How to Understand and Apply Mauchly’s Test of Sphericity

The fundamental purpose of Mauchly’s test is to formally assess whether or not the crucial assumption of sphericity is adequately met within a repeated measures design. This test is a critical component of the diagnostic phase of any repeated measures ANOVA.

The Statistical Foundation: Defining Sphericity

Sphericity is a statistical concept closely related to the assumption of homogeneity of variance, but applied specifically to the differences between levels of a within-subjects factor. More precisely, sphericity requires that the variances of the differences between all pairs of the repeated measurements are equal. For example, if we have three time points (T1, T2, T3), sphericity assumes that the variance of (T1 – T2), the variance of (T2 – T3), and the variance of (T1 – T3) are all statistically equivalent in the population.

While sphericity is often discussed alongside the related condition of compound symmetry (which requires both equality of variances across conditions and equality of covariances between pairs of conditions), sphericity is the less restrictive and more relevant assumption for the validity of the F-ratio in repeated measures ANOVA. If the condition of compound symmetry is met, sphericity is also necessarily met. However, it is possible for sphericity to hold even if compound symmetry does not, making the direct test of sphericity the preferred diagnostic approach.

The failure to satisfy sphericity indicates that the correlation structure among the repeated measures is inconsistent, introducing bias into the standard errors used in the ANOVA calculations. This inconsistency makes the pooled error term, which is central to calculating the F-ratio, inaccurate. Consequently, the resulting p-values derived from the ANOVA may be misleadingly small, increasing the likelihood of committing a Type I error (rejecting a true null hypothesis).

Formulating the Hypotheses for Mauchly’s Test

As with all inferential statistical tests, Mauchly’s Test of Sphericity operates by testing a set of competing hypotheses regarding the population parameters. It assesses whether the observed variances of the differences between related groups deviate significantly from the assumption of equality.

The hypotheses used in Mauchly’s test are formulated as follows:

• H₀: The variances of the differences between all pairs of within-subject groups are equal (i.e., sphericity holds). This is the null hypothesis that researchers hope to retain.
• H_A: The variances of the differences are not equal (i.e., sphericity is violated). This is the alternative hypothesis, suggesting a problem with the data structure.

The goal of the test is not necessarily to reject the null hypothesis, but rather to confirm that the assumption holds. If the statistical evidence strongly suggests that the variances are unequal (leading to the rejection of H₀), researchers must proceed with caution and implement corrective measures.

Interpreting the Results and P-Value

The decision to reject or fail to reject the null hypothesis is based on the p-value generated by Mauchly’s test compared against a predetermined significance level (alpha, typically $alpha = .05$). The interpretation is straightforward but crucial for the subsequent ANOVA analysis.

If the p-value of Mauchly’s test is less than the chosen significance level (e.g., $alpha = .05$), then we have sufficient evidence to reject the null hypothesis. We conclude that the variances of the differences are statistically unequal, meaning the assumption of sphericity has been violated. In this situation, the standard Repeated Measures ANOVA results cannot be trusted without adjustment.

Conversely, if the p-value is greater than or equal to the significance level (e.g., $p ge .05$), we fail to reject the null hypothesis. We conclude that the assumption of sphericity is met, or at least that there is insufficient statistical evidence to suggest a violation. When sphericity is met, the researcher can proceed with the standard Repeated Measures ANOVA using the uncorrected degrees of freedom.

Practical Example: Heart Rate Monitoring Study

To illustrate the application and interpretation of Mauchly’s test, consider a medical researcher studying the impact of a physical training program on resting heart rate. The researcher employs a repeated measures design, collecting data from the same subjects at three distinct time points:

• Time 1 (T1): One month before starting the training program (Baseline).
• Time 2 (T2): In the middle of the training program.
• Time 3 (T3): One month after completing the training program (Follow-up).

