[Propensity score: interest and limits]

Abstract

Propensity score, an indicator of the propensity to get one treatment among two (or more), is encountered in non randomized studies (prospective or retrospective). It is calculated after the research of predictive factors for treatment attribution, and corresponds to the probability to receive one of the treatments conditional to variables observed before treatment. This probability is usually generated thanks to a logistic regression equation. This score sums up by itself a whole set of parameters. It can be used as cofactor in other multivariate models that aim to evaluate with a reduced risk of confusion, the impact of therapeutical modalities on such end-points as survival, morbidity, secondary effects or quality of life. It appears very convenient to realize matching or stratification in order to compare these end-points among resulting subgroups. Despite this advantage that enables to obtain a posteriori similar subgroups, this method cannot pretend to reach the level of evidence of randomized trials, because absence of bias is never guaranteed. Apart from this major methodological weakness, propensity score appears less useful in studies provided with a large population, since in such cases, multivariate models can include enough covariates to produce in a secure way stable conclusions. When samples are small, this score remains interesting although its reliability, once more, depends on sample size and conclusions need nuances. Examples are included to illustrate the topic.