Controlling confounding in observational data

Observational designs handle confounding by measuring it and adjusting for it, in the design, the analysis, or both. A Delhi study of air pollution and children's breathing matched on family income, parental education and neighborhood, then adjusted further with regression. Propensity scores condense many confounders into the probability of exposure, used for matching or weighting. Inverse probability weighting builds a pseudo-population where exposure is independent of measured confounders. TMLE adds machine learning and double robustness. None of them touches a confounder nobody measured.

Also covered: stratification and regression · propensity scores · IPTW · TMLE and target trial emulation

Alongside this step

Slides and transcript — the whole video
Read the chapter — Observational Designs

Lesson 3 · 6 steps

1What descriptive designs can and cannot tell us7:16 2Cross-sectional studies and the problem of timing4:51 3Case-control studies: starting from the outcome6:21 4Cohort studies: following people forward in time6:39 5Controlling confounding in observational data7:49 6Choosing and reading an observational design6:45

← Previous step

Research Designs · lesson 3 of 5