Abstract:

It would seem to be a straightforward journey. We decide the objective of a study; this leads us to estimating something – perhaps some treatment effect; we define that something clearly; then we estimate it. But, as often in statistics, if we gain clarity in one area, we can often lose in another area. A beautifully clearly defined estimand will have unacceptable assumptions; an estimand that is impregnable to midstudy changes is not of interest to stakeholders; an estimand that nicely isolates a statistic to answer a particular question is so focussed that it cannot be applied to most aspects of interest for a new treatment. If we seek to preserve the benefits of randomisation to gain a causal interpretation, that causal inference is of interest to no one in the clinic. If we yearn to estimate something of clinical interest we must join that class of dreamers, of which I am one, who try to persuade others that it may be possible to model statistically that which never happened, using explanatory variables whose adequacy we cannot test, in a population that does not exist. The talk explores these questions with examples
