There are many reasons for the emergence of this new marketing strategy. ..... Physicians receive a quarterly newsletter from Blue Shield's Pharmacy and.
WORKING PAPER This version: June 2001
EFFECTS OF DIRECT-TO-CONSUMER DRUG ADVERTISING ON PRESCRIPTION CHOICE
Marta Wosińska Department of Economics University of California at Berkeley
I would like to thank Benjamin Hermalin, Tülin Erdem, Aviv Nevo, Paul Gertler, Paul Ruud and Kenneth Train for their comments and guidance. I am grateful to Blue Shield of California for sharing their claims data with me. I would also like to acknowledge financial support of the Berkeley Center for Health Research and the Institute for Business and Economic Research.
1. INTRODUCTION Marketing has always played a central role in the budgets of pharmaceutical firms. Because physicians serve as decision-makers on behalf of consumers (patients) and payers (health insurers and patients), this effort has historically been directed at physicians, both through person-to-person promotions and journal advertising. However, over the past decade, a new marketing strategy has emerged. Mass media direct-to-consumer advertising of prescription medicines has increased from $55 million in 1991 to over $1.8 billion in 1999. It is the fastest growing pharmaceutical promotional expenditure and, by 1999, it reached 27% of promotional spending directed at physicians and consumers. There are many reasons for the emergence of this new marketing strategy. This strategy would not have been possible without a change in the attitude of the Food and Drug Administration (FDA), which regulates pharmaceutical advertising. A major shift in the FDA policy came with the August 1997 decision to east the broadcast regulations. On the other hand, the growth of managed care may have inadvertently made advertising to consumers more appealing to pharmaceutical manufacturers. First, managed-care companies began to gain more influence over the physician prescribing process through the implementation of formularies, lists of drugs preferred by the insurer as the first-line treatment. These formularies, if properly structured and successfully implemented, can generate price sensitivity in physicians’ prescribing behavior. Second, the low patient out-of-pocket prescription costs that are characteristic of managed care introduced insurance-based moral hazard. If patients do not bear the cost of their drugs, then perhaps by extension, their physicians cannot be expected to appropriately balance benefits against costs – particularly if benefits are made to seem large through advertising. Largely because direct-to-consumer advertising of prescription drugs is such a recent phenomenon, its impact is not well understood even though it has generated much discussion. In contrast with promotions aimed at the physicians, advertising targeted to patients has been clearly visible, which has fueled a heated policy debate about its social merits. Private and public third-party payers, for instance, often cite advertising to patients as an important but inappropriate factor in prescription drug expenditures. They have called on the FDA to be less lenient in allowing this advertising. A report by the National Institute for Health Care Management (NIHCM, 2000) points out that drugs that contributed at least 1% to the $8 billion
1
total change in drug spending between 1998-1999 were all heavy mass-media advertisers. Among those, drugs introduced prior to 1998 had median sales growth rates of 49% (75% mean growth rates). However, beyond the correlation between sales and direct-to-consumer advertising expenditures and surveys reporting that patients tend to be prescribed the drugs they ask for, a clear relationship between direct-to-consumer advertising and the rise in drug prescriptions and sales has not been established. My paper is the first to estimate the effect direct-to-consumer advertising has on brand-specific demand. One possible effect of direct-to-consumer advertising is that it may encourage patients to seek treatment and, therefore, increase the total market demand for a therapeutic class. But direct-to-consumer advertising may also induce patients to influence their physicians to prescribe the advertised drugs specifically. Other potential effects may include increased therapy compliance. This paper studies the effects of direct-to-consumer advertising on prescription drug choice, while the effects on therapy compliance are studied elsewhere (Wosinska, 2001). Here I address several questions: What is the role of direct-to-consumer advertising vis-à-vis that of physician advertising? To what extent does direct-to-consumer advertising increase the utilization of advertised drugs? Does direct-to-consumer advertising undermine insurers’ efforts to induce price sensitivity in prescription patterns? Is the responsiveness to advertising different for newly diagnosed patients than those who are already being treated for the condition? These questions are addressed using data for one therapeutic category, cholesterollowering drugs. I chose this therapeutic category because of its significant share of prescription expenditures and its lack of over-the-counter substitutes (thus making the choice solely the physician’s). I assume the physician is the primary decision-maker and that his behavior is influenced by objectives of his two principals: the patient and the insurance company. As a result, we may observe price sensitivity in the physician’s prescribing behavior. We may also observe direct-to-consumer advertising having a discernable effect on prescription patterns. My empirical analysis uses a large sample of prescription claims for a large California health insurance plan over the period 1996-1999. These patient-level data are combined with monthly, brand-level direct-to-consumer advertising data from Competitive Media Reporting. The person-to-person detailing expenditures (sales calls to physicians) and free sampling data, also monthly and brand-level, are compiled by IMS Health.
2
I first consider whether the aggregate trend between sales growth and the intensity of direct to consumer advertising holds up in an individual choice model. I find that direct-toconsumer advertising has a significant positive influence on probability of choice. However, this effect is much larger for drugs that are listed on the studied formulary than for those excluded from it. Unless the drugs that are not on the formulary also have less effective advertising campaigns, this differential effect suggests that direct-to-consumer advertising supports, rather than undermines, the formulary. Possibly physicians suggest that advertised formulary drugs be used when patients inquire about ads for drugs that are not listed on the studied formulary. In addition, the effect of direct-to-consumer advertising appears to differ between the initial therapy decision, when the patient has no experience with any cholesterol-lowering drugs, and the subsequent decisions. Namely, the effect of advertising is higher in the first decision, particularly for non-formulary drugs. Strong state dependence is also apparent in these data, but high heterogeneity of preferences over the various brands is also evident. Next, I incorporate physician advertising and free sampling activity, which are largely unexplored in previous analyses of direct-to-consumer advertising. The addition of detailing expenditures lowers the estimate of the impact direct-to-consumer advertising has on choice probabilities. Not only is the effect of consumer advertising diminished, but the estimated impact of detailing is five times as large as that of consumer advertising. The effect on new users also is diminished with the introduction of lagged free-sampling activity. It is also noteworthy that the estimated effect of detailing and free sampling for non-formulary drugs is higher than for formulary medications. This is likely an artifact of the aggregation of the data and suggests that these two forms of advertising may be intensified if a drug is off the formulary, and possibly undermining the formulary efforts. The ability to tailor the detailing and free sampling to individual physicians and the formularies they work with, may suggest why directto-consumer advertising does not appear to undermine the formulary. These results, coupled with the strong state dependence found in the data, present a compelling picture of the pharmaceutical marketing strategy. Given that more than one drug is likely to work for a given patient, pharmaceutical firms aim to have the patients initiate treatment with their brands. Direct-to-consumer advertising plays two roles — it generates foot traffic into doctors’ offices and it has some effect on prescription choice. However, once in the physician’s
3
office, it is the personal sales calls to physicians and the free sampling that have the most significant effect on what drugs are chosen. These preliminary results suggest a change in the way direct-to-consumer advertising should be perceived, at least in the context of the cholesterol-reducing drugs. In particular, the influence of direct-to-consumer advertising on the prescription decision is dwarfed by the influence by physician advertising, contrary to widely held beliefs. But the positive effect of direct-to-consumer advertising, which is undesirable from the perspective of third-party payers, may be partially mitigated by the generated increase in the awareness of the need for lowering elevated cholesterol levels in preventing coronary disease. This kind of cost-benefit analysis of direct-to-consumer advertising is likely to vary across therapeutic categories, however. For example, awareness benefits are far less valuable from the perspective of insurers when it comes to life-style drugs such as hair loss treatments, impotence drugs, and mild allergy problems. In these cases, the insurance-based moral hazard may be enhanced as the patient’s benefits greatly outweigh the insurer’s benefits from drug treatment. Before turning to the details of this research, it is worth considering likely trends directto-consumer advertising and promotions aimed at physicians for the class of drugs considered here. In May 2001, the National Cholesterol Education Program announced new guidelines for cholesterol therapy. These guidelines aim to triple the number of Americans who currently use cholesterol-reducing drug treatment to 36 million, which is more than 18% of American adults. Analysts point out the current market leaders, Lipitor and Zocor, along with a not-yet-launched AstraZeneca’s Crestor, are expected to reap most of the benefits from these new recommendations (Forbes, May 16, 2001). But first, it requires that more patients are diagnosed with cholesterol-levels calling for drug treatment under the new guidelines, which is likely to spur heavy direct-to-consumer advertising.
