Detecting systematic errors in multi-clinic observational data

In multi-clinic studies it is hard to maintain a uniformly high quality of measurement and coding. Systematic errors almost always occur, in spite of the best of intentions and the most rigid protocols. It is the statistician's responsibility to plan for the detection of these errors, as well as to try to avoid them and not be misled by them. The practice of examining the univariate and multivariate sample frequency distributions of the variables under study, with an eye open for anything that looks puzzling, can be very helpful in detecting and trying to correct systematic errors that would bias the analysis. Examples are given from a 21-clinic study on pregnancy and child development.