The Mantel-Haenszel test for grouped failure time data (MHF test) comp
ares the distribution of failure times in two cohorts followed for an
interval of time when the data are collected in discrete subintervals.
This paper derives approximations to the power of the Mantel-Haenszel
test for arbitrary failure time distributions in the presence of cens
oring. The approximations are appropriate for both equal and nonequal
odds ratios in the constituent tables, and can be used for arbitrary s
ubdivisions of time. Four approximations are proposed. They differ fro
m each other according to whether the parameter measuring treatment ef
fect is an odds ratio or a difference in proportions, and whether the
survival distributions are calculated under the null or alternative hy
pothesis. In addition, we demonstrate that when the hazards are consta
nt, increasing the number of subintervals often produces only a neglig
ible increase in the power of the MHF test. On the other hand, for arb
itrary hazards and nonconstant hazard ratios, the choice of frequency
and actual times of measurement can have important effects on power. F
inally, the paper presents simple expressions for power under exponent
ial failure and censoring models.