2. EXISTING RESEARCH Whereas the health industry has been quite attentive to the issue of direct-to-consumer advertising, there is a lack of empirical work on the effects of direct-to-consumer advertising of drugs on physician prescribing behavior (Berndt, 2001). Research has focused on personal sales to physicians, rather than this new marketing strategy. This void is only filled by numerous surveys documenting patient and physician response to advertising. This study bridges a gap in
4
understanding the effects of direct-to-consumer advertising on brand-specific demand for prescription drugs. Empirical work on the competitive effects of pharmaceutical advertising to physicians has focused mainly on the relationship between advertising and new product introductions. The results have been mixed — advertising has been found to have either a positive or negative impact on entry.1 The literature on the effects of detailing, sales calls to physicians, on firmspecific demand is much more limited. Berndt et al. (1994) attempt to distinguish between "industry-expanding" and "rivalrous" marketing efforts. They find that the impact of total marketing on the expansion of overall industry sales declines with the number of products in the market. King (1996) finds that own marketing reduces firms' own price elasticities of demand. He also finds that total industry marketing reduces the degree of product differentiation and raises the cost of entry into the market. Rizzo (1999) similarly finds that both the stocks and flows of detailing expenditures decrease the price elasticity through the development of greater brand loyalty. Manchanda et al. (2000) find that detailing has positive and significant impact on the number of prescriptions written by a physician. They also find diminishing returns to detailing for a majority of physicians in their sample. While there exists a void in the literature with regard to direct-to-consumer advertising of drugs, numerous surveys document the impact of direct-to-consumer advertising on consumer and physician behavior. Wilkes et al. (2000) report that awareness of advertising was strongly associated with having been diagnosed with the advertised condition. They find that better health coverage also increases awareness. In fact, 19% of their respondents who saw an ad actually asked the physician for the drug. A 1998 Prevention magazine survey finds that 29% of consumers who saw a drug ad talked to their physician about it and asked for the drug to be prescribed to them. Coupled with a finding that doctors honored 73% of those consumer requests, the survey estimates that close to 7.5 million consumers in the United States purchase prescription medications specifically because they have been exposed to advertising.2 Market research by Scott-Levin has found that doctor visits for heavily advertised drugs rose, on
1
Older papers (Vernon 1971, Telser et al. 1975, and Leffler 1980) find insignificant or positive impact of advertising on entry. Hurwitz and Caves (1988) find that advertising helped preserve incumbents' market shares. Most recently, Ellison and Ellison (2000) provide evidence that advertising-based entry deterrence appears to be a profitable strategy in medium size pharmaceutical markets only. 2 These findings were reported in the May 1998 issue of Pharmaceutical Executive, pp. 26-30.
5
average, 11% between January and September of 1998, compared to a 2% increase in total office visits (Scott-Levin, 1998).
3. BACKGROUND 3.1 Marketing of prescription medicines Pharmaceutical marketing budgets reach 30 percent of sales, double that of research and development expenditures. These marketing expenditures are largely driven by the large cost of maintaining and utilizing a significant sales force. In 1999, two manufacturers (Pfizer and GlaxoSmithKline) had over 8,000 full-time sales representatives each, and five others had over 4,000 (Financial Times, April 30, 2001). IMS Health (2000b) reports that physician-targeted advertising (in medical journals) and face-to-face selling accounted for $2.7 billion in the first six months of 2000, a 12% rise over first-half 1999 levels. Pharmaceutical companies’ investment in direct-to-consumer advertising reached $1.3 billion, a 44.5% increase over the comparable six-months of the prior year. In addition to these direct promotion expenditures, pharmaceutical manufacturers sponsor conferences, training retreats, and provide physicians with substantial amounts of free samples. The retail value of pharmaceutical sampling — the actual dollar amount of pharmaceutical products distributed to office-based physicians by pharmaceutical sales representatives — totaled a record $3.9 billion for January though June 2000. Direct-to-consumer advertising of drugs is a relatively recent phenomenon as indicated in Figure 1. Until 1997, FDA regulations for pharmaceutical advertising required a summary of contraindications, side effects, effectiveness, and a "fair balance" of risks and benefits. The required "brief summary" often meant the inclusion of one to three additional pages of advertising space; pages which were rarely read or understood by consumers (Wilkes et al., 2000). Moreover, following these rules for broadcast media (radio and TV) was unrealistic. While the industry lobbied for changing the rules, manufacturers experimented with ads that urged consumers to see their doctors if they had a particular health concern, but which did not mention the brand name (e.g., Rogaine's ads were addressed to consumers concerned about hair loss). Alternatively, manufacturers promoted their brands without specifying the indication ("ask your doctor whether brand x is right for you"). This changed in August of 1997 when the FDA temporarily allowed use of broadcast ads without the “brief summary” (with a review scheduled
6
for 18 months later). The discussion leading to this decision focused primarily on safety: whether advertising is leading consumers to improper therapy. However, the argument that the drug would in the end be prescribed by a physician prevailed. These new ads could mention the brand name and treated conditions as long as they included limited in-broadcast disclosure of major side effects and mentioned a supporting national print ad campaign and a website where viewers could turn for more information. These new FDA broadcast rules were made permanent in 1999. Figure 1: Direct-to-consumer advertising over time 1200.00
Expenditure (in $ millions)
1000.00
800.00
600.00
400.00
200.00
0.00 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 Year
total TV advertising
total magazine, newspaper, radio, outdoor
Source: Competitive Media Reporting and IMS Health.
Direct-to-consumer advertising primarily involves products that treat chronic or episodic conditions rather than those that are acute or life-threatening. These products also tend to have widespread incidence. It is heavily concentrated among relatively few drugs — about 50 in 2000. The most heavily advertised therapeutic categories are anti-arthritis medications, cholesterol reducers, antidepressants, antiulcerants, antihistamines, and anti-asthma drugs. While direct-to-consumer advertising still comprises a relatively small, 13% share of the promotional budget (IMS Health, 2000a), it is by far the fastest growing promotional expenditure for pharmaceutical firms and, by 1996, had surpassed professional journal advertising. Within a particular therapeutic category, the newer drugs tend to be the ones that advertise. There are no instances of direct-to-consumer advertising of brands that face same-molecule generic competition.
