THE EFFECT OF PRESCRIPTION DRUG
`
`ADVERTISING ON DOCTOR VISITS
`
`TOSHIAKI IIZUKA
`
`Graduate School of International Management
`Aoyama Gakvin University
`Tokyo, Iapan
`toshi.iizuka@gmail.com
`
`GINGER ZHE JIN
`
`Department of Economics
`University of Maryland
`College Park, MD
`jin@econ.umd.edu
`
`The dramatic increase of direct—to—consumer advertising (DTCA) of prescrip-
`tion drugs created intensive debates on its effects on patient and doctor
`behaviors. Combining 1994-2000 DTCA data with the 1995-2000 National
`Ambulatory Medical Care Surveys, we examine the effect of DTCA on doctor
`visits. Consistent with the proponents’ claim, we find that higher DTCA
`expenditures are associated with increased doctor visits, especially after the
`Food and Drug Administration clarified DTCA rules in August 1997. After
`1997, every $28 increase in DTCA leads to one drug visit within 12 months.
`We also find that the market—expanding effect is similar across demographic
`groups.
`
`1.
`
`INTRODUCTION
`
`The year 1997 witnessed an important change in direct-to-consumer
`advertising (DTCA) of prescription drugs. Prior to 1997, any DTCA that
`contained both brand name and medical claims must disclose a ”brief
`
`summary” of drug effectiveness, side effects, and contraindications.
`Consequently, TV advertising was prohibitively expensive, and DTCA
`was largely limited to newspapers and magazines. A small number
`
`We thank the editors, two anonymous referees, Thomas Hubbard, Scott Stern, Timothy
`Hannan, Iudy Hellerstein, Bill Evans, Seth Sanders, Iohn Rust, Robert Maness, and the
`participants at the NBER 2003 I0 meeting for constructive comments and suggestions.
`We also thank Kathryn Aikon for helping us interpret the FDA rules, and Catherine Burt
`for guiding our usage of the NAMCS data. We are Very grateful to TNS Media Intel-
`ligence/ Competitive Media Reporting (CMR) for generously providing the advertising
`data for this study. All errors remain ours.
`
`© 2005 Blackwell Publishing, 350 Main Street, Malden, MA 02148, USA, and 9600 Garsington Road,
`Oxford OX4 2DQ, UK.
`Journal of Economics & Management Strategy, Volume 14, Number 3, Fall 2005, 701-727
`
`ALL 2068
`MYLAN PHARMACEUTICALS V. ALLERGAN
`IPR2016-01128
`
`1
`
`ALL 2068
`MYLAN PHARMACEUTICALS V. ALLERGAN
`IPR2016-01128
`
`

