Bias in frequently reported analyses of subfertility trials

Randomized controlled trials in subfertility are often designed so that participants in each trial arm are offered the randomized treatment for up to k cycles, until success occurs. In clinical journals, success probabilities are often reported 'per cycle' and inferential procedures carried out as if all cycles were independent. However, this assumption does not hold. In this paper we consider the implications of such an assumption when we have a heterogeneous (Beta) patient population. We conducted a simulation study to assess the bias of estimates of the risk difference, risk ratio and odds ratio calculated using the total number of randomized women or the total number of treatment cycles as the unit of analysis, as well as the coverage of their confidence intervals (CIs). We found that the commonly reported 'per cycle' analyses are more biased and have poorer coverage of the CIs than the analyses based on the total number of women, in heterogeneous populations. Such information will help researchers who wish to interpret publications when the number of randomized participants is not extractable.