7
3.2 Pharmacy benefit cost-containment methods On the other side of the table, insurers have been focusing on containing health care expenditures. The pharmacy benefit has recently become a primary target of cost-containment methods since it is the fastest growing health expenditure, with yearly growth rates reaching 20%. Two primary methods involve the formulary and cost-sharing requirements. Insurers develop formularies, lists of preferred prescription drugs, to guide physicians in prescription writing.3 Formulary drugs are selected according to therapeutic value, side effects, and cost. If successfully implemented, formularies can induce price sensitivity by affecting physician behavior. How formularies get implemented varies greatly across insurers. Formularies range in restrictiveness from closed to open. In closed formularies, if a drug is not listed, the insurer will not cover the cost of the drug and therefore the patient will have to foot the entire bill. On the other extreme, in open formularies, all drugs, whether listed on the formulary or not, will be covered by the insurance company. To aid compliance with formulary recommendations in non-closed formularies, insurers may monitor prescription patterns of physicians. Those physicians whose prescribing patterns tend to deviate from the formulary are contacted and reminded of alternative, more cost-effective treatments. In certain cases, prior authorization or additional supporting documentation are required to prescribe an off-formulary drug. Some insurers also use financial incentives to influence physician prescribing. They may reward formulary compliance, generic substitution, meeting predetermined drug costs per patient, etc. (Aventis, 2000). Conversations with industry experts indicate that formulary compliance varies among organizations and tends to be higher in staff Health Maintenance Organizations (HMOs), which hire their physicians and pharmacists, whereas it can be a mere afterthought in more loosely organized Preferred Provider Organizations (PPOs) or Point-of-Service (POS) plans. In particular, physicians who treat PPO and POS patients deal with multiple plans and, therefore, several formularies. One presumes multiple plans make it more challenging for physicians to remember the formulary placement of a given drug (whether it is listed on the formulary) for a given patient.
3
Many third-party payers outsource management of drug benefits to pharmacy benefit managers (PBMs), but the mechanisms that can be utilized for cost-containment remain the same.
8
Cost-sharing requirements mean consumers pay a portion of the cost of each prescription. Presumably the objective is to raise consumer sensitivity to the real cost of their prescription. There are two forms of cost-sharing: copayments and coinsurance. With copayments, consumers pay an established fixed-dollar amount each time they obtain a prescription. The simplest copayment structure is a uniform copayment (e.g., $5) for all prescriptions. Differential brand/generic copayments specify lower amounts for generic drugs and higher copayment amounts for brand name drugs. It is now common for plans to include an additional third tier of copayment to differentiate the formulary and non-formulary drugs (the data used here has this kind of three-tier structure). The third tier copayment is often substantially higher than the copayment for branded, formulary-based drugs. As such, three-tier copayments can indirectly support the implementation of the formulary. Finally, under coinsurance, consumers pay a percentage of the cost of each prescription dispensed. That percentage typically does not vary by the type of drug dispensed (brand name, generic, or formulary). Coinsurance has been a traditional feature of indemnity-type insurance programs and is often combined with a deductible.
3.3 Cholesterol-lowering drugs This paper focuses on the effects of direct-to-consumer advertising of drugs to treat hyperlipidemia (high cholesterol). In 2000, this therapeutic category ranked first in total sales ($9 billion), second in the number of dispensed prescriptions (96 million), and is among the most heavily advertised to consumers. This category lends itself well to the study of physicians’ therapy decisions, as no over-the-counter medications are available to treat cholesterol. Medical research has associated elevated serum cholesterol levels with coronary heart disease, the most common cause of death in the United States. The American Heart Association Science Advisory and Coordinating committee has developed treatment guidelines for patients with lipid abnormalities. The recommendations relevant for this paper were issued in June 1993, almost three years prior to the studied period. Based on these guidelines, an estimated 13 million adults need lipid-lowering drugs to meet recommended goals for low-density lipoprotein (LDL) levels (Sempos et al., 1993). A certain level of cholesterol is necessary for the production of hormones and Vitamin D. Serum cholesterol levels are determined by the amounts of dietary cholesterol intake and by the cholesterol produced in the liver; thus diet and exercise alone may
9
not be sufficient to lower serum cholesterol levels if there is significant overproduction of cholesterol. Various drug treatments attempt to block particular steps in the body’s synthesis of cholesterol. Drugs that target the earlier stages of synthesis can more effectively lower the production of cholesterol.4 There are three components that make up total cholesterol: lowdensity lipoprotein (LDL or “bad” cholesterol), high-density lipoprotein (HDL or “good” cholesterol) and triglycerides (TG). Several subcategories of drugs treat hyperlipidemia: HMG-CoA reductase inhibitors (often called statins), bile acid sequestrants, fibric acid derivatives, and, in a class of its own, niacin. Statins are the newest and, already, the most commonly prescribed lipid-lowering agents. They are generally effective and are supported by favorable outcome studies. There are no clinically appreciable differences in the safety profiles across these drugs and they have similar side-effect profiles. Currently, there are six statins available: Baycol, Lescol, Lipitor, Mevacor, Pravachol, and Zocor, all under patent protection. As a group, statins decrease total and LDL cholesterol levels. Although all of the statins decrease triglyceride levels, not all are labeled by the FDA for this use. All statins have a minimal effect in raising HDL levels, but, again, not all are labeled for this indication. At least two studies (Pfizer funded) that compared all the statins (except Baycol) have shown Lipitor to be the most effective in reducing LDL levels. However, unlike the more extensively studied Pravachol and Zocor, Lipitor has not been proven to reduce total morbidity and mortality. Furthermore, even if Lipitor truly is the most “powerful,” that does not necessarily make it the best choice for every patient. People with mildly or moderately elevated cholesterol might just as easily reach their target levels with any of the statins. Table 1 summarizes basic information about these statin drugs. Table 1: Characteristics of statins Brand name Baycol Lescol Lipitor Mevacor Pravachol Zocor
Molecule name cerivastatin fluvastatin atorvastatin lovastatin pravastatin simvastatin
Marketed by Bayer Novartis Pfizer Merck Bristol-Myers Merck
Available since 1/1998 4/1994 1/1997 9/1987 11/1991 2/1992
Approved by FDA for* ↓LDL, ↓TG ↓LDL ↓LDL, ↓TG, ↑HDL ↓LDL ↓LDL ↓LDL, ↓TG, ↑HDL
↓=decreases, ↑=increases
* The lack of approval by the FDA for a particular problem does not preclude the drug from being prescribed for that problem. 4
Source: personal communication with Gabriela G. Loots, Ph.D., Lawrence Berkeley Laboratories.
