MerQuR+Statistics Glossary

Statistics Glossary → Inverse probability weighting (IPW)

Inverse probability weighting (IPW)concept

A method for observational studies that weights each individual by the inverse of their probability of receiving the treatment they actually received (the propensity score), so that treatment groups are balanced on the observed confounders.

The propensity score is the probability of being in the treatment group given an individual's observed characteristics, usually estimated with logistic regression. In IPW, treated individuals receive weight 1/e and untreated individuals 1/(1 - e), where e is the propensity score. Individuals who were unlikely to be treated but were treated therefore gain more weight, and the weighted sample represents a pseudo-population in which treatment is assigned independently of the confounders. The mean difference computed with these weights estimates the average treatment effect (ATE).

The method relies on all confounders being measured (no unmeasured confounding), on a correctly specified propensity model, and on positivity (every individual having a non-zero probability of being in either group). When propensity scores approach 0 or 1, weights become extreme and the estimate unstable; stabilized weights (multiplying the weights by the marginal treatment proportion) or trimming the weights ease this problem. The standard error should account for the fact that the propensity score is itself estimated; robust (HC0) errors that treat the weights as fixed are often conservative. Balance should be checked after weighting with standardized differences.

An everyday example

A hospital examines the effect of a new rehabilitation programme, but younger and healthier patients were more likely to join it. The few 80-year-olds who did join get a high weight so that they 'stand in' for the many similar patients who did not. After weighting, the two groups have similar distributions of age and health, and the comparison becomes fairer.

In MerQur

It is used in Propensity Score Analysis with the 'Stabilized weights' option and the 'IPW standard error (M-estimation / Robust (HC0))' parameter.

Analyses where this concept appears

Propensity Score Analysis

Related concepts

Propensity score matching (nearest neighbour, caliper)Doubly robust estimation