One of my blog articles on fharrell.com frames this in terms of mean squared error of the effect estimate. Inclusion of differential effects can increase MSE if strength of interacting factor is low.

That’s well put and has certainly been my experience with many smart clinical colleagues who fall into the trap of looking for HTE when there is no evidence for an average treatment effect. Effective treatments should reduce disease, creating anti-heterogeneity (treated pts -> more alike) #StatsSky

It uses change from baseline as a patient response, which is known to be very problematic + includes regression to the mean. And if the method identified HTE that could not be detected by Tx x covariate interactions, the resulting HTE is not actionable anyway, right?

There must have been many clue but I'm not sure.

Yes that can occur if there is in interaction between the marker of “sicker” and treatment effect. But given that both trials study the same disease etiologically speaking it is most common for one to merely observe risk magnification in action.

I just put a comment on www.sensible-med.com/p/the-pisces... about it's silly interpretation of an "insignificant" overall mortality difference.

It’s hard in general for a risk ratio to be as transportable as an odds ratio, e.g. a risk ratio of 2 cannot apply to a base risk of 0.51. hbiostat.org/rmsc/lrm

I’m not seeing a direct effect but would like your thoughts of how the patient population or interpretation of a trial might change because of this.

Other than the usual DAG-related challenges of aiming for a causal interpretation of effect of patient management choice, this needs to use a flexible model (allowing for ties at zero time) that accounts for death e.g. multistate model. Key estimand is expected time alive in a good state.

Yes I’d like to see more of such comparisons. Sometimes I do that in statistical reports using R/Quarto with different graphical attempts in different tabs, allowing easy switching for comparison but leaving my preferred one as the default open tab. #RStats #StatsSky #graphics

And here is an example using interactive graphics (hover text to get more detail) that combines confidence interval information with raw data points by using 1/2 width confidence intervals. These have the property of the 2 points being connected if and only if p > 0.05 hbiostat.org/rflow/descri...

Here’s the example: hbiostat.org/bbr/images/a...

Michael - this one doesn’t do it for me. I’d rather see a horizontal dot chart with vertical categories (countries) sorted in descending order of overall life expectancy, with horizontal dots for F, M augmented by CIs for difference in life expectancy, as in hbiostat.org/bbr/descript... #RStats