Ecin_215.dvi

THE EFFECT OF DIRECT TO CONSUMER TELEVISION ADVERTISING
ON THE TIMING OF TREATMENT
W. DAVID BRADFORD, ANDREW N. KLEIT, PAUL J. NIETERT and STEVEN ORNSTEIN∗ We examine how direct to consumer advertising affects the delay between diagnosis and pharmacological treatment for patients suffering from a common chronic disease.
The primary data for this study consist of patients diagnosed with osteoarthritis(N = 18,235) taken from a geographically diverse national research network of 72primary care practices with 348 physicians in 27 states over the 1999–2002 timeperiod. Brand-specific advertising data were collected for local and network televisionat the monthly level for the nearest media markets to the practices. Results of durationmodels of delay to treatment suggest advertising does affect the length of time thatpatients and physicians wait to initiate therapy. This evidence suggests that these effectsmay be welfare enhancing in that advertising tends to encourage more rapid adoptionamong patients who are good clinical candidates for the therapy and leads to lessrapid adoption among some patients who are poor clinical candidates. (JEL D12, I11) the disease and the drug brand name (as long asa brief list of side effects was mentioned and a In August 1997, the Food and Drug Adminis- 1–800 number or World Wide Web site was pro- tration (FDA) relaxed the rules governing tele- vided with more detailed information). Spend- vision advertising of prescription pharmaceuti- ing for direct to consumer advertising (DTCA) cal products. Before that time, broadcast ads for prescription drugs went from $985 million were permitted only to mention either the name in 1996 to approximately $4.2 billion for 2005 of a drug or a disease against which a drug (Donohue, Cevasco, and Rosenthal 2007). This was effective but not both. After August 1997, has led to a great deal of debate in the med- pharmaceuticals were allowed to mention both ical profession and among health care insurersand managed care organizations and an ongo- *The authors would like to thank Jack Calfee, John Caw- ley, and Terry Steyer for their comments and assistance on ing review of advertising rules by the FDA.
this paper as well as participants at the Sloan Program Con- However, very little is actually known about the ference on Consumers, Information and the Evolving Health- effects of DTCA for the efficient allocation of care Market Place, the 2005 Annual Health Economics Con- ference (State College, Pennsylvania), and the 2005 Inter-national Health Economics Biannual Congress (Barcelona).
This paper was funded by grants from the Agency for cian prescribing patterns and courses of care Healthcare Research and Quality (1 R01 HS011326-01A2) for patients suffering from a representative and from the National Heart, Lung and Blood Institute(1 R01 HL077841-01).
chronic condition, osteoarthritis (OA). Thiscondition is of special significance, as one Bradford: Busbee Chair in Public Policy, 201C Bald- win Hall, Department of Public Administration and of the major products in this area, Merck’s Policy, University of Georgia, Athens, GA 30602.
“Vioxx” cyclooxygenase-2 (COX-2) inhibitor, Phone 706-542-2731, Fax 706-583-0610, E-Mail brad- was forced to withdraw from the market in Kleit: Professor of Energy and Environmental Economics, MICASU Faculty Fellow, Department of Meteorology,Department of Energy and Mineral Engineering, Penn-sylvania State University, 507 Walker Building, Univer- ABBREVIATIONS
Nietert: Associate Professor, Department of Medicine, Med- ical University of South Carolina, 135 Cannon St., Room 403J, Charleston, SC 29425. Phone (843) 876-1204, Ornstein: Professor, Department of Family Medicine, Medi- cal University of South Carolina, Charleston, SC 29425.
Phone (843) 876-1213, E-mail ornstesm@musc.edu Online Early publication September 24, 2009 2009 Western Economic Association International October 2004 due to side effects on patients pharmaceutical, making it the 39th most adver- with heart conditions. The side effects of Vioxx tised brand of any kind in 2000 (“Top brands have caused considerable criticism of Merck’s in network primetime –2000” 2001). Over the advertising strategy for Vioxx (see, e.g., editori- same time periods, Pharmacia and Pfizer jointly als from the New England Journal of Medicine; spent $78 million in DTCA supporting Cele- The primary goal of this paper was to deter- DTCA in the context of the market for COX- mine what effect local and national television 2 inhibitors. During the period spanned by our advertising on behalf of the two main COX-2 data (1999–2002), the two available COX-2 inhibitors had on the treatment decisions that inhibitors were Vioxx and Celebrex. These drugs patients made in collaboration with their physi- are appropriate subjects of study for a number cians. In particular, we will examine the impact of reasons. First, they have been heavily adver- of DTCA on the time patients wait after diag- tised. Second, they are significantly more expen- nosis with OA and before initiating treatment sive than alternative pharmacologic treatments with a COX-2 inhibitor. The paper will pro- for chronic pain. Third, data published approxi- ceed by first reviewing the literature on DTCA mately halfway through our sample period indi- in Section II. Section III will present a the- cated that these products may carry a signifi- oretical model of optimal delay to treatment.
cant risk of adverse cardiac side effects. Conse- Section IV will present details of the data quently, from mid-2001, many physicians began and empirical model we implement. Empiri- to question the wisdom in their use (Topol cal results are presented in Section V, and 2004). Additionally, given the widespread sus- Section VI concludes with a discussion about picion of advertising for prescription drugs, clinicians and policy makers have questionedwhether DTCA for COX-2’s might be caus-ing fatalities. Our results will speak directly tothis issue in that if DTCA had a deleterious effect (in terms of matching high-risk patientsto COX-2 use) then we will get one sign pat- A. Advertising for COX-2 Inhibitors tern on interaction terms between advertising Historically, pharmaceutical advertising was and patient comorbidities. If, however, DTCA done largely through “detailing” —promotion is having a positive effect (by matching patients directly from the manufacturer to the physician, most suited for COX-2 inhibitors with that ther- either through visits by representatives or con- apy) then we will get the opposite sign pat- tacts by pharmacists or through advertisements tern on the advertising-comorbidity interaction in professional journals. Since the mid-1980s, however, drug companies in the United States have increasingly turned their marketing strate- television advertising for both Celebrex and gies directly toward the consumer. This advertis- Vioxx at the national (network) level and for ing largely takes place through television media the top 75 local media markets in the United and in newspapers. This change in advertis- States. These data are aggregated to the monthly ing approach has its share of both critics and level. Celebrex was approved by the FDA in The pharmaceutical industry in the United May 1999. Thus, at least one of the products States is large —accounting for more than $132 was available for use over the entire 1999–2002 billion in retail sales in 2000 alone (NIHCM time period of our analysis. Figure 1 presents the 2001). In 2000, Celebrex (celecoxib), the lead- 2000–2002 trend for the spending on national ing COX-2 inhibitor, had sales of approxi- television advertising for Vioxx and Celebrex mately $2.6 billion (“Pharmacia has setback taken from our advertising database (described for parecoxib” 2001)—while Vioxx (rofecoxib) sold more than $1.2 billion in the first half of2000 (Knight Ridder News 2001). In supportof Vioxx, Merck spent almost $161 million in B. Literature on the Impact of DTCA In economic terms, we would expect adver- the most spent on DTCA for any prescription tising for prescription drugs to have three National Celebrex and Vioxx Advertising Dollars 2000.012000.032000.052000.072000.092000.112001.012001.032001.052001.072001.092001.112002.012002.032002.052002 2002.092002.11 possible effects. First, advertisement for a par- been published to date have yielded conflict- ticular prescription product will provide infor- ing results. There is an arm of this literature mation regarding the symptoms and regarding that is generally supportive of advertising in the fact that effective treatments are available this market, such as work by Telser (1975) and about the medical condition that the drug treats.
Leffler (1981). Keith (1995) finds that patient This may be labeled a “public good” effect, suggestions regarding pharmaceuticals (aspirin as the information conveyed by advertisements for cardiovascular disease) are important deter- for one brand may aid sales of all brands sell- minants in prescription decisions and that adver- ing competing products. Second, advertisements tising tends to lead to more appropriate care as may provide important information regarding a consequence. In this, Keith is advancing an side effects, contraindications, and the like that argument made earlier by Masson and Rubin may prompt patients to consult with their physi- that posits several mechanisms that could leadto positive impacts from advertising on the effi- cian regarding a treatment modality. This com- ciency of the pharmaceutical market (includ- ponent of the advertising may be labeled as a ing that it might encourage people to associate “matching” effect since it would assist patients symptoms with a disease and seek care or that it and physicians in matching treatment regimes.
might alert people to treatments they were previ- Third, advertising may simply lead patients to ously unaware of, which would encourage them demand a product because of the aesthetic or to seek care) (Masson and Rubin 1985). For a persuasive characteristics of the ad, or the rep- survey of the more optimistic literature in this utational impact of the ad, rather than the effi- area, see Rubin (1991) and Kleit (1998).
cacy of the drug. This effect may be labeled as a “brand” effect. Since health care markets, guine about the prospects of positive wel- including pharmaceuticals, are often character- fare effects from prescription pharmaceutical ized by moral hazard (as patients do not gen- advertising. Hurwitz and Caves (1988) find erally pay the full cost of the medications they that—on net—promotional activities by phar- consume), the welfare implications of this third maceutical firms tend to have the effect of preserving market share for existing products The studies on the impact of advertising in the prescription pharmaceutical market that have pounds in the market. King (1996) uses monthly sales data in the ulcer drug market to test to initiate treatment, the patient will evaluate the effect that marketing efforts have on the her utility with treatment and her utility without industry. He finds that marketing by a firm treatment and pick whichever path yields the causes the demand for the firm’s own prod- highest expected value. If the instantaneous ucts to become more inelastic. Similarly, Rizzo utility associated with therapy is not higher (1999) finds that DTCA significantly reduces than the utility associated with no therapy then price elasticity in the market. A reduction in the patient will choose to delay. Each period price elasticity would increase opportunities for thereafter, the patient will undertake the same utility calculus, to determine whether to begin The post-1997 era has presented an opportu- nity for examination of the new policy regime This model is similar in form to real option for DTCA, and much of the literature has been theories that have been applied to financial focused on the FDA policy shift. As Zachry and instruments (see, e.g., McDonald and Siegel Ginsburg point out, however, there is a paucity 1986; Merton 1973), to the timing of land of studies that examine the actual impacts of development (see, e.g., Arnott and Lewis 1979; Capozza and Li 1994, 2002; Titman 1985), and In one of the few such studies, Dubois (2003) even to the timing of initial public offerings (see, examines the impact of DTCA through the lens e.g., Benninga, Helmantel, and Sarig 2005). An of variation in procedure and drug use. He notes interior optimum for the delay to an action can previous evidence that there is a wide geo- exist when there is a cost to undertaking some graphic variety in the use of various medications action and when the benefits from the action and suggests that such variations imply under- served populations. Dubois cites several sources We will motivate our empirical research by that indicate that geographic variations in pre- exploring a simple model of the timing of treat- scriptions have declined since the relaxation ment for a stylized chronic disease. Consider of DTCA regulations, perhaps implying that a patient who has been diagnosed with a con- DTCA is conveying important medication infor- dition, which reduces her health. Furthermore, mation to previously underserved populations.
assume that the impact on health is cumula- Calfee, Winston, and Stempski (2002) study tive —delaying treatment implies that the health whether the August 1997 policy changes at FDA state continues to decay. The progression of increased the demand for the statin class of the disease can be countered by pharmacolog- drugs using monthly data from IMSHealth and ical therapy. (Assume for the sake of simplic- Scott-Levin for a 58-mo period. The authors, ity that there is only one viable therapeutic however, found that advertising did not have a option.) The value of that therapy is unknown, statistically significant impact on aggregate pre- but the patient has expectations about the treat- scriptions filled. According to the authors, “it ment effect—expectations that can be affected may only be possible to detect the effect of by information, such as physician advice, tes- timonials from friends, or DTCA. Thus, the disaggregated data that link a patient’s choles- terol treatment history with the timing of DTCAexpenditures.” ht = H − δt, if no treatment H − δt + θ(a), with treatment, where H is the base level of health at diseaseonset, δ is the impact of the disease on health during each time period, t represents time, and Consider a patient who has been diagnosed θ(.) is the expectations around the pharmaco- with some chronic disease. On diagnosis, the logic treatment effect, which can be affected patient faces a choice —either initiate or delay by advertising, a. Furthermore, θa > 0 and treatment. The benefits of treatment are potential improvements in symptoms of the underlying Utility is defined across consumption of some disease. The costs are monetary (the treatment numeraire good, xt , and health, ht . Consumers generally must be purchased at some positive may consume 1 unit of pharmaceutical treatment market price) and potentially psychic (people per period, at price P, or may continue to often resist taking medication) as well as the cost delay therapy. Once a person chooses to initiate of any potential