Detection of atypical data in multicenter clinical trials using unsupervised statistical monitoring

A risk-based approach to clinical research may include a central statistical assessment of data quality. We investigated the operating characteristics of unsupervised statistical monitoring aimed at detecting atypical data in multicenter experiments. The approach is premised on the assumption that, save for random fluctuations and natural variations, data coming from all centers should be comparable and statistically consistent. Unsupervised statistical monitoring consists of performing as many statistical tests as possible on all trial data, in order to detect centers whose data are inconsistent with data from other centers.

We conducted simulations using data from a large multicenter trial conducted in Japan for patients with advanced gastric cancer. The actual trial data were contaminated in computer simulations for varying percentages of centers, percentages of patients modified within each center and numbers and types of modified variables. The unsupervised statistical monitoring software was run by a blinded team on the contaminated data sets, with the purpose of detecting the centers with contaminated data. The operating characteristics (sensitivity, specificity and Youden's J-index) were calculated for three detection methods: one using the p-values of individual statistical tests after adjustment for multiplicity, one using a summary of all p-values for a given center, called the Data Inconsistency Score, and one using both of these methods.

The operating characteristics of the three methods were satisfactory in situations of data contamination likely to occur in practice, specifically when a single or a few centers were contaminated. As expected, the sensitivity increased for increasing proportions of patients and increasing numbers of variables contaminated. The three methods showed a specificity better than 93% in all scenarios of contamination. The method based on the Data Inconsistency Score and individual p-values adjusted for multiplicity generally had slightly higher sensitivity at the expense of a slightly lower specificity.

The use of brute force (a computer-intensive approach that generates large numbers of statistical tests) is an effective way to check data quality in multicenter clinical trials. It can provide a cost-effective complement to other data-management and monitoring techniques.