Comment on "the diffusion of a medical innovation: is success in the stars?".

Huesch, Marco D.

JEL Classification: UO, UI, O33

1. Introduction

Burke, Fournier, and Prasad (2007), henceforth BFP, recently presented in these pages an important and path-breaking exploration of social learning models, noting potential welfare consequences of heterogeneous diffusions of innovations in medicine.

Unfortunately, while I am broadly sympathetic to the central hypothesis of social learning in medical practice, there remains a series of questions regarding the conceptual model of BFP. (1) Establishing the claim that social influences may significantly affect the timing and rate of technological adoption among physicians would be easier if these key issues were addressed differently. Most prominent of these issues is the possibility of omitting key variables that might independently explain diffusion of medical innovations.

In this light, it is instructive to consider the backstory to the classic study of Coleman, Katz, and Menzel (1966), which investigated the dynamics of physician adoption of tetracycline antibiotics. As BFP point out, Coleman, Katz, and Menzel found that peer effects mediated by influential physicians drove adoption. Unfortunately, when Van den Bulte and Lilien (2001) reanalyzed Coleman's data and controlled for pharmaceutical company marketing mix variables, the conclusion was that peer effects were no longer significant drivers of adoption.

Notwithstanding these comments, I conjecture that the heterogeneous adoption observed in practice is' due to serious structural issues (principally the unequal distribution of higher volume teaching and for-profit hospitals, rather than unequal distributions of particular star physicians). BFP's central conclusion on the resultant unequal adoption/utilization entailing welfare costs is thus likely to remain an important and urgent one, albeit one even less amenable to policy changes.

2. A Series of Model Questions

In BFP's conceptual model, star physicians (defined as having completed residency at top medical schools) systematically influence the adoption hazard and utilization rate (conditional on adoption) of nonstar physicians. But several difficulties underlie this prima facie natural hypothesis of social influence. These may be summarized as (i) the proposed mechanism of influence; (ii) the absence of hospital time-based effects; (iii) the discarding of the distinction between attending and operating physician; and (iv) the inclusion of nonadopting stars in the model.

First, it is unfortunately not made clear, nor is it intuitive, how a star's residency school prestige might actually drive opinion leadership and social influence. Moreover, no other potential locus of social influence, apart from residency school (on average, a dozen years antecedent in the study's sample, and up to 25 years distant), seems to have been considered by BFP. (2) Incorporating age (closely correlated with tenure and hierarchical status within a hospital, and inferable from graduation year, though unlikely to survive fixed effects), cumulative angioplasty volume (inferable, though top censored, from the panel), and periodic performance, such as mortality and morbidity rates (available from the discharge status data), as well as their interactions with residency status, ought to strengthen their social influence model. (3)

In a related vein, one wonders how much or how little residency status is correlated with other possible mediators of social influence. If star status were correlated with, say, a position as chief of cardiology in a large hospital, then might that position and place not be the actual mechanism of influence? If not correlated, what would such positional authority add in terms of model predictive power?

The second area in which the model ought to be strengthened is by incorporating the time-varying influence of the overarching institution embedding the physicians--the hospital. This is restricted by the authors to a time-invariant fixed effect. In Burke, Fournier, and Prasad (2007), it is pointed out that "if hospitals implement updated practice guidelines for stent usage at irregular intervals, fixed effects for hospitals are not sufficient" (Burke, Fournier, and Prasad 2007, p. 601). Interestingly, they do include quarterly time dummies, "[which] may proxy for the influence of the lagged statewide adoption rate" (Burke, Fournier, and Prasad 2007 p. 594). An additional highly plausible time-based effect is cumulative collocated total hospital experience with stents and angioplasties in general. As the whole operating room team and the hospital's various departments become more comfortable with stent use, local spillover dynamics becomes a very feasible mechanism driving increased adoption and utilization. Analogously, in Huckman and Pisano (2006) it is shown that the mortality benefits of experience gained by a physician through current volume in one hospital are not portable to her work elsewhere: The institution matters.

Of course, as Ho (2002) pointed out in a related analysis of stent diffusion, the autocollinearity occasioned by the positive serial correlation in some of these data series can be serious. For example, time series of hospital stent caseloads (period or cumulative) will necessarily display strong positive serial correlation. Nonetheless, even point estimates and upper bounds on confidence intervals may still serve the authors well.