side effects. To decide whether therapy, she will continue to receive treatment until her death at time T. (Given the degenerative Since diminishing returns in health imply that nature of the disease assumed in this model, the higher levels of health resulting from treat- once it becomes optimal to purchase treatment, it ment reduce the marginal utility of health, and will necessarily be optimal in every time period the complementarity between income and health will imply further reductions in Uh with the Consequently, each patient will maximize the lower net income resulting from the requirement present value of lifetime utility by selecting the that some portion of any consumed therapy is optimal delay for treatment onset according to: Comparative static analysis of the impact of advertising on the optimally selected delay of U (I, H − δt)e−rt dt treatment is straightforward. Inserting the opti-mal delay time functional, d∗(.) into Equa- tion (2) and differentiating with respect to the advertising parameter, a, and rearranging + U(I, H −δt +θ(a))e−rtdt, where U (.) is patient utility defined across ≡ ∂V 2 · ∂d∗(.) − Uhθa income (I ) and health. The optimal delay toonset, d∗, is defined by taking the first partial ⇒ ∂d∗(.) = Uhθa < 0. of Equation (1) with respect to d and setting it The sign of this effect depends on the sign d = U (I , H − δd) a . If a patient believes himself to be a good candidate for the treatment then greater expo- −U(I − P, H − δd + θ(a)) = 0. sure to advertising, or other positive informa-tion about the efficacy of the treatment, will The usual convexity assumptions require that reduce the optimally chosen delay to therapy the second partial of Equation (1) with respect initiation. However, if the advertising conveys information that leads the patient to believe thathe is a poor candidate for the treatment—by emphasizing adverse events or highlighting a ≡ Vdd = −δUh(I, H − δd) contraindication from which the patient suf- fers —then the expected treatment effect will h(I − P , H − δd + θ(a)) < 0. be reduced, which will lengthen the optimal Given these assumptions, the optimal time to delay. Thus, whether the information contained treatment exists, d∗ = d∗(I, H, δ, r, a). Note, of in the ad is “positive” or “negative” will depend course, that one condition for an interior solu- on patient expectations, other diagnoses, and tion is that at time = 0, the value of utility preferences across the characteristics of the without therapy must be greater than the value of utility with therapy. That is, if U (I − P , There are several implications from the theo- H − δt + θ(a)) > U (I, H − δt)∀t retical model for the empirical estimation. First, patient will choose to initiate therapy immedi- patients select the optimal delay from diagnosis ately and d∗ = 0. Similarly, if U (I −P , H − to therapy. Thus, we will calculate below this δt + θ(a)) < U(I, H − δt)∀t then the patient delay for each person in the data by estimat- ing a parametric duration model. Second, since pharmacological treatments are better suited tosome patients than others, the optimal delay ∂U (I, H − δd) = −δ will be a function of the patient’s other clin- ical diagnoses. Finally, the model implies thatthis optimal delay will be a function of income, −δUh|with T x = ∂U(I − P, H − δd + θ(a)). health state, opportunity cost of treatment, and advertising exposure —measured at the point of therapy initiation. We will employ either direct Survival Curves for Delay between Diagnosis with OA and Initiating Treatment, Survival Curves for Delay between Diagnosis with OA and Initiating Treatment, measures or proxy measures for each of these nonuse) having persisted to time t. One com- mon specification for this hazard rate is toassume it is the product of some baseline pop-ulation average hazard and an individual spe-cific term (which may or may not depend on The problem with this proportional hazard model is that it requires the population aver- Before specifying an empirical implementa- age effect to be constant over time. A more tion, we note that the decision we will model flexible approach for duration models is pre- is one that is joint between the patient and sented in Kalbfleisch and Prentice (1981) and the physician. Thus, our question will be “whatimpact does DTCA have on the likelihood that in Ridder (1990) and known as accelerated fail- the physician/patient interaction will result in a ure time (AFT) models. These models permit the prescription?” We will model the length of the baseline hazard to increase, decrease, or remain spell between diagnosis and initiation of COX-2 constant over time. Figure 2 below presents the inhibitor therapy. There is a long econometric empirical hazards for the delay to treatment literature on duration modeling, which presents data we explore. The instantaneous failure rates us with various modeling options. Options range are not constant for our data, as the propor- from nonparametric methods that impose few tional hazard assumption is strongly rejected distributional restrictions to parametric meth- for nearly all covariates in our models. Con- ods where the distribution of the “time to fail- sequently, we will estimate an AFT version of ure” (to use the language of much of the lit- a duration model. In addition, the decreasing erature) must be explicitly specified. All ver- hazard we observe is consistent with a Weibull sions of the model are based on estimating the distribution—which is the distribution we will hazard function or the likelihood of a transi- tion (in our case, between nonuse of COX-2 In addition to nonconstancy in the baseline inhibitors and use of COX-2 inhibitors) between hazard, we have one other characteristic of the data to accommodate in our model choice.
Patients may exit their delay spell in one of agnostic with regard to what the actual mech- two ways: by choosing Vioxx as the treatment to initiate or by choosing Celebrex. These exit ever, to assess whether such selection issues strategies are not likely to be random and may are driving the results, we reestimated, ex-post, consequently affect the length of time that the both versions of the model presented below delay spell itself lasts. Thus, we are faced with and included proxy variables for the probabil- the possibility of underling heterogeneity in the ity that the patient seeks a visit and the prob- duration data, which arises due to unobservable ability that the patient receives a prescription characteristics of the patient. Competing risk during the month when therapy is initiated.
versions of the AFT model can account for These proxy variables are the percent of all this data-generating process. For applications of the physician’s monthly office visits that are to the competing risk models in health care, see OA patients and the percent of all the physi- Hamilton (1997), Cutler (1995), and Picone, cian’s monthly visits to OA patients that are Wilson, and Chou (2003). These models are associated with a COX-2 inhibitor prescription, based on the generalized method of Heckman respectively. These variables represent na¨ıve and Singer (1984), which model heterogeneous estimates of the probability of a visit and pre- transition frailty as a multiplicative term with a scription for each patient during the month in which the patient actually began therapy. Since Thus, we will model the delay to COX-2 ini- these are practice average measures, they will tiation as an AFT hazard function with heteroge- be exogenous to the individual patient’s char- neous failures in a competing risk model where acteristics but capture general influences at the the instantaneous hazard follows a Weibull dis- practice level on the likelihood that the patient tribution, and the unobservable heterogeneity is progresses to treatment. The estimated DTCA modeled as a gamma distribution. The hazard parameters and the parameters on the DTCA interaction effects were highly robust to inclu-sion or exclusion of these variables —thus sug-gesting that selection effects are not playing alarge role in generating the results we present λi(t, xi, β, σi, αj) = exiβαjtαj−1α , below. For the sake of simplicity, therefore, we do not include these models in Table 3; how-ever, they are available on request from the characteristics, t is the duration of the individualdelay spell, and αj capture the group-levelheterogeneity in exit (to Vioxx or Celebrex). For some patients, it will be optimal to delay longer than we observe them, and for some, it will be television advertising for both Celebrex and optimal to delay indefinitely. Given that we only Vioxx at the national (network) level and for have data spanning the 1999–2002 time period, the top 75 local media markets in the United these two populations will be observationally States. These data are aggregated to the monthly equivalent—requiring that we take right-hand level. Figure 1 presents the 2000–2002 trend censoring into account. The model is estimated for the dollars spent on national television advertising for Vioxx and Celebrex taken from Finally, one remaining issue to be addressed is whether any selection effects are at work As Figure 1 indicates, at the national level, in the models. It seems plausible that advertis- monthly advertising exposure for the two brands ing could bring more people to the physician is roughly comparable over the entire 2000– office —which is what we find in prior research 2002 time period. Interestingly, a very differ- (Bradford et al. 2006). While this may lead ent picture emerges at the local level, where to more diagnoses of OA, it is not immedi- television advertising spots for Celebrex are ately clear that this would bias our estimates always greater than that of Vioxx for the entire on how long individuals wait before adopting period. The value of ads purchased at the local therapy. (Even if the shift in patient popula- level is much lower in dollar terms than at tion did lead to a change in delay, our esti- the national level. Since patients do not gen- mates are reduced form in nature and thus erally distinguish between the payment source of a television ad (whether the local station or national network receives the revenue), we will include measures of the total spending on tices indicates that the patient population is television advertising in our empirical mod- approximately 57% female, with a mean age els. Consequently, we will have both cross- (±SD) of 43.1 ± 21.1 yr and has a racial break- sectional and across time variation in DTCA African American, and 4.3% Hispanic. PPR-Net practices are located in urban, suburban,and rural areas. Based on a four-tiered classi- fication scheme for the Rural-Urban Commut- Practice Partner, Inc. (Seattle, Washington) ing Area codes (Morrill, Cromartie, and Hart has marketed a commercial electronic medi- 1999), PPRNet practices are highly represen- cal record (i.e., Practice Partner) to physician tative of the distribution of the U.S. popula- practices for more than a decade. This prod- tion. Among the 114 practices active as of uct is intended to replace paper charts and December 2005, there are 462 physicians, 51 has been widely adopted—largely by prac- physician assistants, and 63 nurse practitioners.
tices that are community based—for clinical Eighty-nine (78%) are family medicine prac- reasons and not for research purposes (nor tices; four of these are family medicine res- because the practice has any affiliation with idency programs. Twenty (18%) are internal a research group or institution). The Medi- medicine practices and five (4%) are combined cal University of South Carolina collaborates practices of primary care physicians. Thus, the with the vendor to gain access to the record PPRNet practices appear reasonably representa- extracts of practices that were willing to have tive of other primary care practices in the United their data used for research purposes. This led to the development of a geographicallydiverse national research network of ambula-tory, mostly primary care practices that use thissingle electronic medical record system (known as PPRNet). We will examine data on prac-tices from 1999 through 2002. As of 2002 (the end of the study period), 72 practices in the length of time each patient waits before 25 states, with 348 physicians, were network initiating therapy with one of the two COX- 2 inhibitors after being diagnosed with OA.
Each quarter, participating practices run a This is calculated as the difference (in days) computer program, developed and maintained between the diagnosis date for OA (recorded by the electronic medical record vendor, to in the clinical data) and the prescribing date extract patient activity of the previous quar- for the first instance of a Vioxx or Celebrex ter from the electronic medical record system.
prescription. However, COX-2 inhibitors are These data are taken from the patient’s med- used for many complaints in addition to OA; ical record—similar to chart abstraction. The additionally, patients may use a COX-2 for data capture all diagnoses, medications, patient occasional arthritis pain even before a formal characteristics (weight, blood pressure, etc.), lab diagnosis is made. In these cases, the “delay” tests ordered, and lab results. Currently, theentire research network database has informa- would be negative. For such negative delays, tion on 604,111 patients, including 3.6 mil- the duration between the diagnosis and the lion patient contacts, 3.8 million prescription prescription is set to 1 d. In addition, most of the records, 10.1 million vital signs, 12 million lab- patients with a diagnosis of OA do not receive oratory records, and 1.3 million preventive ser- a prescription for Vioxx or Celebrex as long as vices records. We extract a subset of this data they are observed in the data. Thus, the duration on 22,011 patients who had ever been diag- models we estimate will control for this right nosed with OA and who had physician visits in the years 1999–2002. (The time period of analysis is dictated by the availability of adver- eral categories, such as advertising information, tising data and not the availability of clinical patient individual clinical information, and mar- We obtained national and local advertising patient is female. We also include variables information from Competitive Media Report- that capture whether the patient has ever been ing, Inc., which collects data on media adver- diagnosed with (or treated for) other relevant tising for all products, including pharmaceuti- comorbidities. These include indicator variables cals, at the market (e.g., city) level. The data for if the patient has ever been diagnosed with are specific to the brand name of the prod- heart disease (coronary disease or hyperten- uct. Consequently, it is possible to determine sion), depression, diabetes, hyperlipidemia, or which products were advertised, which month ever treated for gastrointestinal difficulties with they were advertised, how many times they proton-pump inhibitors, H2 blockers, or related were advertised, and how many dollars were spent on the ads. Patients and physician prac- Imputations of additional descriptive vari- tices were assigned to the nearest media mar- kets separately by two of the authors. When imputed the price of an intermediate length a practice was close to multiple media mar- of physician’s office visit with an established kets, they were assigned to the one that was patient from the American Chamber of Com- nearest (by driving miles). In addition, we merce Research Association’s Quarterly Price excluded all practices that are unusually far Reports (http://www.coli.org/). These quarterly from the nearest media market (more than 100 reports contain average prices for 50 commodi- straight-line miles) to avoid any bias from mis- ties (including physician office visits) for around matching of practice and local media market 300 metropolitan areas. The linking between average physician visit price and the patient We measure advertising exposure as the total was accomplished by using the average price (national and local) dollars spent on ads broad- in the metropolitan area nearest the primary cast for each brand advertised. We add the care practice site. Average county per capita separate measures for national advertising and income, the percent of the county population local advertising into a single total monthly covered by Medicare, the percent of the county spending variable. While it is the case that employed in the labor force, the percent of local ads tend to be shown during different the county population that is Caucasian and times of the day and during different program- African American, the county population, and ming, it is unclear that this difference mat- the number of physicians per 10,000 popu- ters empirically. We have estimated versions lation were also merged onto the data from of our models with the national and localad spending included separately, and the net the Area Resource File. Counties were iden- results are not meaningfully different. Addi- tified as the county in which the practice is tionally, the parameters on national advertis- located. This information is available on an ing are essentially unchanged in magnitude and significance if we exclude local advertis- In addition to the impact of advertising, ing; so, while there may some concern that another source of information that may affect local advertising could be endogenous (a con- physician prescribing is medical journals. The cern that is ameliorated by the use of individ- late 1990s and early 2000s was a period when ual data, typically from only one practice in a significant amount of research was being con- each media market), the practical biases appear ducted on the efficacy and side effects of COX-2 inhibitors. We will control for clinical knowl- Additionally, since the theoretical model sug- edge in two ways. First, over the period of gests that patient characteristics at the time of our study (1999–2002), there were more than the switch are the important factors in initiating 900 publications in English-language medical therapy, and since we have not captured pre- journals about COX-2 inhibitors. Of those, 132 cisely when in the month therapy begins, we were specifically in the area of OA. To con- will measure potential advertising exposure as trol for the effect of this research on clinical the dollars spent in the month preceding the ini- providers, we created a data series, that mea- sures the number of publications in each month The patient data contain limited demographic that had the keywords: rofecoxib, celecoxib, and detailed clinical information. For patient Vioxx, Celebrex, and osteoarthritis. We further demographic information, we include patient refined the measure by dividing it into three age and an indicator variable for whether the series: the number of publications each month that focused on Celebrex, the number of publica- for Celebrex (Vioxx) in the separate models tions each month that focused on Vioxx, and the we run for delay to initiation for each brand number of publications each month that focused says little about the social welfare impact sen, and Topol (2001) published an influen- of DTCA. We can, however, learn something tial article in a major medical journal where about the welfare effects of advertising by they reviewed data available from a major examining its effect on delay to treatment for clinical trial, which indicated serious statisti- patients that are likely to benefit from, or be cally significant concerns about the cardiovas- poor candidates for, COX-2 inhibitor use. In cular risk associated with Vioxx (rofecoxib).