The primary goal is to perform a Repeated Measures ANOVA to determine if there is a significant change in the mean resting heart rate across these three longitudinal measurements. Before interpreting the F-ratio for the main effect of time, the researcher must first verify the sphericity assumption using Mauchly’s test.

The raw data collected across the three time points would be used to calculate the variances of the differences: Var(T1-T2), Var(T2-T3), and Var(T1-T3). If these variances are observed to be unequal in the sample, Mauchly’s test determines if this inequality is significant enough to suggest a population-level violation of sphericity.

The following table represents hypothetical results from the data collection, showing the subject measurements across the three time points. Although we often observe differences in variances in sample data, the statistical test determines if these differences warrant rejecting the null hypothesis.

While a visual inspection of the calculated differences might suggest that the variances are not perfectly equal, we rely on sophisticated statistical software (such as R, SPSS, or Python libraries) to perform Mauchly’s test of sphericity and generate the precise p-value necessary for a formal decision.

Analyzing Statistical Software Output

When running the analysis, the statistical software will produce a dedicated output section for Mauchly’s test. This output typically includes the test statistic ($chi^2$), the corresponding degrees of freedom (df), and the critical p-value (Sig.).

The output might look similar to this example:

In this specific example, the accompanying text block summarizes the finding:

Mauchly’s test of sphericity indicates that the assumption of sphericity has not been violated, X²(2) = 1.867, p = .356.

Based on this output, since the p-value ($p = .356$) is significantly greater than the standard alpha level ($alpha = .05$), we fail to reject the null hypothesis. Therefore, the researcher can confidently conclude that the assumption of sphericity is met for this heart rate data set, and the subsequent ANOVA results can be interpreted without needing corrective factors.

Addressing Violations: Sphericity Corrections

A more challenging scenario arises when the p-value of Mauchly’s test falls below the significance level (e.g., $p < .05$), compelling us to reject the null hypothesis. When sphericity is violated, the standard F-ratio calculation is biased, leading to an inflated test statistic and an increased risk of Type I error.

In such cases, the established statistical procedure requires applying a correction factor (often denoted by epsilon, $epsilon$) to the degrees of freedom used in the F-test calculation. This adjustment effectively penalizes the test by making it more difficult to find significance, thereby controlling the Type I error rate back to the nominal alpha level.

The correction factor, epsilon, estimates the degree to which sphericity has been violated. The corrected degrees of freedom are calculated by multiplying the original degrees of freedom by this epsilon value. This correction results in a more conservative F-test, which is necessary to maintain the integrity of the statistical conclusions.

Overview of Correction Methods

Statisticians have developed several well-recognized methods for adjusting the degrees of freedom when Mauchly’s test indicates a violation. The choice among these methods often depends on the severity of the violation and the desired level of statistical conservatism. The three most commonly applied corrections are:

1. Huynh-Feldt Correction: This method provides the least conservative adjustment. It is generally preferred when the estimated epsilon ($hat{epsilon}$) is close to 1 (indicating only a slight violation). The Huynh-Feldt epsilon is calculated to avoid over-correcting, especially in smaller samples.
2. Greenhouse–Geisser Correction: This is the most widely reported and generally accepted correction method. It applies a more substantial penalty than Huynh-Feldt. If the estimated epsilon is less than .75, the Greenhouse–Geisser correction is typically recommended as the standard procedure for correcting the inflated F-ratio.
3. Lower-bound Correction: This method represents the most conservative adjustment possible. The lower-bound correction sets epsilon equal to $1/(k-1)$, where $k$ is the number of repeated measures. This is essentially the maximum possible reduction in degrees of freedom, useful as a benchmark or when the violation is extreme.

Each of these corrections tends to increase the p-values in the output table of the Repeated Measures ANOVA to properly account for the fact that the assumption of sphericity is violated, ensuring robust and accurate statistical inference.

Further Resources on Repeated Measures ANOVA

The following tutorials provide additional information on how to perform a Repeated Measures ANOVA: