Design Suggestion

In an observational comparative effectiveness study, assignments to treatment and control groups are determined by the choices of patients and providers, choices that are potentially affected by numerous covariates (i.e., characteristics of the patient, provider, and therapeutic context as of the time the choice is made). Importantly, any covariate that is associated with both the choice of treatment or control and a study outcome is potentially confounding and will bias estimation of the effect of treatment relative to control on that outcome unless an appropriate method is used to disentangle the effect of the covariate from the effect of the treatment.

The approach most often used in an attempt to disentangle the effects of covariates from the effect of treatment is simple regression adjustment, with the observed, potentially confounding covariates included as predictors in the regression model. Unfortunately, this approach assumes that the effects of covariates can be represented equally well by the same fitted model in both treatment and control groups, an assumption unlikely to hold unless the distribution of covariates is quite similar in each group. For regression adjustment to perform reliably, the distributions of observed covariates must first be balanced between the groups.

In a study with a binary treatment choice, the propensity score is the probability of choosing the treatment rather than the control, conditioned on the values of observed covariates, and can be estimated for each patient via logistic regression. Propensity score matching, i.e., constructing treatment and control groups by matching patients on estimated propensity scores, is a very effective means of balancing the distributions of observed covariates between the two groups. Furthermore, if after matching, all observed, potentially confounding covariates are well-balanced between treatment and control groups, regression adjustment is not only more reliable, but also usually no longer necessary to reduce confounding. Thus, with propensity score matching, the treatment effect can generally be estimated using the same methods that would have been used had group assignments been made randomly.