🎻 One-Way ANOVA

1. Why not just run a bunch of t-tests?

Say you have three groups. To compare them with t-tests you'd need three separate runs: A-vs-B, A-vs-C, B-vs-C. Four groups? That jumps to six runs. Five groups? Ten. The comparisons multiply fast.

Here's the trap. Every t-test carries a small risk of a false alarm — calling a difference "real" when it's actually just chance. That's a Type I error, and we usually set the risk at 5% per test (α = .05). Run one test, you live with 5%. But run many tests and those little risks pile up. The chance of at least one false alarm across the whole family of tests — the familywise error rate — climbs well past 5%.

The arithmetic is unforgiving. With three independent tests each at 5%, the chance of avoiding every false alarm is 0.95 × 0.95 × 0.95 ≈ 0.857, so the chance of at least one false positive is about 14% — nearly triple what you signed up for. With ten comparisons it's about 40%. Run enough t-tests and you're almost guaranteed to "find" something that isn't there.

2. How ANOVA actually works — partitioning the spread

The deep idea behind ANOVA is hiding in its name: Analysis of Variance. We take the total spread in the whole dataset and split it into two piles — two sources that, added together, account for all the variation. This split is called partitioning the variance, and it's the heart of the method.

• Between-group variability — how far apart the group means sit from each other (and from the grand mean of everyone combined). If the groups truly differ, their averages land far apart. (big = interesting — this is the signal)
• Within-group variability — the everyday noise inside each group. People in the same group still aren't identical; they scatter around their own group's mean. (this is just static — the noise)

Picture three groups as three clouds of dots. Between-group spread is the distance between the centers of the clouds. Within-group spread is how puffy each cloud is on its own. ANOVA's whole job is to ask: are the clouds farther apart than they are puffy?

3. The F ratio — signal over noise

To turn the two piles of spread into one number, ANOVA first expresses each as a variance — in ANOVA-speak, a mean square (MS). (A mean square is just a variance: a sum of squared deviations divided by its degrees of freedom — same recipe you met back in Unit 4.) There's a mean square for each pile:

• MS_between — the variance among the group means. The signal.
• MS_within — the average variance inside the groups. The noise. (You'll also see it called MS_error.)

Then the test statistic is simply one divided by the other:

F = MS_between ÷ MS_within
  = (variance BETWEEN groups) ÷ (variance WITHIN groups)
  = signal ÷ noise

A big F means the gap between groups is much larger than the random noise — so the difference is probably real. An F near 1 means the groups are no more different than ordinary within-group noise would produce all by itself. F can't go below 0 (it's a ratio of two variances, which are never negative), so the whole action is on the high side: the bigger F climbs above 1, the stronger the evidence.

4. Degrees of freedom — the two numbers riding with F

F always travels with two degrees-of-freedom values, one for each pile of spread. With k = number of groups and N = total people across all groups:

So three diet groups (k = 3) with 200 students total (N = 200) give between-df = 3 − 1 = 2 and within-df = 200 − 3 = 197. Reported in full, the result reads:

F(2, 197) = 32.2, p < .001

The two numbers in parentheses are always the between-df then the within-df, in that order — and notice they encode your design: the first tells a reader how many groups you had (k − 1), the second how big your sample was (N − k). The decision rule is the one you already know: if p < .05, reject the null — at least one group mean stands apart from the others.

5. The ANOVA source table

Every ANOVA bookkeeps itself in a source table — the standard layout SPSS, Sheets templates, and journals all use. It walks left to right from raw spread (SS, sums of squares) through the df, to the variances (MS), and finally the F. Here's the skeleton, with the formula in each cell:

Read it as a recipe: divide each SS by its own df to get an MS, then divide the between-MS by the within-MS to get F. The "Total" row is just a checksum — the two SS rows add to it, and the two df rows add to it, proving the variance was partitioned exactly. Let's fill one in with real numbers.

📋 The fine print — ANOVA's three assumptions

Like the t-tests it grew out of, ANOVA leans on a few conditions. They don't have to hold perfectly — the test is famously sturdy — but you should be able to name them and know which one actually bites:

• Independent observations. Each score comes from a separate, unrelated case. Same non-negotiable rule as Units 10–12 — no formula rescues linked observations. (Same people measured in every condition? That's a repeated-measures design, and it needs a different ANOVA.)
• Roughly normal populations. Each group's scores should come from an approximately normal population. Thanks to the Central Limit Theorem, with decent group sizes ANOVA barely notices moderate skew — worry only when groups are small and obviously lopsided.
• Homogeneity of variance. The groups should have roughly similar spreads — one puffy cloud and two tight ones make MS_within a misleading average of the noise. Rule of thumb: trouble starts when one group's SD is more than about twice another's, especially when the group sizes are also unequal.

6. A significant F → run post-hoc tests

ANOVA has one honest limit. A significant result is an omnibus finding — it says "somebody is different" — but not who. Was it Vegans beating both? Or Vegetarians edging out Omnivores only? The F ratio can't tell you; it lumps all the groups together.

To find out, you run post-hoc tests — small follow-up comparisons after a significant ANOVA that pinpoint which pairs of groups actually differ, while still protecting the familywise error rate (common ones: Tukey's HSD, Bonferroni, Scheffé). Think of ANOVA as the smoke alarm and post-hoc tests as walking room to room to find the fire. The order is fixed: significant ANOVA first, then post-hoc — never the reverse.

