Calculating Corrected F Tests From ANOVA Output

From an ANOVA Table to a Corrected Test

Repeated-measures ANOVA software may report an uncorrected F statistic alongside several corrected p-values. When an epsilon value is supplied, you can calculate the corresponding corrected test directly from the uncorrected table. The essential distinction is that the correction changes the degrees of freedom used to find the reference distribution; it does not change the observed F statistic.

This calculation matters when the usual univariate repeated-measures F reference distribution is not appropriate because sphericity is not satisfied. The Greenhouse–Geisser and Huynh–Feldt procedures, for example, use an epsilon to adjust the reference degrees of freedom. Here the task is to apply a supplied epsilon, not to estimate it or choose among correction methods.

Definition: Let \(df_1\) be the uncorrected numerator degrees of freedom for a repeated-measures effect and \(df_2\) its uncorrected denominator degrees of freedom. For a supplied correction value \(\epsilon\), the corrected degrees of freedom are \(df_1^*=\epsilon df_1\) and \(df_2^*=\epsilon df_2\). The corrected p-value is the upper-tail probability of the observed F statistic under an F distribution with \(df_1^*\) and \(df_2^*\) degrees of freedom.

What the Correction Changes—and What It Does Not

For a univariate repeated-measures effect, the observed statistic is formed from the effect mean square divided by its error mean square. In a standard uncorrected ANOVA table, the numerator and denominator degrees of freedom determine those mean squares from their sums of squares. Once the table has provided the observed F statistic, the epsilon correction uses that same statistic and changes the degrees of freedom for its reference distribution.

F_{\text{obs}}=\frac{MS_{\text{effect}}}{MS_{\text{error}}} \qquad (df_1^*,df_2^*)=(\epsilon df_1,\epsilon df_2)

The test’s p-value is then calculated as \(P(F_{df_1^*,df_2^*}\geq F_{\text{obs}})\). The degrees of freedom may be fractional; they are parameters of the reference F distribution and do not need to be rounded to integers. Use the unrounded products when entering values into software or a calculator, then report a sensible number of significant figures.

The calculation assumes that the supplied epsilon is applicable to the repeated-measures effect and that the uncorrected table provides the relevant effect and error degrees of freedom and F statistic. As with a repeated-measures F analysis generally, observations from different participants must be independent, the same participants contribute measurements across conditions for the analysis shown, and the model’s normality conditions should be reasonably appropriate. Epsilon correction adjusts the reference degrees of freedom for departure from sphericity; it does not make dependent participants independent or cure other model problems.

A Short Justification for Keeping F Fixed

The correction is a change to the reference distribution, not a rescaling of the observed effect-to-error ratio. This follows directly from how the statistic is calculated: the observed F is already the ratio of the two mean squares reported in the uncorrected table. The correction uses epsilon to reduce both degrees of freedom by the same factor, then evaluates that unchanged ratio against the adjusted F distribution.

F_{\text{obs}} =\frac{SS_{\text{effect}}/df_1}{SS_{\text{error}}/df_2} \quad\longrightarrow\quad F_{\text{obs}}\text{ is retained, while } (df_1,df_2)\text{ become }(\epsilon df_1,\epsilon df_2)

In other words, epsilon is not multiplied by F, and it is not applied to the sums of squares. Its role in this calculation is to adjust both degrees of freedom. The corrected p-value can differ from the uncorrected p-value because the F distribution’s shape and upper-tail probability depend on its degrees of freedom.

Calculating the Corrected Test

A reliable calculation follows the same sequence each time: record the uncorrected values, adjust both degrees of freedom, and find the upper-tail probability for the unchanged F statistic. Keep the hypotheses and the interpretation tied to the repeated-measures effect being tested.

Read the uncorrected output.
Record \(F_{\text{obs}}\), \(df_1\), and \(df_2\) for the effect of interest.

Apply epsilon to both degrees of freedom.
Calculate \(df_1^*=\epsilon df_1\) and \(df_2^*=\epsilon df_2\), keeping fractional results.

Find the corrected upper-tail probability.
Calculate \(P(F_{df_1^*,df_2^*}\geq F_{\text{obs}})\) using an F-distribution calculator or statistical software.

Report the test in context.
State the unchanged F statistic, corrected degrees of freedom, p-value, and conclusion relative to the chosen \(\alpha\).

Worked Example: Correcting a Four-Condition Test

Suppose a hypothetical repeated-measures study compares a quantitative response across four conditions for 15 participants. The uncorrected ANOVA table gives \(SS_{\text{effect}}=24\), \(df_1=3\), \(SS_{\text{error}}=140\), and \(df_2=42\). The supplied epsilon is \(\epsilon=0.68\). Calculate the corrected test.

First check the reported F statistic from the sums of squares. The effect mean square is \(24/3=8\), and the error mean square is \(140/42=3.3333\). Thus,

F_{\text{obs}}=\frac{24/3}{140/42} =\frac{8}{3.3333} \approx 2.400

Now multiply each uncorrected degree of freedom by \(0.68\):

df_1^*=0.68(3)=2.04, \qquad df_2^*=0.68(42)=28.56

The corrected p-value is the upper-tail probability for \(F_{\text{obs}}=2.400\) with \(2.04\) and \(28.56\) degrees of freedom. An F-distribution calculator gives approximately \(p=0.108\). At \(\alpha=0.05\), this is not statistically significant: the data do not provide sufficient evidence to reject the null hypothesis that the four condition means are equal. The corrected degrees of freedom are fractional, and the F statistic remains \(2.400\).