particular, patients who have required treat- To a lesser, and not statistically significant, ment for gastrointestinal problems using proton- extent, the paper raised concerns about Cele- pump inhibitors, H2 blockers, or other simi- brex (celecoxib). This was the first publica- lar treatments are more likely to suffer gastric tion in a major outlet to raise issues about irritation from nonsteriodal anti-inflammatory increased risk of myocardial infarction associ- drugs (NSAIDs) and so are the good candi- ated with COX-2 inhibitors in general and Vioxx dates for COX-2 inhibitors. In this case, we in particular. These concerns were later to be identified patients with gastrointestinal comor- bidities as those who have ever had a pre- the market in October 2004. We will include scription for ranitidine (e.g., Zantac), famoti- an indicator variable, which equals 1 after dine (e.g., Pepcid), cimetidine (e.g., Tagamet), August 2001 and 0 otherwise to test whether omeprazole (e.g., Prilosec), esomeprazole (e.g., the practicing clinical community responded Nexium), lansoprazole (e.g., Previcid), rabepra- zole (e.g., Aciphex), pantoprazole (e.g., Pro- of significant DTCA in favor of Vioxx and tonix), sulcralfate (e.g., Carafate), misoprostol (e.g., Cytotec), and Helidac, Prevpac, or meto-clopromide (e.g., Reglan). We interacted theadvertising measures with the indicator vari- E. Possible Impacts of Advertising able for gastrointestinal problems. If advertis- There are several variables that are cen- ing improves patient matching then the param- tral to understanding the impact of advertising eters on those interactions will be negative on patient treatment delay decisions. The first and significant—indicating shorter delays to set, obviously, are the measures of advertis- ing spending on behalf of Vioxx and Celebrex.
In contrast, there are a number of conditions In general, one would expect the impact of that make a person a poor candidate for use Vioxx or Celebrex brand television DTCA to of a COX-2 inhibitor. These include a diag- shorten the time that patients wait before initi- nosis of heart disease (hypertension and other ating any COX-2 inhibitor treatment—in which coronary diseases). Individuals with these con- case the parameter on the advertising measures ditions should initiate therapy with a COX-2 will be negative and significant. It is possible inhibitor less frequently, which translates into that the only effect of advertising is class level.
longer delay times. However, this set of con- That is, it may not matter which drug is adver- traindications was not widely discussed in the tised—any advertising for a COX-2 inhibitor clinical community until the publication of the may affect the demand for both approximately Mukherjee, Nissen, and Topol (MNT) article in equally. To evaluate this, we will estimate two August 2001. Consequently, we will test for versions of the model. The first will include total the informational components of the DTCA for dollars spent (as both a linear and a squared heart disease by including a three-way interac- term) in the previous month on both Celebrex tion term between the advertising measures, the and Vioxx as one variable. The second will heart disease indicator variable, and the indi- include separate measures (again, both linear cator variable for whether the treatment began and squared terms) for Celebrex and Vioxx ad after August 2001. If the advertising is con- spending. If television advertising has a pure veying clinically useful information then the class effect then advertisement for Vioxx and parameter on this interaction will be positive Celebrex would have roughly equal effects when and significant (indicating that advertisement entered individually on the delay to therapy induces individuals with those comorbidities to Data Description (N = 18, 235) Standard
Variable
Deviation
Average delay between diagnosis with OA and first use of COX-2 inhibitor (for 6,494 Total dollars in COX-2 advertising squared, month preceding therapy (in $100,000s) Total dollars in COX-2 advertising squared, month preceding therapy (in $100,000s) Total dollars in Vioxx advertising, month preceding therapy (in $100,000s) Total dollars in Vioxx advertising squared, month preceding therapy (in $100,000s) Total dollars in Celebrex advertising, month preceding therapy (in $100,000s) Total dollars in Celebrex advertising squared, month preceding therapy (in $100,000s) Patient has received gastrointestinal treatment Patient has been diagnosed with heart disease Interaction between gastrointestinal treatment and total COX-2 inhibitor advertising Interaction between heart disease and total COX-2 inhibitor advertising dollars during Interaction between heart disease and total COX-2 inhibitor and advertising dollars Number of journal publications discussing COX-2 inhibitors in month preceding Number of journal publications discussing Vioxx in month preceding therapy Number of journal publications discussing Celebrex in month preceding therapy Patient has been diagnosed with depression Patient has been diagnosed with hyperlipidemia Number of physicians per 1,000 population in county Percent of county residents who are employed Percent of county population that is Caucasian Percent of county population that is African American Average price for intermediate length physician visit wait longer to begin treatment after the con- including the average delay to treatment (con- traindication for heart disease was published in ditional on initiating therapy). More informa- tion on the raw delay measures is presented inTable 2, which shows the average delay timebetween the diagnosis and the first long-term use of each COX-2 inhibitor as well as the numberof people who ultimately adopt each. We find that patients who first adopt Celebrex tend to deviations of the variables used in our model, adopt more rapidly than those who first adopt Delay between OA Diagnosis and First COX-2 Use (Number of Patients) Average Delay for
Average Delay
Patients Who Do
for Patients Who
Not Adopt Vioxx
Adopt Vioxx
