What is the endogeneity problem that the authors are facing?The endogeneity problem that McClellan et al. (1994) are facing is that based on Medicare claims data, it is difficult to discern the main cause driving longevity after AMI. More intense treatment? Or better physical health and the choice of a better doctor in general? To correct for this selection bias, the authors are using the distance of the patient from the primary care facility as instrumental variable. They found that the further away the patient lives from the hospital, the more intense the treatment he or she receives. Thus, this distance – which is accessible from medical records – instruments the treatment intensity. The overall health of the patient should be independent from his distance to a hospital.

You're lucky! Use promo "samples20"
and get a custom paper on
"Public Health Economics"
with 20% discount!
Order Now

Is the instrument valid?
According to Pearl (2000), an Instrumental Variable z has to satisfy several conditions. If there is an independent variable x – here, the intensity of treatment – and a dependent variable y – in this case, survival after AMI – which are connected via y = fp · x + fe. The entity fp is a proportionality factor and fe an error factor containing the sum of all confounding influences (e.g. general patient health) that affect y when x stays constant; z must be independent of fe, z is not independent of x, and if x is held constant, z will not affect y. The authors do comment that the “distribution of unobserved health characteristics” is independent from the physical distance of patients from the hospital, thus suggesting that z is indeed independent from fe. Furthermore, the closer to the hospital the patient lives, the more likely it is he received instant care and catheterization, suggesting that z is not independent of x. Thus, ths instrument seems to be valid.

Which predictors are most sensitive to the inclusion of the instrument?
Table 5 (McClellan et al., 1994) compares ANOVAs under adjustment for observable patient characteristics, adjustment for characteristics and hospital types and inclusion of the IV “distance”. In general, adjusting for both characteristics and hospital type together makes the ANOVA relatively similar to the IV inclusion. This is not surprising, since large urban hospitals are more likely to administer intense treatments. The difference between ANOVA with adjustment for observable characteristics alone and IV – vs. adjustment for both hospital type and characteristics and IV – is most pronounced when looking at the probability of catheterization and rural residence, and negligible when looking at female or black race effects. This could suggest that females and blacks tend to live closer to large hospitals in city areas, such that distance from hospital is already included. It also means that the IV “Distance” has its biggest impact when looking at elderly male patients.

Are Medicaid and commercial patient claims different from Medicare claims?
Medicare is a Federal single-payer healthcare insurance, while Medicaid offers assistance. Similar to commercial insurers, there is no standard of coverage. This might introduce bias – in contrast to Medicare, Medicaid and private insurers sometimes provide only minimum spending. As a result, less coverage negatively affects survival rates for AMI patients.

Who are the compliers in this study?
According to Imbens and Rubin (1997), if Di(0) and Di(1) are the values of the treatment for individual i that would be obtained given the instrument Zi = 0 and Zi = 1, then i is a complier when Di(0) = 0 and Di(1) = 1. For the study here, this means that compliers are all those individuals that live close to the hospital and survive better. If one uses catheterization within 90 days as a proxy for close distance to the hospital, then there is a slight correlation between lower mortality likelihood and probability of catheterization (see table 6, graphic in McClellan et al., 1994). There is not a clear binary distinction between compliers and treatment, but more gradual transition between poor and intense treatment vs. survival.

  • Imbens, G.W., Rubin, D.B. (1997). Estimating outcome distributions for compliers in
    instrumental variables models. Rev. Econ. Stud. 64: 555-574.
  • McClellan, M., McNeil, B.J., Newhouse, J.P. (1994). Does more intensive treatment of acute myocardial infarction in the elderly reduce mortality? J. Amer. Med. Assoc. 272(11): 859-866.
  • Pearl, J. (2000). Causality: Models, reasoning, and inference. New York: Cambridge University Press.