Inverse Probability Weighting (IPW)

Inverse probability weighting estimates causal treatment effects from observational data by weighting each subject by the inverse of their probability of receiving the treatment they actually received.

역확률가중법은 각 대상자를 실제로 받은 처치를 받을 확률의 역수로 가중하여 관찰자료에서 처치의 인과효과를 추정하는 방법입니다.

How do I adjust for confounding when estimating a treatment effect from observational data?

관찰자료로 처치효과를 추정할 때 교란요인은 어떻게 보정하나요?

Inverse probability weighting uses propensity scores to weight subjects so that the treated and untreated groups resemble each other on measured covariates, creating a pseudo-population in which treatment is independent of those covariates. It fits observational studies with a binary treatment and many confounders when all important confounders are measured and propensity scores overlap between groups.

When to use Inverse Probability Weighting (IPW)

• Observational data comparing treated and untreated groups on an outcome
• Many measured confounders and a wish to keep all subjects rather than matching
• Estimating the average treatment effect for the whole population or the treated

When to use something else

• Important confounders are unmeasured but a valid instrument exists: use instrumental variable estimation
• Protection against misspecifying one model is wanted: use a doubly robust estimator such as augmented IPW

Assumptions and requirements

• No unmeasured confounding
• Positivity: every subject has a non-zero probability of each treatment
• Correctly specified propensity score model
• Extreme weights checked and stabilized or trimmed

What the analysis reports

• Propensity score distributions by group
• Weight summary statistics
• Covariate balance before and after weighting (standardized mean differences)
• Average treatment effect with robust or bootstrap confidence interval

How to report it

Using stabilized inverse probability weights (all weighted standardized mean differences < 0.1), the estimated average treatment effect of [treatment] on [outcome] was [value], 95% CI [[value], [value]].

Related methods

• Propensity Score Matching
• Augmented Inverse Probability Weighting (AIPW)
• Instrumental Variable Estimation (IV)

Run Inverse Probability Weighting (IPW) on Quorant