10
In contrast, non-statin drugs are older and, in general, less effective in lowering LDL levels. Niacin is the oldest lipid-lowering agent and, until the recent introduction of extendedrelease tablets, it has been plagued by low compliance rates because of side effects. Fibric acid derivatives are used primarily to treat hypertriglyceridemia. They, however, are often associated with gastrointestinal intolerance and other side effects. A similar side effect profile affects bile acid sequestrant drugs. Given that these drugs have been on the market longer than the statins, both generic and branded versions of these drugs exist. It should also be mentioned that combination (often a statin with a non-statin) regimens can be considered for use in patients who fail to meet target cholesterol levels with one drug (but is recommended only for patients that are compliant with their current therapy). Table 2: Characteristics of primary non-statins Brand name Niaspan Lopid Tricor LoCholest Colestid
Generic name niacin gemfibrozil fenofibrate cholestyramine colestipol
Drug type niacin fibrate fibrate bile acid bile acid
Approved by FDA for ↓LDL, ↓TG ↓TG ↓TG ↓LDL ↓LDL
4. CONCEPTUAL FRAMEWORK Typically, the choice, payment, and consumption of a product fall to the consumer. However, in the pharmaceutical market, three distinct parties are involved: patients, physicians and insurers. Patients entrust the prescription decision to the physician because they lack the knowledge needed to choose the proper treatment. It should be noted that this agency is legislated by giving physicians an exclusive ability to prescribe certain drugs. Prescription insurance, if used, shields patients from part or all of the cost of the drug. The decision as to which drug should be used remains the physician's; however, his objectives may conflict with the objectives of the patient and the insurance company. The patient seeks to maximize the health outcome but she balances it against the cost of treatment (and thus her cost-sharing arrangement with her health plan). The managed care insurer, on the other hand, has to balance the effectiveness of a treatment against its cost.5 5
The cost of other health expenditures generally increases with lower quality of pharmaceutical therapy. Pharmaceuticals are often the cheaper preventive option relative to other forms of treatment such as surgeries, which
11
When choosing a prescription, the physician’s primary objective is, presumably, maximizing the health outcome for his patient as a function of drug and patient characteristics. Yet, the perceived costs and benefits of various drugs are likely to vary across physicians, even for the same patient. It is highly relevant that prescription drugs have both experience and credence good characteristics. The patient and her physician rarely know what drug she will respond well to or what drug will induce side effects, unless she actually tries it. This underscores the importance of the patient’s drug treatment history and potentially of the initial choice of therapy. But even after discovering that there are no appreciable side effects, and that the drug accomplishes therapy targets (such as lowering total cholesterol), the true benefits, such as lowered mortality, may be assessed differently across physicians. Finally, just as physicians learn what works for a particular patient, they use this information to form beliefs about what is likely to work well for other patients. They also form beliefs about drug characteristics from the medical literature, their colleagues, pharmaceutical literature and sales people. The physician never incurs expenditures for the prescribed drug and therefore price sensitivity is strictly a result of influence by the patient and the insurance company. This indirect price sensitivity can be divided into two parts, referred to here as residual consumer price sensitivity and hassle costs. The residual consumer price sensitivity results from the physician’s desire that his patient follows the prescribed therapy. However, the consumer's full price sensitivity may not be completely internalized by the physician. Hassle costs, on the other hand, are the mechanism through which the patient and insurer attempt to control physician behavior. Patients may directly hassle their physicians to prescribe a particular drug they consider the best choice. Most relevant for this study, they may pressure their physicians to choose an advertised drug over a non-advertised drug. Insurers can choose from a more complex set of incentive mechanisms for implementing the formulary described in Section 3.2, such as prior authorization or supporting documentation requirements. Physicians are likely to act in ways to minimize these hassles. Alternatively, insurers may tie the copayment structure to the formulary status of a drug, and therefore attempt to induce price sensitivity through the patient. The above discussion can be succinctly summarized in the following indirect utility U that physician j derives from prescribing medication m to patient i on occasion t:
are capital and labor intensive. By trying to mitigate the incentives within the prescription insurance alone through the introduction of risk sharing, there is a potential spillover effect to other health expenditures.
12
Uijmt = αm + β1 DOCmt + β2 HISTimt
perceived quality of health outcome
+ CPimt (β3 + β4 DTCmt)
residual consumer price sensitivity
+ β5 DTCmt + OFFmt (β6 + β7 DTCmt + β8 DOCmt)
hassle costs
+ εijmt The parameter αm is an alternative specific constant that captures unobserved, timeinvariant characteristics, and presumably quality of the drug. The term DOCmt is the aggregate physician advertising of brand m during time period t, and HISTimt is a drug-therapy history construct for patient i at time t. These variables, as well as direct-to-consumer advertising can be constructed in various ways. While it is likely that current and past advertising matter, I initially conduct my analysis with current advertising expenditures. Some very preliminary results from estimating the wear-out rate of advertising are also presented. I also use a simple history variable indicating whether the last prescription the patient received was for the same drug. More complex history constructs will be explored in future work. The variable CPimt is patient i’s copayment for medication m prescribed on occasion t. As described in Section 3.2, the copayment may be a function of the retail pharmacy price (coinsurance) or a function of formulary status of a drug (copayment). In the latter case, the patient price does not vary with the price of individual drugs, but rather with their formulary/generic status. The hassle cost constructs are as follows: DTCmt is direct-to-consumer advertising of brand m during the time period t and is constructed similarly as physician advertising, first as current level, and then as a discounted sum of current and past advertising. The variable OFFmt is an indicator variable equal to 1 if drug m is off the formulary at time t. The price and formulary effects may not be identified separately if copayment structures are a function of the formulary status, unless there is sufficient variation in patient prices over time and/or across patients.6 Consumer and physician advertising might influence the ability to implement the formulary. This is captured here by two interaction terms of the off-formulary constant (OFFmt) with direct-to-consumer advertising (DTCmt) and with detailing (DOCmt). Finally, the stochastic component of utility that is known to the physician, but not the econometrician, is captured by εijmt. 6
Note that all Blue Shield patients face the same formulary.
13
This utility maximizing model lends itself well to logit-based choice models if we assume that the
ijmt
stochastic error term is an i.i.d. extreme value random variate independent of .
Assuming that physicians choose the alternative that yields the highest utility, the probability of choice for alternative k by physician j for patient i from a choice set J at time t: exp(x ikt . ∑ exp(x imt
Prob ijt (k) =
m∈ J t
where ximt=[1, DOCmt, HISTimt ,…, OFFmt*DOCmt] and
> 1« 8]’.
A general model allows for tastes to vary in the population. In other words, each decision makers’ preference coefficients, , can be represented as the average preference in the population and the individual’s deviation from that mean: =
ij + β
.7 However, as such this probability
cannot be computed because we do not have enough information to separately estimate the random coefficients for all the patients. Instead, we may choose to assume that there is no heterogeneity of tastes, which results in the standard multinomial logit analysis. Alternatively, we can integrate out the unknown individual deviations.8 Let function f(β_ describe the density with which tastes vary in the population, where are the parameters of this multivariate distribution such as the mean and standard deviation of tastes in the population. If we assume that the variate independent of
ij,
ijmt
stochastic error term is an i.i.d. extreme value random
then the choice probability that physician j will choose drug k (out of a
set of J drugs at time t) for patient i is modified to: Prob ijt (k) = ∫
exp(x ikt
∑ exp( x
ij
)
imt
ij
)
f( | G .
m∈J t
ijmt,
Assuming independence in
the probability of each physician making a sequence of
prescription decisions for each patient, conditional on βij, is then S ij (
ij ) = ∏ t
exp(xikt
∑ exp(x
ij
imt
) ij
)
.
m∈J t
Since the values of βij are not known, the actual probability is the integral over the sequence Sij(βij) over all values of βij: 7
Note that the individual deviations do not have a time subscript. By fixing the random coefficient to be constant for each patient, patient-specific time correlation is generated. 8 The discussion of mixed logit borrows heavily from Revelt and Train (1998).
