Unmeasured confounding almost always exists in observational studies—and it can bias our analysis of the effects we see in patients. Instrumental variable methods are popular choices to combat this phenomenon. The treatment preference of clusters (e.g., physician practices) are the most frequently used instrumental variables in comparative effective studies. In this study, however, the authors demonstrate that, depending on the nature of unmeasured confounding, alternative methods may yield less biased estimates than preference-based instrumental variable analyses.

For example, the authors looked at the administration of erythropoietin-stimulating agents to raise essential hemoglobin levels among patients receiving dialysis for end-stage kidney disease. They aimed to quantify the bias of treatment-effect estimates they obtained from several statistical methods—including instrumental variable analysis, ordinary least squares regression, fixed effect regression, and linear mixed models—when unmeasured confounding occurs. They examined unmeasured confounding both within clusters (for example, individual patient response to medication) and between clusters (for example, care quality of various dialysis facilities).  

They derived bias formulae, conducted simulations and illustrated their findings through data analyses. Bias formulae showed that preference-based instrumental variable analyses could provide unbiased estimates when unmeasured within-cluster confounding exists, but not when between-cluster confounding exists. On the other hand, they found that fixed effect models and linear mixed models could provide unbiased estimates when there is unmeasured between-cluster confounding, but not when there is within-cluster confounding. So do instrumental variable analyses better reduce bias than do fixed effect models and linear mixed models? That depends on the extent of unmeasured within-cluster confounding relative to between-cluster confounding, the authors say.

These findings can point researchers in various situations toward the best statistical methods to combat dominant types of unmeasured confounders, and can facilitate the interpretation of statistical analysis. The results stand to improve the robustness and validity of findings from observational studies, providing strong evidence to inform clinical care and to improve patient outcomes.