`

`702
`
`Iournal of Economics <9’ Management Strategy
`
`of the prescription drug ads that aired on TV included only brand
`names without describing their indications. This tradition changed
`drastically after August 1997, when the Food and Drug Administration
`(FDA) clarified that pharmaceutical firms can use DTCA on TV that
`contain both brand name and indications without a ”brief summary.”
`Following the clarification, DTCA expenditures increased from $800
`million in 1996 to $2.5 billion in 2000. As of 2000, DTCA accounted for
`2.5% of the overall mass media ad spending in the United States. The
`top promoted drug—Vioxx—spent $146 million in DTCA, beating Pepsi
`Cola, Budweiser Beer, and most automobile manufacturers (NIHCM,
`2001).
`
`The effects of prescription drug advertising are controversial.
`Proponents argue that DTCA primarily has a market-expanding effect:
`the ads inform consumers of new treatment options and, therefore,
`generate new doctor visits. If true, this could improve patient welfare,
`because many diseases are underdiagnosed. Opponents argue, how-
`ever, that DTCA has a business-stealing effect that misleads patients
`into demanding heavily advertised drugs, leading to inappropriate drug
`use and the unnecessary purchase of expensive drugs. Not surprisingly,
`pharmaceutical firms support the former position, while insurers and
`medical providers generally agree with the latter view.2 Clearly, the heart
`of the debate is the distinction between the market-expanding versus
`business-stealing effects of advertising, a familiar issue in economics
`literature (e.g., Roberts and Samuelson, 1988; Gasmi et al., 1992).
`This paper contributes to the growing literature that investigates
`the effects of DTCA on the demand for prescription drugs.3 We focus our
`analysis on one type of market-expanding effect, namely, the extent to
`which DTCA affects patients’ visits to the doctor. For this study, we use
`nationally representative, patient-level data that cover all classes, which
`allows us to generalize the effect of DTCA beyond specific categories
`studied by previous papers (e.g., Berndt et al., 1995; Calfee et al., 2002;
`Wosinska 2002; Rosenthal et al., 2003).4 In addition, we exploit a rich,
`
`1. DTCA still needs to include a ”major statement” of the most important risks and
`refer consumers to other sources for more comprehensive information.
`2. Both sides of the debates are well documented. See I-Iolmer (1999, 2002) for a
`summary of the proponents’ position, and Hollon (1999) and Wolfe (2002) for a summary
`of the opponents’ position. See, also, the debate on the role of DTCA by several authors
`published in the February 26, 2003, issue of Health Affairs. In response to these debates,
`the FDA held a public hearing in September 2003 to review its policy on DTCA.
`3. In addition to the economics literature we discuss here, a number of surveys have
`been conducted in order to understand consumer and doctor responses to DTCA. For
`example, the FDA conducted surveys on DTCA in 1999 and 2002. Prevention Magazine
`(1998-2000) has also conducted surveys on DTCA annually since 1998. Gonul et al. (2000)
`analyze one of those surveys conducted by Scott-Levin, a pharmaceutical information
`company, and find that consumers and doctors value DTCA differently depending on
`ongoing needs for health care, degree of experience, and exposure to DTCA.
`4. We discuss these papers in more detail in the next section.
`
`2
`
`

`

`Prescription Drug Advertising
`
`703
`
`patient-level dataset and provide new insights into the heterogeneous
`responses to DTCA. The question of whether DTCA has a business-
`stealing effect is addressed in a companion paper (Iizuka and ]in, 2004).
`Combining 1994-2000 monthly DTCA data with the 1995-2000
`National Ambulatory Medical Care Surveys (NAMCS), we estimate
`the effect of DTCA on doctor visits using a nonlinear least-squares
`regression with drug-class-fixed effects and allow DTCA to depreciate
`over time. We find that higher DTCA expenditures are associated with
`increased doctor visits and that this relationship is stronger after the
`1997 clarification. Specifically, after the clarification, every $28 increase in
`monthly DTCA expenditures leads to one patient visit within 12 months,
`and the effect concentrates on the visits that result in prescription drugs.
`In terms of heterogeneous responses to DTCA, we find that the market-
`expanding effect does not vary across demographic groups.
`The rest of the paper is organized as follows. Section 2 discusses
`the background and reviews the literature. After a data description
`in Section 3, we set up the empirical model in Section 4 and report
`estimation results in Section 5. Our conclusion is offered in Section 6.
`
`2. BACKGROUND AND RELATED LITERATURE
`
`The large increase of DTCA after the 1997 FDA clarification has created
`a controversy over the effects of DTCA. From a social planner’s per-
`spective, DTCA will improve consumer welfare if its benefits outweigh
`the costs. One benefit that proponents suggest is the market-expanding
`effect of DTCA. For example, DTCA may inform untreated patients of
`existing or new drug treatments and encourage them to seek medical
`help via office visits. This effect could be substantial because a number
`of leading diseases, such as diabetes, high cholesterol, and high blood
`pressure, are underdiagnosed and undertreated (Holmer, 1999). Holmer,
`who represents the pharmaceutical industry, further asserts, ”DTCA
`merely motivates patients to learn more about medical conditions and
`treatment options and to consult their physicians, but once the dialogue
`is started, the physician's role is preeminent” (p. 381).
`On the other hand, most opponents of DTCA worry about the
`business-stealing effects of DTCA. They are concerned that, once un-
`derinformed patients watch DTCA, they may demand inappropriate
`therapies from doctors and increase the cost of treatment. For example,
`Hollon (1999), who provides a doctor's perspective, argues that ”by
`creating consumer demand, [DTCA] undermine the protection that
`is a result of requiring a physician to certify a patient’s need for a
`prescription drug” (p.382). Cohen (1990) also argues that DTCA may
`encourage people to try more expensive drugs though cheaper, but
`equally effective, drugs may be available.
`
`3
`
`