Probij Sij(βij) f(β_ Gβ. The log likelihood function is //
ij
ln(Probij but since Probij does not have a
closed form, it has to be approximated numerically through simulation (see e.g., Hajivassiliou, 1993, and Hajivasiliou and Ruud, 1994). For a given value of the parameters , a value of drawn from its distribution. Using this draw of
ij,
ij
is
the probability of a sequence of decisions,
Sij(βij), is calculated.9 This process is repeated for many draws to calculate the simulated probability of a sequence of choices made by physician j for consumer i. The average of the resulting Sij(βij) is calculated and taken as the approximate choice probability Probij : 1 S ij ( ijr| ) ∑ R r =1,...,R where R is the number of draws of SPij (
ij
)=
ij
and
ij
U_
is the rth draw from f( _ We can then estimate
parameters that maximize the simulated log-likelihood function defined as 6//
ij ln(SPij
5. DATA The primary data come from insurance prescription claims under Blue Shield of California PPO plans. These insurance claims are for the individual and family plan as well as employer group plans for the time period 1996 through 1999. During this time period, 38,358 patients filled a total of 349,129 prescriptions for cholesterol reducing drugs, of which 224,676 were refills. In its PPO plans, Blue Shield utilizes virtually no direct methods of formulary implementation. Physicians receive a quarterly newsletter from Blue Shield’s Pharmacy and Therapeutics Committee with updated formulary information. Instead, BSC relies on the physician’s sensitivity to the patient’s price, and links copayments with the formulary status of drugs. These copayments vary among employers, depending on the contracted benefits package and premiums. The observed prescription claims are a result of a series of decisions by the physician and the patient. First, the physician chooses the form of therapy. Next, the patient chooses whether to fill the prescription, what type of pharmacy to use (mail-in or retail) and finally whether to pick up the prescription. Presumably, the physician’s decision depends on the patient’s fill
9
I actually calculate the sequence of probabilities for each patient, rather than each patient-physician. Therefore, the analysis I conduct really allows for patient specific i random coefficients.
decision, a hypothesis reflected by allowing for price sensitivity in physician’s utility function.10 However, I make an assumption that the physician’s choice of brand does not depend on the patient’s choice of pharmacy. This is a reasonable assumption since the same drugs are available in retail and mail-in pharmacies, and the ranking of prices is the same. This assumption allows me to use the retail pharmacy copayments only in analyzing the physician’s choice of drug.11 There are several other characteristics of claims data that are relevant. First, only the copayment for the chosen alternative is present, so the patient prices for non-chosen alternatives have to be imputed (discussed below). In addition, claims data restrict the analysis to be conditional on choosing a drug therapy, and thus may not capture the full effect that advertising has on demand. In particular, the effect of direct-to-consumer advertising on demand will be underestimated if advertising has a positive effect on the prescription decision for all drugs. Furthermore, I cannot distinguish between prescriptions written with and without a doctor visit (the latter called in by the pharmacist). Finally, claims include minimal demographic information: age and gender only. Until 1997, Blue Shield’s PPO patients paid a lower copayment for generic drugs and a higher one for branded drugs. Then throughout 1997, Blue Shield began phasing in the three-tier cost-sharing structure: the lowest copayment for is generic drugs, a higher one for branded drugs that are on the formulary, and the highest one for off-formulary branded drugs. The actual copayments vary by employer. I construct the prices (copayments) faced by the patient for nonchosen alternatives using all the claims submitted by the patient and his family. First, using family data in a given month, I identify unique copayments for the three drug groups. This way I can impute the benefits in the following percentage of family-months: 27% for generics, 40% for formulary brand name drugs, and 17% for non-formulary brand-name drugs. These family time series can then be filled in using change-of-benefits dates. This requires one prescription of each of three types between two change-of-benefits dates to construct the complete benefits structure. At this point, however, the change-of-benefits data are not yet available to me. Therefore, I fill in family copayment time series using the monthly data imputed in the first stage. More specifically, I look at runs of missing copayments for each family’s time series and 10
I do not observe if a patient chose not to submit the initial prescription or the refill to a pharmacy. However, I do observe whether the patient filled but did not pick up the prescription. I delete these claims from the analysis. 11 Mail pharmacy copayments are higher per prescription, but lower per daily dosage. The most commonly used BSC mail copayment is a three-month supply (instead of one month) for double the price to the patient.
16
fill in these missing values with the end values for the run if the end values are equal to each other. This method is inferior to using change-of-benefits data as it does not allow for imputing any missing values at the beginning or the end of time series, as well as for any missing runs where the end values are not equal. Using this method, I am able to complete 60% of familymonths for generics, 70% for brand name formulary-based drugs, but only 30% for nonformulary drugs. I am then able to construct a sample of 4,728 hyperlipidemic patients (with 11,520 prescription decisions, not including refills), for whom these two steps generate100% complete benefits data. The choice set constructed for this study also requires some explanation. I consider the choice of drug, rather than the choice of a drug of a specific dosage. While the choice of strength is an important part of the physician’s decision, the price sensitivity mechanisms (formulary and copayments) do not vary across dosages and advertising is not dosage specific. Instead, the dosage considerations will be a factor of the severity of the patient’s hyperlipidemia. However, I do break the Zocor alternative into two separate alternatives because the 5mg dose of Zocor was being taken on and off the formulary, while the Zocor dosages of 10mg and higher were on the Blue Shield formulary throughout the studied period (see Table 3 for more information).12 Besides the six statins, there were a total of 21 various non-statin cholesterollowering drugs utilized during the study period. Some of these drugs were used less than a dozen times, while several had utilization rates comparable to those of less utilized statins. In a cross-sectional analysis, it would be advisable to exclude the rarely utilized drugs, however in a panel setting, it is not obvious what to do with patients that have switched between these and more heavily utilized drugs. While the treatment of non-statin drugs can be modified, here I create three price-based alternatives that incorporate these drugs: generic non-statins, brand name formulary non-statins, and brand name non-formulary non-statins. This results in a choice set of ten alternatives: seven statins and three non-statins. Table 3 provides more specific information on the formulary status for each of these alternatives. The prescription data are matched with monthly, brand-level direct-to-consumer advertising expenditures compiled by Competitive Media Resources. These data are broken down into ten media categories such as magazine, newspaper, and various forms of TV, radio 12
No statin tablets are scored. While some higher dosage tables can be cut in half accurately with a pill splitter, the 10-mg tablet of Zocor cannot be halved easily or consistently (Crouch, 2001).
