IPWerte

IPWerte, short for inverse probability weights, are statistical weights used to adjust for confounding and informative sampling in observational data. They are based on the estimated probability of receiving the observed treatment given observed covariates and aim to recreate a pseudo-population in which treatment assignment is independent of these covariates.

Calculation typically begins with estimating propensity scores, p(X) = P(T=1 | X), where T denotes treatment status and X represents covariates. Weights are then assigned as follows: for treated units, weight is 1/p(X); for control units, weight is 1/(1-p(X)). To reduce variance, stabilized weights are often used, multiplying by the marginal treatment probability P(T=1) for treated and P(T=0) for controls. Weights can be normalized or trimmed to avoid extreme values that inflate variance.

IPWerte are commonly employed to estimate average treatment effects in causal inference, particularly within marginal structural models. They enable unbiased estimation under the assumption of no unmeasured confounding and correct model specification for the propensity scores, given sufficient overlap between treatment groups.

Practical considerations include assessing the distribution of weights, diagnosing balance after weighting, and checking overlap or common support. Extreme weights can distort results, making trimming or alternatives like doubly robust methods advisable.

Historically, inverse probability weighting gained prominence in causal inference through the work of Robins, Hernán, Brumback, and colleagues in the 1990s. It remains a core tool in epidemiology, economics, and social sciences for addressing confounding in observational studies. See also propensity scores, marginal structural models, TMLE, and causal inference.