Two-Way ANOVA Examples: Step-by-Step Calculations and Interpretation

What Is a Two-Way ANOVA Example?

What Is Two-Way ANOVA?

The three null hypotheses tested are as follows. First, that there is no main effect of Factor A (the group means for Factor A, averaged across all levels of Factor B, are all equal). Second, that there is no main effect of Factor B (the group means for Factor B, averaged across all levels of Factor A, are all equal). Third, that there is no interaction between Factor A and Factor B (the effect of Factor A does not change across levels of Factor B).

Each hypothesis gets its own F-statistic and p-value. You do not need the main effects to be significant for the interaction to be significant, and vice versa.

Two-way ANOVA requires multiple observations per cell (each combination of Factor A and Factor B levels). That structure is called a replicated factorial design. When there is only one observation per cell, a different model applies and the interaction cannot be estimated separately from error. This distinction matters and is covered in the without-replication section below.

Interactive Two-Way ANOVA Calculator

Example 1: Complete 2 x 2 Two-Way ANOVA

This example works through every calculation by hand. The dataset is small enough to verify each step, but large enough to produce a meaningful analysis.

Step 1: Cell Means

Method 1 Quiet cell mean: (70 + 74 + 72 + 68) / 4 = 284 / 4 = 71.0. Method 1 Group: (65 + 67 + 63 + 61) / 4 = 256 / 4 = 64.0. Method 2 Quiet: (82 + 78 + 80 + 76) / 4 = 316 / 4 = 79.0. Method 2 Group: (75 + 73 + 77 + 71) / 4 = 296 / 4 = 74.0.

Marginal means: Method 1 = (71 + 64) / 2 = 67.5. Method 2 = (79 + 74) / 2 = 76.5. Quiet = (71 + 79) / 2 = 75.0. Group = (64 + 74) / 2 = 69.0. Grand mean = (71 + 64 + 79 + 74) / 4 = 72.0.

Step 2: Degrees of Freedom

a = 2 levels of Factor A (Method), b = 2 levels of Factor B (Environment), n = 4 observations per cell, N = 2 x 2 x 4 = 16.

Step 3: Sum of Squares

All SS calculations below use the balanced replicated factorial formulas. Every observation is denoted y_ijk where i = Factor A level, j = Factor B level, k = replicate index.

Verification: SS_Total = SS_A + SS_B + SS_AB + SS_Error = 324 + 144 + 4 + 80 = 552.0. Direct calculation of SS_Total (all 16 observations minus grand mean of 72) confirms 552.0. ✓

Step 4: Complete ANOVA Table

MS_A = 324/1 = 324.0. MS_B = 144/1 = 144.0. MS_AB = 4/1 = 4.0. MS_Error = 80/12 = 6.667. F_A = 324/6.667 = 48.60. F_B = 144/6.667 = 21.60. F_AB = 4/6.667 = 0.60.

Partial eta squared for A = 324 / (324 + 80) = 0.802. Partial eta squared for B = 144 / (144 + 80) = 0.643. Partial eta squared for A x B = 4 / (4 + 80) = 0.048.

Example 2: Significant Interaction Effect

This example constructs a dataset where the effect of teaching method on test scores differs depending on student ability level. Looking at main effects alone would give an incomplete and potentially misleading picture.

Example 3: Two Significant Main Effects, No Significant Interaction

When the interaction is not significant and both main effects are significant, main effect interpretations are more direct. The following manufacturing example illustrates this.

The differences are perfectly additive. Switching from Low to High setting adds 20 units regardless of material type. Switching from Standard to Premium adds 10 units regardless of setting. This is exactly what zero interaction looks like.

Example 4: Significant Interaction with Non-Significant Main Effects

This scenario surprises many students: the interaction can be statistically significant even when both main effects are not. It occurs when factor effects reverse direction across levels of the other factor, a pattern called a crossover interaction.

Both marginal means for Factor A are 50.0. Both marginal means for Factor B are 50.0. The grand mean is 50.0. Averaged across everything, neither factor appears to matter. Yet the cell means show a dramatic reversal: when Factor B is Low, Factor A High scores 20 points higher than Factor A Low. When Factor B is High, Factor A Low scores 20 points higher than Factor A High.