17
and outdoor advertising. Here I use total dollar expenditures without disentangling different media impacts. This appears to be a reasonable assumption because the potentially more effective TV advertising is also more expensive, but future work will address the validity of this assumption. Direct-to-consumer advertising is national in nature, but the actual exposure may exhibit regional variation because of heterogeneity in consumer characteristics such as TV viewing versus magazine reading habits. Because advertising is not observed on an individual level, the direct-to-consumer advertising coefficients reflect the patient's exposure to advertising and her responsiveness to the ads, which she has seen. The separation of these two effects, exposure and responsiveness, is not important to the paper, however lack of individual-level advertising data lowers the dimensions that allow for identification of the advertising coefficients. Brand-level monthly medical journal advertising and office promotions from IMS Health are used to control for promotions aimed directly at physicians. These large national survey data include total minutes (the sum of what the physician perceives as minutes for various contacts with the detail person), total dollars, percentage of visits with a free sample and the amount of free samples left (or mailed in), and the percentage minutes by the type visit (full discussion, brief contacts, educational-onsite, service visits, telephone and other). In contrast to direct-toconsumer advertising, using aggregate direct-to-physician advertising (except for journal advertising) is problematic because pharmaceutical companies can tailor their detailing to the formulary of a particular insurer. If a particular drug is not on the formulary, a salesperson may spend more time discussing that drug and may be more likely to provide free samples. If that is the case then direct-to-physician spending measures will be overstated for formulary drugs and understated for non-formulary drugs. This effect can, however, be captured by brand fixed effects and the interaction term of physician advertising with the off-formulary indicator.
6. RESULTS The results presented in this section rely on a smaller sample of 4,728 patients (11,520 prescriptions) for whom the full copayment structure could be constructed without the knowledge of change-of-benefits dates. It should be pointed out that patients included in the sub-sample are more likely to have been in the panel for a short time, likely face fewer changes in their benefits, or more likely have used non-formulary drugs. Some basic comparisons across
18
the full sample and the sub-sample, presented in Figures 2 and 4, are briefly discussed below. Moreover, the data set does not include refills as the decision to refill a prescription belongs to the patient, rather than physician. Finally, because I construct several variables based on the patient’s history, I discard the first four months of the panel. I also delete the last three months of the panel because not all prescriptions had been reported by pharmacies by the end of the year.
6.1 Trends Before turning to estimation of choice probabilities, it may be instructive to describe the trends in the prescribing patterns for the patients in the sample. Figures 2A and 2B indicate that the utilization of cholesterol-lowering drugs among Blue Shield patients is increasing rapidly. Figures 4A and 4B present market shares (calculated as a percentage of total prescriptions) for the sub-sample. These indicate that Lipitor has largely driven the increased utilization of cholesterol-reducing drugs seen in Figures 2A and 2B. Also note the increased Pravachol market share in the sub-sample. Mevacor has had the most significant drop in the market share. This is driven by the fact that Merck, the manufacturer of Mevacor, has been phasing out that brand and increasingly promoting its other brand, Zocor. Tables 4A and 4B show the incidence of switching behavior among patients. It is apparent that a very high percentage of patients stay with only one therapy, a pattern that will be confirmed in the regression analysis. The bias towards Pravachol is also apparent. Not displayed in the table is the fact that close to 50% of patients only have one prescription.
6.2 Analysis using direct-to-consumer advertising promotions Prior to estimating the complete behavioral model presented in Section 4, I restrict my analysis to those promotional activities that are aimed directly at consumers. With that I determine whether the aggregate correlation between sales and direct-to-consumer advertising is reflected in the data. These results are compared later with those that incorporate the physicianspecific promotional methods. At this stage, the history of treatment is captured by an indicator variable (LASTUSED), which is one if the last prescription was for the same brand. Current direct-to-consumer advertising is used as an explanatory variable. The addition of lagged consumer advertising expenditures presents similar but weaker impact on choice probabilities. However, an addition of any lag to the full model (Regressions 4 through 6) requires the
19
estimation of four additional variables. Section 6.4 discusses how a weighted cumulative sum of advertising can be incorporated, thus adding only one parameter to the model.13 Regression 1 is the baseline mixed logit specification in which I allow for normally distributed random coefficients on price.14 The estimates on variables other than the copay are similar to those in the multinomial logit, but the addition of the random copay effect improves the log-likelihood from -14,131 to -14,115. Direct-to-consumer advertising is found to have a significant positive impact on choice probability, however this varies by formulary status. In particular, the effect of direct-to-consumer advertising is estimated to be almost three times as large for formulary drugs than for those drugs that are off the Blue Shield formulary (0.027 versus the net coefficient of 0.01). One possibility for this differential effect is that the advertising strategies for drugs that happen to be off the Blue Shield formulary are not as effective as for the drugs that are on the Blue Shield formulary. What is more likely the case, the physician will suggest an advertised formulary drug (Lescol, Lipitor or Zocor) when a patient asks for the off-formulary Pravachol. This suggests that direct-to-consumer advertising might generate foot traffic into physician offices, however the formulary might play a role which drugs are prescribed. Note also that the off-formulary constant is highly significant supporting this hypothesis.15 While the off-formulary constant is the primary source of copayment variation, the lowered effectiveness of direct-to-consumer advertising across formulary status appears to be driven by non-price factors. More specifically, when an interaction term of direct-to-consumer advertising and patient copay is included in the specification (not presented), the estimated effect is virtually zero and the estimate of the effect of direct-to-consumer advertising for nonformulary drugs is unaffected. The log-likelihood ratio test cannot reject the hypothesis that both models are equivalent. This suggests that direct-to-consumer advertising does not modify the patients’ price sensitivity. 13
Because current advertising is correlated with past advertising, it will be overestimated, as it also proxies for past advertising. Therefore, the interpretation of this coefficient should not be limited to current advertising only. 14 A similar model with random effects on direct-to-consumer advertising yields estimates of standard deviations that are not statistically significant from zero. The log likelihood ratio test reject such a model is an improvement over the one with copay random effects only. 15 Note that because of the alternative-specific constants, the off-formulary constant is only identified by the first four months after Lipitor’s introduction (before it is added to the formulary) and by the months where the 5mg Zocor was taken off the formulary. On the other hand, the coefficient on the interaction term of direct-to-consumer advertising with off-formulary indicator is identified by all drugs that are off the formulary and advertise (thus including Pravachol).
