causal-inference-in-R
causal-inference-in-R copied to clipboard
malcolmbarrett
Missing data
Lucy's slides https://onlinelibrary.wiley.com/doi/full/10.1002/sim.8355:
Results indicated that the Within approach produced unbiased estimates with appropriate confidence intervals, whereas the Across approach produced biased results and unrealistic confidence intervals
https://journals.sagepub.com/doi/abs/10.1177/0962280210395740?journalCode=smma https://pubmed.ncbi.nlm.nih.gov/29165572/ https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5860553/ https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5860387/ https://academic.oup.com/aje/article/190/4/663/5923802?searchresult=1
Articles on when complete case is ok and when it is not:
https://statisticalhorizons.com/listwise-deletion-its-not-evil/ https://arxiv.org/pdf/1801.03583.pdf https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7705610/ http://people.csail.mit.edu/jrennie/trg/papers/rubin-missing-76.pdf Figure 4 https://researchonline.lshtm.ac.uk/id/eprint/1198/1/research_online_1198.pdf
it seems like null effect vs non-null effect might be relevant, perhaps by changing the dag
Imputation of some sort allows you to calculate a marginal ATE instead of using a conditional with the predictors of y
More notes after Lucy's in-depth analysis of this problem. We'll use her writing as a basis for this but a few big picture ideas:
- Stochastic (single imputation,
norm.predictin mice) without Y works because the ratio of the covariance and variance is correct - Full on MICE with Y also works but is more complex
- Ignorable vs nonignorable for causal. Does it induce confounding bias? Dags useful here
- If ignorable, imputation can improve precision because it increases N but model mentioned above does add variance, too, so it's kind of a wash. A little improvement
callout box for why we're not using terms MAR, MNCAR, etc. Not clear if talking about causal mechanism or missingness mechanism
MB Question: does propensity score model fitted with variables from deterministic model without Y need any different approaches?
Lucy doesn't think so but will investigate
exclusion criteria + missingness: https://onlinelibrary.wiley.com/doi/10.1002/sim.9685
checks for partially missing covariates: https://janickweberpals.gitlab-pages.partners.org/smdi/articles/smdi.html course materials from a class on causal effects and missingness: Missing Data Theory and Causal Effects
MI + gcomp: https://thestatsgeek.com/2023/01/31/g-formula-for-causal-inference-via-multiple-imputation/
https://bmcmedresmethodol.biomedcentral.com/articles/10.1186/s12874-024-02157-x