Example 5: 3 x 2 Factorial Design (Fertilizer x Density)

Cell Means and Marginal Means

Degrees of Freedom

a = 3, b = 2, n = 4. df_A = 3 - 1 = 2. df_B = 2 - 1 = 1. df_AB = (3-1)(2-1) = 2. df_Error = 3 x 2 x (4-1) = 18. df_Total = 24 - 1 = 23. Check: 2 + 1 + 2 + 18 = 23. ✓

ANOVA Table

Two-Way ANOVA With and Without Replication

When there is only one observation per cell, the classical approach cannot distinguish the interaction from pure random error. The no-replication model is often appropriate for randomized block designs where one factor is a "blocking" variable rather than a treatment factor of primary interest.

No-Replication Example (Randomized Block Structure)

In this structure, the residual after removing block and treatment effects serves as the error term. No separate interaction term can be extracted. Excel's "Two-Factor Without Replication" tool handles this structure. The analysis tests treatment effects after accounting for block-to-block variation, but it assumes no interaction between blocks and treatments.

Balanced vs Unbalanced Designs

A balanced design has the same number of observations in every cell. All five examples above are balanced. The formulas for SS_A, SS_B, and SS_AB simplify considerably in balanced designs and the three effects are orthogonal (statistically independent of each other).

An unbalanced design has different cell counts. This happens in observational studies and when data is lost. In unbalanced designs, the three effects are no longer orthogonal, so different sums-of-squares conventions can give different answers. The choice between Type I, Type II, and Type III SS matters and should be made deliberately based on the research question.

Sum of Squares and Degree of Freedom Formulas

How to Interpret the Interaction Effect

An interaction exists when the effect of Factor A on the dependent variable changes depending on which level of Factor B you are looking at. The interaction test asks whether the cell means can be explained by the two main effects alone, or whether additional structure is present.

Reading an Interaction Plot

An interaction plot shows Factor A or B on the x-axis, the mean of the dependent variable on the y-axis, and a separate line for each level of the other factor. Three patterns are common.

Approximately parallel lines: The effect of Factor A is similar across all levels of Factor B. Visual parallelism does not guarantee a non-significant interaction test, but it is encouraging evidence.

Non-parallel but non-crossing lines: The effect of Factor A differs in magnitude across levels of Factor B, but the direction is the same. One level of B consistently produces higher values, just by different amounts. The interaction test may or may not be significant.

Crossing lines: The direction of the Factor A effect reverses across levels of Factor B. This is a crossover interaction. The main effects averaged across this reversal can be misleading or meaningless. The interaction is typically the primary finding.

What to Do When the Interaction Is Significant

Post-Hoc Tests After Two-Way ANOVA

The right follow-up analysis depends on which effect is significant and the nature of the research question. There is no single universal rule like "always run Tukey HSD when ANOVA is significant."

When only main effects are significant and the interaction is not, post-hoc comparisons on Factor A marginal means and Factor B marginal means are reasonable. Tukey HSD is appropriate for all pairwise comparisons among levels of a factor when no directional hypothesis was specified in advance.

When the interaction is significant, simple-effects analysis is more informative than overall main-effect comparisons. Compare Factor A levels within each level of Factor B separately. Apply a multiple comparison correction to the full set of follow-up tests, not just within each simple-effects analysis.

Planned comparisons (contrasts specified before data collection) do not require the same correction as unplanned pairwise tests. If your research question was specifically "does fertilizer B outperform fertilizer A at high density?", that planned contrast can be tested directly without a family-wise correction for all other possible pairs.

Assumptions of Two-Way ANOVA

One-Way ANOVA vs Two-Way ANOVA

Running two separate one-way ANOVAs instead of one two-way ANOVA has two problems. It inflates the family-wise error rate (two tests at alpha = 0.05 gives a combined false-positive risk above 5%), and it cannot detect whether the factors interact. A factorial design with two-way ANOVA answers both main-effect questions and the interaction question in a single, controlled analysis.

