Missing Data Problems for Two Samples on a Dichotomous Variable.

The problem of comparing proportions when some data are missing is investigated, and determination is made of what statistical techniques are appropriate under each of several probability models describing the observations likely to be missing. Monte Carlo methods were used to investigate the properties of standard estimators under each of the missing data models. Applying standard techniques which ignore the occurrence of missing observations may yield misleading conclusions. Some tests and estimators are fairly robust to the model for missing data, but others may be seriously affected. If the model for missing observations is complex, the sample information may be insufficient for adequate data analysis. This report is useful since the problem of missing data is recurrent in educational research and may present serious difficulties even for simple problems. (Author/LH)