Remarks about #m-bias; bigger picture: why #longitudinal data are generally needed for #causalinference ( #causaldiagrams are not enough)

https://go-bayes.github.io/b-causal/posts/m-bias/m-bias.html

My #introduction:

I repurpose observational #RealWorldData into scientific evidence for the prevention and treatment of human disease. At #CAUSALab, we often do so by explicitly emulating a #TargetTrial. Other times we analyze #RandomizedTrials.

I teach #causalinference methods at the #Harvard T.H. Chan School of #PublicHealth. My online course #CausalDiagrams and “Causal Inference" #WhatIfBook (with James Robins) are free. See my profile.

#epidemiology #epiverse
#statistics
#datascience
#AI

4/
The key point:

Explicit #TargetTrial emulation prevents mistakes in observational data analyses that we wouldn't make in analyses of randomized trials.

That we can do something with the data doesn't imply that the result is causally interpretable.

More here:
https://www.nejm.org/doi/10.1056/NEJMp2113319

If you're looking for a gentle introduction to #causaldiagrams, #selectionbias, and colliders
enjoy our free course
"Causal Diagrams: Draw Your Assumptions Before Your Conclusions"
https://www.edx.org/course/causal-diagrams