Your ANOVA is significant. One of your groups differs from another — but which ones? The overall F-test never answers that question. To find out exactly which groups differ, you need post-hoc tests: the follow-up comparisons that run after a significant ANOVA.
SPSS offers a long menu of post-hoc tests — Tukey, Bonferroni, Games-Howell, Scheffé, and more — and choosing the wrong one can cost you either power or credibility. This tutorial explains the three you will actually use (Tukey, Bonferroni, and Games-Howell), shows you how to run them in SPSS click-by-click with paste-ready syntax, and walks through interpreting and reporting the results in APA style (7th edition).
What are post-hoc tests, and why do you need them?
An ANOVA tests the omnibus hypothesis that all group means are equal. When it rejects that hypothesis, all you know is that at least one group differs from at least one other. With three groups (A, B, C), there are three pairwise comparisons to make: A vs. B, A vs. C, B vs. C.
The problem with running several comparisons is the family-wise error rate: every additional test is another chance of a false positive. With three uncorrected t-tests at α = .05, the chance of at least one false positive climbs to about 14% — and with ten groups it exceeds 40%. Post-hoc tests control this error rate while comparing every pair of groups.
You only need post-hoc tests when two conditions hold: (1) the overall ANOVA was significant, and (2) a factor has more than two levels. With only two groups, the ANOVA result already tells you everything — the two groups differ.
Tukey vs Bonferroni vs Games-Howell: how they differ
Tukey’s HSD (Honestly Significant Difference)
Tukey’s test is the default recommendation for pairwise comparisons. It controls the family-wise error rate at exactly α across all pairwise comparisons and has good statistical power when two conditions hold: equal sample sizes (or nearly equal) and equal variances across groups. When those conditions hold, Tukey is the test reviewers expect to see.
Bonferroni correction
Bonferroni is the simplest correction: divide your alpha level by the number of comparisons (with 3 comparisons, test each at .05/3 ≈ .017). It works with any design and is easy to explain, but it is conservative — it sacrifices power, and with many groups it can fail to detect real differences. Use it when you have a small number of planned comparisons, not as your general-purpose post-hoc test.
Games-Howell
Games-Howell is the test to reach for when Levene’s test is significant — it does not assume equal variances or equal sample sizes. It is more powerful than Bonferroni-style corrections under unequal variances and is increasingly the expected choice when the homogeneity assumption fails. Its only downside: it can be slightly liberal with very small samples.
Quick decision guide
- Equal variances + roughly equal group sizes → Tukey’s HSD (the standard choice).
- Equal variances + very unequal group sizes → Hochberg’s GT2 or Gabriel’s test (both in the same SPSS dialog).
- Unequal variances (Levene’s test significant) → Games-Howell (no equal-variance assumption).
- Only a few specific comparisons you planned in advance → Bonferroni, or better, planned contrasts specified before running the ANOVA.
How to run post-hoc tests in SPSS: click-by-click
Option A: One-Way ANOVA dialog (single factor)
- Click Analyze > Compare Means > One-Way ANOVA.
- Move your continuous dependent variable into the Dependent List box and your grouping factor into the Factor box.
- Click Post Hoc. Under “Equal Variances Assumed”, tick Tukey (or Bonferroni). If Levene’s test suggests unequal variances, tick Games-Howell under “Equal Variances Not Assumed” instead. Click Continue.
- Click Options and tick Homogeneity of variance test (Levene’s test) and Descriptive. Click Continue.
- Click OK.
Option B: GLM Univariate (factorial designs)
- Click Analyze > General Linear Model > Univariate.
- Move your dependent variable into Dependent Variable and your factors into Fixed Factor(s).
- Click Post Hoc. Move each factor with more than two levels into the Post Hoc Tests for box. Tick Tukey under equal variances assumed, or Games-Howell under equal variances not assumed. Click Continue.
- Click Options, tick Homogeneity tests and Estimates of effect size. Click Continue, then OK.
Paste-ready SPSS syntax
Paste into a Syntax window and press Ctrl+A, Ctrl+R (Run All). Replace the variable names with your own.
One-way ANOVA with Tukey:
ONEWAY score BY group /STATISTICS DESCRIPTIVES HOMOGENEITY /POSTHOC = TUKEY ALPHA(.05).
One-way ANOVA with Games-Howell (unequal variances):
ONEWAY score BY group /STATISTICS DESCRIPTIVES HOMOGENEITY /POSTHOC = GH ALPHA(.05).
Bonferroni and Tukey together in a factorial GLM:
GLM score BY method classsize /METHOD = SSTYPE(3) /POSTHOC = method (TUKEY BONFERRONI) /PRINT = DESCRIPTIVE ETASQ HOMOGENEITY /DESIGN = method classsize method*classsize.
Note: post-hoc tests in GLM apply to main effects only. To follow up a significant interaction, use estimated marginal means with a Bonferroni adjustment (/EMMEANS = TABLES(method*classsize) COMPARE(method) ADJ(BONFERRONI)) rather than the Post Hoc dialog.
