Bias in analytic research.

CASE-CONTF~OL studies are highly attractive. They can be executed quickly and at low cost, even when the disorders of interest are rare. Furthermore, the execution of pilot case-control studies is becoming automated; strategies have been devised for the ‘computer scanning’ of large files of hospital admission diagnoses and prior drug exposures, with more detailed analyses carried out in the same data set on an ad hoc basis [l]. As evidence of their growing popularity, when one original article was randomly selected from each issue of The New England Journal of Medicine, The Lancet, and the Journal of the American Medical Association for the years, 1956, 1966 and 1976, the proportion reporting case-control analytic studies increased fourfold over these two decades (2-8”;) whereas the proportion reporting cohort analytic studies fell by half (30-159); incidentally, a general trend toward fewer study subjects but more study authors was also noted [2]. If an ebullition of case-control studies is in progress, a review of their merits and shortcomings is of more than academic interest, and this symposium was well-timed. Because this meeting also coincided with the completion of some work we had been doing on biases in analytic research (Appendix 3), I offered to summarize a portion of this work for presentation and discussion here. A first draft of a catalog of biases which may distort the design, execution, analysis and interpretation of research appears as an appendix to this paper (additions, corrections and citations of examples would be welcomed by the author). For this paper, I have considered those biases which arise in analytic studies and have focused on two subsets which affect the specification and selection of the study sample and the measurement of exposures and outcomes, since these attributes most clearly distinguish the case-control study from its relatives. * Furthermore. I have included occasional discussions of cohort analytic studies because they represent a common, alternative, subexperimental approach to determining causation. Finally, after describing the prospects for the prevention (or at least the measurement) of these biases in these two forms of analytic studies, this paper closes with suggestions for further methodologic research.