Causal inference

Causal inference is the process of estimating whether changes in one factor produce changes in another, using assumptions and evidence to distinguish cause from association.

Potential outcomes framework: A framework that defines causal effects by comparing outcomes under different possible treatments for the same units. It formalizes effects as contrasts between outcomes under treatment and control.
Counterfactual: A hypothetical outcome describing what would have happened under a condition that did not occur. Causal effects compare observed outcomes with the unobserved alternatives for the same units.
Correlation: A statistical measure of how two variables vary together. Correlation alone does not establish that changing one variable changes the other.
Difference-in-differences: A quasi-experimental method that compares outcome changes over time between treated and comparison groups. It estimates effects when treatment timing differs and parallel trends is plausible.
Sensitivity analysis: Methods that assess how conclusions change under variations in assumptions or unmeasured biases. It quantifies how strongly hidden confounding would need to act to alter an estimate.
Directed acyclic graph: A directed graph with no directed cycles, often used to represent assumed causal relationships. Its arrows encode causal assumptions that determine which variables to adjust for.
Confounding: Bias in an estimated relationship caused by a variable associated with both the exposure and outcome. It creates noncausal associations that can be mistaken for treatment effects.
Prediction: The task of estimating unknown or future outcomes from available information. A variable can predict an outcome without being a cause of it.
Propensity score: The probability of receiving a treatment given observed covariates. Matching or weighting by this score can balance measured confounders between treatment groups.
External validity: The extent to which findings from a study apply to other populations, settings, or times. A well-identified effect in one sample may not transfer to a different target population.

Linked from 132 pages