Two-Way ANOVA in R

# Load data frame with columns: method, environment, score
# (Replace with your actual data frame)
data <- data.frame(
  method      = rep(c("Method1", "Method2"), 8),
  environment = rep(c("Quiet","Quiet","Quiet","Quiet",
                          "Group","Group","Group","Group"), 2),
  score = c(70,74,72,68, 82,78,80,76,
             65,67,63,61, 75,73,77,71)
)

# Fit the two-way ANOVA model with interaction
model <- aov(score ~ method * environment, data = data)
summary(model)

# The formula method * environment expands to:
#   method + environment + method:environment

# Interaction plot
interaction.plot(data$method, data$environment, data$score,
                  xlab = "Study Method",
                  ylab = "Mean Score",
                  trace.label = "Environment")

# Effect sizes (partial eta squared)
# install.packages("effectsize")  if needed
library(effectsize)
eta_squared(model, partial = TRUE)

# Post-hoc comparisons (if main effects significant and no interaction)
TukeyHSD(model, "method")
TukeyHSD(model, "environment")

Two-Way ANOVA in Python

import pandas as pd
import statsmodels.api as sm
from statsmodels.formula.api import ols
import matplotlib.pyplot as plt

# Build the data frame
df = pd.DataFrame({
    'method': ['M1']*4 + ['M2']*4 + ['M1']*4 + ['M2']*4,
    'env':    ['Quiet']*8 + ['Group']*8,
    'score':  [70,74,72,68, 82,78,80,76,
               65,67,63,61, 75,73,77,71]
})

# Fit the model: C() tells statsmodels these are categorical variables
model = ols('score ~ C(method) * C(env)', data=df).fit()

# Type II SS is appropriate for balanced designs (default for main effects)
# Use typ=3 for Type III SS (requires effects coding / sum coding)
anova_table = sm.stats.anova_lm(model, typ=2)
print(anova_table)

# Interaction plot
means = df.groupby(['method', 'env'])['score'].mean().reset_index()
for env, grp in means.groupby('env'):
    plt.plot(grp['method'], grp['score'], marker='o', label=env)
plt.ylabel('Mean Score')
plt.legend(title='Environment')
plt.title('Interaction Plot')
plt.show()

Two-Way ANOVA in Excel

Excel's Data Analysis ToolPak includes two two-way ANOVA procedures. Go to Data > Data Analysis to find them.

ANOVA: Two-Factor With Replication

Use this when you have multiple observations per cell. Arrange your data with Factor A categories as column headers and Factor B categories as row labels (or use a stacked layout). In the dialog, specify the number of rows per sample (this is n, the number of replicates per cell). Excel reports SS, df, MS, F, P-value, and F critical for Factor A, Factor B, and the interaction.

ANOVA: Two-Factor Without Replication

Use this for a one-observation-per-cell structure such as a randomized block design. This procedure cannot estimate an interaction term. If your data has multiple observations per cell, do not use this option; the results will be misleading.

Effect Size in Two-Way ANOVA

Statistical significance tells you whether an effect likely exists given the sample size. Effect size tells you how large that effect is, independently of sample size. Both are needed for a complete analysis.

Partial eta squared (partial eta^2) is the most common effect size for two-way ANOVA. For each effect, it is SS_effect / (SS_effect + SS_Error). It answers the question: what proportion of the variability in the outcome, that is not explained by other effects in the model, is explained by this one effect? Partial eta^2 ranges from 0 to 1 and is always tied to the specific model you fitted.

Eta squared (eta^2) is SS_effect / SS_Total. It is a less common choice in factorial designs because SS_Total changes depending on which effects are in the model, making comparisons across studies difficult.

Common reference points from the literature treat partial eta^2 of 0.01 as small, 0.06 as medium, and 0.14 as large, but these benchmarks are rough conventions from a specific context and should not be applied mechanically across all research fields.

Common Two-Way ANOVA Mistakes

How to Report a Two-Way ANOVA

A complete APA-style report includes the F-statistic, numerator and denominator degrees of freedom, p-value, and effect size for each effect. Below is a template using values from Example 1.

Practice Problems

Sample Dataset

The CSV below matches Example 5 (Fertilizer x Density). Copy it into a spreadsheet or paste it into the calculator above using the "Load 3 x 2 Example" button.

Fertilizer,Density,Yield
A,Low,42
A,Low,45
A,Low,43
A,Low,46
A,High,51
A,High,49
A,High,52
A,High,50
B,Low,47
B,Low,46
B,Low,45
B,Low,48
B,High,55
B,High,57
B,High,53
B,High,55
C,Low,38
C,Low,40
C,Low,37
C,Low,41
C,High,44
C,High,43
C,High,45
C,High,44

Frequently Asked Questions