Combining Information from Similar Experiments: I. Statistical Issues

A general method for combining information from similar experiments is illustrated in the case where two independent experiments are designed to estimate a dose‐response curve. By accounting for variability across experiments, the proposed method avoids inferential pitfalls such as extended forms of Simpson's paradox. The validity of the method is supported by seven fundamental assumptions about data from replicated experiments. In contrast, an example indicates that failing to reject a preliminary test of equal distributions is inadequate justification for pooling data from two experiments. Methods that account for the variability across experiments in terms of known covariates are also discussed.