Comparing imperfect measures of exposure.

With the implication of dietary factors in the origin of many common diseases, there has been renewed interest in methods of dietary enquiry suitable for epidemiologic research (1-14). Validation of both new and old approaches proceeds, but under severe constraints. In the development of questionnaires, for example, the need may arise to compare two alternative, simple methods each to a reference which is taken to be accurate though impractical for large scale use. Only rarely can the two alternatives and the reference be tested together in a single individual, since subjects may remember previous answers in responding to new questions (7). This note begins with a consideration of statistical techniques by which two alternative questionnaires can be compared indirectly to one another when only pairwise administration is feasible. Concentrating on the estimation of a single score (e.g., total intake of a given nutrient), we briefly review techniques for the comparison of two correlation coefficients, and then explore some of the implications of imperfect correlation for epidemiologic study design. The areas of study design covered are changes in the power of the Mantel-Haenszel extension procedure (15) to detect a trend in ordered categorical responses under misclassification, sample sizes for pilot comparisons of imperfect questionnaires, and estimation of the validity of questionnaires from repeated administration. Some previous relevant work in these areas includes that of Quade et al. (16), who con-