Applying the Greenhouse–Geisser Correction

Why Apply the Greenhouse–Geisser Correction?

A univariate repeated-measures ANOVA compares condition means using an F statistic. Its usual reference distribution relies on sphericity: the variances of all pairwise differences between conditions are equal. When that condition is doubtful or not supported, the usual degrees of freedom can make the test too liberal, increasing the chance of a Type I error. The Greenhouse–Geisser correction provides a way to adjust the reference distribution.

The correction is deliberately focused. It does not change the observed F statistic or the mean squares that produced it. Instead, it reduces both degrees of freedom by multiplying them by the Greenhouse–Geisser epsilon, then uses the resulting F distribution to calculate the p-value. This tutorial focuses on carrying out that adjustment and interpreting the resulting test.

Definition: Let \(k\) be the number of repeated conditions and \(n\) the number of participants with complete measurements. In the standard univariate repeated-measures test, the uncorrected degrees of freedom for the condition effect and its error are \(df_1=k-1\) and \(df_2=(n-1)(k-1)\). The Greenhouse–Geisser correction multiplies each by the same epsilon, \(\hat{\epsilon}_{GG}\), to obtain \(df_{1,GG}=\hat{\epsilon}_{GG}df_1\) and \(df_{2,GG}=\hat{\epsilon}_{GG}df_2\). The corrected p-value is the upper-tail probability from an F distribution with these corrected degrees of freedom.

Model and Conditions for the Test

Suppose \(Y_{ij}\) is the measured outcome for participant \(i\) under condition \(j\), where \(i=1,\ldots,n\) and \(j=1,\ldots,k\). The condition effect tests whether the population means are equal:

H_0:\mu_1=\mu_2=\cdots=\mu_k \qquad\text{versus}\qquad H_a:\text{not all condition means are equal.}

The usual univariate repeated-measures F test assumes independent participants, a suitable normal model for the within-participant measurements (or contrasts), and sphericity for its exact uncorrected reference distribution. Sphericity concerns the variances of the differences between condition measurements; it is not the same as requiring every condition to have the same variance. The Greenhouse–Geisser adjustment is used when sphericity may not hold. It adjusts the reference distribution, but it does not repair dependence between different participants, severe departures from the normal model, or other problems with the study design.

The epsilon used here is an estimate from the within-subject covariance information. It reflects how far the covariance pattern is from sphericity: values closer to 1 indicate less adjustment, while smaller values indicate a greater reduction in the reference degrees of freedom. For \(k\) conditions, it lies between \(1/(k-1)\) and 1. Its estimation is a separate calculation; once an epsilon is supplied, applying the correction requires only multiplication and an upper-tail probability calculation.

What Changes—and What Does Not

For a complete repeated-measures design, the uncorrected condition F statistic is the condition mean square divided by the error mean square:

F_{\text{observed}}=\frac{MS_{\text{condition}}}{MS_{\text{error}}}.

The correction does not change either mean square, so \(F_{\text{observed}}\) stays the same. It changes the distribution used to judge how unusual that statistic would be under \(H_0\). In particular, multiply both uncorrected degrees of freedom by the same epsilon. The corrected p-value is then the probability, under that corrected F distribution and assuming \(H_0\), of obtaining an F statistic at least as large as the observed one.

Key result: If the uncorrected test has \(df_1\), \(df_2\), and observed statistic \(F\), then the Greenhouse–Geisser test uses \(F(df_{1,GG},df_{2,GG})\), where \(df_{1,GG}=\hat{\epsilon}_{GG}df_1\) and \(df_{2,GG}=\hat{\epsilon}_{GG}df_2\). Its upper-tail p-value is \(P(F_{df_{1,GG},df_{2,GG}}\geq F)\). The observed \(F\) is unchanged; its reference distribution is changed.

This is an approximate correction: the adjusted degrees of freedom need not be whole numbers. Statistical software evaluates the F distribution for fractional degrees of freedom as well as integer ones. Do not round the corrected degrees of freedom to whole numbers before calculating the p-value.