Average delay for patients who do not adopt Celebrex Average delay for patients who adopt Celebrex Vioxx, with a delay of 163 d for the Cele- is only one practice per media market, our brex users, compared with a delay of 199 d practice-level fixed effects correspond largely for the Vioxx users. Interestingly, for all the to media market fixed effects. Finally, we rely attention paid to the introduction of COX-2 on the fixed effects to correct for clustering inhibitors, and the large expenditures on promo- (repeated observations) at the physician practice tion on their behalf, 64% of the sample (11,741 level. The coefficient estimates are presented in In the first model, where the COX-2 tele- We estimate two versions of the hazard func- vision advertising dollars for both brands are tion presented in Equation (6). Model 1 contains summed together, we find that television adver- the advertising measures (as combined national tising does have a significant effect on the delay and local dollars spent) for Vioxx and Celebrex in treatment. Total television spending has a pos- together, along with the set of independent vari- itive first derivative and negative second deriva- ables listed above. Model 2 includes separate tive —both significant at better than the 1% measures for total Vioxx and Celebrex monthly level. The parameter estimates in Table 3 are advertising, along with the other independent not direct measures of the net effect in terms of variables. Thus, we will estimate the hazard rate days of delay. Table 4 presents these net effects for the variables of primary interest and con-verts the raw effects to days of delay inducedby a 1-unit increase in the variable of inter- xiβ ≡ β1 · T Vi + β2 · T V 2 + β est. When calculated at the mean of the data, a $100,000 increase in all COX-2 advertising 4 · CV Di · T Vi · T N Mi + β5 · CV Di during the month of diagnosis has the effect ·T Vi(1 − T NMi) + β6 · GIi + β7 of lengthening the time patients wait to begin COX-2 therapy after being diagnosed with OA CV Di + β · Zi + i by about 30 d (evaluated at the mean of thedata). Note that this is in addition to average and where TV is either the combined Vioxx and delay periods of between 160 and 200 d delay Celebrex national and local advertising dollars or brand-specific national and local advertising dollars, GI is the indicator for a gastrointestinal vision advertising spending into separate mea- comorbidity, CVD is the indicator for heart dis- sures for total monthly spending on behalf of ease comorbidities (cardiovascular disease and Vioxx and total monthly spending on behalf hypertension), TNM is the indicator variable of Celebrex. We find that the results are sim- for the post-August 2001 period (which corre- ilar—in that spending on behalf of both brands sponds to the publication of the Topol, Nissan, tends to increase the delay initially and then and Mukerjee article that first alerted the med- lead to a decrease —again, with statistical sig- ical profession to the increased risk for heart nificance in excess of the 1% level. There is, attack associated with the use of some COX- however, a significant difference in terms of 2 inhibitors), and Z represents the remainder
where on this quadratic function the two brands of the explanatory variables discussed above.
The vector Z also includes individual practice
In the case of Vioxx, the relationship between fixed effects. Since in nearly all cases there total spending and delay of therapy becomes AFT Duration Models of Time between Diagnosis and Treatment for OA Patients Model 1 Combined
Model 2 Separated
Vioxx and Celebrex
Vioxx and Celebrex
Variable
Total Advertising
Total Advertising
Total dollars in COX-2 advertising, month preceding therapy (in 0.0730 (17.300) $100,000s)Total dollars in COX-2 advertising squared, month preceding therapy −0.0002 (−11.91) (in $100,000s)Total dollars in Vioxx advertising, month preceding therapy (in 0.0482 (4.17) $100,000s)Total dollars in Vioxx advertising squared, month preceding therapy −0.0010 (−9.12) ($100,000s)Total dollars in Celebrex advertising, month preceding therapy (in 0.0687 (14.79) $100,000s)Total dollars in Celebrex advertising squared, month preceding −0.0001 (−2.39) therapy (in $100,000s)Patient has received gastrointestinal treatment −0.0182 (−0.16) 0.0970 (0.87) Patient has been diagnosed with heart disease −0.3110 (−2.51) −0.3271 (−2.63) Interaction between gastrointestinal treatment and total COX-2 −0.0029 (−2.17) −0.0037 (−2.69) inhibitor advertising dollarsInteraction between heart disease and total COX-2 inhibitor advertis- −0.0451 (−22.87) −0.0044 (−2.16) ing dollars during pre-TNM periodInteraction between heart disease and total COX-2 inhibitors adver- 0.0554 (32.82) 0.0132 (7.03) tising dollars during post-TNM periodNumber of journal publications discussing COX-2 inhibitors in month 0.1709 (5.39) −0.1664 (−4.79) preceding therapyNumber of journal publications discussing Vioxx in month preceding −0.0542 (−1.77) 0.1468 (4.31) therapyNumber of journal publications discussing Celebrex in month pre- 1.2403 (26.41) 0.9796 (22.66) 1.9057 (5.42) −0.3832 (−1.03) 2.5651 (8.18) 1.6536 (5.44) −2.9144 (−9.41) −2.4908 (−8.10) 0.0157 (5.87) 0.0163 (6.23) −0.0118 (−0.16) −0.0283 (−0.39) Patient has been diagnosed with depression −0.0762 (−0.90) −0.0919 (−1.11) 0.1694 (1.85) 0.1244 (1.38) Patient has been diagnosed with hyperlipidemia 0.1950 (2.51) 0.2036 (2.66) −0.00003 (−5.99) −0.00003 (−6.15) Number of physicians per 1,000 population in county 0.0021 (1.18) 0.0034 (1.95) 0.00003 (1.52) 0.00001 (0.48) −0.1345 (−1.39) −0.3372 (−3.47) Percent of county residents who are employed −0.0871 (−5.30) −0.0838 (−5.30) Percent of county population that is Caucasian −0.0159 (−1.30) −0.0108 (−0.93) Percent of county population that is African American −0.2625 (−4.44) −0.2604 (−4.53) Average price for intermediate length physician visit 0.0686 (6.35) 0.0517 (4.85) 9.8063 (4.40) 12.7530 (5.89) Likelihood ratio test of overall significance p value 11, 957.97 (<0.0001) 13, 164.55 (<0.0001) Note: t statistics are given in parentheses.
