An Empirical Evaluation of the Predictive Mean Matching Method for Imputing Missing Values

This article reports empirical explorations of how well the predictive mean matching method for imputing missing data works for an often problematic variable—income—when income is used as an explanatory variable in a substantive regression model. It is found that the performance of the predictive mean method varies considerably with the predictive power of the imputation regression model and the percentage of cases with missing data on income. In comparisons of single-value with multiple-imputation methods, it also is found that the amount of bias and the loss of precision associated with single-value methods is considerably less than that associated with a weak imputation model. Situations in which using imputed data can lead to seriously biased estimates of regression coefficients (and related statistics) and situations in which the bias is so minimal as to be nonproblematic are identified.