A third and closely related conceptual issue is the distinction between the two types and roles of the physician delivering treatment: the "operating physician" and the "attending physician." The latter has complete legal and clinical responsibility for the patient's care during the admission, while the former performs the intervention in the catheterization lab. The latter is moreover obliged to marshal the hospital's entire resources in support of the patient's care, and thus would appear to be sensitive to the hospital's experience and preferences in regard to stent usage. BFP apparently used just the operating physician as the unit of analysis; although, approximately half of all stent admissions have different physicians in the two roles, and the relative social influences of these two roles are thus important. Since the dataset will contain both data fields, it would be instructive to understand the magnitude and direction of the attending physician's social influence, especially given her overarching legal responsibility for her admitted patient.

A final conceptual problem is that BFP acknowledge that star physicians neither adopt earlier nor utilize the stent innovation more than nonstars. (4) Indeed BFP hypothesize simply that "...innovation will spread faster among nonstars in places where influential physicians [stars] operate" (Burke, Fournier, and Prasad 2007, p. 597). On the one hand, if this implies that stars who don't adopt stents somehow contemporaneously influence nonstars to adopt, then a deeper problem of logic seems to exist in this model. (5)

On the other hand, if such nonadopting stars do not have such an unlikely type of influence, why are they part of the right-hand side variable [[S.sub.rt]] in the model's equations 1 and 2? Shouldn't such relatively dull stars have been isolated and removed from their brighter star peers in the adoption model (as they seem to have been from the utilization model)? The presence of these nonadopting stars in [S.sub.rt] risks biasing the coefficient on the focal variable--the impact of the number of stars encountered upward.

3. Conclusions

BFP deserve recognition and praise for venturing into an important and underresearched area. The conceptual questions raised here do, however, impinge on the overall persuasiveness of their study and may render some of their more marginal results more likely to be due to chance alone. Further investigations are necessary to buttress such important social influence results. Even more unfortunately (if it turns out to be the unequal distribution of large hospitals that drives heterogeneity in the diffusion of such innovations, as opposed to unequal distributions of stars), it is again likely that ensuing aggregate social welfare costs are highly resistant to easy policy fixes.

References

Burke, Mary A., Gary M. Fournier, and Kislaya Prasad. 2007. The diffusion of a medical innovation: Is success in the stars? Southern Economic Journal 73:588-603.

Coleman, James S., Elihu Katz, and Herbert Menzel. 1966. Medical innovation: A diffusion study. Indianapolis, IN: Bobbs-Merrill.

Ho, Vivian. 2002. Learning and the evolution of medical technologies: The diffusion of coronary angioplasty. Journal of Health Economics 21:873-85.

Huckman, Rob, and Gary Pisano. 2006. The firm-specificity of individual performance: Evidence from cardiac surgery. Management Science 52:473-88.

Van den Bulte, Cristophe, and Gary L. Lilien. 2001. Medical innovation revisited: Social contagion versus marketing effort. The American Journal of Sociology 106:1409-35.

(1) My experience as a practicing hospital physician leads me to be biased toward the acceptance of the importance of social learning and peer influence in hospital delivery of medical and surgical services.

(2) Consider our own academic sources of influence: number of papers, endowment budgets, and teaching ratings position. Anecdotally, the earned doctorate often figures more in negative analyses of the form "despite coming from a top school, she is not publishing enough."

(3) In a hospital's monthly Mortality and Morbidity meetings, caseload, outcomes, and summaries of performance are available for all gathered cardiologists. Similar interactions between status and realized performance would be observed in cardiologist practice groups because these physicians prefer to practice in groups.

(4) An open question, tangentially related: The stars' impact on the profound adoption decision is only marginally significant, while the impact on utilization is considerably less likely to be due to chance alone. But changes in utilization could be much more a continuous function of individual skill and cumulative experience (i.e., own learning), and arguably less susceptible to peer influence.

(5) Because "Do as I say, not as I do" seems a particularly unlikely summary of a mechanism of opinion leadership.

I thank the referee and the coeditor for valuable comments and suggestions, l am also grateful to Mary Burke for helpful answers to earlier questions. Usual caveats apply.

Marco D. Huesch, The Anderson School of Management at UCLA Anderson, D513, 110 Westwood Plaza, Los Angeles, CA 90095, USA; E-mail mhuesch@anderson.ucla.edu.

Received August 2007; accepted April 2008.