What Is a Two-Way ANOVA?

A two-way ANOVA (also called a factorial ANOVA) tests the effects of two categorical independent variables (called factors) on a continuous dependent variable. It simultaneously answers three questions:

1. Main effect of Factor A: Does the dependent variable differ across the levels of Factor A, averaging over Factor B?
2. Main effect of Factor B: Does the dependent variable differ across the levels of Factor B, averaging over Factor A?
3. Interaction effect (A x B): Does the effect of Factor A depend on the level of Factor B (and vice versa)?

The interaction effect is what makes the two-way ANOVA uniquely powerful. It allows you to detect situations where the combination of two factors produces an outcome that neither factor alone would predict.

Example: Suppose a new drug lowers blood pressure in young adults but has no effect in older adults. Neither age alone nor drug alone tells the whole story — you need the interaction to understand the pattern.

When to Use It

Use a two-way ANOVA when:

• You have one continuous dependent variable (e.g., exam scores, reaction time, sales revenue)
• You have two categorical independent variables (e.g., teaching method and class size, drug type and dosage level)
• The groups are independent (different participants in each cell of the design)

Examples:

• Teaching method (lecture vs. active learning) x class size (small vs. large) on exam scores
• Gender (male vs. female) x treatment (drug vs. placebo) on symptom reduction
• Advertising type (humor vs. emotional vs. informational) x medium (TV vs. social media) on brand recall

Design notation: A 2 x 3 ANOVA has two levels of Factor A and three levels of Factor B, producing 6 cells. A 2 x 2 design has four cells and is the simplest factorial design.

Understanding Interaction Effects

The interaction is often the most important finding in a factorial ANOVA. An interaction means the effect of one factor changes depending on the level of the other factor.

No interaction (additive effects): The lines in an interaction plot are roughly parallel. Active learning is better than lecture regardless of class size, and the advantage is about the same in both small and large classes.

Interaction present: The lines in an interaction plot cross or diverge. Active learning might outperform lecture in small classes but show little advantage in large classes. The effect of teaching method depends on class size.

When the interaction is significant, interpret it first. The main effects can be misleading because averaging across levels of the other factor hides the conditional pattern. Report simple effects — the effect of Factor A at each level of Factor B — rather than relying on marginal means alone.

Assumptions

1. Independence of observations. Each participant appears in only one cell of the design.

2. Normality. The dependent variable should be approximately normally distributed within each cell. The two-way ANOVA is robust to moderate violations when cell sizes are equal and at least 15-20 per cell.

3. Homogeneity of variances. The variances should be roughly equal across all cells of the design. Test with Levene's test. If violated, consider a more robust method or transform the data.

4. No significant outliers. Extreme values within cells can distort the F-tests. Check boxplots within each cell.

Formula

Decomposing Variance

The total variance is partitioned into four components:

$$SS_{total} = SS_A + SS_B + SS_{A \times B} + SS_{within}$$

Main effect of Factor A:

$$SS_A = \sum_{j=1}^{a} bn_c (\bar{X}_{j\cdot} - \bar{X}_{grand})^2$$

Main effect of Factor B:

$$SS_B = \sum_{k=1}^{b} an_c (\bar{X}_{\cdot k} - \bar{X}_{grand})^2$$

Interaction effect:

$$SS_{A \times B} = n_c \sum_{j=1}^{a}\sum_{k=1}^{b}(\bar{X}_{jk} - \bar{X}_{j\cdot} - \bar{X}_{\cdot k} + \bar{X}_{grand})^2$$

Within-groups (error):

$$SS_{within} = \sum_{j=1}^{a}\sum_{k=1}^{b}\sum_{i=1}^{n_c}(X_{ijk} - \bar{X}_{jk})^2$$

Where $a$ = number of levels of Factor A, $b$ = number of levels of Factor B, $n_c$ = number of participants per cell.

Mean Squares and F-Ratios

Each effect has its own F-test:

$$F_A = \frac{MS_A}{MS_{within}} = \frac{SS_A / (a-1)}{SS_{within} / (N - ab)}$$

$$F_B = \frac{MS_B}{MS_{within}} = \frac{SS_B / (b-1)}{SS_{within} / (N - ab)}$$

$$F_{A \times B} = \frac{MS_{A \times B}}{MS_{within}} = \frac{SS_{A \times B} / (a-1)(b-1)}{SS_{within} / (N - ab)}$$

Effect Size: Partial Eta-Squared

For each effect:

$$\eta_p^2 = \frac{SS_{effect}}{SS_{effect} + SS_{within}}$$

Worked Example

Scenario: An educational researcher investigates the effects of teaching method (Lecture vs. Active Learning) and class size (Small: 20 students vs. Large: 80 students) on final exam scores. This is a 2 x 2 between-subjects factorial design with 8 students per cell ($N = 32$).

Cell means and standard deviations:

Step 1: State the hypotheses.

• Main effect of Method: $H_0$: $\mu_{\text{lecture}} = \mu_{\text{active}}$
• Main effect of Class Size: $H_0$: $\mu_{\text{small}} = \mu_{\text{large}}$
• Interaction: $H_0$: The effect of teaching method does not depend on class size

Step 2: Calculate sums of squares.

