Missing (Completely?) At Random: Lessons from Insurance Studies

A dilemma frequently faced by empirical researchers is whether they should keep observations without complete information in the analysis. Assuming missingness is not biased in any perceivable direction, most studies use a complete case analysis approach, whereby only observations with complete information are kept for empirical estimation. However, the literature on statistics (e.g., Little and Rubin 2002) suggests that potential biases may arise from such practice, especially if missing data are not missing completely at random (MCAR). When there are missing data, Little’s MCAR test (1988) can be performed to reveal whether imputation methods are necessary to minimize the problems arising from incomplete data. We take two recently studied insurance data sets as examples to show that missing data issues can be better handled.