS by n−1, not n, because sample variability runs biased downward, and the n−1 fix makes the sample variance an unbiased estimate of the population variance.

   The idea

   Variability is not an afterthought to the mean; it is what makes the mean trustworthy or meaningless. A mean of 50 means something different when scores cluster within 2 points than when they spread across 20. The n−1 correction is the book's quiet masterstroke: it admits that a sample underestimates the spread of its population, and it pays a small price in the denominator to buy an honest estimate.

   Try this: Compute the standard deviation of your daily screen time for a week and see how far a 'typical' day strays from your average.

4. 4Chapter 5

   z-scores: putting every distribution on the same ruler

   You score 70 on one exam and 85 on another. Which performance is better? Raw numbers cannot say, because the two exams may have different means and different spreads. The book's fix is the z-score: subtract the mean and divide by the standard deviation, z = (X−μ)/σ. The sign tells you whether the score sits above or below the mean; the number tells you how many standard deviations it sits from it. The stated purpose is to transform X values into z-scores so that the resulting distribution has a mean of 0 and a standard deviation of 1, making different distributions comparable. Your 70 might be 2 standard deviations up while your 85 is barely half a standard deviation up—and now the comparison is honest.

   The idea

   A z-score is a translation, not a transformation of the person. It relocates every score onto one universal scale where zero means average and each unit means one standard deviation. The surprise is how much information the sign and size carry: location within the distribution, not the raw number, is what makes scores comparable across tests, courses, or measures.

   Try this: Convert one of your test scores to a z-score using the class mean and standard deviation, and see where you actually stand.

5. 5Chapter 6

   Probability as a proportion

   Before any hypothesis test can work, the book needs a definition of probability that arithmetic can handle. Its answer is disarmingly plain: probability is a fraction or a proportion of all the possible outcomes. The probability of A equals the number of outcomes classified as A divided by the total number of possible outcomes. No metaphysics, no long-run philosophy—just counting. With that definition in hand, the proportions under a distribution of scores become probabilities of drawing particular values, and the bridge from descriptive curves to inferential decisions is built.

   The idea

   Defining probability as a proportion is what lets a curve of scores do double duty as a statement about chance. If you know the proportion of outcomes in a range, you know the probability of landing there. Every p-value you will ever read descends from this simple act of division: favorable outcomes over all outcomes.

   Try this: For one yes/no decision this week, write out the possible outcomes as a fraction and see whether your gut odds match the arithmetic.

6. 6Chapter 7

   The distribution of sample means and the Central Limit Theorem

   Draw every possible random sample of size n from a population, compute each sample's mean, and collect all those means: that collection is the distribution of sample means, which the book calls a central concept of inferential statistics. The Central Limit Theorem then describes its shape: for any population with mean μ and standard deviation σ, the distribution of sample means will have mean μ, standard deviation σ/√n, and will approach normal as n approaches infinity. It will be normal if the population itself is normal or the sample size is relatively large—at least n ≥ 30. The mean of this distribution equals the population mean, and its standard deviation—the standard error, σM = σ/√n—is a measure of the average distance between a sample mean and the population mean.

   The idea

   This chapter explains why sampling error from Chapter 1 is not chaos but structure. Sample means pile up around the population mean in a predictable, mostly normal pattern, and bigger samples squeeze that pile tighter—divide σ by √n and watch the error shrink. The standard error turns 'samples are imperfect' into a number you can compute, which is exactly what hypothesis tests will spend.

   Try this: Flip 10 coins, average the heads per 10 flips, repeat five times, and watch your five averages cluster near 0.5.

7. 7Chapter 8

   The jury trial of hypothesis testing

   The book stages hypothesis testing as a courtroom. The defendant is presumed innocent, just as the null hypothesis is presumed true until the data argue otherwise, and the jury must decide between two possible verdicts. Two mistakes are possible: convicting an innocent person, which is a Type I error—rejecting a null hypothesis that is actually true—and freeing a guilty person, a Type II error—failing to reject a null that is really false. The alpha level, α, is the probability that the test leads to a Type I error; the beta symbol represents the probability of the Type II error. Every 'significant' result in a journal is a verdict reached under exactly these risks.

   The idea

   The alpha level is not a magic threshold but a pre-set tolerance for convicting an innocent null. Lower α and you convict fewer innocent hypotheses—while freeing more guilty ones, because the two errors trade off. Reading research with this analogy in mind changes the question from 'is it significant?' to 'which error did this study choose to risk, and at what price?'

   Try this: Before reading a study's results, note its alpha level and ask which error—false alarm or missed effect—the researchers chose to minimize.

8. 8Chapter 9

   The t statistic: testing means when σ is unknown

   The z-test of the previous chapter demands something real research rarely provides: the population's standard deviation. In practice you have one sample and no σ. The t statistic fills the gap: it tests hypotheses about an unknown population mean μ when σ is unknown, using the formula t = (M−μ)/sM, where sM is the estimated standard error s/√n. The sample's own standard deviation stands in for the population's, and the estimated standard error stands in for the true one. The logic of the test is unchanged—compare your sample mean against the null's claimed mean—but the yardstick is now estimated from the data you actually have.

