search

explodecomputer / prs-vs-mr

1 stars 0 forks source link

Possible things to explore #1

Open explodecomputer opened 5 years ago

explodecomputer commented 5 years ago

Outstanding questions:

  1. Is PRS same MR (turns out same as fixed effects IVW)
  2. Can we show this mathematically?
  3. Is PRS and IVW FE the same with weak instruments?
  4. Is summary based PRS and PRS the same?
  5. Does sample overlap bias PRS and IVW FE the same?
  6. What about 2SLS vs PRS?

Dichotomy of PRS in prediction and IVW in inferencde:

Does IVW FE have any role in prediction? Probably not, because it doesn't create a value for an individual, it just gives a population average.

Does PRS have a role to play in causal inference? It gives the same thing as IVW FE, but fails to let you test for violations e.g. heterogeneity etc. So, it has no place there.

mightyphil2000 commented 5 years ago

Maybe in a hypothesis-free context it may make sense to use IVW FE to maximise power? The latter results get interpreted as possible associations but no conclusions about causality. Then follow-up prioritised results with rigorous sensitivity analyses and replication before making causal claims?