In recent weeks I have encountered the same issue relating to two papers I
worked on submitted to medical journals. Both of these papers involved post
hoc analyses of a large randomised clinical trial. Both involved
non-randomised comparisons of some variable of interest. Obviously such
comparisons have limitations. Some form of multivariate adjustment for
baseline characteristics is usually employed (either standard models or
propensity scores). I know Stephen Senn has been lecturing recently on some
of the pitfalls of the latter, but it is the former my query relates to. I
have argued that the purpose of this multivariate analysis is to explain as
much of the differences between the non-randomised groupings for the outcome
of interest. Hence, I have developped comprehensive models for the outcome
using stepwise procedures. However, journals have questioned this approach
and suggested using an "epidemiological" approach with a smaller set of well
known covariates. What are people's views on this?
|