Each cell has $n_c = 8$. Factor A (Method) has $a = 2$ levels, Factor B (Class Size) has $b = 2$ levels.

$SS_A$ (Method):

$$SS_A = bn_c \sum_{j=1}^{a}(\bar{X}_{j\cdot} - \bar{X}_{grand})^2 = 2 \times 8 \times [(71.0 - 75.25)^2 + (79.5 - 75.25)^2]$$

$$= 16 \times [18.0625 + 18.0625] = 16 \times 36.125 = 578.0$$

$SS_B$ (Class Size):

$$SS_B = an_c \sum_{k=1}^{b}(\bar{X}_{\cdot k} - \bar{X}_{grand})^2 = 2 \times 8 \times [(78.5 - 75.25)^2 + (72.0 - 75.25)^2]$$

$$= 16 \times [10.5625 + 10.5625] = 16 \times 21.125 = 338.0$$

$SS_{A \times B}$ (Interaction):

$$SS_{A \times B} = n_c \sum_{j}\sum_{k}(\bar{X}_{jk} - \bar{X}_{j\cdot} - \bar{X}_{\cdot k} + \bar{X}_{grand})^2$$

Compute the residual for each cell:

• Lecture, Small: $72 - 71.0 - 78.5 + 75.25 = -2.25$
• Lecture, Large: $70 - 71.0 - 72.0 + 75.25 = 2.25$
• Active, Small: $85 - 79.5 - 78.5 + 75.25 = 2.25$
• Active, Large: $74 - 79.5 - 72.0 + 75.25 = -2.25$

$$SS_{A \times B} = 8 \times [(-2.25)^2 + (2.25)^2 + (2.25)^2 + (-2.25)^2] = 8 \times 4 \times 5.0625 = 162.0$$

$SS_{within}$ (Error):

Using $SS_{within} = \sum (n_c - 1) \times s_j^2$ for each cell:

$$SS_{within} = 7(36) + 7(49) + 7(25) + 7(36) = 252 + 343 + 175 + 252 = 1022$$

Step 3: Calculate mean squares and F-ratios.

$MS_{within} = 1022 / 28 = 36.5$

Step 4: Determine p-values.

Using the F-distribution with $df_1 = 1$, $df_2 = 28$:

• Method: $F(1, 28) = 15.84$, $p < .001$
• Class Size: $F(1, 28) = 9.26$, $p = .005$
• Interaction: $F(1, 28) = 4.44$, $p = .044$

All three effects are statistically significant.

Step 5: Calculate effect sizes (partial eta-squared).

$$\eta_p^2(\text{Method}) = \frac{578}{578 + 1022} = .361$$

$$\eta_p^2(\text{Class Size}) = \frac{338}{338 + 1022} = .249$$

$$\eta_p^2(\text{Interaction}) = \frac{162}{162 + 1022} = .137$$

Step 6: Interpret the interaction.

Because the interaction is significant, we examine the simple effects:

• In small classes: Active learning ($M = 85$) outperformed lecture ($M = 72$) by 13 points
• In large classes: Active learning ($M = 74$) outperformed lecture ($M = 70$) by only 4 points

The advantage of active learning shrinks considerably in large classes. This is the interaction — teaching method and class size jointly influence exam scores in a way that cannot be explained by either factor alone.

Interpretation

The two-way ANOVA revealed significant main effects and a significant interaction:

• Main effect of teaching method: Active learning ($M = 79.5$) produced significantly higher exam scores than lecture ($M = 71.0$), $F(1, 28) = 15.84$, $p < .001$, $\eta_p^2 = .36$.
• Main effect of class size: Small classes ($M = 78.5$) produced significantly higher scores than large classes ($M = 72.0$), $F(1, 28) = 9.26$, $p = .005$, $\eta_p^2 = .25$.
• Interaction: The advantage of active learning over lecture was significantly larger in small classes (13 points) than in large classes (4 points), $F(1, 28) = 4.44$, $p = .044$, $\eta_p^2 = .14$.

The interaction suggests that active learning is most effective in small class environments. In large classes, the benefit of active learning is reduced — possibly because large group sizes limit the student engagement and participation that active learning relies on.

Common Mistakes

1. Interpreting main effects when the interaction is significant. Main effects represent average differences across levels of the other factor. When an interaction is present, these averages are misleading. Focus on simple effects instead.

2. Ignoring unbalanced designs. When cell sizes are unequal, the sums of squares for different effects are no longer independent. Use Type III sums of squares (the default in most software) to correctly partition variance.

3. Not plotting the interaction. An interaction plot (with one factor on the x-axis, the DV on the y-axis, and separate lines for the other factor) is essential for understanding the pattern. Non-parallel lines indicate an interaction.

4. Confusing ordinal and disordinal interactions. In an ordinal interaction, one group is always higher but the gap changes in size (the lines do not cross). In a disordinal (crossover) interaction, the direction of the effect reverses (the lines cross). Disordinal interactions are typically more theoretically interesting.

5. Running separate one-way ANOVAs instead. Analyzing each factor separately ignores the interaction and wastes statistical power. The two-way ANOVA tests both factors and their interaction simultaneously.

6. Not checking assumptions within cells. Normality and homogeneity of variances must hold within each cell (combination of factor levels), not just within each level of a single factor.

How to Run It