Observational Causal Inference
In observational studies, treatment is not assigned by the researcher. Causal identification therefore requires assumptions about how treatment relates to potential outcomes.
See
-
Unconfoundedness and Overlap
- The identification assumptions behind adjustment on observed covariates.
-
Propensity Scores
- The conditional probability of treatment and its balancing role.
-
Inverse Probability Weighting
- Reweights observed outcomes to estimate an average treatment effect.
-
Matching for Causal Inference
- Constructs comparisons between treated and control units with similar covariates or propensity scores.
Identification chain
define the causal estimand
-> measure pretreatment confounders
-> assume unconfoundedness and overlap
-> estimate an adjustment object or outcome model
-> check balance and common support
-> estimate the effect
-> conduct sensitivity and uncertainty analysis
Central limitation
Propensity scores, weighting, and matching can balance observed covariates. They cannot by themselves remove bias from unmeasured confounders. The credibility of the result therefore depends on subject-matter knowledge, measurement, design choices, diagnostics, and sensitivity analysis.
Worked problems and practice
Observational Causal Inference Exam Workshop develops standardization, inverse weighting, matching, overlap diagnostics, and covariate selection through realistic multi-part cases.
Experimental benchmark
In Randomized Assignment, the assignment mechanism is known and independent of potential outcomes by design. Observational methods try to recover comparable groups conditionally, but the required independence is an assumption rather than a consequence of researcher-controlled randomization.