negative at $2.4 million per month. During of Vioxx; thus, the marginal dollar spent for Vioxx advertising led to a reduction in the aver- approximately $3.5 million per month on behalf age delay before initiating any COX-2 therapy AFT Duration Models of Time between Diagnosis and Treatment for OA Patients. Net Marginal Effects: Change in Days of Delay to Treatment for a 1-Unit Change in the Explanatory Variable Model 1 Combined
Model 2 Separated
Vioxx and Celerex Total
Vioxx and Celebrex
Variable
Advertising
Total Advertising
Total dollars in COX-2 advertising, month preceding therapy (in $100,000s)Total dollars in Vioxx advertising, month preceding therapy (in $100,000s)Total dollars in Celebrex advertising, month preceding therapy (in $100,000s)Interaction between gastrointestinal treatment and total COX-2 inhibitor advertising dollarsInteraction between heart disease and total COX-2 inhibitor adver- tising dollars during pre-TNM periodInteraction between heart disease and total COX-2 inhibitor adver- of approximately 14 d (Table 4). In contrast, in that they have exhibited the sorts of gas- the relationship between spending for Cele- trointestinal sensitivities for which Vioxx and brex and the delay to therapy adoption does Celebrex were designed. Other patients have not become negative until Pfizer would spend hypertension and coronary disease comorbidi- more than $34.5 million per month. During ties that make them poor candidates for using 1999–2002, Pfizer spent $6.5 million per month COX-2 inhibitors because these conditions have on average—and never more than $13.8 million been recognized as contraindications for COX- per month. This implies that the marginal dollar 2 use since the publication of the TNM arti- spent for Celebrex by Pfizer actually lengthens cle. As discussed above, we estimate models the delay in initiating any COX-2 treatment by with interactions between total television adver- approximately 41 d (Table 4). Thus, the positive tising (local and national) and the GI indicator effect of joint advertising spending on delay to variable, the heart disease and pre-TNM indi- adoption in the first model is driven completelyby Celebrex ads. This curious reverse associ- cators, and the heart disease and post-TNM ation between Celebrex television advertising and the use of Celebrex is consistent with other Table 3 presents the relevant parameters.
recent research we have conducted that exam- We find strong evidence for welfare-enhancing ined the association between television adver- informational effects. The interaction between tising and the rate of prescribing for Vioxx and total COX-2 inhibitor advertising and the indi- Celebrex at the physician practice level (Brad- cator variable for gastrointestinal sensitivity is negative and highly significant in both mod- Given that we find that television adver- els. In terms of the magnitude of the effect, tising on behalf of the COX-2 inhibitors has patients with GI comorbidities reduce their some effect on the delay to use, the next ques- wait time to adopt COX-2 inhibitor treatment tion to address is whether this effect is welfare by between 2 and 2.4 d for every additional enhancing or not. As discussed above, some $100,000 in monthly COX-2 inhibitor advertis- patients are better candidates for COX-2 use The effect of advertising on patients with 1. We also estimated versions of the duration of delay heart disease comorbidities (cardiovascular dis- for Vioxx and Celebrex separately, including spending on ease and hypertension) depends on whether both brands in each equation. In those models, we foundthat the general effects demonstrated in the second column we observe the therapy being initiated prior of Table 3 hold: Vioxx spending generally has a net effect of to the publication of TNM or after. Prior to reducing the delay to therapy adoption for both Vioxx andCelebrex, while Celebrex has a net effect of lengthening the TNM, patients with cardiovascular comorbidi- delay for both brands. Those results are available on request.
ties actually adopted COX-2 inhibitor therapy more rapidly when exposed to increased adver- tising. Given the information about the potential Rural/Urban Distribution of PPRNet Practices cardiac dangers associated with COX-2 inhibitor use (especially Vioxx), this is contrary to whatwould improve social welfare. However, prior to the publication of the TNM paper, this risk Practices
Patients
Population
was poorly appreciated in the clinical commu- nity. However, once the TNM paper was pub- lished, we find a strong positive impact on delay to therapy initiation for patients with diagnosed heart disease when exposed to greater levels of COX-2 inhibitor advertising. This is the direc-tion one would expect if the ads provide realinformation that assists patients and physiciansto more optimally match therapies. In fact, the COX-2 inhibitors (either generically or focus- post-TNM interaction effect is larger than the ing on a specific brand) in the context of pre-TNM interaction effect by approximately care for patients with OA. While the measures of clinical publication rates are uniformly sta- Recall that we proposed relatively strong tistically significant, there is little consistency tests of whether DTCA for COX-2 inhibitors in the parameter values (except for publica- leads to welfare-enhancing or-reducing effects.
tions involving only Celebrex—which tended If increasing DTCA is associated with reduc- to always increase delay times). We take this tions in the time patients who are good can- as evidence that clinical publications are mea- didates for the therapy wait before initiating surably important for the prescribing patterns treatment and is associated with increases in of primary care clinicians in community prac- the time patients who are poor candidates wait tice; we cannot characterize how the effect is before initiating treatment then the advertising felt. This is likely because some publications must be providing useful information to the are favorable toward the effectiveness (and cost- clinical matching process. In this case, DTCA effectiveness) of COX-2 inhibitors and some are has at least some welfare-enhancing character- pessimistic. However, our data do not currently istics. We find exactly this pattern in our inter- characterize the tenor of the publication. More actions between patients with previously treated research on how clinical information and DTCA GI difficulties (good candidates) and with pre- information interact appears warranted.
viously diagnosed heart disease (poor candi-dates). In addition, the expected positive effect from the heart disease only shows up after thefirst important clinical publication to demon- The increased use of television advertising strate that patients with heart disease are, in by manufacturers of prescription pharmaceuti- fact, at increased risk from COX-2 inhibitor use.
cal has been a controversial development over Taken together, this evidence strongly supports the past 5–10 yr in the United States. While the the neoclassical view of advertising as infor- use of such advertisement has grown dramati-cally since the early 1990s, to date, there have mation—and throws the strong criticism that been few studies which have empirically exam- pharmaceutical DTCA has recently received into ined the effect of these ads on patient care. The primary goal of this paper was to determine what Finally, Table 3 presents parameters which effect local and national television advertising test whether patients and clinicians respond to on behalf of the two main COX-2 inhibitors had information from the clinical literature. As dis- on the treatment decisions that patients made in cussed above, we included measures of gross collaboration with their physicians. In particular, publication rates in the month preceding ini- the treatment decision we studied is the time tiation of COX-2 use for papers that discuss patients choose to wait before initiating treat-ment with either Vioxx or Celebrex. Using dataon 18,235 patients from a set of geographically 2. Note that while the absolute magnitudes of the pre- and post-TNM interaction effects are quite different between dispersed community-based primary care prac- the two models, the net effects are quite similar.
tices, we have measured the determinants of the delay between diagnosis for OA and onset of Bradford, W. D., A. Kleit, P. Neitert, T. Mcilwain, and S.
COX-2 inhibitor therapy. To accomplish these Ornstein. “The Effect of Direct to Consumer TelevisionAdvertising on Physician Prescribing Behavior for the goals, we estimated a series of competing risk Treatment of Osteoarthritis.” Health Affairs, 25(5), duration regressions using an AFT model.
