Abstract:
|
In a multi-center clinical, the estimand of interest is typically a weighted combination of the treatment effects in the individual centers. Suppose we knew the distribution of subject responses to both treatments. While common sense would suggest that it would be straightforward to estimate an overall treatment effect from this information, in practice the choice of estimand and the associated center weights is still far from obvious. Weights based upon minimizing variance are not unique, since any choice of weights leads to an estimand that can be "estimated" exactly. An estimand based upon the frequencies the different types of centers and subjects occur in a postulated super population would yield a rational choice of estimand, but this information is rarely available in practice. Moreover, clinical trials are often designed to exclude a large fraction of the subjects who comprise the target patient population, so no estimand yields the population value. This talk will discuss how a practicing statistician can choose an estimand, and evaluate the impact of the choice of estimand on inference, whether or not the treatment effects in the individual centers are known.
|