The Propensity Score is the estimated probability that an individual receives a specific treatment/exposure, given observed covariates (e.g., age, comorbidities, severity). It is used in observational studies to reduce confounding and to make groups more comparable when randomization is not present.
Core Idea
In non-randomized data, treatment groups often differ systematically. The Propensity Score condenses many covariates into a single metric. When individuals with similar Propensity Scores are compared, the groups should be “balanced” with respect to the covariates considered – similar to randomization (but only concerning observed variables).
How is the Propensity Score estimated?
Typically with a model that has treatment as the target variable, e.g.:
- logistic regression,
- Gradient Boosting / Random Forests,
- other machine learning models (with a focus on balance, not just predictive accuracy).
Careful selection of covariates is important: Variables that can influence both treatment and outcome (potential confounders) are included. Variables that occur after treatment initiation are generally not part of the PS model.
Typical Applications
- Matching (e.g., 1:1 matching): Comparison of individuals with similar scores.
- Stratification: Division into quantiles (e.g., quintiles) and comparison within strata.
- Weighting (IPTW): Weighting by the inverse of the probability of treatment to create a “pseudo-population.”
- Adjustment: Including the Propensity Score as a covariate in an outcome model.
What to consider?
- Balance diagnostics: After matching/weighting, check whether covariates are balanced between groups (e.g., standardized mean differences).
- Positivity/Overlap: There must be sufficient overlap of scores between groups; otherwise, effects are difficult to identify.
- Unobserved confounding: Propensity Scores only address observed confounders. Unmeasured confounders remain a risk.
- Model choice: A very flexible model can improve balance but may also carry overfitting risks; balance is crucial, not AUC.
FAQ
Does the Propensity Score replace randomization?
No. It is a tool for confounding control in observational data but cannot eliminate unobserved confounders.
When is IPTW useful?
When there is good overlap and one wants to use the entire sample. Weighting can then be efficient but requires stability checks (extreme weights).
Why is “adjustment in the regression model” sometimes not enough?
Direct adjustment can work, but the Propensity Score approach separates the design phase (establishing balance) and the analysis phase (estimating effect) and often makes assumptions more transparent.