Detection of Outlying Correlation Coefficients in Multicenter Clinical Trials.

Abstract

Central statistical monitoring aims at finding centers whose data distribution differs significantly from the other centers in multicentric clinical trials. Such differences may point to data quality issues due to negligence, misconduct, or fraud. Data distributions can be compared across centers in many different ways, depending on the type of data (e.g., numerical or categorical), whether a univariate or a multivariate comparison is performed, and so on. In that framework, we present two methods aimed at detecting centers with outlying bivariate Pearson correlation coefficients. One of the methods directly compares the correlations across centers. The other method conditions the test on one of the marginal standard deviations, which makes the test on correlation independent of the centers' standard deviations. Both methods are shown to perform equally well on simulated data. They are also applied on real world data, where they identify centers with outlying correlations. The findings of the two tests are compared, showing that they concord for centers with average standard deviations, but differ for centers with extreme standard deviations. While the focus here is on central statistical monitoring, the methods are general and can be used in other settings.