`

`704
`
`Iournal of Economics 8 Management Strategy
`
`This paper contributes to the debate by providing a detailed
`analysis of the market-expanding effects of DTCA in outpatient office
`visits. We not only examine an aggregate market-expanding effect,
`but also examine the distribution of this effect among patient groups.
`Understanding the heterogeneous effects of DTCA is important because
`not all market-expanding effects are welfare improving. For example,
`moral hazard may encourage insured patients who watched DTCA to
`visit doctors ”too often,” because they do not bear all the costs of the
`visit (and the costs of resulting treatment). In such cases, DTCA may
`or may not improve welfare even if DTCA has a market-expanding
`effect.
`
`Our paper complements the few academic studies on DTCA. On
`the demand side, the earliest paper examining the effect of DTCA
`on prescription drugs is Berndt et al. (1995). They used the data for
`antiulcer drugs for 1977-1994, which precedes the surge of DTCA in
`the late 1990s.5 Calfee et al. (2002) estimated a monthly time-series
`regression of total statin drug prescriptions on advertising expenditures
`during 1995 and 2000. They found that advertising had no statistically
`significant effect on new statin prescriptions or renewals, but television
`advertising increased the proportion of cholesterol patients who had
`been successfully treated. Rosenthal et al. (2003) investigated the effects
`of DTCA and detailing on the aggregate sales of prescription drugs,
`using monthly data for five therapeutic classes between August 1996
`and December 1999. They found that DTCA has a significant effect on
`total class sales, but does not have any significant impact on market
`shares within each class. Our study builds upon these studies by using
`nationally representative, patient—level data that cover substantially
`larger number of therapeutic classes. Because of the advantage of the
`data, the conclusion of our paper is more applicable to a broader class
`of prescription drugs.
`Wosinska (2002) also examined the effect of DTCA on the demand
`
`for cholesterol-reducing drugs, using individual prescription claim data
`between 1996 and 1999. She finds that DTCA may affect the demand for
`an individual brand positively, but only if that brand is on the third-party
`payer's formulary. Similarly, using the NAMCS data as in the current
`paper, Iizuka and ]in (2004) examined the business-stealing effect of
`DTCA in nonsedating antihistamines. This paper is different from those
`papers because, while the above references are concerned about the
`
`5. In related research, Ling et al. (2002) examined the spillover of DTCA between
`prescription and over—the-counter (OTC) segments. Using data for antiulcer drugs, many
`of which switched from prescription to the OTC market in the late 1990s, they found small
`but significant spillovers from prescription to the OTC market for some brands, but not
`vice versa.
`
`4
`
`

