Sign In
ANOVA: One-Way, Two-Way, and Repeated Measures
When a research question involves comparing more than two group means, a t-test is no longer the right tool. Running multiple t-tests inflates the risk of a false positive, and the more comparisons a researcher runs, the more likely at least one will appear statistically significant by chance alone. Analysis of variance, universally known as ANOVA, solves this problem by testing all group differences simultaneously within a single framework.
ANOVA is one of the most flexible and widely used tools in inferential statistics. It applies to experimental and observational research across nearly every discipline, from psychology and education to agriculture and clinical medicine. This guide explains the three most common forms of ANOVA (one-way, two-way, and repeated measures), when to use each, how to interpret the results, and how to report them in academic papers.
What Is ANOVA?
ANOVA compares the means of three or more groups by partitioning the total variability in the data into two components: variability between groups and variability within groups. The test statistic, called F, is the ratio of these two variances. If the between-group variance is much larger than the within-group variance, the F-statistic will be large and the p-value small, suggesting that at least one group mean differs from the others.
The core insight of ANOVA is that "analysis of variance" is a somewhat misleading name. ANOVA does not directly analyze the variance of the outcome variable. It compares the means, using variance as the mathematical machinery to do so.
Every ANOVA tests a null hypothesis that all group means are equal against an alternative hypothesis that at least one mean differs. A significant F-test indicates that a difference exists somewhere among the groups but does not identify which specific pairs of groups differ. That determination requires post hoc tests.
Why Not Just Run Multiple T-Tests?
Comparing three or more groups pairwise using multiple t-tests inflates the family-wise error rate. If each individual test uses α = 0.05, the probability of at least one false positive across multiple tests grows quickly.
With three groups requiring three pairwise comparisons, the family-wise error rate rises to approximately 0.14. With five groups requiring ten comparisons, the rate approaches 0.40. ANOVA controls this problem by conducting a single omnibus test at the chosen alpha level. Post hoc procedures then apply corrections to keep the overall error rate at the intended level when pairwise comparisons are needed.
The Three Types of ANOVA Covered Here
| ANOVA Type | Design Structure | Example Research Question |
|---|---|---|
| One-way ANOVA | One independent variable with three or more levels; different participants in each group | Do students in three different teaching conditions score differently on a final exam? |
| Two-way ANOVA | Two independent variables, each with two or more levels; different participants in each cell | Do teaching method and class size independently or interactively affect exam scores? |
| Repeated measures ANOVA | Same participants measured across three or more conditions or time points | Do participants' anxiety scores change across pretest, midtest, and posttest? |
One-Way ANOVA
The one-way ANOVA is the simplest and most commonly used form. It examines the effect of a single independent variable (called a factor) with three or more levels on a continuous outcome. Each participant appears in only one group.
The F-statistic is calculated as:
F = MSbetween / MSwithin
where MSbetween is the mean square between groups (variability attributable to the group differences) and MSwithin is the mean square within groups (variability that cannot be explained by group membership).
Example. A researcher compares final exam scores across three teaching methods, with 20 students in each group. Group means are 74 (traditional lecture), 79 (flipped classroom), and 82 (project-based learning). The within-group standard deviation is approximately 8 in each group.
Step 1. Calculate the overall grand mean: (74 + 79 + 82) / 3 = 78.33
Step 2. Calculate the sum of squares between groups: SSbetween = 20 × [(74 − 78.33)² + (79 − 78.33)² + (82 − 78.33)²] = 20 × (18.75 + 0.45 + 13.47) = 653.4
Step 3. Calculate the sum of squares within groups: SSwithin = (19 × 64) × 3 = 3,648
Step 4. Calculate mean squares: MSbetween = 653.4 / 2 = 326.7; MSwithin = 3,648 / 57 = 64.0
Step 5. Calculate F: F = 326.7 / 64.0 = 5.10
Step 6. With df = (2, 57), the critical F-value at α = 0.05 is approximately 3.16. Because 5.10 exceeds 3.16, we reject the null hypothesis. The p-value is approximately 0.009.
At least one teaching method produces different mean scores from the others. Post hoc tests are needed to determine which specific pairs differ.
Common post hoc procedures include Tukey's HSD (a good default for equal group sizes), Bonferroni correction (conservative but simple), and Scheffé's method (flexible but less powerful). The choice depends on the number of comparisons and the balance between controlling Type I error and preserving statistical power.
Two-Way ANOVA
The two-way ANOVA extends the logic of one-way ANOVA to designs with two independent variables. It produces three F-tests: one for each main effect and one for the interaction between the two factors.