Worked Example: A Significant Corrected Test

In a second hypothetical study, 16 participants each complete five conditions. The uncorrected table reports \(SS_{\text{effect}}=46.2\), \(df_1=4\), \(SS_{\text{error}}=180\), and \(df_2=60\). A supplied epsilon is \(0.55\). Calculate and interpret the corrected test at \(\alpha=0.05\).

The uncorrected mean squares are \(46.2/4=11.55\) for the effect and \(180/60=3.000\) for error. Their ratio is

F_{\text{obs}}=\frac{11.55}{3.000}=3.850

The corrected degrees of freedom are

df_1^*=0.55(4)=2.20, \qquad df_2^*=0.55(60)=33.00

Using the F distribution with numerator degrees of freedom \(2.20\), denominator degrees of freedom \(33.00\), and \(F_{\text{obs}}=3.850\), the upper-tail probability is approximately \(p=0.028\). Since \(0.028<0.05\), the corrected test is statistically significant at the 5% level. In this hypothetical setting, the result provides evidence that not all five population condition means are equal. It does not, by itself, identify which conditions differ.

Worked Example: Fractional Degrees of Freedom Near the Significance Threshold

A third hypothetical output concerns four repeated conditions measured on 16 participants. The uncorrected table has \(SS_{\text{effect}}=28.8\), \(df_1=3\), \(SS_{\text{error}}=135\), and \(df_2=45\). Suppose the supplied epsilon is approximately \(0.7167\). Find the corrected degrees of freedom and p-value.

The uncorrected effect and error mean squares are \(28.8/3=9.6\) and \(135/45=3.0\). Therefore, \(F_{\text{obs}}=9.6/3.0=3.200\). Applying the supplied epsilon gives

df_1^*=0.7167(3)\approx 2.15, \qquad df_2^*=0.7167(45)\approx 32.25

The corrected p-value is

P(F_{2.15,32.25}\geq 3.200)\approx 0.051

At \(\alpha=0.05\), \(0.051\) is just above the significance level, so the corrected test is not statistically significant at 5%. The result is close to the threshold, but it should not be rounded in a way that changes the conclusion: report the p-value as approximately \(0.051\), rather than treating it as \(0.05\) or declaring it significant.

Reading and Reporting the Result

A clear report separates the observed statistic from its corrected reference distribution. For example: “The repeated-measures effect had \(F(2.20,33.00)=3.850\), corrected \(p=0.028\).” The values in parentheses are the corrected numerator and denominator degrees of freedom. Stating that they are corrected can prevent confusion if the output also lists the original degrees of freedom.

A p-value is calculated assuming the null hypothesis and the test’s model conditions; it is not the probability that the null hypothesis is true. When comparing the corrected p-value with \(\alpha\), describe the evidence against the null hypothesis, rather than claiming that the correction proves the null or alternative. If the p-value is above \(\alpha\), the appropriate conclusion is that the test does not provide sufficient evidence to reject \(H_0\), not that the condition means have been shown to be identical.

Common Mistakes

Changing the F statistic. Multiplying \(F_{\text{obs}}\) by epsilon or recomputing a “corrected F” is not the degrees-of-freedom correction. Retain the F value from the uncorrected table.
Adjusting only one degree of freedom. Multiply both the numerator and denominator degrees of freedom by the supplied epsilon.
Rounding fractional degrees of freedom to integers. The corrected F distribution can use fractional degrees of freedom. Rounding prematurely changes the p-value.
Using a lower-tail probability. The p-value for this F test is the upper-tail area, \(P(F\geq F_{\text{obs}})\).
Reporting the uncorrected p-value as corrected. A corrected p-value must use the corrected degrees of freedom, even though the F statistic remains the same.
Overstating a near-threshold result. A p-value of \(0.051\) is greater than \(0.05\); under a 5% rule it is not statistically significant.

Key takeaway: Given an uncorrected repeated-measures F statistic and a supplied epsilon, calculate \(df_1^*=\epsilon df_1\) and \(df_2^*=\epsilon df_2\), then find the upper-tail probability for the unchanged \(F_{\text{obs}}\). Keep fractional degrees of freedom in the calculation and state the corrected test clearly.

Check Your Understanding

Answer each question using the corrected-test procedure.

An uncorrected table reports \(F_{\text{obs}}=2.75\), \(df_1=4\), and \(df_2=48\). If \(\epsilon=0.60\), what are the corrected degrees of freedom?
For the test in Question 1, should the observed F statistic be multiplied by \(0.60\)? Explain.
Why is the corrected p-value an upper-tail probability rather than a lower-tail probability?
A corrected test has \(F_{\text{obs}}=3.10\), corrected degrees of freedom \(2.40\) and \(28.80\), and \(p=0.047\). What is the decision at \(\alpha=0.05\), stated in context-neutral statistical language?
A student rounds corrected degrees of freedom \(2.15\) and \(32.25\) to 2 and 32 before finding the p-value. What is the problem with this approach?