`

`Prescription Drug Advertising
`
`705
`
`effect of advertising once patients arrived at doctor offices, that is, the
`business-stealing effect of DTCA, this paper examines whether DTCA
`brings potential patients to doctor offices, that is, the market-expanding
`effect of DTCA.
`
`On the supply side, Rosenthal et al. (2002) analyzed the industry-
`wide trends for DTCA and found that DTCA is highly concentrated on
`a subgroup of products and the spending fluctuates over time. Iizuka
`(2004) examined the determinants of DTCA and found that DTCA tends
`
`to concentrate in classes that involve fewer competitors. He also found
`that drugs that are new, of high quality, and for undertreated diseases
`are more frequently advertised. Our finding that DTCA of prescription
`drugs has a market-expanding effect on the demand side complements
`their findings.
`This paper also contributes to the body of literature that em-
`pirically distinguishes the market-expanding effect from the business-
`stealing effect of advertising [see Bagwell (2001) for a broad overview
`of classic papers on the economics of advertising, and King (2003) for
`a study on the disagglomeration and growth of the US advertising-
`agency industry]. An ad is viewed as market expanding when it purely
`increases total market size and business-stealing when it solely shifts
`market share among brands. Roberts and Samuelson (1988), for exam-
`ple, found that cigarette advertising has a significant market-expanding
`effect, but not a business-stealing effect. In contrast, Gasmi et al. (1992)
`found that advertising in the carbonated soft-drink industry is primarily
`characterized as business stealing.
`Finally, we recognize that the demand effect of direct-to-doctor
`advertising (i.e., detailing promotion) has been examined in earlier
`literature. Hurwitz and Caves (1988) looked at a cross-section of 56
`
`off-patent drugs and found that detailing promotion has a positive
`effect on the market shares between branded and generic drugs. Rizzo
`(1999) looked at the demand for antihypertension drugs for 1988-1993
`and found that detailing promotion lowers price sensitivity. Gonul
`et al. (2001) showed that detailing and free samples affect physician
`prescription behavior for an undisclosed therapeutic class. Azoulay
`(2002) found that, in addition to detailing promotion, scientific evidence
`from medical literature affected the diffusion pattern of antiulcer drugs.
`However, none of these papers looked at the effect of advertising
`directed to consumers. To be sure, this is mainly because DTCA in-
`creased its significance only recently, after the FDA clarification in
`1997. Moreover, because we are interested in the patient’s decision
`to visit a doctor rather than the doctor's decision to choose a specific
`drug, it is natural to focus on drug advertising that is oriented toward
`consumers.
`
`5
`
`

`

`706
`
`Iournal of Economics <9’ Management Strategy
`
`3 . DATA
`
`We combine individual-level data from the 1995-2000 NAMCS with
`
`the 1994-2000 monthly DTCA data from the TNS Media Intelli-
`gence/ Competitive Media Reporting (CMR). Each year since 1993,
`NAMCS has collected a national representative sample of individual
`visits to office-based physicians. For each office visit, NAMCS provides
`patient demographics, insurance status, physician specialty, time spent
`with the patient, diagnoses, dispositions, and prescription choices, if
`any.6 Although NAMCS has been constructed by stratified sampling
`each year, the Centers for Disease Control and Prevention only provide
`detailed sampling information for 1995 and beyond. To make sure the
`aggregate counts of office visits are nationally representative, we focus
`on 1995-2000.
`
`In comparison, the DTCA data provides the total DTCA expen-
`ditures for every prescription drug advertised via direct-to-consumer
`channels. Specifically, CMR monitors advertising outlays in units and
`dollars for several different media, including network TV, cable TV,
`newspapers, and magazines. The DTCA dollars reflect the typical costs
`of buying such elements as television time and print space.7 Our DTCA
`data covers 1 year longer than the period covered by the NAMCS data,
`so we can estimate the long-lasting effect of DTCA on patient visits.
`The DTCA and NAMCS data are matched by drug names and the
`month during which the advertising and physician office visits took
`place.8
`Our unbalanced panel data contain a total of 7,824 observations
`covering 151 drug classes over 72 months. Defining class-month as
`the unit of observation, we include a class-month in the sample if at
`least one visit (either drug or nondrug) occurred in that class and that
`month. To keep the sample stable, this sample construction applies to
`all regressions.9 A drug class is defined by the four-digit National Drug
`Code (NDC).1° Some classes do not appear in all months because some
`diseases are seasonal, and the NDC has added or deleted a few four-digit
`class codes between 1995 and 2000.
`
`6. See Cherry et al. (2001) and Burt (2002) for more detailed description of NAMCS.
`7. However, they may not reflect the discounts typically given to large buyers who
`bundle various products’ ads with one advertising agency.
`8. In rare cases where NAMCS assigns different drug class codes to the same drug
`across years, we use the 2000 NDC definition.
`9. As a result, some class-months may have zero drug visits in our sample, because
`we only observe nondrug visits in that class-month and vice versa.
`10. For example, hyperlipidemia (which includes cholesterol reducing drugs), ace
`inhibitors, and calcium channel blockers belong to separate four-digit NDC categories.
`
`6
`
`