A main effect is the effect of one independent variable averaged across the levels of the other. An interaction effect occurs when the effect of one variable depends on the level of the other. Interactions are often the most theoretically interesting result of a two-way ANOVA because they reveal conditional relationships that main effects alone would obscure.
Example. A researcher examines whether teaching method (traditional vs. project-based) and class size (small vs. large) affect exam scores. The design creates four cells, each with 15 students. Mean scores across the cells are:
- Traditional, small class: 78
- Traditional, large class: 76
- Project-based, small class: 85
- Project-based, large class: 74
A two-way ANOVA would test three effects: the main effect of teaching method (do project-based classes score differently on average?), the main effect of class size (do small classes score differently on average?), and the interaction (does the effect of teaching method depend on class size?).
In this example, the interaction is likely to be the most interesting result because project-based learning appears more effective in small classes than in large ones, while traditional teaching shows a smaller class-size effect.
Interpreting a two-way ANOVA requires care. When a significant interaction exists, main effects should be interpreted cautiously, if at all, because the interaction indicates that the effect of one variable depends on the level of the other. Simple effects analyses (examining one factor at each level of the other) usually replace main-effect interpretation in the presence of a significant interaction.
Repeated Measures ANOVA
The repeated measures ANOVA compares three or more measurements taken from the same participants. Common applications include studies measuring participants across time points (pretest, midtest, posttest) and studies where each participant experiences all conditions of an experiment.
Repeated measures designs offer two advantages over between-subjects designs. First, they typically require fewer participants because each person contributes multiple data points. Second, they reduce error variance by removing between-participant differences from the error term, which increases statistical power.
The tradeoff is a stricter assumption called sphericity, which requires that the variances of the differences between all pairs of conditions be approximately equal. When sphericity is violated (as it often is), a correction to the degrees of freedom is applied. The two most common corrections are the Greenhouse-Geisser and Huynh-Feldt adjustments.
Example. A researcher measures anxiety in 25 participants at three time points during a semester-long intervention: baseline, midpoint, and end. Mean anxiety scores decline from 45 at baseline to 39 at midpoint to 33 at end. A repeated measures ANOVA tests whether these differences are statistically significant.
The F-test compares the variability across time points to the variability within participants across time points. If the F is significant, follow-up analyses (such as pairwise comparisons with Bonferroni correction) determine which specific time points differ.
In a typical result, the analysis might report F(2, 48) = 12.4, p < .001, ηp² = 0.34, indicating a large effect of time on anxiety scores. Pairwise comparisons would then reveal whether the drop from baseline to midpoint, midpoint to end, or both, drives the overall effect.
Assumptions of ANOVA
All three types of ANOVA share a common set of assumptions, with variations specific to each design.
- Continuous dependent variable. The outcome should be measured on an interval or ratio scale.
- Independence of observations. Observations should be independent within each group. Repeated measures designs relax this within participants but require independence between participants.
- Normality. The dependent variable should be approximately normally distributed within each group. ANOVA is robust to modest violations with large samples.
- Homogeneity of variance. For between-subjects ANOVAs, group variances should be similar. Levene's test evaluates this assumption.
- Sphericity. For repeated measures ANOVAs, the variances of the differences between conditions should be approximately equal. Mauchly's test evaluates this, and violations are corrected with Greenhouse-Geisser or Huynh-Feldt adjustments.
When assumptions are seriously violated, non-parametric alternatives include the Kruskal-Wallis test (replaces one-way ANOVA) and the Friedman test (replaces one-way repeated measures ANOVA). No fully general non-parametric equivalent exists for two-way ANOVA, though several specialized procedures cover common cases.
Effect Sizes for ANOVA
Effect sizes for ANOVA describe the proportion of variance in the outcome that is attributable to the factor or factors being tested. The two most common are:
- Eta squared (η²) represents the proportion of total variance explained by a factor. It is simple to calculate but tends to overestimate effect size in complex designs.
- Partial eta squared (ηp²) represents the proportion of variance explained by a factor after accounting for other factors in the model. It is the preferred effect size for factorial and repeated measures ANOVAs and is reported by most statistical software.
Cohen's conventions for partial eta squared are approximately 0.01 for small, 0.06 for medium, and 0.14 for large effects, though field-specific benchmarks often apply. Report effect sizes for every F-test in your ANOVA, not only for the significant ones.
Reporting ANOVA Results in APA Format
APA style specifies the format for reporting ANOVA results. The general template is:
F(df1, df2) = F-value, p = p-value, ηp² = effect size
Complete examples for each ANOVA type:
One-way ANOVA: "A one-way ANOVA revealed a significant effect of teaching method on final exam scores, F(2, 57) = 5.10, p = .009, ηp² = 0.15. Tukey's HSD post hoc tests indicated that project-based learning (M = 82, SD = 8) produced significantly higher scores than traditional lecture (M = 74, SD = 8), p = .007, but not significantly higher than flipped classroom (M = 79, SD = 8), p = .424."