20
The baseline model in Regression 1 also presents interesting results with respect to the price variable (copayment). While the mean of price does not significantly change from that estimated with a multinomial logit (results not presented), the estimate of the standard deviation is surprising. A standard deviation of 0.04, with a mean of -0.005 (imprecisely estimated), implies that 45% of patients have positive price coefficients. This might be driven by the fact that I am restricting the distribution of random coefficients to be symmetric around the mean. However, this pattern may be indicative of price signaling quality. In particular, it is noteworthy that when I control for quality by incorporating brand random effects, the estimate of standard deviation on price decreases substantially, while the magnitude of the mean increases. The remaining regressions in Table 5 incorporate a mixed logit specification that allows for random coefficients on alternative specific constants. Identification of such a model requires the scale of the model to be set, i.e. one of the alternative specific constants needs to be normalized, just as in the multinomial logit. In addition, there is the issue of normalizing one of the variance parameters. A recent paper by Ben-Akiva et al. (2001) shows that this normalization is not arbitrary. They recommend normalizing the lowest variance parameter.16 Applying this strategy to my data reveals that Lipitor is the alternative with the lowest variance. This is problematic because it is not in the choice set at all times. The next lowest variance was that of generic non-statins. Therefore instead of excluding the generic constant (which is equivalent to setting the mean and variance to zero), I fix its mean to zero and its standard deviation to one (thus not forcing the Lipitor variance below zero). The addition of random brand effects allows me to account for heterogeneity in preferences over the brand specific quality and to obtain a cleaner estimate of the level of state dependence.17 The estimated coefficient on the treatment history variable (LASTUSED) in Regression 1 is overstated because I am not properly controlling for consumer heterogeneity in this specification, and thus I am confounding it there with state dependence. With the addition of brand random effects, the coefficient on LASTUSED decreases by about 25% as the random effects on brand dummies now incorporate the time invariant, patient-specific mean utilities. 16
Following the recommendations of that paper, estimating a model with all variances specified can identify the lowest variance parameter. This model “will pseudo-converge to a point that reflects the true covariance structure of the model. The heteroskedastic term with minimum estimated variance in the unidentified model is the minimum variance alternative” (Ben-Akiva, Bolduc and Walker, 2001). 17 More specifically, the random effects on alternative-specific constants properly account for heterogeneity only if the random coefficients are fixed across choice situations for each person, which is the approach I take.
21
The addition of the 9 brand random effect parameters improves the log likelihood by 308, which is a significant improvement of the model as suggested by the log likelihood ratio test. The addition of these brand variances has a significant impact on the copay and advertising variables. First, note that the estimate of the copayment mean practically doubles in absolute value, while the estimate for its standard deviation decreases significantly. Next, the estimated effects of direct-to-consumer advertising change with the addition of brand-specific random coefficients. Because the latter model better separates true state dependence from spurious state dependence, this suggests that the effectiveness of direct-to-consumer advertising may be modified by the patient’s experience. Regressions 3 and 4 consider the possibility that the effect of direct-to-consumer advertising is modified by state dependence. Specifically, I consider the possibility that the effect of direct-to-consumer advertising varies between the first choice decision, when the patient has no experience with any cholesterol drugs, and subsequent decisions. I construct an indicator variable, NEWUSER, equal to 1 for the first choice situation for each patient in the panel. Because this variable is incorrectly overstated for many patients at the beginning of the panel, I exclude the first 4 months from the analysis. Still, this variable will be overstated for patients who have had drug therapy under a different insurance plan or who have discontinued therapy before 1996 and then returned to it after April 1996. Keeping this in mind, I add the interaction of new user with direct-to-consumer advertising. The net effect of advertising for new users is more than twice as large than for subsequent decisions (0.023+0.032=0.055 versus 0.023). The effect of direct-to-consumer advertising of non-formulary drugs in the first prescription decision is 0.023+0.032-0.024=0.031. This effect is significant and larger than for the formulary-based drugs in subsequent decisions. However, the effect of advertising on prescription choice for the subsequent decisions for non-formulary drugs is virtually zero (0.0230.024= -0.01). The specification in Regression 5 attempts to explore further the interaction of advertising, formulary status, and first choice situation. The resulting estimates indicate that the coefficient on the interaction term between direct-to-consumer advertising and new user status in Regression 3 was primarily driven by users of non-formulary drugs. Specifically, the estimate of the marginal effect for new users of formulary drugs is insignificant from zero, while the analogous estimate for new users for non-formulary drugs is highly significant and positive. The
22
net effect of advertising in the first choice decision for off-formulary drug is then 0.051 (0.0460.073-0.015+0.093), while the net effect for subsequent decisions is negative (0.046-0.073= -0.027). This large difference in the effectiveness of advertising for formulary drugs requires further investigation.
6.3 Analysis using consumer and physician promotions Now I consider the addition of physician advertising — the element missing from other analyses of the impact of direct-to-consumer advertising on demand. Table 6 presents some general statistics on physician advertising and free sampling. It is apparent that the three major brands that advertise heavily also promote heavily to physicians. However, some differences among the top brands exist. Namely, Pravachol has relied more heavily on direct-to-consumer advertising relative to detailing and free sampling than has Zocor or Lipitor. Regressions 5 and 6 in Table 5 are the result of incorporating promotional activities directed at physicians. The addition of detailing expenditures in Regression 5 lowers the estimate of the main effect of direct-to-consumer advertising. A comparison of Regressions 4 and 5 reveal that the main direct-to-consumer estimate is significantly lowered (0.046 versus 0.03) when I control for physician advertising. Additionally, the main effect of physician advertising (DOC) is five times as high as the main direct-to-consumer advertising effect, 0.152 versus 0.03. Moreover, the addition of lagged free sampling activity in Regression 6 virtually eliminates the significance of new user effects. This suggests the possibility that these “new” users might have been exposed to these drugs through free samples, effectively making them into experienced customers. Notice that the interaction terms of off-formulary constant with detailing and also with lagged free sampling are both significant and positive. This suggests that the detailing and free sampling of off-formulary drugs is more effective than that for formulary drugs. This may be, however, an artifact of the aggregation of the detailing and free sampling data, because pharmaceutical companies can tailor their detailing to the formulary of a particular insurer. If a particular drug is not on the formulary, a salesperson may spend more time discussing that drug and may be more likely to provide free samples, which would be reflected in a positive marginal effect of detailing and free sampling for off-formulary drugs.18 It also suggests that these two 18
Another possibility is that the detailing efforts for the drugs on the Blue Shield are of lower quality.
23
forms of promotions are powerful in their ability to undermine the formulary. This contrasts with the negative sign of the interaction term of off-formulary dummy with direct-to-consumer advertising, which indicates that direct-to-consumer advertising is somehow less effective for non-formulary drugs. This beneficial effect of direct-to-consumer advertising is however overpowered by the presumed increased detailing effort for off-formulary drugs (-0.052 versus 0.187).
6.4 Estimating the wear-out of advertising effectiveness The results presented above incorporate only current advertising expenditures. However, the effects of past advertising are unlikely to dissipate within a month. When I incorporate previous month’s advertising into regressions, I find that it does have a positive and significant effect on choice probability. However, the addition of each additional lag to the MNL specification as in Regression 5.8 requires estimation of four additional parameters. Alternatively, one can use a single advertising variable that is the weighted cumulative sum of past advertising, and estimate the rate at which the advertising impact wears out. One way to structure such a variable is to assume the following form for discounted direct-to-consumer advertising (DDTCt): DDTC t =
n i =0
δ i DTC t −i
It might be reasonable to limit the number of periods used in this construct, for example to the advertising within the last year. Alternatively, I can construct an exponentially smoothed weighted average of past advertising, as in Guadagni and Little (1983). In this case, advertising is a weighted sum of current and past advertising (WDTCt): WDTCt = αDTCt +(1-α)WDTCt-1 = αDTCt +(1-α)[αDTCt-1 + (1-α) WDTCt-2] = …. In this scenario as well, there is a computational advantage to limit the number of time periods incorporated in the weighted sum, so I choose to use a full year of advertising expenditures (current plus eleven lags). Similar constructs would apply to physician advertising. I estimate a separate discount factor on consumer advertising and detailing.19 The estimated discount factors are surprisingly high — they are both precisely estimated to be around 0.96. This would imply that there is no wear-out effect of advertising within a year. Estimation 19
At this point, I have only estimated the discount parameters in a multinomial logit.
