Cautions on the reanalysis of epidemiologic databases

The statistical and epidemiologic literatures contain many examples of the reanalyses of medical databases. Some are confirmatory, such as the many reanalyses of the Stanford heart transplant data. Others, however, concern issues that differ from those that the data originally intended to address. We demonstrate in this paper that the second kind of reanalysis, if undertaken uncritically, can produce biased results. Utilizing a log-linear model, we show analytically that the reanalysis of a designed study without adjustment for the design variable can produce biased results. We express the magnitude of this bias as a function of second- and third-order interactions in the data. We demonstrate further that preliminary testing for the existence of second- and third-order interactions can provide substantial bias reduction.