Comparison of tests for trend

We compared the size, power and shape of the power function of nine two-sided exact, mid-p and asymptotic tests for trend in an ordered 2xK table with a sample size not that large. The twice-the-smaller-tail and the conditional score mid-p tests emerged as the tests of choice with the Cochran-Armitage test, a plausible substitute for them. The exact power functions of all tests displayed aberrant features, like diminishing power with an increasing effect size, in design configurations in which sparse data outcomes were not rare. When planning a dose-response study, it is therefore prudent to determine exact power over a broad range of parameter values