Worked Applications

Worked Example: Correcting a Three-Condition Test

Suppose 12 hypothetical participants each complete three repeated conditions. Their condition means are 10, 12, and 14 units. A repeated-measures ANOVA summary gives \(SS_{\text{condition}}=96\), \(SS_{\text{error}}=211.2\), and a computed Greenhouse–Geisser epsilon of 0.50. Apply the correction at \(\alpha=0.05\).

Check the uncorrected degrees of freedom. Here \(k=3\) and \(n=12\), so \(df_1=k-1=2\) and \(df_2=(n-1)(k-1)=11(2)=22\). The condition mean square is \(MS_{\text{condition}}=96/2=48\), and the error mean square is \(MS_{\text{error}}=211.2/22=9.6\).

Find the observed statistic.

F_{\text{observed}}=\frac{48}{9.6}=5.00.

Correct both degrees of freedom.

df_{1,GG}=0.50(2)=1, \qquad df_{2,GG}=0.50(22)=11.

The corrected test uses \(F(1,11)=5.00\). Its upper-tail probability is \(P(F_{1,11}\geq 5.00)\approx 0.04702\). Since \(0.04702<0.05\), reject \(H_0\) at the 5% significance level. The data provide evidence that not all three population condition means are equal. This omnibus test does not identify which means differ.

The summary is consistent with the stated means: for 12 participants, the condition sum of squares based on means 10, 12, and 14 is \(12[(10-12)^2+(12-12)^2+(14-12)^2]=96\). The degrees of freedom and mean squares therefore produce the reported observed statistic.

Worked Example: A Stronger Adjustment With Five Conditions

In a second hypothetical study, 12 participants each complete five repeated conditions. The condition means are 18, 19, 20, 21, and 22 units. The ANOVA summary has \(SS_{\text{condition}}=120\) and \(SS_{\text{error}}=264\), and the estimated Greenhouse–Geisser epsilon is 0.25. Calculate the corrected test at \(\alpha=0.05\).

The uncorrected degrees of freedom are \(df_1=5-1=4\) and \(df_2=(12-1)(5-1)=44\). Thus, \(MS_{\text{condition}}=120/4=30\) and \(MS_{\text{error}}=264/44=6\), giving

F_{\text{observed}}=\frac{30}{6}=5.00.

Now multiply each degree of freedom by 0.25:

df_{1,GG}=0.25(4)=1, \qquad df_{2,GG}=0.25(44)=11.

The corrected reference distribution is again \(F(1,11)\), so the corrected p-value is \(P(F_{1,11}\geq 5.00)\approx 0.04702\). At \(\alpha=0.05\), reject \(H_0\); the test provides evidence that at least one of the five population condition means differs from the others.

The condition sum of squares agrees with the listed means: their grand mean is 20, and \(12[(18-20)^2+(19-20)^2+(20-20)^2+(21-20)^2+(22-20)^2]=120\). The adjustment substantially reduces the degrees of freedom from \((4,44)\) to \((1,11)\). It does not reduce the observed F statistic. This example also shows why the number of conditions alone does not determine the corrected degrees of freedom: the epsilon matters.

Worked Example: A Corrected Test That Is Not Significant

Consider 12 hypothetical participants measured under three repeated conditions, with means of 8, 9, and 10 units. Suppose \(SS_{\text{condition}}=24\), \(SS_{\text{error}}=88\), and the estimated epsilon is 0.50. Determine the corrected p-value and decision at \(\alpha=0.05\).

The uncorrected degrees of freedom are \(2\) and \(22\). Therefore, \(MS_{\text{condition}}=24/2=12\) and \(MS_{\text{error}}=88/22=4\), so

F_{\text{observed}}=\frac{12}{4}=3.00.

