This program is for the research community.
In this seminar I use the setting of coronary stents to illustrate the use of instrumental variables and sensitivity analyses for hidden biases, methods first pioneered by epidemiologists but not in widespread use. Large observational studies (e.g. Douglas et al., 2009) find that drug-eluting stents (DES) were associated with significantly better clinical outcomes than bare metal stents (BMS); however, unmeasured patient selection biases cannot be ruled out. I study all 38,000 patients receiving either DES or BMS as an index procedure in Pennsylvania between 2004-2005. DES use was associated with a significantly lower hazard of mortality through 3 years in stratified and shared frailty Cox models. Physician preference was used as an instrument for the stent type assignment, operationalized as the preceding quarter’s rate of DES use for each patient’s cardiologist in a two stage least squares model. Instrumented risk differences were consistent with the hazard rates. Greenland’s method (1996) showed that implausibly large odds ratios of 3 or more, between a hypothetical unmeasured confounder and both the outcome as well as the use of a particular stent type, would be required to fully explain the observed association between DES use and long-term mortality. Unusually rich data on observed pre-treatment differences in this setting appeared sufficient to allow a causal interpretation. The approaches discussed here can help to inform the conclusions of comparative effectiveness research using observational studies.