[Inverse probability weighting (IPW) for evaluating and "correcting" selection bias].

INTRODUCTION: the Inverse probability weighting (IPW) is a methodology developed to account for missingness and selection bias caused by non-randomselection of observations, or non-random lack of some information in a subgroup of the population.
OBJECTIVES: to provide an overview of IPW methodology and an application in a cohort study of the association between exposure to traffic air pollution (nitrogen dioxide, NO₂) and 7-year children IQ.
METHOD: this methodology allows to correct the analysis by weighting the observations with the probability of being selected. The IPW is based on the assumption that individual information that can predict the probability of inclusion (non-missingness) are available for the entire study population, so that, after taking account of them, we can make inferences about the entire target population starting from the nonmissing observations alone.The procedure for the calculation is the following: firstly, we consider the entire population at study and calculate the probability of non-missing information using a logistic regression model, where the response is the nonmissingness and the covariates are its possible predictors.The weight of each subject is given by the inverse of the predicted probability. Then the analysis is performed only on the non-missing observations using a weighted model.
CONCLUSIONS: IPW is a technique that allows to embed the selection process in the analysis of the estimates, but its effectiveness in "correcting" the selection bias depends on the availability of enough information, for the entire population, to predict the non-missingness probability. In the example proposed, the IPW application showed that the effect of exposure to NO2 on the area of verbal intelligence quotient of children is stronger than the effect showed from the analysis performed without regard to the selection processes.