`

`Prescription Drug Advertising
`
`707
`
`4. EMPIRICAL MODEL
`
`4.1
`
`SPECIFICATION
`
`Suppose all drugs in therapeutic class k treat disease k. If DTCA has
`raised consumer awareness of treatment options, greater exposure to
`DTCA in class k should encourage more consumers at the risk of disease
`k to visit doctors. However, it is quite possible that a DTCA of a specific
`drug motivates an individual to visit the doctor, but he or she ends up
`getting a different drug within the same therapeutic class. Therefore, we
`look at the effects of DTCA at the class level rather than at the individual
`
`drug level. We use therapeutic class and drug class interchangeably.
`Following this logic, the ideal model will link an individual’s
`exposure to DTCA in drug class k with his or her decision to visit
`the doctor's office for disease k. Unfortunately, the DTCA data are not
`individual specific, and the NAMCS data only record those patients who
`choose to visit doctors. To overcome these difficulties, we use NAMCS’
`sampling weights to calculate the total number of patient visits by class
`and time. Specifically, we estimate the following model:
`
`VISITk[ = (Xk + [313 + QKT + }\.1(’q[r
`
`+ ybSUMDTCAkt(1 — A1-"TER;)
`
`+ y,, SUMDTC/lkt X AFTER;
`
`+ ektl
`
`where VISIT,“ stands for the number of outpatient office visits related
`to drug class k at month t, and SLIMDTCAM stands for the discounted
`sum of DTCA of class k up to month t. AFTER, is a dummy equal to
`1 if month t is after the FDA clarification (August, 1997), 0 otherwise.
`The key coefficients, )4, and ya, denote the marginal effect of S LIMDTCA
`on VISIT before and after the clarification. ark and fit are drug-class and
`time-fixed effects, respectively. The other terms, OKT and A.K,qfr, further
`control time trends and are discussed in Section 4.2. 6k, is the error term.
`The definition of S LIMDTCA and VISIT requires more discussion.
`First, it may be reasonable to expect that the effect of DTCA would last
`for more than 1 month but depreciate over time. To capture the long-
`lasting effect of advertising, we define SUMDTCA in the following way:
`t
`
`SUMDTC/lk, = Z 8“iDTCAkt,
`'=0
`
`where 6 denotes the monthly depreciation rate to be estimated in the
`empirical analysis, and DTCA“ is the total DTCA expenditures reported
`for class k in month t. Because our DTCA data starts in January 1994,
`
`7
`
`

