`Vol. 29, No. I, Spring 1998
`pp. 108-136
`
`The importance of the physician in the
`generic versus trade-name prescription
`decision
`
`Judith K. HeUerstein*
`
`I examine the importance of physicians in the process by which patients receive either
`trade-name or generic drugs. Using a dataset on physicians, their patients, and the
`multisource drugs prescribed, I find that almost all physicians prescribe both types of
`drugs to their patients, but some physicians are more likely to prescribe generic drugs
`while other physicians are more likely to prescribe trade-name drugs. Very little of the
`prescription decision can be explained by observable characteristics of individual pa(cid:173)
`tients, but all of the evidence indicates that physicians are indeed an important agent
`in determining whether patients receive either trade-name or generic drugs.
`
`1. Introduction
`•
`In 1989, over 70% of pharmaceutical prescriptions were written for multisource
`drugs, that is, drugs for which both generic and trade-name versions are available. Yet
`of these multisource prescriptions, fewer than 30% specified the generic version of the
`drug. Since generics are generally priced 30-60% lower than their trade-name coun(cid:173)
`terparts (Grabowski and Vernon, 1992), substantial cost savings could be realized in
`this $40-billion-per-year market if generics captured greater market share. Possible
`explanations for the paucity of generic prescriptions include the existence of infor(cid:173)
`mation imperfections that limit the physician's knowledge, and agency problems arising
`from the physician acting as agent for the patient and for the patient's insurance com(cid:173)
`pany.
`In this article I examine whether the seemingly small market share of generics can
`be attributed at least partially to the behavior of physicians. Using data from a survey
`of physicians, their patients, and the drugs prescribed, I examine whether physicians
`vary their prescription decisions on a patient-by-patient basis or whether they system(cid:173)
`atically prescribe the same versions (trade name or generic) to all patients. I test whether
`
`-----------------------~-
`* University of Maryland and NBER; hellerst@econ.umd.edu
`This is a substantially revised version of parts of my Ph.D. dissertation at Harvard University. An
`earlier version circulated under the title "The Demand for Post-Patent Prescription Pharmaceuticals." I am
`grateful for the comments provided by Eli Berman, Ernie Berndt, Tim Bresnahan, Gary Chamberlain, lain
`Cockburn, David Cutler, Sara Ellison, Ed Glaeser, Shane Greenstein, Zvi Griliches, Hank Farber, Guido
`Imbens, Larry Katz, Bruce Meyer, Ariel Pakes, Gary Solon, Phillip Swagel, the referees, and participants of
`seminars at the NBER and numerous universities. In addition, I thank Sandra Decker and the National Centers
`for Health Statistics for providing data. Financial support was generously provided by the Sloan Foundation.
`
`108
`
`Copyright © 1998, RAND
`
`SENJU EXHIBIT 2200
`INNOPHARMA v SENJU
`IPR2015-00903
`
`PAGE 1 OF 5
`
`
`
`130
`
`I THE RAND JOURNAL OF ECONOMICS
`
`from those for the full sample, indicating that measurement error in the mean-charac(cid:173)
`teristics variables is not causing large biases in the results.
`There are two potential reasons to be concerned about the inclusion of state laws
`in the estimation in Tables 4-6. As explained in Section 5, the inclusion of only linear
`terms for state substitution laws may not adequately capture their effect on the pre(cid:173)
`scription choice if they change the effective price differenti~l of the drug. Moreover,
`estimating a coefficient on an interaction between mandatory substitution laws and the
`drug dummies might arguably allow for separate identification of baseline differences
`across drugs from the proportion of the drug's cost to the patient that is internalized
`by physicians (the term ')'1 in the model). Rather than attempting to estimate and in(cid:173)
`terpret regression coefficients from a model allowing for full interactions between drug
`dummies, insurance dummies, and the two types of state prescription laws, I indirectly
`account for this possibility by simply estimating the model with only the subsample of
`4,334 observations from states with permissive substitution laws and one-line prescrip(cid:173)
`tion pads. The results, reported in Tables 7-9, are quite similar to those for the full
`sample reported in Tables 4-6. This means that there is little difference in the treatment
`of patients across different regimes of state substitution and prescription-pad laws (so
`that ')'1 is essentially zero). Given that much of the variance in the prescription decision
`is unexplained, it is not surprising that differences across states in laws that may be
`poorly understood by physicians and poorly adhered to by pharmacists have little or
`no direct effect on prescription behavior. This result is nonetheless consistent with the
`conclusion that physicians internalize little of any differential costs to different patients.
