Small Sample Properties of Some Estimators of a Common Hazard Ratio

The simulated small‐sample behaviours of several estimators of a common hazard ratio are compared. In most circumstances, a modified version of the empirical logit (Haldane 1955; Anscombe 1956) seems to be the analytic technique of choice. The performance of the Standardized Mortality Ratio (SMR) is indistinguishable from that of the ML estimator when the denominator experience is much larger than that of the numerator experience. When many cells have small expectations, then the empirical logit becomes biased, and a variant of the Mantel–Haenszel estimator (Mantel and Haenszel, 1959) is the most widely reliable non‐iterative estimator.