`

`708
`
`Iournal of Economics 8 Management Strategy
`
`we treat all the DTCA before January 1994 as zero. Although this
`treatment is motivated by our limited data, given the small amount
`of DTCA before 1994, the omission of the data prior to 1994 is unlikely
`to affect estimation results. In fact, results do not change if we impute the
`before-1994 DTCA from the aggregate DTCA amount reported by
`CMR.“ The discount factor 8 enters the model nonlinearly, so we
`estimate the whole model by nonlinear least squares.
`Second, we define our dependent variable, VISIT, in five different
`ways: drug visits, RX visits, OTC Visits, nondrug visits, and all visits. A
`Visit is counted as an RX visit of class k if it results in any prescription in
`class k. If a visit results in no prescription drug but at least one over-the-
`counter drug in class k, it is referred to as an OTC visit. Drug Visits are
`the sum of RX and OTC visits. If a Visit leads to no treatment or nondrug
`treatment, it is categorized as a nondrug visit. All visits are the sum of
`drug visits and nondrug visits. If a visit involves more than one class,
`we count it as one visit for each relevant class.
`
`Drug visits, RX visits, and OTC visits are well defined as NAMCS
`provides a drug class code for each drug that was prescribed or men-
`tioned by a doctor during each visit. In contrast, the definition of non-
`drug visits (and all visits, accordingly) involves a technical challenge:
`no therapeutic class code exists for nondrug visits. To address this issue,
`using drug visits observations, we create a mapping between a diagnosis
`and the most common drug class associated with the diagnosis. Specif-
`ically, we construct the mapping in the following way. First, we create a
`subsample of NAMCS visits that have a single diagnosis and at least one
`prescription or OTC drug. We focus on single-diagnosis visits to ensure
`that we can establish the link from a specific diagnosis to drug classes.
`Then, using the subsample, we identify the most common drug class
`for each diagnosis. This is done for each year separately, allowing drug
`treatment of specific disease to change over time. After this procedure,
`we are able to associate each diagnosis with a specific drug class and
`count the number of nondrug visits by drug class.”
`Because the purpose of this paper is to examine the effect of DTCA
`on physician office visits, it is theoretically correct to use all visits (i.e.,
`drug visits + nondrug Visits) as the dependent Variable. However, the
`nondrug visit definition as described above generates a fair amount of
`
`11. The imputation is implemented in the following way: if drug class k advertised at
`dollars in a specific month of 1994, we define the DTCAk in the corresponding month
`of 1993 as x - TOTALDTC93 /TOTALDTC94. The same imputation applies to any year
`between 1989 and 1993. The aggregate DTCA numbers are taken from "Prescription Drug
`Advertising Soars through Third Quarter; Expected to Top $1 Billion in 1998,” PR Newswire,
`December 30, 1998.
`12. NAMCS provides up to three diagnosis codes for each visit, so a nondrug visit
`may be linked to, at most, three drug classes.
`
`8
`
`

`

`Prescription Drug Advertising
`
`709
`
`noise and, therefore, results for nondrug visits are merely suggestive.”
`For this reason, we report the results of drug Visits as our main results.
`Nonetheless, main results stay the same even when all visits are used
`as the dependent variable. We show these results in Section 5.4.
`Finally, note that both VISIT and S LIMDTCA enter the specification
`linearly. We choose the linear—linear specification over alternatives due
`to the following reasons. First, as we show in Section 5.4, the linear-
`linear specification dominates the linear—log specification. Similarly, the
`log—linear specification dominates the log—log specification (by a small
`margin). Second, we prefer linear—linear over log—linear because log-
`linear forces us to drop or artificially modify Visit counts when these
`numbers are zero. This happens frequently when we break down the
`Visits by nondrug, RX, and OTC, or by patient groups. To avoid the
`sample selection problem, we use the linear—linear specification as our
`primary specification. To further address the concern of zero advertising,
`we rerun the linear—linear regression on advertising classes only. As a
`robustness check for the nonlinear estimate of depreciation, we also
`rerun the linear—linear regression using the depreciation rate estimated
`in Berndt et al. (1995). As showed in Section 5.4, basic results are robust
`regardless of specification.
`
`4.2 IDENTIFICATION
`
`Aggregate VISITH and SLIMDTCAM raise several econometric concerns.
`For example, a class with a large number of potential patients naturally
`has more patient visits. In the meantime, drug companies may also
`allocate large advertising budgets to large drug classes. Therefore, a
`high correlation between DTCA and patient Visits would not necessarily
`imply a causal effect. Moreover, as manifested by the concentration of
`DTCA in a small number of drug classes, drug companies may intention-
`ally select which classes to advertise. To address these concerns, we use
`class-fixed effect oak to control for time-invariant cross-class differences,
`a full set of month dummies fit to account for over-time fluctuation
`common for all drug classes, HKT for class-specific time trends, Where
`T denotes the number of months between the visit month and January
`1994, and A4K’qfr for class-specific seasonality. 6 and A are specific to an
`aggregated drug class K, defined by two-digit NDC classes.”
`
`13. NAMCS data not only record drug mentions that are directly related to diagnosis,
`but also include drug refills that may be irrelevant to the current diagnosis. To minimize
`the noise, we restrict the reference group to patients with a single diagnosis in the same
`calendar year.
`14. There are 21 two-digit NDC classes, such as ”cardiovascular-renal drugs” and
`"gastrointestinal agents.” In theory, the model is still implementable if we define 9 and A
`by four-digit NDC class. In reality, the model with four-digit trends yields qualitatively
`similar results in the coefficients of SLIMDTCA, but the depreciation rate 6 becomes
`
`9
`
`

