Selection bias in studies of major depression using clinical subjects

Selection bias may systematically distort estimates deriving from psychiatric studies using clinical subjects. Such bias may impact on cross-sectional studies using samples of convenience and also on clinical case-control studies. The objective of this report is to describe examples of such bias, and to identify probably mechanisms underlying it. A series of cases was recruited from among inpatients at a general hospital in Calgary, Canada. This case-series consisted of consenting subjects with current episodes of major depression according to a structured diagnostic interview. Comparison subjects consisted of non-depressed (according to the structured interview) individuals admitted to the same units and a sample of community subjects scoring negatively on a major depression predictor. Bayesian calculations using ancillary census and national survey data were used to estimate the selection probabilities underlying bias apparent in several of the odds ratio estimates. Neither a cross-sectional analysis incorporating all of the clinical subjects, the use of a community comparison group nor a case-control analysis using a subset of the clinical subjects resulted in valid estimation. This study confirms that the probability of selection of clinical subjects can be conditionally dependent on diagnosis and other variables in ways that create a substantial vulnerability to selection bias.