In most observational studies, treatments or other “exposures” (in an epidemiologic sense) do not occur at random. Instead, treatments or other such interventions depend on several patient-related and patient-independent characteristics. Such factors, associated with the receipt vs nonreceipt of treatment, may also be—independently—associated with outcomes. Thus, confounding exists making it difficult to ascertain the true association between treatments and outcomes. Propensity scores (PS) represent an intuitive set of approaches to reduce the influence of such “confounding” factors. PS is a computed probability of treatment, a value that is estimated for each patient in an observational study and then applied (in a variety of ways such as matching, stratification, weighting, etc.) to reduce distortion in the true nature of the association between treatment (or any similar exposure) and outcomes. Despite several advantages, PS-based methods cannot account for unmeasured confounding, ie, for factors that are not being included in the computation of PS.
Observational research using propensity scores