`

`710
`
`Iournal of Economics 8 Management Strategy
`
`After all these controls, the effect of DTCA is mainly identified
`from over-time variations within each drug class. Still two reasons to
`suspect endogenous SLIMDTCAM exist. First, if the advertising budget
`is proportional to sales revenue and somehow patient visits correlate
`with sales revenue, reverse causality will reinforce a positive correlation
`between VISIT and S LIMDTCA and, therefore, overestimate y. Second,
`as the Pharmaceutical Manufacturing and Research Association claimed
`(Holmer, 1999, 2002), drug companies may devote much DTCA to
`underdiagnosed classes. Although class-fixed effects partially control
`for such selection bias, it is still possible that, for a specific drug class
`over time, drug companies commit to high DTCA expenditures when
`the actual number of visits is relatively low. This negative correlation
`will imply a downward bias in y.
`To address the endogeneity problem, we use the same drug com-
`panies’ DTCA expenditures in other ”unrelated” drug classes DTCA_kt
`as an instrument for DTCA“. We define class —k ”unrelated” to class k
`if (1) k and —k do not belong to the same two-digit NDC class and (2)
`the correlation of patients getting any drug (including over the counter)
`in the two-digit NDC classes K and —K at the same time is smal1.15 The
`latter ensures that we do not include the DTCA of complementary drugs
`in the instrument.
`
`For example, cholesterol-reducing drugs involve four major drug
`companies—Bristol-Myers Squibb (for Pravachol), Merck (for Zocor),
`Pfizer (for Lipitor), and Novartis AG (for Lescol). These four companies
`also produce and advertise prescription drugs in other classes; for in-
`stance, Novartis’s Habitrol targets smoking, Bristol-Myer’s Zerit targets
`HIV, Merck's Fosamax targets Osteoporosis, and Pfizer's Viagra targets
`erectile dysfunction. If class k refers to cholesterol reducing, DTCA_kt
`is defined as the sum of DTCA that these four drug companies spent
`on all the other classes excluding those under the same two-digit NDC
`class (metabolic/ nutrients) or under related two-digit NDC classes (e.g.,
`cardiovascular—renal drugs).
`We argue that DTCA across classes is correlated within the same
`company, either because the company pursues a particular marketing
`strategy for all products or because different drugs are subject to a
`common constraint in the advertising budget. After controlling for drug-
`class-fixed effects and time trends, we assume unobserved factors that
`drive changes in patient visits are uncorrelated across two classes, unless
`both belong to the same two-digit NDC class or they are under different
`two-digit NDC classes but often used on the same patients. By this
`
`unstable, often reaching the boundary of 1 or 0. We suspect this is because the four-digit
`trends absorb too many variations.
`15. After examining the distribution of correlation across two-digit classes, we use
`correlation = 0.1 as the cut-off point.
`
`10
`
`10
`
`

