Analysis of proportions from clustered data with missing observations in a matched-pair design

Typically, methods for the estimation of differences in proportions from clustered data are based on complete cases with no missing information. In this paper we propose an extension to the method of Rao and Scott and Obuchowski to allow for the explicit computation of the variance of the estimator for the difference in presence of incomplete cases.

We divided the full analysis set into a set of complete cases and a set of incomplete cases. The differences in proportions of correct diagnoses were estimated for each set by taking into consideration the clustering effect for both sets and the correlation between the procedures in the set with complete cases. Then the estimates of the two parts were combined by appropriate weights, which then allowed the explicit calculation of the variance. The performance of the extension as compared to the original method and generalized estimation equations model (GEEs) was examined by simulations.

The results of the examples suggest that the extended approach is superior to the complete-case method and is therefore appropriate when all data are to be used. In comparison to GEEs, the extended method appears to be slightly inferior, when the number of observations per patient is high, but of similar efficiency with a low number of observations per patient.

With the extension of the method by Rao and Scott and Obuchowski we make use of all available data. Therefore, we follow the intent-to-treat principle as close as possible.