   The idea

   The t statistic is the workhorse version of the z-score test, built for the common case where the population is a stranger. Substituting s for σ costs certainty, which is why t distributions are fatter-tailed than the normal—but it buys applicability to nearly every real experiment. One sample, one mean, one honest estimate of error: that is the whole machinery.

   Try this: Find a one-sample t value in a paper you are reading and check that its standard error came from the sample, not a known population.

9. 9Chapters 10–11

   Two samples, two designs: independent or repeated

   Two ways exist to compare two treatments, and the choice changes the test. The independent-measures t test uses two separate samples—a separate group for each treatment condition—to evaluate the mean difference between two populations, with a null hypothesis stating μ1 − μ2 = 0. The repeated-measures t test instead puts each individual in both treatment conditions, or observes each person at two different times, and works with difference scores, D = X2 − X1, evaluating the mean of those differences. Same question—did the treatments differ?—but the data arrive shaped differently: two independent piles of scores versus one column of within-person changes.

   The idea

   Design determines the arithmetic. Separate groups force the test to compare two sample means against each other; the same people in both conditions let the test track each person's change directly. The repeated-measures route quietly removes person-to-person differences from the comparison, which is why researchers who can test participants twice often prefer it.

   Try this: For your next comparison question, decide first whether one group can experience both conditions—it may halve the noise.

10. 10Chapter 12

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

    A researcher with three or four treatments could run t test after t test—but each additional test rolls the dice again, inflating the experimentwise alpha level and raising the risk of a Type I error. Analysis of variance solves this with a single test: ANOVA evaluates mean differences between two or more treatments by computing one F-ratio, F = variance between treatments divided by variance within treatments, or MSbetween over MSwithin. If the treatments truly differ, the variance between them swells beyond the background variance within them, and the ratio grows. One number, one decision, no compounding error.

    The idea

    The F-ratio is a signal-to-noise argument in a single division: differences between treatments on top, ordinary variability among participants underneath. The insight behind ANOVA is that stacking pairwise tests multiplies your Type I error risk, so one omnibus test protects the alpha level you set. Only after F speaks do pairwise comparisons earn their turn.

    Try this: When you see a study comparing three or more groups, check that it reports one overall F before trusting any pairwise claims.

11. 11Chapters 13–14

    Repeated-measures and two-factor ANOVA: removing people, adding interactions

    When the same participants serve in every treatment condition, the repeated-measures ANOVA—presented as an extension of the repeated-measures t test—can strip individual differences out of the error term entirely, because each person acts as their own baseline. Its F-ratio becomes MSbetween treatments over MSerror, with the person-to-person variance subtracted from the denominator. Then the book widens the lens: a two-factor ANOVA evaluates mean differences from two independent variables at once, computing three separate F-ratios—one for the main effect of factor A, one for the main effect of factor B, and one for the A × B interaction. An interaction occurs when the effect of one factor depends on the level of another factor.

    The idea

    These chapters show ANOVA as a flexible architecture rather than a single test. Repeated measures buy power by deleting the variance that comes from people simply being different. Two-factor designs buy realism by testing whether effects hold across conditions—or bend, which is what an interaction reveals and what a main effect alone can never show.

    Try this: In your next data comparison, ask whether the effect of one variable changes across levels of another—that is the interaction worth plotting.

12. 12Chapters 15, 17

    Correlation and chi-square: measuring relationships without means

    Not every question is about mean differences. The Pearson correlation measures the degree and the direction of the linear relationship between two variables, computed as r = SP/√(SSx·SSy) from the sum of products of deviations. Its values run from −1.00, a perfect negative relationship, to +1.00, a perfect positive one, with 0 meaning no linear relationship at all. And when your data are not scores but counts—how many people fall in each category—the chi-square goodness-of-fit test compares the frequency distribution from a sample with a hypothesized distribution stated by the null hypothesis, using χ² = Σ(fo − fe)²/fe, where fo is the observed frequency and fe the expected frequency.

    The idea

    Correlation answers 'do these two variables move together, and how tightly?' while chi-square answers 'do my observed counts fit what the null predicted?' Both abandon means entirely—one works with covariation, the other with frequencies.

    Try this: For one paired set of numbers in your life—hours slept versus mood rating—compute or plot the correlation and check its sign and strength.

13. 13Chapter 12

    Estimation: from yes-or-no to a range of values

    Hypothesis testing gives a verdict; estimation gives a measurement. Instead of only asking whether a treatment worked, you can estimate how much of the population parameter lies where. A point estimate is a single number, like a sample mean; an interval estimate is a range of values—a confidence interval. For an unknown population mean μ, the interval is computed as M ± t(sM): the sample mean at the center, the estimated standard error scaled by t setting the width. The result is a range that states, with a chosen level of confidence, where the population value plausibly sits.

    The idea

    A confidence interval converts a binary decision into a range with width, and the width itself is information: wide intervals confess uncertainty, narrow ones claim precision. Because the interval is built from the same estimated standard error that powers the t statistic, estimation and hypothesis testing are two readings of one computation—reject or not, and by how much.

    Try this: Next time you read a reported mean, look for its confidence interval and judge the study by the interval's width, not just the mean.

Students in psychology, education, nursing, or any behavioral science course that uses Gravetter and Wallnau's text, plus self-learners who want the classical statistics toolkit explained without proofs. Researchers returning to the basics will find the jury-trial and seesaw analogies worth the refresher.

FAQ