Barnard–Rubin Small-Sample Degrees Of Freedom

Why Adjust the Degrees of Freedom?

Rubin’s large-sample degrees of freedom reflect uncertainty due to missing information and the finite number of imputations. But they do not, on their own, account for limited degrees of freedom in the complete-data analysis. When that analysis has a small sample or a model with many estimated parameters, treating its variance estimate as if it came from unlimited data can make inference too optimistic.

The Barnard–Rubin adjustment combines the large-sample degrees of freedom with a second quantity that reflects the complete-data degrees of freedom. This tutorial focuses on calculating and interpreting that adjustment for a scalar estimate. It does not change the pooled point estimate or the pooled variance; it adjusts the degrees of freedom used to describe uncertainty.

Here, \(\gamma\) is an adjusted estimate of the fraction of missing information for the scalar estimate. The term \(\nu_{\text{obs}}\) is the observed-data degrees-of-freedom component: it accounts for the fact that the complete-data analysis itself has finite degrees of freedom. The final value \(\nu_{\text{BR}}\) combines that component with \(\nu_{\text{old}}\).

The adjustment is used when the scalar estimate and its variance are appropriately pooled across imputations, and the complete-data analysis provides a suitable \(\nu_{\text{com}}\). The imputed data sets should arise from an adequate imputation procedure, and the estimate should have a distribution for which the usual approximate \(t\)-based degrees-of-freedom calculation is reasonable. The same scalar estimand must be analyzed in every imputation. If the complete-data degrees of freedom are not well defined or do not represent the analysis at hand, the adjustment cannot repair that problem.

How the Combination Works

The combination is easiest to understand by writing the formula for \(\nu_{\text{BR}}\) in reciprocal form:

This identity follows directly by taking the reciprocal of the product-over-sum formula. Since both degrees of freedom are positive, their reciprocal sum is larger than either reciprocal individually. Therefore, \(\nu_{\text{BR}}\) is smaller than either input. The adjustment does not simply choose whichever input is smaller; it combines both.

The observed-data component includes two corrections. The factor \((\nu_{\text{com}}+1)/(\nu_{\text{com}}+3)\) modifies the finite complete-data degrees of freedom. The factor \(1-\gamma\) reduces that quantity according to the estimated missing information. If complete-data degrees of freedom increase without bound while the other quantities remain fixed, \(\nu_{\text{obs}}\) increases without bound, and the harmonic combination approaches \(\nu_{\text{old}}\). Thus, in the large complete-data limit, the adjustment returns to Rubin’s large-sample degrees of freedom. For finite \(\nu_{\text{com}}\), it generally yields a smaller value.

A practical calculation follows a fixed order. First find \(r\) and \(\nu_{\text{old}}\). Next use \(\nu_{\text{old}}\) to calculate \(\gamma\), then calculate \(\nu_{\text{obs}}\) using \(\nu_{\text{com}}\), and finally combine the two degrees of freedom. Retaining unrounded intermediate values avoids avoidable discrepancies.

Worked Calculations Under Different Conditions

Checks, Alternatives, and Limitations

A useful diagnostic is to check the direction and scale of the result. For positive finite inputs, \(\nu_{\text{BR}}\) should be positive and below both \(\nu_{\text{old}}\) and \(\nu_{\text{obs}}\). When one input is much smaller than the other, the result should be close to the smaller input, but slightly below it. If a calculation violates either check, inspect the arithmetic and the inputs.

The most direct alternative is to use \(\nu_{\text{old}}\) alone. That is the large-sample degrees-of-freedom approximation discussed in the earlier material on large-sample degrees of freedom. It can be adequate when complete-data degrees of freedom are effectively large, but it may overstate the degrees of freedom in a small complete-data analysis. Using \(\nu_{\text{com}}\) alone also omits the contribution of missing information and the finite number of imputations. Barnard–Rubin pooling incorporates both components, but it remains an approximation: it does not correct a misspecified imputation model, an inappropriate analysis model, or a poorly chosen complete-data degrees-of-freedom value.

The adjustment also depends on stable estimates of the inputs. The number of imputations \(m\) must be at least two for \(\nu_{\text{old}}\) to be defined. The calculation of \(r\) requires a meaningful positive \(\bar U\). When quantities are extremely close to zero or very large, direct computation can be numerically delicate; keep adequate precision and verify the result against the reciprocal form. Problems at a variance boundary require separate care and should not be handled by mechanically applying a formula outside its assumptions.

Common Mistakes

• Using only \(\nu_{\text{old}}\). A large \(\nu_{\text{old}}\) does not imply that the complete-data analysis has abundant degrees of freedom. Include \(\nu_{\text{com}}\) through \(\nu_{\text{obs}}\).
• Using \(\nu_{\text{com}}\) as the final answer. The complete-data value is an input, not the adjusted pooled degrees of freedom.
• Omitting the finite correction in \(\gamma\). The term \(2/(\nu_{\text{old}}+3)\) is part of the stated adjustment; do not replace \(\gamma\) with \(r/(r+1)\).
• Rounding too early. Carry unrounded values through \(\gamma\), \(\nu_{\text{obs}}\), and the final combination, then round the reported result.
• Assuming the adjustment repairs poor modeling. It combines degrees-of-freedom information; it does not make an inadequate imputation or analysis model adequate.