This JAMA Guide to Statistics and Methods discusses the use of propensity scores as a way to reduce bias in estimates of treatment effect when randomized trials are not feasible.
Many observational studies analyze data to estimate the effect of a treatment on patient outcomes. For example, in a study by Rozé et al,1 a large observational data set was analyzed to estimate the relationship between early echocardiographic screening for patent ductus arteriosus and mortality among preterm infants. The authors compared mortality rates of 847 infants who were screened for patent ductus arteriosus and 666 who were not. The 2 infant groups were dissimilar; infants who were screened were younger, more likely female, and less likely to have received corticosteroids. The authors used propensity score matching to create 605 matched infant pairs from the original cohort to adjust for these differences. In another study by Huybrechts et al,2 the Medicaid Analytic eXtract data set was analyzed to estimate the association between antidepressant use during pregnancy and persistent pulmonary hypertension of the newborn. The authors included 3789330 women, of which 128950 had used antidepressants. Women who used antidepressants were different from those who had not, with differences in age, race/ethnicity, chronic illnesses, obesity, tobacco use, and health care use. The authors adjusted for these differences using, in part, the technique of propensity score stratification.
Why Were Propensity Methods Used?
Many considerations influence the selection of one therapy over another. In many settings, more than one therapeutic approach is commonly used. In routine clinical practice, patients receiving one treatment will tend to be different from those receiving another, eg, if one treatment is thought to be better tolerated by elderly patients or more effective for patients who are more seriously ill. This results in a correlation—or confounding—between patient characteristics that affect outcomes and the choice of therapy (often called “confounding by indication”). If observational data obtained from routine clinical practice are examined to compare the outcomes of patients treated with different therapies, the observed difference will be the result of both differing patient characteristics and treatment choice, making it difficult to delineate the true effect of one treatment vs another.
The effect of an intervention is best assessed by randomizing treatment assignments so that, on average, the patients are similar in the 2 treatment groups. This allows a direct assessment of the effect of the intervention on outcome. In observational studies, randomization is not possible, so investigators must adjust for differences between groups to obtain valid estimates of the associations between the treatments being compared and the outcomes of interest.3 Multivariable statistical methods are often used to estimate this association while adjusting for confounding.
Propensity score methods are used to reduce the bias in estimating treatment effects and allow investigators to reduce the likelihood ...