`
`7. Conclusion
`•
`This article examines the importance of physicians in the process by which patients
`get either trade-name or generic drugs. The central result is that the physician is an
`important agent in the prescription decision. This should be a key focus of future
`research, since the reasons for why some physicians are more likely than others to
`prescribe generic drugs is largely left unexplained by the empirical analysis presented
`here. Identifying the sources of heterogeneity in behavior across physicians is an im(cid:173)
`portant part of understanding how the market for prescription drugs operates and, more
`generally, how physicians behave when faced with different information and incentives.
`One avenue for future research should focus on differences across drugs in generic
`prescription rates. A formal treatment of information diffusion would be a useful start(cid:173)
`ing point for thinking about this issue. One possibility for examining diffusion empir(cid:173)
`ically is to gather data on the length of time generics and trade-name drugs have been
`marketed and to incorporate such information into the model of prescription choice.
`Another element in the examination of the diffusion of generics would be to combine
`data from the NAMCS surveys in other years. At the time this article was written, the
`NCHS would not release to me physician-identifying data for year~ other than 1989.29
`It would also be useful to consider other dimensions of differences across drugs, such
`as their use in treating chronic versus acute conditions, or life-threatening versus mild
`conditions.
`On the policy side, it is clear that there are potentially large social costs due to
`the habitual prescription of trade-name drugs. When physiciui1s make prescription de(cid:173)
`cisions based on incomplete information combined with agency problems, they do not
`make cost-effective decisions. Even state legislation that encourages generic substitu(cid:173)
`tion does not seem to have had an impact on physician prescription decisions. Changes
`in the structure of the health care system, however, may dramatically alter the market
`
`29 As of the 1991 NAMCS. the NCHS has included physician identifiers in the public-use data. but
`there is no information on the state in which the physician practices.
`
`PAGE 2 OF 5
`
`
`
`HELLERSTEIN
`
`I 131
`
`TABLE 7
`
`Estimated Coefficients on Demographic Variables, Geographic
`Variables, and Average Characteristics for Subsample from States
`with Permissive Substitution and One-Line Prescription Pads
`
`Variable
`
`Random-Effects Random-Effects
`Pro bit
`Pro bit
`t-Statistic
`Coefficient
`
`%Change in
`Generic
`
`Constant
`
`Age
`
`Female
`
`Nonwhite
`
`Hispanic
`
`Specialist
`
`Mean age
`
`Percent female
`
`Percent black
`
`Percent Hispanic
`
`Percent Medicaid
`
`Percent Medicare
`
`Percent private
`
`Percent HMO/prepaid
`
`Midwest
`
`South
`
`West
`
`p
`
`-.31
`
`-.01
`
`-.10
`
`.07
`
`-.02
`
`.00
`
`.00
`
`-.41
`
`-.13
`
`-.60
`
`.21
`
`-.10
`
`.08
`
`.20
`
`-.28
`
`-.30
`
`-.05
`
`.25
`
`-2.13
`
`-3.60
`
`-2.31
`
`.77
`
`-.17
`
`.02
`
`.87
`
`-2.00
`
`-.52
`
`-1.85
`
`.77
`
`-.31
`
`.50
`
`.92
`
`-2.08
`
`-.09
`
`-.36
`
`8.98
`
`-.15%
`
`-3.08%
`
`2.29%
`
`-.67%
`
`.05%
`
`.08%
`
`-12.n%
`
`-4.02%
`
`-19.77%
`
`6.59%
`
`-3.24%
`
`2.60%
`
`6.09%
`
`-9.00
`
`-9.56
`
`-1.66
`
`Note: The dependent variable is one if the generic is prescribed, zero otherwise. The
`sample size is 4,334. The mean and percent variables refer to the mean and percent char(cid:173)
`acteristics of the physician whom the patient visits. The omitted region category is Northeast.
`The omitted insurance categories are self-pay and percent self-pay. The percent changes in
`generic prescription are calculated as the average over the sample of the percent change in
`the probability of receiving a generic. For example, the percent change in the probability
`of generic prescription for age is the average percentage change for a marginal increase in
`age. For the dummy variables, the percent change in generic prescription represents the
`average percentage change that occurs when a person moves into the category represented
`by the dummy variable. The parameter p is the estimated variance of the random physician
`effect.