The corrected degrees of freedom are \(df_{1,GG}=0.50(2)=1\) and \(df_{2,GG}=0.50(22)=11\). The corrected p-value is \(P(F_{1,11}\geq 3.00)\approx 0.110\). Because \(0.110>0.05\), do not reject \(H_0\). These data do not provide sufficient evidence at the 5% level that the population condition means differ. This conclusion is not proof that the means are equal.

The condition sum of squares is consistent with the means: their grand mean is 9, and \(12[(8-9)^2+(9-9)^2+(10-9)^2]=24\). The test statistic and corrected degrees of freedom therefore follow from the stated hypothetical summary.

Why the Same Statistic Can Have a Different P-Value

The examples used the same observed statistic, \(F=5.00\), in two designs, yet the corrected degrees of freedom depended on the uncorrected degrees of freedom and epsilon. An F distribution is determined by both degrees of freedom. Changing them can change the upper-tail probability even when \(F\) stays fixed. Thus, reporting the observed F statistic without the degrees of freedom and p-value leaves out essential information.

The reason the correction scales both degrees of freedom can be seen directly from the standard repeated-measures test. Under its usual setup, the condition effect has \(k-1\) degrees of freedom and the error has \((n-1)(k-1)\). Applying the same estimated epsilon to each gives

df_{1,GG}=\hat{\epsilon}_{GG}(k-1), \qquad df_{2,GG}=\hat{\epsilon}_{GG}(n-1)(k-1).

The two values describe the adjusted reference distribution for the same observed ratio of mean squares. There is no new observed numerator or denominator mean square to calculate, and hence no new F statistic. The correction instead reduces the degrees of freedom used to assess that ratio, accounting approximately for the loss of information about independent within-subject contrasts when sphericity does not hold.

Common Mistakes

Adjusting only the numerator degrees of freedom. The Greenhouse–Geisser correction multiplies both the numerator and denominator degrees of freedom by the same epsilon.
Multiplying the F statistic by epsilon. Epsilon adjusts the reference degrees of freedom, not the ratio of mean squares. The observed F statistic remains unchanged.
Using the uncorrected degrees of freedom to find the corrected p-value. Once the correction is applied, calculate the upper-tail probability with \(df_{1,GG}\) and \(df_{2,GG}\).
Rounding fractional degrees of freedom to integers. Retain the corrected values as calculated; F distributions are defined for fractional degrees of freedom.
Interpreting a nonsignificant result as proof of equality. A large corrected p-value means the data do not provide sufficient evidence against \(H_0\) at the chosen \(\alpha\); it does not establish that \(H_0\) is true.

Start with the repeated-measures ANOVA output.
Record the observed condition F statistic, the uncorrected degrees of freedom, and the estimated Greenhouse–Geisser epsilon.

Apply epsilon to both degrees of freedom.
Calculate \(df_{1,GG}=\hat{\epsilon}_{GG}df_1\) and \(df_{2,GG}=\hat{\epsilon}_{GG}df_2\), without rounding fractional results.

Calculate and interpret the corrected p-value.
Find the upper-tail probability for the observed F statistic using the corrected degrees of freedom. Compare it with \(\alpha\), then state the conclusion about the condition means in context.

Key takeaway: Applying the Greenhouse–Geisser correction means keeping the observed repeated-measures F statistic, multiplying both of its degrees of freedom by the estimated epsilon, and finding the upper-tail p-value from that corrected F distribution.

Check Your Understanding

Use the correction steps above to answer these questions.

A repeated-measures test has uncorrected degrees of freedom \((3,27)\) and estimated epsilon 0.60. What are the corrected degrees of freedom?
When applying the Greenhouse–Geisser correction, should the observed F statistic be multiplied by epsilon? Explain.
A test has corrected degrees of freedom \((1,11)\), observed \(F=5.00\), and corrected \(p=0.04702\). State the decision at \(\alpha=0.05\) in the context of a test of repeated condition means.
Why is it incorrect to use the uncorrected degrees of freedom when calculating a Greenhouse–Geisser p-value?
A corrected p-value is greater than the chosen significance level. What conclusion is justified, and what stronger conclusion is not justified?