A common problem in epidemiologic studies is the inability to match all study subjects on all potential matching factors. I examined the effects of partial matching relative to full matching and no matching on the efficiency of case-control designs. For the case of two binary matching covariates, I compared five types of design: unmatched, matched on only one of the two factors, fully matched on both factors, partially matched on both factors, and a design matched on themarginal distribution of both factors. I further compared the unmatched, partially matched, and fully matched designs using a matching factor with five interval-scaled categories. Marginal matching always gave an efficiency nearly equal to full matching. Although the efficiency of partial matching often fell between that of unmatched and fully matched designs, partial matching could sometimes be more or less efficient than the extremes of no matching or full matching, as could matching on a nonconfounding correlate of an unmatched confounder. Furthermore, even when partial matching was not the most efficient of the designs it often came close to the best efficiency among the designs when only 60% of the subjects were matched. These results indicate that, in practice, complete matching is unnecessary to produce most of the efficiency benefits (or penalties) of matching. I also compar.ed various analysis options for the study designs, including the option of controlling the matching factor by entering it directly into an unconditional analysis. The results indicate that the latter strategy may be particularly worthwhile for partially matched designs involving small sample sizes. (Author/publisher)
Abstract