Huynh-Feldt Correction

Why Use the Huynh–Feldt Correction?

A univariate repeated-measures F test uses a reference F distribution whose degrees of freedom depend on the number of within-subject contrasts. When sphericity does not hold, those nominal degrees of freedom can make the test too liberal. The Greenhouse–Geisser correction addresses this by multiplying both degrees of freedom by an estimated epsilon. The Huynh–Feldt correction uses a finite-sample adjustment to that estimate, often producing a less conservative adjustment.

The key distinction is operational: both corrections leave the observed F statistic unchanged and alter its reference degrees of freedom. The Huynh–Feldt estimate is not automatically preferable; its value, its upper bound, and the resulting degrees of freedom should be understood before using it.

Greenhouse–Geisser Epsilon and Its Estimate

Suppose each of \(n\) independent subjects is measured at \(k\) occasions, and the within-subject effect is represented by \(p=k-1\) linearly independent, orthonormal contrast coordinates. Let \(\Sigma_C\) be the population covariance matrix of those \(p\) coordinates. The population Greenhouse–Geisser epsilon summarizes how evenly variation is distributed across them.

Here \(\operatorname{tr}\) denotes the trace, and \(\operatorname{tr}(S_C^2)\) means the trace of the matrix product \(S_CS_C\). Epsilon equals one when the covariance is proportional to the identity matrix, and is smaller when variation is less evenly spread across contrast directions. The earlier material on sphericity in contrast coordinates gives the relevant covariance setting; the trace expression provides a scalar summary for the degrees-of-freedom adjustment.

For a nonzero positive semidefinite matrix, the estimate lies between \(1/p\) and \(1\). The lower endpoint corresponds to variation concentrated in one direction; the upper endpoint corresponds to equal eigenvalues. The following derivation establishes these bounds.

Proof. The eigenvalues are nonnegative and not all zero. Since trace is invariant under orthogonal diagonalization, \(\operatorname{tr}S_C=\sum_i\lambda_i\) and \(\operatorname{tr}(S_C^2)=\sum_i\lambda_i^2\). Cauchy–Schwarz gives \((\sum_i\lambda_i)^2\leq p\sum_i\lambda_i^2\), so \(\hat{\epsilon}_{GG}\leq1\). Also, nonnegativity gives \(\sum_i\lambda_i^2\leq(\sum_i\lambda_i)^2\), so \(\hat{\epsilon}_{GG}\geq1/p\). \(\square\)

The Huynh–Feldt Estimate

The Huynh–Feldt (HF) estimate adjusts the Greenhouse–Geisser estimate using the number of subjects and contrast dimensions. In the complete, one-group repeated-measures setting, let \(n\) be the number of subjects and \(p=k-1\) the within-subject numerator degrees of freedom before correction. The formula below is used when \(n>p+1\), ensuring a positive denominator.

The upper cap matters because a correction factor above one would increase the nominal degrees of freedom. The raw HF estimate can exceed one, so report and use the capped value for the correction. Under the stated sample-size condition, its denominator is positive: since \(\hat{\epsilon}_{GG}\leq1\), \(n-1-p\hat{\epsilon}_{GG}\geq n-1-p>0\).

Proof. Write \(g=\hat{\epsilon}_{GG}\). The difference between the raw HF estimate and \(g\), on the positive common denominator \(p(n-1-pg)\), has numerator \[ np g-2-pg(n-1-pg)=p^2g^2+pg-2. \] Set \(x=pg\). The Greenhouse–Geisser bound \(g\geq1/p\) implies \(x\geq1\). Thus \(p^2g^2+pg-2=x^2+x-2=(x-1)(x+2)\geq0\). Dividing by the positive denominator proves \(\hat{\epsilon}_{HF}^{\,raw}\geq g\). Capping at one does not make the operational HF estimate less than \(g\), because \(g\leq1\). \(\square\)

This comparison explains the usual description of HF as less conservative: before the cap, it is no smaller than GG under these conditions, so it retains at least as much of the nominal degrees of freedom. It does not guarantee that the two estimates differ substantially, nor that HF is a better choice in every application.

Adjusting the Repeated-Measures Test

For a one-group repeated-measures effect with \(p\) uncorrected numerator degrees of freedom and \(p(n-1)\) uncorrected denominator degrees of freedom, an epsilon correction multiplies both by the same estimate. The observed univariate F statistic is computed as usual. Its p-value is then evaluated against an F distribution with the corrected degrees of freedom, which need not be integers.

The two corrected degrees of freedom are related to the uncorrected ones by the same multiplier. Therefore their ratio is unchanged: \(\{\hat{\epsilon}p(n-1)\}/(\hat{\epsilon}p)=n-1\), for positive \(\hat{\epsilon}\) and \(p\). The F statistic itself is not multiplied by epsilon. The correction changes the reference distribution used to assess that statistic.

Interpretation and a Common Mistake

An epsilon correction is useful because it provides a scalar adjustment to the univariate repeated-measures reference degrees of freedom when the covariance structure does not support the uncorrected test calibration. GG and HF make the adjustment comparable: both rescale numerator and denominator degrees of freedom, and the observed F ratio remains unchanged. The main practical difference is the estimate used as the multiplier.

A common mistake is to calculate an HF estimate greater than one and use it directly, or to multiply the observed F statistic by epsilon. The first would increase the degrees of freedom beyond their nominal values; the second changes the statistic rather than its reference distribution. The correction is to cap the estimate at one and apply it to both degrees of freedom only. Another error is to confuse the raw and operational HF estimates: show the raw calculation, then state whether the cap applies.