Red Door Analytics
Resources · Methods

Causal inference

Everything Red Door Analytics has published on causal inference — the packages that fit it, the tools that show it, the tutorials that work it through, and where it decided something.

Estimating what would happen under one intervention rather than another, from data where people were not randomised. The estimand is a contrast between two versions of the same cohort, and it rests on assumptions the model cannot check.

Also called target trial emulation, g-methods.

Causal inference: Both curves describe the same people, twice: the model is fitted conditional on covariates, then everyone is predicted as if treated and again as if untreated, and each set averaged over the cohort's own covariate distribution.
Both curves describe the same people, twice: the model is fitted conditional on covariates, then everyone is predicted as if treated and again as if untreated, and each set averaged over the cohort's own covariate distribution. The estimand is the vertical gap, not a hazard ratio — and whether it means what you want turns entirely on why the treated and untreated differed to begin with, which is an assumption about the data and not something the model can check.
Worked through2
What is immortal time bias? What is landmark analysis?
The evidence2
From our publications.
Where it decided something1
Does the procedure really improve survival — or is it patient selection?

Working on causal inference?

We do this every week — as a collaborator on your study, or as the people who wrote the software you are using.

Get in touch Start a project