Analysis of differences in proportions from clustered data with multiple measurements in diagnostic studies

In diagnostic studies, proportions such as sensitivities are often to be calculated and to be compared between different diagnostic procedures. As statistical unit of analysis, multiple observational units may be assessed within each patient, i.e., multiple lesions in an organ. As a requirement, these are to be assessed by multiple blinded readers. In this paper we propose a method to cover correlations between units within patients, correlations between procedures and correlations between different raters assessing each observational unit.

The proposed approach is a two-step method to analyze clustered data with multiple measurements to compare diagnostic procedures in a paired modality design and the correlation between the readers in a paired reader design. The performance of the approach was compared to a generalized estimation equations model (GEEs) by power simulations.

Power simulations suggest, that the two-step approach is not inferior to GEEs with regard to the single readers as well as with regard to the average reader.

An intuitive approach was developed next to established methods to analyze "paired modality, paired reader" and "unpaired modality, paired reader" studies with binary endpoints when estimating proportions and differences in proportions for clustered data with multiple measurements. This two-step approach is an alternative method to cover routine designs of diagnostic studies where the difference of proportions is to be estimated directly along with confidence intervals.