Propensity scores: help or hype?

In this issue of Nephrology Dialysis Transplantation, Kazmi et al. report an evaluation of the association between late nephrologist referral and mortality in a cohort of incident renal replacement therapy (RRT) patients [1]. After multivariable adjustment, they found that patients who reported having first been seen by a nephrologist <4 months prior to RRT had a nearly 50% higher risk of 1 year mortality compared to those patients who had their first nephrologist referral earlier in relation to their first RRT [hazards ratio (HR) 1.44; 95% confidence interval (CI): 1.15–1.80]. In addition to standard multivariable regression adjustment, the authors used propensity score (PS) analysis to control for confounding and argued that this approach was a more robust method to balance covariates, and that it helped in their study to overcome confounding and selection bias compared with the traditional approach. However, after adjusting for quintiles of PS, their findings were virtually unchanged (HR1⁄4 1.42; 95% CI: 1.12–1.80). In recent years, PS analyses have become a fashionable tool and its use is increasing particularly in pharmacoepidemiological studies [2]. It seems that lately, some journals and reviewers are in favour of this approach in observational outcomes research. However, it appears that there is much uncertainty among researchers regarding what PS can or cannot accomplish, or in which cases this technique is of no use. It is the purpose of this editorial to shed some light on these issues. What is the propensity score?