Two-way ANOVA: "A 2 × 2 ANOVA revealed a significant main effect of teaching method, F(1, 56) = 6.42, p = .014, ηp² = 0.10, and a significant interaction between teaching method and class size, F(1, 56) = 8.71, p = .005, ηp² = 0.13. The main effect of class size was not significant, F(1, 56) = 2.30, p = .135, ηp² = 0.04."
Repeated measures ANOVA: "A repeated measures ANOVA showed a significant effect of time on anxiety scores, F(2, 48) = 12.4, p < .001, ηp² = 0.34. Mauchly's test indicated that the assumption of sphericity was not violated, χ²(2) = 3.14, p = .208. Pairwise comparisons with Bonferroni correction revealed significant reductions from baseline to midpoint (p = .012) and from midpoint to end (p = .003)."
Common Mistakes to Avoid
Interpreting main effects in the presence of a significant interaction. A significant interaction means the effect of one factor depends on the other. Interpreting main effects in isolation misleads readers about what the data show.
Skipping post hoc tests after a significant omnibus F. A significant ANOVA tells you that at least one difference exists but not which pairs differ. Post hoc tests or planned contrasts are required to identify the source of the effect.
Running multiple ANOVAs on related outcomes without correction. Testing several dependent variables separately inflates the family-wise error rate. Consider MANOVA or apply a Bonferroni-type correction across the ANOVAs.
Ignoring assumption checks. Homogeneity of variance and sphericity affect the validity of ANOVA results. Report assumption checks and, when violations occur, report the corrected analysis.
Reporting only p-values. APA style requires F, degrees of freedom, p-value, and effect size at minimum. Descriptive statistics for each group (means and standard deviations) should also appear in the results.
Before You Submit: A Self-Audit
Work through this checklist before submitting a paper reporting ANOVA:
- Have you chosen the correct ANOVA type for your design (one-way, two-way, repeated measures)?
- Are the assumptions (normality, homogeneity of variance, sphericity) reported and addressed?
- Are means and standard deviations reported for each group or condition?
- Is each F-statistic reported with degrees of freedom in parentheses?
- Are effect sizes (partial eta squared) reported for every F-test?
- Are post hoc tests reported with the correction method identified?
- If a significant interaction is present, does the interpretation focus on simple effects rather than main effects?
ANOVA results appear in nearly every experimental and quasi-experimental study across the social and behavioral sciences, and peer reviewers scrutinize this section closely. Small reporting errors in ANOVA writeups (missing effect sizes, misinterpreted interactions, absent assumption checks) are among the most common reasons manuscripts receive requests for revision. If you have finished your analysis and want a professional editor to review your methods, results, and statistical reporting before submission, Editor World offers journal article editing by editors with subject-matter backgrounds. Clients browse editor profiles and choose the editor whose expertise best matches their field. A free sample edit on the first 300 words is available for every project, and a certificate of editing is available as an optional add-on for publishers or committees that require one.
Frequently Asked Questions
What is the difference between one-way and two-way ANOVA?
One-way ANOVA tests the effect of a single independent variable with three or more levels on a continuous outcome. Two-way ANOVA tests two independent variables simultaneously and also examines whether they interact. The choice depends on how many factors your design manipulates.
When should I use repeated measures ANOVA instead of a between-subjects ANOVA?
Use repeated measures ANOVA when the same participants provide data across all conditions or time points. Between-subjects ANOVA is appropriate when different participants are assigned to different conditions. Repeated measures designs typically require fewer participants and have greater statistical power but require the sphericity assumption to hold.
Do I need post hoc tests after a significant ANOVA?
If the omnibus F-test is significant and you want to know which specific group pairs differ, post hoc tests are required. A significant ANOVA tells you that at least one difference exists among the groups but does not identify which pairs. Common post hoc procedures include Tukey's HSD, Bonferroni correction, and Scheffé's method.
What is an interaction effect in a two-way ANOVA?
An interaction effect occurs when the effect of one independent variable depends on the level of another. For example, a teaching method might work well in small classes but poorly in large ones. When an interaction is significant, main effects should be interpreted with caution because the interaction indicates that the factors' effects are conditional on each other.
What effect size should I report for ANOVA?
Partial eta squared (ηp²) is the most commonly reported effect size for ANOVA and is provided by most statistical software. Cohen's conventions are approximately 0.01 for small, 0.06 for medium, and 0.14 for large effects. Report effect sizes for every F-test, not only significant ones, so readers can judge the magnitude of both detected and undetected effects.