`
`for prescription drugs. Information from IMS America Inc., a market research firm,
`shows that managed-care payments (both private managed care such as HMOs and
`Medicaid HMOs) accounted for 58.5% of dollar revenues for pharmaceutical retail
`sales in 1996, up from less than 30% in 1990 (IMS America, 1996). Given the emphasis
`on cost containment in HMOs, the continued growth of managed care may increase
`the market share of generl.:;, drugs, or may cause the price differential between trade(cid:173)
`name and generic drugs to fall as HMOs negotiate with trade-name manufacturers for
`price discounts. Other information from IMS (IMS America, 1995) indicates that there
`is some evidence that changes are already occurring. As of 1995, pharmacists substi(cid:173)
`tuted generics in approximately half of all cases where physicians wrote a new pre(cid:173)
`scription for a trade-name drug for which a generic was available. This is up from less
`than 30% in 1989. Interestingly, however, while 44% of all new prescriptions (including
`
`PAGE 3 OF 5
`
`
`
`132
`
`I THE RAND JOURNAL OF ECONOMICS
`
`TABLE 8
`
`Estimated Coefficients for Drug-Class Dummy Variable for
`Subsample of States
`
`Drug Class
`
`Antimicrobials
`
`Cardiovascular/renals
`
`Central nervous system
`
`Hormones/hormonal
`mechanisms
`
`Skin/mucous membrane
`
`Ophthalrnics
`
`Respiratory tract
`
`Random-Effects
`Probit
`Coefficient
`
`Random-Effects
`/-Statistic
`
`%Change in
`Generic
`
`.97
`
`.21
`
`.76
`
`1.05
`
`-.77
`
`-.26
`-.66
`
`5.53
`
`.88
`
`3.85
`
`5.78
`
`-2.59
`
`-.74
`
`-.88
`
`30.42%
`
`5.02%
`
`22.54%
`
`33.61%
`-10.72%
`
`-4.89%
`
`-9.80%
`
`Note: The omitted drug category is pain relief.
`
`both single- and multisource drugs) were filled with generics in 1995, only 42.4% of
`prescriptions paid for by private managed care were filled generically. This may suggest
`that managed-care groups have successfully bargained for price discounts from trade(cid:173)
`name drug manufacturers.
`There is one important caveat to the potential social benefits of increased generic
`prescription. Reducing the returns to trade-name drugs may have an adverse effect on
`
`TABLE 9
`
`Tests of Moral Hazard for the Subsample of States Equality of Individual Insurance
`Variables with Self-Payment Random-Effects Probit Results
`
`Insurance
`Variable
`
`Medicaid
`Coefficient
`
`t-statistic
`
`Skin/
`Mucous
`Mem- Opthal-
`Cardio-
`Anti-
`microbials vasculars Metabolics Hormones branes mics
`
`Pain
`Relief
`
`Respir-
`atory
`Tract
`
`-.06
`
`-.45
`
`.56
`
`2.18
`
`-.48
`
`-1.67
`
`.37
`
`1.37
`
`.61
`
`1.45
`
`-2.31%
`
`17.97%
`
`-15.37%
`
`14.30%
`
`7.11%
`
`-.04
`
`-.13
`
`-.17
`
`-.21
`
`-.90%
`
`-1.24%
`
`-.02
`
`.37
`
`-.16
`
`.56
`
`.13
`
`%change
`Medicare
`Coefficient
`
`!-statistic
`
`-.10
`
`1.97
`
`-.71
`
`.36
`
`2.01
`
`.32
`
`.71
`
`1.49
`
`.46
`
`%change
`Private
`Coefficient
`
`t-statistic
`
`%change
`HMO/prepaid
`Coefficient
`t-statistic
`
`-.61%
`
`10.91%
`
`-5.60%
`
`14.13%
`
`2.87%
`
`12.07%
`
`2.89%
`
`-.01
`
`-.13
`
`-.01
`
`-.07
`
`-.14
`
`-.91
`
`-.20
`
`-1.28
`
`.13
`
`.40
`
`.20
`
`.45
`
`.12
`
`.55
`
`.10
`
`.13
`
`-.51%
`
`-.36%
`
`-5.06%
`
`-7.50%
`
`1.03%
`
`3.42%
`
`2.62%
`
`.94%
`
`-.04
`-.32
`
`.17
`
`.55
`
`-.17
`
`-.65
`
`-.12
`
`-.59
`
`.43
`1.26
`
`.28
`
`.53
`
`-.06
`
`-.21
`
`-.05
`
`-.03
`
`%change
`
`-1.38%
`
`4.68%
`
`-5.83%
`
`-4.54%
`
`4.31%
`
`5.21%
`
`-1.13%
`
`-.41%
`
`Note: The percent change row represents the average percent change over the sample of patients in the
`probability of receiving a generic prescription when the patient's insurance status changes from self-pay to
`the appropriate insurance category. Sample sizes in empty cells are too small to estimate coefficients.
