Comparing the averages of two separate groups — and deciding whether the gap is real.
Up to now you've asked, "is this one group different from a number?" Now comes a question you'll use for the rest of your life: do these two groups differ? Treatment vs. placebo. Commuters vs. residents. One menu vs. another. People paid a little vs. people paid a lot. The independent-samples t-test is the tool for the most common comparison in all of social science — two separate groups of people, one mean apiece — and Google Sheets does the arithmetic in a single line.
T.TEST formula.An independent-samples t-test compares the means of two separate groups — different people in each one. Nobody is in both. You already know how to find one mean. Here you find two means and ask the key question: are they far enough apart to be real, or is this just random wiggle?
Picture the most ordinary study you can: a researcher wants to know whether commuters report more stress than residents who live on campus. Two clearly separate sets of people. Each person sits in exactly one group, gets measured once, and contributes one stress score. We average each group, look at the gap between the two averages, and ask whether that gap is bigger than the everyday noise you'd expect from grabbing two random handfuls of students. That is the whole job of the independent t-test.
| Term | Plain English |
|---|---|
| Independent samples | Two groups, different people in each — nobody is in both |
| t | How far apart the two means are, measured in "noise units" |
| df | Degrees of freedom — here, n₁ + n₂ − 2; sets which t-curve to use |
| p | The chance you'd see a gap this big if the groups were really the same |
| Cohen's d | How big the difference is, in plain size terms |
commuters versus residents — two groups made of different students. Which test fits?Before you run anything, ask one question: are the two sets of scores from different people, or from the same people twice? Everything hinges on it. Two separate groups (commuters vs. residents; $1 group vs. $20 group) call for the independent t-test. The same people measured twice (mood before vs. after a workshop; each student's score in spring vs. fall) call for the paired t-test — the subject of the next unit. The math differs because paired scores come in linked couples, and the independent test treats every score as its own free-floating observation.
| Independent-samples t | Paired-samples t | |
|---|---|---|
| Who's in each column? | Different people in each group | The same people, measured twice |
| Scores linked? | No — two unrelated piles | Yes — each row is one person's pair |
| Classic example | Commuters vs. residents; $1 vs. $20 | Mood before vs. after a treatment |
| Equal group sizes? | Not required (n₁ ≠ n₂ is fine) | Always equal — every score is paired |
| This unit? | ✓ Yes | Unit 12 |
Could a single person appear in both columns of your data? If no (a person is either a commuter or a resident, never both) → independent. If the two columns are the same people at two times → paired. That one question settles the choice.
Here's the whole idea in one fraction:
t = (difference between the two group means) ÷ (standard error of that difference). The top is the gap between the averages — the signal. The bottom is the noise within the groups — the standard error, which measures how much two group means would bounce around by chance alone.
So t is a signal-to-noise ratio. The same one-point gap means very different things depending on the spread inside the groups. If everyone in each group clusters tightly around their average, a one-point gap is huge — clearly bigger than the wiggle. If the scores are wildly scattered, that same one-point gap could easily be random. The standard error in the denominator captures exactly that scatter (and shrinks as your samples get larger, because big samples give steadier means).
You don't grind through the standard-error formula by hand — Sheets and the calculator below do it. Just hold onto the intuition: bigger t = stronger evidence the two groups truly differ.
The t-test asks one visual question: is the gap between the two means large compared with the spread of scores within each group? Tightly packed curves with a clear gap → big t. Wide, heavily overlapping curves → small t. The overlap is the noise; the gap is the signal.
To turn your t value into a p value, the test needs to know which t-curve to look at — and that's set by the degrees of freedom (df). For an independent-samples t-test with two groups, the rule is short:
df = n₁ + n₂ − 2. Add the two group sizes, then subtract 2 — one for each group mean you had to estimate. With 20 people per group: df = 20 + 20 − 2 = 38.
Why subtract two? Each group's mean "uses up" one degree of freedom (once you know a group's mean and all but one of its scores, the last score is locked in). Two groups, two means, so you lose two. This is the small but important contrast with the one-sample test from Unit 10, where you estimated just one mean and so subtracted only one (df = n − 1). Degrees of freedom feed straight into the critical value: more df → a slightly more forgiving cutoff, because larger samples pin the means down better.
The independent t-test earns its trustworthy p value only if a few background conditions roughly hold. They're not fussy hurdles — they're just the situations where the test's math makes sense.
Break independence and the whole p value is wrong — that's non-negotiable. Mild violations of normality or equal variance usually just nudge the result, and good software can correct for unequal spread. Check the assumptions before you trust the verdict, not after.
The p value is the chance of seeing a gap this big if the two groups were actually identical. A small p means "this gap would be a freak accident if nothing were going on," so we believe the difference is real. The line everyone uses is p < .05. Below it, we call the difference statistically significant and we reject the null hypothesis (the claim that the two group means are equal).
One precise habit worth building now: when p is not below your cutoff, you fail to reject the null — you do not "accept" it or "prove the groups are the same." A non-significant result just means you didn't gather enough evidence to rule out chance; the groups could still differ in a way your study was too small to catch. "Fail to reject" keeps you honest about what the test can and cannot show.
A significant result tells you the difference is real — not how big it is. For that we use Cohen's d, the gap between the means measured in standard-deviation units. It strips out sample size, so it answers a different question than p: not "is there an effect?" but "how much of one?"
| Cohen's d | Size |
|---|---|
| around 0.2 | small |
| around 0.5 | medium |
| around 0.8+ | large |
Two questions, two answers. p asks is it real? Cohen's d asks how much? A study can be statistically significant yet tiny (huge samples make even trivial gaps significant), so always report both. p says real; d says how big.