Despite the importance of our study, there are limitations. First, without monthly data for Cholesterol-Reducing Drugs.” Journal of Law and pharmaceutical detailing, we were unable to Economics,45(2, pt 2), 2002, 673–690.
account for the impact and interaction of detail- Capozza, D., and Y. Li. “The Intensity and Timing of ing directly. Such physician-based marketing Investment: The Case of Land.” American Economic remains a larger component of pharmaceuti- Capozza, D., and Y. Li. “Optimal Land Development cal marketing efforts than DTCA advertising.
Decisions.” Journal of Urban Economics, 51, 2002, It is possible that DTCA and detailing efforts were coordinated; if so, the DTCA effects mea- Cutler, D. “The Incidence of Adverse Medical Outcomes Under Prospective Payment.” Econometrica, 63, 1995, sured here might have included some detail- ing effect. Personal communication with phar- Donohue, J., M. Cevasco, and M. Rosenthal. “A Decade maceutical representatives, however, suggested of Direct to Consumer Advertising of Prescription that since both Vioxx and Celebrex were such Drugs.” New England Journal of Medicine, 357, 2007,673–80.
important products to the manufacturers, the rep- Dubois, R. “Pharmaceutical Promotion: Don’t Throw the resentatives discussed these products at every Baby Out with the Bath Water.” Health Affairs Web opportunity. This implies that there would be Exclusive, W3, 2003, 96–103.
little correlation between the levels of detail- Hamilton, B. H., and V. H. Ho. “Estimating Surgi- cal Volume-Outcome Relationships Applying Survival ing and local or national variation in advertising, Models: Accounting for Frailty and Hospital Fixed which in turn would minimize the potential for Effects.” Health Economics, 6, 1997, 383–95.
bias in the results. Second, practices also gener- Heckman, J., and B. Singer. “A Method for Minimizing the ally have a supply of pharmaceutical samples on Impact of Distributional Assumptions in EconometricModels for Duration Data.” Econometrica, 52, 1984, hand to give patients when they write a prescrip- tion. The availability of samples may influence Hurwitz, M., and R. Caves. “Persuasion or Information? which product is prescribed. Again, personal Promotion and the Shares of Brand Name and GenericPharmaceuticals.” Journal of Law and Economics, 31, communication with physicians and pharmaceu- tical representatives suggested that most physi- Kalbfleisch, J., and R. Prentice. “Estimation of the Average cians would typically have had a stock of sam- Hazard Ratio.” Biometrika, 68, 1981, 105–112.
ples of both on hand, implying that the poten- Keith, A. “Regulating Information About Aspirin and the Prevention of Heart Attack.” American Economic tial for omitted variables bias is limited. Addi- tionally, the practice-level fixed effects included King, C. “Marketing, Product Differentiation and Competi- in the models capture any general tendency to tion in the Pharmaceutical Industry.” Working Paper, MIT Department of Economics, 1996.
Kleit, A. N. “Using Advertising to Generate Information In summary, we find that increases in tele- and Signals for Product Quality: Lessons for Biotech- vision advertising for Vioxx is associated with nology in Canada from Pharmaceutical Markets in shorter wait times between diagnosis and use.
the United States,” in Biotechnology and the Con- On the other hand, Celebrex television adver- sumer, edited by K. A. Mathios. Boston: Kluwer, 1998,257–75.
tising is associated with longer delays to COX- Knight Ridder News. “Drug Maker Merck reports 5% 2 adoption. Finally, we also present evidence Increase in Profits.” Tribune Business News, July 21, that the effect of DTCA may tend to improve Leffler, K. “Persuasion or Information? The Economics of economic efficiency in that advertising tends Prescription Drug Advertising.” Journal of Law and to shorten the delays to adoption for patients Economics, 24, 1981, 45–74.
who are better candidates for COX-2 use and Masson, A., and P. Rubin. “Matching Prescription Drugs lengthens the delay to adoption for patients who and Consumers: The Benefits of Direct Advertis-ing.” New England Journal of Medicine, 313, 1985, are worse candidates for the use of COX-2 McDonald, R., and D. Siegel. “The Value of Waiting to Invest.” Quarterly Journal of Economics, 101, 1986, REFERENCES
Merton, R. C. “An Intertemporal Capital Asset Pricing Arnott, R., and F. Lewis. “The Transition of Land to Urban Model.” Econometrica, 41, 1973, 867–87.
Use.” Journal of Political Economy, 87, 1979, 161–69.
Morrill, R., J. Cromartie, and G. Hart. “Metropolitan, Benninga, S., M. Helmantel, and O. Sarig. “The Timing Urban, and Rural Commuting Areas: Toward a Better of Initial Public Offerings.” Journal of Financial Eco- Descriptions of the United States Settlement System.” Urban Geography, 20, 1999, 727–48.
Mukherjee, D., S. Nissen, and E. Topol. “Risk of Car- Rubin, P. “Economics of Prescription Drug Advertising.” diovascular Events Associated with Selective COX-2 Journal of Research in Pharmaceutical Economics, Inhibitors.” Journal of American Medical Association, Schumann, M. “Top 100 Megabrands: Chevrolet Leads National Institute for Health Care Management. “Prescrip- tion Drugs and Mass Media Advertising, 2000.” keters Hit 2001 Potholes.” Advertising Age, 72, Washington, DC: NIHCM Foundation, November Telser, L. “The Theory of Supply with Application to the “Pharmacia Has Setback for Parecoxib.” Chemical Market Ethical Pharmaceutical Industry.” Journal of Law and Economics, 18, 1975, 449–78.
Picone, G., R. Wilson, and S. Chou. “Analysis of Hospi- Titman, S. “Urban Land Prices Under Uncertainty.” Ameri- tal Length of Stay and Discharge Destination Using can Economic Review, 75, 1985, 505–14.
Hazard Functions with Unmeasured Heterogeneity.” “Top Brands in Network Primetime–2000.” Brandweek, Health Economics, 12, 2003, 1021–34.
Ridder, G. “The Non-Parametric Identification of Gener- Topol, E. “Failing the Public Health-Rofecoxib, Merck and alized Accelerated Failure-Time Models.” Review of the FDA.” New England Journal of Medicine, 351, Economic Studies, 57, 1990, 167–81.
Rizzo, J. “Advertising and Competition in the Ethical W. M., S. J., Coons, and Ginsburg, D. B.
Pharmaceutical Industry: The Case of Hypertensive “Patient Autonomy and the Regulation of Direct-to- Drugs.” Journal of Law and Economics, 42, 1999, Consumer Advertising.” Clinical Therapy, 23, 2001, Copyright of Economic Inquiry is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission.
However, users may print, download, or email articles for individual use.

Source: http://afertig.myweb.uga.edu/health/bradford_dtc.pdf