Tukey's HSD — the workhorse

The most-used post-hoc test is Tukey's HSD — Honestly Significant Difference. The "honestly" is the whole sales pitch: it compares every pair of means while keeping the familywise false-alarm rate at 5% across the whole set of comparisons. Its engine is a cousin of t called the studentized range statistic, q. For any two group means:

q = (larger mean − smaller mean) ÷ SE_pair

where  SE_pair = √[ (MS_within / 2) × (1/n₁ + 1/n₂) ]

Look at what's inside: the top is the familiar gap between two means, and the bottom recycles MS_within — the noise estimate you already computed for the ANOVA — sized to the two groups being compared. (With equal group sizes this simplifies to the classic √(MS_within/n); the version above handles unequal ns, which our diet groups certainly are.) Then the decision rule is the same rhythm as every test this year: compare your q to a critical q, which depends on k (how many groups — more groups, higher bar) and df_within. For our design — k = 3 groups, df_within = 197 — the critical q at α = .05 is about 3.34. Beat it and that pair "honestly" differs.

Tukey vs. Bonferroni — picking your protection

The other name you'll meet constantly is the Bonferroni correction — less a test than a blunt, honest rule: run your pairwise t-tests, but divide α by the number of comparisons. Three comparisons? Each must beat α = .05 ÷ 3 ≈ .017 to count. (Equivalently: multiply each p by 3 and see if it's still under .05.)

7. Try it yourself — the One-Way ANOVA calculator

Here's the real thing, embedded right on the page. Hit “Load class data (compassion by diet)” to drop in our three diet groups, then Run and read the plain-English result. When the F comes back significant, the calculator now walks the rooms for you — it prints the full Tukey HSD pairwise table underneath, exactly the follow-ups from §6. This is the exact analysis from our worked example below.

🐮 Worked example — on OUR data

Open the SJNY Student Life dataset. We're looking at one of your instructor's own research questions: do people who eat differently also feel differently about animals? Each of our 200 students reported their diet and filled out the Compassion for Animals scale (1 = low, 7 = high):

• Omnivore — eats meat
• Vegetarian — no meat, but eats some animal products
• Vegan — no animal products at all

This is exactly an ANOVA design: three groups (k = 3), one quantitative outcome. The three group means tell the story — they climb steadily:

In Sheets (column X = diet, column Y = animal_compassion), grab each mean with one function:

=AVERAGEIF(X2:X201, "Omnivore",   Y2:Y201)   → 3.91
=AVERAGEIF(X2:X201, "Vegetarian", Y2:Y201)   → 4.92
=AVERAGEIF(X2:X201, "Vegan",      Y2:Y201)   → 5.67

Then make a bar chart of the three means. The bars climb steadily — a clear step up at each level, with Vegan towering over Omnivore. Now feed those numbers into the calculator above and you get the full result, laid out in a source table:

✍️ Report it in APA style

Every stats result you'll ever publish, present, or hand in gets written in the same standardized sentence — the APA (7th ed.) results format. It packs the test, the df, the statistic, the p, and the effect size into one line a reader anywhere can decode. The template for a one-way ANOVA:

F(df_between, df_within) = value, p = value, η² = value

The house rules, once, since they apply to every test from here on: statistical letters (F, t, p, M, SD) are italicized; statistics get two decimals; p-values get two or three decimals with no leading zero (.032, not 0.032) because p can't exceed 1; and p below .001 is written p < .001, never "= .000". Report the means and SDs so readers see the actual pattern, and always include the effect size. Our full result, written for a journal:

Read the anatomy of that sentence: what test ("a one-way ANOVA"), the omnibus verdict with its full statistics, then the post-hoc story — which pairs, in which direction, with the group means anchoring each claim. That sentence structure — omnibus first, follow-ups second — mirrors the order of the analysis itself. To the extent that a reader only skims one sentence of your results, this is the sentence.

💻 In Google Sheets

Sheets has no one-click one-way-ANOVA button in the base app, so for the full source table you have two honest routes:

• The add-on route. Install the free XLMiner Analysis ToolPak add-on, then Extensions ▸ Add-ons ▸ open it ▸ choose Anova: Single Factor, highlight your columns of group data, and it prints the whole source table (SS, df, MS, F, p) — the same one SPSS gives. (In Excel proper it lives under Data ▸ Data Analysis ▸ Anova: Single Factor.)
• The shortcut formula. If you only need the p-value for a quick check, =FTEST(range1, range2) compares the variances of two groups — useful, but it's a two-group tool, not the full multi-group source table. For a real three-group ANOVA, use the add-on above (or the calculator on this page).

Either way, Sheets nails the part that builds your intuition — the group means and the chart:

=AVERAGEIF(X2:X201, "Omnivore",   Y2:Y201)   → each group mean
=COUNTIF(X2:X201, "Vegan")                    → each group's n

Highlight your little table of three means, then Insert ▸ Chart and pick a column chart. You'll see the staircase climb with your own eyes — then read the F(2, 197) = 32.2, p < .001 from the add-on (or this page's calculator) to confirm the steps are real, not noise.

Post-hoc follow-ups in Sheets. Sheets has no Tukey button, so the honest route is Bonferroni — run the three pairwise t-tests with T.TEST and hold each to α = .05 ÷ 3 ≈ .017:

=T.TEST(omnivore_range,   vegetarian_range, 2, 3)   → compare to .017
=T.TEST(omnivore_range,   vegan_range,      2, 3)   → compare to .017
=T.TEST(vegetarian_range, vegan_range,      2, 3)   → compare to .017

(The final 3 = unequal-variances type, the safe default when group sizes differ this much.) For real Tukey values, use this page's calculator — after a significant F it prints the full pairwise table — or SPSS's one-click Post Hoc ▸ Tukey.

Key terms