Let's run a clean one start to finish. Group A (a campus-life workshop) has mean wellbeing x̄ = 7.0 with n = 20; Group B (no workshop) has mean x̄ = 6.0 with n = 20. Software reports the standard error of the difference ≈ 0.40. Predict each step before you reveal it.
Notice what each piece did: the gap (1.0) was the signal, the standard error (0.40) was the noise, their ratio was t (2.50), and comparing it to the critical value for our df (38) gave the verdict. Had the standard error been larger — say 0.60 — the same one-point gap would have given only t ≈ 1.67, which falls short of 2.02, and we'd fail to reject. Same gap, more noise, different conclusion.
Here's the live Independent-Samples t-Test calculator. Hit Load class data to drop in the $1-vs-$20 numbers, then Run and read the plain-English result. Paste your own two groups any time — it computes t, df, p, and Cohen's d for you.
Unit 11. Compare two separate groups. Demo loads Festinger & Carlsmith's classic $1-vs-$20 dissonance study, hidden in our data.
One of the most famous experiments in psychology is hiding in our dataset. Festinger and Carlsmith (1959) had people do a deadly-boring task, then paid them to tell the next person it was fun. Half got $1; half got $20. Afterward everyone rated how much they actually enjoyed the boring task. Common sense says the $20 group should like it more — more money, better mood. The opposite happened. (Note the design: two separate groups of people, one payment apiece — a textbook independent-samples setup.)
In our data, task_enjoyment by payment group came out like this:
| Group | Mean enjoyment | n |
|---|---|---|
| Paid $1 | 1.42 | 100 |
| Paid $20 | 0.04 | 100 |
t(198) = 5.96, p < .001, Cohen's d = 0.84 (large).
Notice the df: two groups of 100, so df = 100 + 100 − 2 = 198 — exactly the rule from §4. The $1 group rated the boring task as MORE enjoyable, and the gap is far too large to be chance. This is classic cognitive dissonance: people paid only a dollar had no good external reason for lying, so to ease the discomfort they convinced themselves the task was genuinely fun. The $20 group had a fat excuse ("I did it for the money"), so they felt no need to change their minds. Two separate groups, one mean apiece, and an independent t-test tells the whole story.
The APA (7th ed.) template for an independent-samples t adds one thing to Unit 10's: both groups' descriptives, so the reader sees who scored what before the statistics arrive:
t(df) = value, p = value, d = value"Participants paid $1 rated the task as more enjoyable (M = 1.42) than participants paid $20 (M = 0.04), t(198) = 5.96, p < .001, d = 0.84."
Same house rules as always: italicized letters, two decimals on statistics, no leading zero on p, "p < .001" rather than "= .000," and the effect size always rides along. One sentence carries the direction (who scored higher), the means, and the full inferential verdict — a reader who skims only this line still gets the whole finding.
The independent t-test isn't just for lab studies. Researchers walked into a real British university cafeteria and made one tiny change to the menu (Murray et al., 2026): next to each meat dish they slipped in a small photo of the living animal — a cow beside the beef, a pig beside the pork. Same food, same prices, same line. Two separate sets of diners — one seeing the photos, one not — make this another independent-samples comparison.
Then they counted what people actually ordered. Across 3,674 meal sales, the odds of choosing a vegetarian dish were 22% higher when the animal photos were on display. Most of us never connect the meat on the plate to the animal it came from — psychologists call that gap the meat paradox. The photo breaks the spell: for a moment you see the animal, and your willingness to eat it dips.
Our dataset reproduces the study. Each row is a campus dining day; column menu_condition tags it as Animal Photos or Standard, and plant_based_choice logs the percent of meals chosen plant-based:
| Group | Mean % plant-based | n |
|---|---|---|
| Animal Photos | 45% | 100 |
| Standard | 37% | 100 |
The result: t(198) = 5.39, p < .001, Cohen's d = 0.76 (large). The animal-photo menu pulled more plant-based orders, the gap is far too big to be chance, and the effect is large — a reminder of where meat comes from can quietly shift behavior, no lecture required. Same test, real-world stakes.
You don't compute the standard error by hand — Sheets does it. The two groups usually live in one column, tagged by another column that says which group each row belongs to. You use FILTER to split them apart, then T.TEST to compare. In any empty cell:
=T.TEST(FILTER(AC2:AC201, AB2:AB201="Animal Photos"),
FILTER(AC2:AC201, AB2:AB201="Standard"), 2, 2)The FILTER pieces pull out each group. The middle 2 means two-tailed (we'd accept a difference in either direction). The last 2 is the one that makes it an independent test — type = 2 tells Sheets these are two independent samples. (Type 1 would be paired, and type 3 the unequal-variance version.) The formula returns the p value directly; if you also want t and df, the =TTEST/Data ▸ Statistics add-ons or the calculator above will print them.
In =T.TEST(range1, range2, tails, type), the final argument is the test switch: type = 2 = two independent samples. That single digit is what separates this unit's test from the paired test next door (type 1).
When you run a t-test, the number that comes back sometimes has a minus sign. Students panic — did I break it? No. A negative t just means the first group you entered had the lower mean. Swap the two groups and the sign flips, but the size of t and the p-value stay exactly the same.
The lesson: read the magnitude of t (how far from zero), not its sign. The sign only tells you which group was bigger — which you already knew from the means. Always look at the group averages first, so the sign tells you a story you can already see.
Independent = different people in each group. If the same people are measured twice (before/after), that's a different test — the paired t-test, next unit. df = n₁ + n₂ − 2, not n − 1. t can be negative (the first group's mean was lower); the size is what matters, not the sign. Significant ≠ big — always report Cohen's d too. When p is high you fail to reject, you don't "accept" or "prove" the groups equal. And the groups must be tagged: you need a column saying which group each row is in.
Ten quick true/false items with instant explanations. (Different from your Canvas quiz — real practice.)