`

`Prescription Drug Advertising
`
`711
`
`assumption, the only way for DTCA_kt to influence VISIT“ is through
`DTCA”. Berndt et al. (1995) pursued a similar identification strategy.
`To further justify the validity of the instrument, we note that 90%
`of all the classes with positive DTCA“ also have positive DTCA_k¢.
`Moreover, the correlation of DTC/lkt and DTCA_kt are significant and
`positive in all years. If we regress DTCA,‘; on class-fixed effects and
`time trend (i.e., all exogenous variables), including DTCA-” on the
`right-hand side would improve the within-class R2 from 11% to 14%.
`The coefficient of DTCA_kt is also positive and highly significant in the
`first stage regression.” These statistics suggest that DTCA-,“ is a valid
`instrument for DTCA“. Because SLIMDTCA is defined as a discounted
`sum of current and past DTCA, we use DTCA_kt as instrument for
`DTCA“, DTCA_k(,_1) as instrument for DTCAk(t_1), and so on.
`The extent to which our instrument solves the endogeneity prob-
`lem depends on the validity of assumptions. Should one of the assump-
`tions fail, our results are better interpreted as a statistical association
`rather than a causal relationship between doctor visits and DTCA.
`Bearing this in mind, we proceed to the next section, which reports
`our results.
`
`5. RESULTS
`
`We start this section by describing the data and showing the results for
`drug visits. Then, we report the effect of DTCA on drug visits by patient
`groups. The final subsection conducts robustness checks and provides
`additional estimation results.
`
`5.1 DESCRIPTIVE STATISTICS
`
`Table I summarizes the dataset. As noted before, we have an unbal-
`
`anced panel dataset with 7,824 observations covering 151 classes over
`72 months between 1995 and 2000. A unit of observation is drug class-
`month. The first block of Table I summarizes DTCA data. As Rosenthal
`
`et al. (2003) and Iizuka (2004) showed, DTCA often concentrates in a few
`drug classes. In our data, on average, only 20.8% of classes advertise in a
`typical month." Conditional on positive advertising, the average DTCA
`expenditures are $3.84 million per class per month.
`The second block of Table I shows the number of visits by different
`visit definitions. By far the majority of NAMCS visits are drug visits,
`especially RX visits. Nondrug visits and OTC visits account for smaller
`
`16. P-statistics is 236.00, with p-Value equal to 0.
`17. This number has increased over time. In 1995, only 10 classes advertised via DTC
`channels. This number increased to 30-35 in 2000. During the 6 years from 1995 to 2000,
`67 classes have ever advertised through DTC channel(s).
`
`11
`
`11
`
`

`

`712
`
`Ioarnal of Economics 8 Management Strategy
`
`TABLE I.
`
`SUM MARY STATISTICS
`
`DTCA data
`
`Dummy =1 if DTCA>0
`DTCA ($ million)
`DTCA ($ million) conditional on DTCA>0
`Visit data
`
`Drug visits (million)—visits that lead to any drug mention(s)
`RX visits (million)—visits that lead to any prescription drug
`(RX) mention(s)
`
`OTC visits (million)—visits that lead to any otc drug
`mention(s) and no RX
`
`Nondrug visits (million)—visits that lead to no drug mention
`All visits (million) ( = drug visits + nondrug visits)
`Drug visits by demographics and doctor types
`Drug visits (million)—belongs to an HMO?
`Drug visits (million)—patient age265?
`Drug visits (million)—who pays for the visit?
`Self
`
`Govemment-sponsored program
`Private insurer
`
`Total OBS
`Total number of classes
`
`Mean
`
`Std. Dev.
`
`0.208
`0.799
`3.836
`
`0.778
`0.641
`
`0.406
`2.960
`5.516
`
`0.984
`0.875
`
`0.137
`
`0.406
`
`0.417
`1.275
`
`0.296
`0.390
`
`0.104
`
`0.413
`0.593
`
`0.226
`1.004
`
`0.202
`0.258
`
`0.051
`
`0.293
`0.433
`
`7,824
`151
`
`Note: A unit of observation is four—digit NDC class—month. For each observation, the count of visits takes into account
`the NAMCS sampling weights. The data in an unbalanced panel show some classes do not exist in all years due to
`seasonality or definition changes in the NDC.
`
`percentages of NAMCS visits. In the third block of the table, we provide
`a breakdown of drug visits by patient demographics. This shows that
`the majority of patients are non-HMO members, younger than 65 years
`old, and i