How to interpret the SPSS output
1. Homogeneity of variance (Levene’s test)
Check this first. If Sig. is above .05, equal variances are assumed and Tukey’s results are the ones to read. If Sig. is below .05, read the Games-Howell row instead.
2. Multiple Comparisons table
This is your main result table. Each row compares one pair of groups (I vs. J) and reports:
- Mean Difference (I−J) — the size and direction of the difference.
- Std. Error — the precision of that estimate.
- Sig. — the adjusted p-value for that pair. Values below .05 (marked with an asterisk *) indicate a significant difference.
- 95% Confidence Interval — the range of plausible values for the difference; if it excludes zero, the pair differs significantly.
Read down the Sig. column: every asterisked row is a pair of groups that genuinely differ. Non-asterisked rows are pairs you cannot distinguish statistically.
3. Homogeneous Subsets table
SPSS also prints a compact summary: groups placed in the same column (subset) do not differ significantly from each other, while groups in different columns do. It is the fastest way to see the pattern — e.g., groups A and B in subset 1, group C alone in subset 2, means C differs from both A and B.
A worked example
Suppose a one-way ANOVA on test scores across three teaching methods is significant, F(2, 87) = 7.34, p = .001, and Tukey’s HSD shows: Lecture vs. Workshop, mean difference = −9.4, p = .004; Lecture vs. Online, mean difference = −2.1, p = .612; Workshop vs. Online, mean difference = 7.3, p = .031. Conclusion: the Workshop group scored significantly higher than both the Lecture and Online groups, which did not differ from each other.
Reporting in APA style (7th ed.)
Name the post-hoc test you used and why — the choice is part of your method. Report the significant pairwise comparisons with mean differences, adjusted p-values, and confidence intervals where space allows. Template:
A one-way ANOVA revealed a significant effect of teaching method on test scores, F(2, 87) = 7.34, p = .001, η² = .14. Post-hoc comparisons using Tukey’s HSD indicated that the Workshop group (M = 84.6, SD = 8.1) scored significantly higher than the Lecture group (M = 75.2, SD = 9.4), p = .004, 95% CI [3.1, 15.7], and the Online group (M = 77.3, SD = 8.8), p = .031, 95% CI [0.6, 13.9]. The Lecture and Online groups did not differ significantly, p = .612.
Formatting notes per APA 7th edition:
- Italicize F, p, M, SD, and η²; report exact p-values (e.g., p = .004), except p < .001.
- State the post-hoc procedure in the method or results (“Post-hoc comparisons were conducted using Tukey’s HSD”) so readers know how the error rate was controlled.
- If you used Games-Howell because Levene’s test was significant, say so: “Because Levene’s test indicated unequal variances, Games-Howell post-hoc tests were used.”
- Report group means and standard deviations alongside the comparisons — a significant p-value without the means tells the reader nothing about the direction of the effect.
Key Takeaways
- Post-hoc tests answer the question the ANOVA leaves open: which specific groups differ.
- Only run them after a significant ANOVA, and only for factors with more than two levels.
- Tukey’s HSD is the standard choice when variances and group sizes are roughly equal.
- Bonferroni is simple but conservative — best for a small number of planned comparisons.
- Games-Howell is the right choice when Levene’s test is significant (unequal variances).
- Read the Multiple Comparisons table’s Sig. column and the Homogeneous Subsets table to see the pattern at a glance.
- In APA reporting, name the test, justify the choice, and report means, p-values, and confidence intervals.
FAQs
Do I always need post-hoc tests after a significant ANOVA?
Not always. If every factor has only two levels, the ANOVA result is already the complete answer. You need post-hoc tests only when a significant factor has three or more levels and you want to know which pairs differ.
Which post-hoc test should I choose?
Use Tukey’s HSD when group sizes and variances are roughly equal — it is the expected default. Use Games-Howell when Levene’s test shows unequal variances. Use Bonferroni when you have only a handful of specific comparisons. When in doubt, match the test to your assumption checks, not to habit.
Can I use post-hoc tests with a two-way ANOVA?
Yes, but only for interpreting main effects. The Post Hoc dialog in GLM does not follow up interactions. For a significant interaction, request estimated marginal means with a Bonferroni adjustment for simple comparisons instead.
What is the difference between planned contrasts and post-hoc tests?
Planned contrasts test specific hypotheses you stated before seeing the data, so they need less severe error correction and have more power. Post-hoc tests explore all possible pairwise comparisons after the fact, so they apply stricter family-wise error control. If you knew which comparisons mattered before running the study, contrasts are the stronger choice.
My ANOVA was significant but no post-hoc comparison is. How is that possible?
The omnibus F-test pools evidence across all groups, so it can detect an overall pattern that no single pairwise comparison captures on its own — especially with conservative corrections like Bonferroni or small samples. Report the significant ANOVA honestly, note that no pairwise difference reached significance, and consider whether you have enough power.
Need this done on your data? Get a free quote → /services/