A general estimator for the variance of the Mantel-Haenszel odds ratio.

This commentary presents a new estiof matching variables, unmatched varimator of the variance of the natural log of ables, or both. This notation encompasses the Mantel-Haenszel odds ratio and illusthe full spectrum of matched and untrates its application in typical case-control matched designs. For example, in the pairdesigns. In contrast with any single previmatched study, MI and Mo are always one; ously proposed estimator (1-7), this estiin a design with J controls uniquely mator is easily computed and can be used matched to each case (1: J matching), MI in the analysis of data from any sampling is always one and Mo is always J; in an design (e.g., individually matched, category unmatched design, MI and Mo will usually matched, or unmatched). Furthermore, in vary in an arbitrary fashion across the a series of Monte Carlo simulations, this strata. estimator yielded confidence limits as good Suppose now that the parameter of inas or better than those obtained using preterest is the odds ratio and the odds ratio vious estimators (8). has a common (constant) value across strata. If R = AD/T and S = BC/T, the BACKGROUND Mantel-Haenszel estirnator2f the common We will use the following notation to odds ratio is defined as ORMH = R+/S+, where R+ = C, R, S+ = C, S, and the sumrepresent data values in a basic stratum mations are acrzss all strata. from an epidemiologic study: Valid use of O R M ~ requires that the study ' Exposed Unexposed Totals sample be "large," but there are at least two