To guide their complex work, healthcare policymakers and healthcare system leaders and managers all depend on well-established, convincing evidence of plausibly causal relationships between inputs and outputs, actions and consequences. There is not much in public health that can hope to meet such a high standard. In particular, establishing causality in non-randomized, real-world observational studies is an intractable problem due to unmeasured confounding. Little mitigation exists beyond hoping that controlling for observable variables (e.g. propensity score) also controls for unmeasured confounders or hoping for the existence of plausibly exogenous variables (i.e. instruments) not correlated with those confounders.
I will illustrate this problem and such aspirational solutions in the setting of long-term coronary stent outcomes. Drug-eluting coronary stents (DES) were shown to be more effective compared with the older bare metal stents (BMS) in initial RCTs and subsequently in large observational studies with. Yet if patients who receive BMS are sicker in ways not captured by observed confounders, then propensity score approaches may fail. The resulting selection bias may distort the comparative effectiveness of DES. For example, physicians may have assessed the patient's ability to afford or to comply with subsequent anti-platelet therapy in ways that were unobserved by the analyst. BMS may then be used in patients independently more prone to adverse outcomes.
I use propensity score techniques as well as an instrument variable approach relying on physician inertia in prescribing and choosing treatments, but also present a little known but simple external adjustment technique presented by Cornfield (1954) in the context of establishing the causal effect of smoking on lung cancer and refined by Greenland (1996), which can provide additional hope that observed associations are, in fact, causal.
der1 [at] columbia [dot] edu