`
`PAGE 4 OF 5
`
`
`
`HELLERSTEIN
`
`I 133
`
`pharmaceutical R&D investment and new drug development. There is little evidence
`on the magnitude of this effect, which suggests another important avenue for future
`research. Nonetheless, if the private returns to pharmaceutical R&D need to be sup(cid:173)
`plemented to promote more efficient levels of drug discovery, the best mechanism to
`subsidize private drug development is probably not the indirect subsidies provided by
`market imperfections in the demand for prescription drugs.
`
`Appendix A
`
`The data for this article are taken from three versions of the 1989 NAMCS: the publicly available
`•
`NAMCS for patient visits; a version of the NAMCS for patient visits with additional confidential identifying
`information; and the publicly available NAMCS for drug mentions. The NAMCS is a survey of approximately
`1,200 office-based physicians and a subsample of their patients, conducted not-quite annually by the National
`Center for Health Statistics (NCHS). It is a three-stage sample of primary sampling units (PSUs), physician
`practices within a PSU, and patient visits within practices. A PSU is a county, group of counties, or standard
`metropolitan statistical area. After the first and second stages of the sample, selected physicians were ran(cid:173)
`domly assigned to two consecutive weeks of the year beginning in February 1989, and they filled out detailed
`questionaires on a random subsample of patient visits during those two weeks. The average physician re(cid:173)
`corded data for approximately 30 patients, although there is a lot of variability in the number of patients per
`physician. The sampling scheme was designed so that physicians with larger practices recorded data for more
`patients, although not in fixed proportions to the overall sizes of the practices. (Physicians who saw fewer
`than ten patients filled out questionaires for all patients they saw.)
`These questionaires contain data on demographic characteristics of the patient (age, sex, race, ethnicity)
`as well as data pertaining to his or her medical condition and details about what occurred during the visit
`such as duration of the visit, procedures performed, and diagnosis. In addition, the physician recorded for
`each patient the expected source(s) of payment for the visit: self-pay, Medicare, Medicaid, Blue Cross/Blue
`Shield, other commercial insurance, HMO/prepaid plans, no charge, or other. If the patient paid for the visit
`but was to be reimbursed by a third-party payer, the physician was told to only consider the third-party payer
`as the source of payment. Most importantly, the physician was instructed to list up to five medications
`ordered for the patient and to record .. the same specific drug name (brand or generic) . . . used on any
`prescription." The definition of medications was interpreted broadly and included both prescription and
`nonprescription pharmaceuticals.
`All three versions of the data contain the results of the questionaires as well as information identifying
`the specialty of the physician and the region of the country in which the physician practices (North, South,
`East, Midwest). In the publicly available 1989 NAMCS for patient visits and its confidential counterpart, the
`unit of observation is a patient visit, and patient-specific sampling weights are included in the data. The
`confidential NAMCS also links the patients of each physician together via a physician identification number
`and contains information on the U.S. state in which the physician practiced. The state identifiers allow
`prescriptions to be classified according to state laws about generic substitution, and the physician identifiers
`allow for the inclusion of physician-specific effects into the model. Although the public version of the
`NAMCS is available for other years as well, the confidential version of the data has only been prepared for
`1989.
`In the NAMCS for drug mentions, the unit of observation is an ordered medicine. Therefore, infor(cid:173)
`mation is included only for those patients for whom a drug was ordered; for patients for whom multiple
`medicines were ordered, multiple observations appear (and these observations cannot be linked in these data).
`Because the drug-mentions data focus on medicines, drug-specific sampling weights are attached to each
`observation. In addition, these data contain information matching each ordered medicine to a unique trade(cid:173)
`name drug code as well as a corresponding generic drug code. The data also include other information about
`the drug ordered such as the generic name, manufacturer (either generic or trade-name), prescription status
`(over-the-counter or prescription), and drug class code (one of 20 major classes such as opthalmics or
`neurologics).30 In conversations with representatives at the NCHS, it became clear that the manufacturer
`codes for each drug are not entirely reliable. I therefore verified each manufacturer codt uv:ng the 1991 Drug
`Facts .and Comparisons, a comprehensive pharmaceutical industry source for drug information.
`The two sources of sampling weights in the data are the patient weights from the NAMCS for patient
`visits and the drug weights from the drug-mentions data. Experimentation with these two sets of sampling
`weights yielded very little difference between unweighted and weighted estimates of any of the results in
`this article. All results reported here are derived without sampling weights.
`
`30 See Table 3 for a list of the drug codes used in the empirical analysis.
`
`PAGE 5 OF 5