24
using the Guadagni and Little construct, similarity suggests that the history of advertising expenditures matter greatly (estimated α=0.13). The advertising coefficients now decrease by a factor of 5 to 10, which is not surprising given that the low wear-out factor effectively implies that a cumulative sum of 12 monthly advertising expenditures is used in estimation. The high discount factors estimates clearly require further investigation. Possibly, the discount factors are overestimated as a result of wrong functional form for the wear-out effect. Suppose that the real effect is such that there is no wear-out in the first three months, and then the effectiveness of advertising dies out completely. In such a scenario, the exponential discount factor I have estimated will over-emphasize the importance of the more distant advertising expenditures in order to provide a better fit for the more recent three months of advertising outlays.
7. CONCLUDING REMARKS This research is an attempt to bridge the gap in understanding the effect of mass media advertising of prescription drugs. The highly visible direct-to-consumer advertising of drugs is often cited an important influence of physician’s choice of drug therapy. The preliminary results suggest that this indeed is the case if physician detailing and free sampling activity are not controlled for. However, the inclusion of these less public forms of promotion lowers the estimates direct-to-consumer advertising has on choice probability. Not only does the effect of consumer advertising become diminished, the estimated impact of detailing is five times as large as that of consumer advertising. The positive effect on new users also becomes diminished with the introduction of lagged free-sampling activity into the model. This paper also highlights the potential role of the formulary in affecting prescription choice. The findings of this paper indicate that while own direct-to-consumer advertising increases the probability that the drug will be chosen, this effect is significantly lower for drugs that are not listed on the Blue Shield formulary. In fact, an increase in direct-to-consumer advertising by an off-formulary drug decreases the probability that it will be chosen by experienced users. This does not necessarily mean that the makers of the drugs not listed on Blue Shield’s formulary are not behaving optimally, because Pravachol and Zocor 5mg are listed on formularies of many other insurance plans. For example, currently Pravachol is listed on 24
25
formularies among a sample of 37 California health plans.20 As a result, the negative impact of the formulary is mitigated across plans. But the role of the formulary is nevertheless of great interest to both sides of the issue — pharmaceutical managers and insurers — as it identifies an important component in affecting prescription choice. The data used here does not provide cross-formulary variation, which means that the differential direct-to-consumer advertising effectiveness for formulary and nonformulary drugs found here may be not be driven by formulary status per se. Instead, the formulary status may be proxying for effects not explicitly captured in the model. For example, patients reacting to direct-to-consumer advertising of Pravachol may be put on Lipitor not because the latter has preferential formulary status, but because Lipitor advertises heavily to physicians. This hypothesis can be tested by including an interaction term between offformulary status, own direct-to-consumer advertising, and detailing expenditures for formularybased drugs. In addition, drug specific advertising coefficients may aid in identifying whether the formulary-advertising effects are driven by particular brands. Finally, a comparison of U.S. national market shares and Blue Shield’s market shares can be used to discern the effect of the formulary. These analyses will be incorporated in future work.
20
Source: Citizens for the Right to Know website (http://ca.mcodrugs.com).
26
8. REFERENCES Aventis Pharmaceuticals (2000), Managed Care Trends Digest, Parsippany, NJ. Ben-Akiva, Moshe, Denis Bolduc, and Joan Walker (2001), “Specification, Identification and Estimation of the Logit Kernel (or Continuous Mixed Logit) Model,” unpublished manuscript. Berndt, Ernst, Linda Bui, David Reily, and Glen Urban (1994), "The Roles of Marketing, Product Quality and Price Competition in the Growth and Composition of the U.S. AntiUlcer Drug Industry," National Bureau of Economic Research, Working Paper No. 4904. Berndt, Ernst (2001), "The U.S. Pharmaceutical Industry: Why Significant Growth in Times of Cost Containment?" Health Affairs, vol. 20(2), pp. 110-114. Crouch, Michael (2001), “Effective Use of Statins to Prevent Coronary Hearth Disease,” American Family Physician, January 15, (www.aafp.org/afp/20010115/309.html). Ellison, Glenn and Sara Fisher Ellison (2000), "Strategic Entry Deterrence and the Behavior of Pharmaceutical Incumbents Prior to Patent Expiration," unpublished manuscript, Massachusetts Institute of Technology, (web.mit.edu/gellison/www/drugs20.pdf). Financial Times (2001) “Relentless Rise in Role of Reps and Big Launches,” April 30. Guadagni, Peter and John Little (1983), "A Logit Model of Brand Choice Calibrated on Scanner Data," Marketing Science, vol. 2(Summer), pp. 203-238. Hajivassiliou, Vassilis (1993) "Simluation Estimation Methods for Limited Dependent Variable Models," in Handbook of Statistics, eds. G. Maddala, C. Rao, H. Vinod; Amsterdam: Elsevier Science Publishers, vol. 11, pp.519-543. Hajivassiliou, Vassilis and Paul Ruud (1994), "Classical Estimation Methods for LDV Models Using Simulation," in Handbook of Econometrics, eds. R. Engle and D. McFadden; Amsterdam: Elsevier Science Publishers, vol. 4. Hurwitz, Mark and Richard Caves (1988), "Persuasion or Information? Promotion and the Shares of Brand Name and Generic Pharmaceuticals," Journal of Law and Economics, vol. 31(2), pp. 299-320. IMS Health (2000a), " IMS Health Reports U.S. Pharmaceutical Promotional Spending Reached Record $13.9 Billion in 1999," http://www.usimshealth.com IMS Health (2000b), “Product Samples Account for Nearly Half of Total Promotional Investment” press release, October 19, http://www.usimshealth.com
27
King, Charles (1996), "Marketing, Product Differentiation, and Competition in the Pharmaceutical Industry," Program on the Pharmaceutical Industry Working Paper No. 39-97, MIT Sloan School of Management. Leffler, Keith (1980), "Persuasion or Information? The Economics on Prescription Drug Advertising," Journal of Law and Economics, vol. 24, pp. 45. Manchanda, Puneet, Pradeep Chintagunta and Susan Gertzis (2000), “Responsiveness of Physician Prescription Behavior to Salesforce Effort: An Individual Level Analysis,” University of Chicago, mimeo. NIHCM (National Institute for Health Care Management) (1999), Factors Affecting the Growth of Prescription Drug Expenditures, (http://www.nihcm.org/FinalText3.pdf). NIHCM (2000), Prescription Drugs and Mass Media Advertising, (http://www.nihcm.org/ DTCbrief.pdf). Revelt David and Kenneth Train (1998), "Mixed Logit with Repeated Choices," Review of Economics and Statistics, Vol. 80(4), pp. 647-657. Rizzo, John (1999), "Advertising and Competition in the Ethical Pharmaceutical Industry: The Case of Antihypertensive Drugs," Journal of Law and Economics, vol. 42(1), pp. 89-116. Scott-Levin Inc. (1998), "Patient Visits Up for DTC Conditions," press release, November 6. Sempos CT, Cleeman JI, Carroll MD, Johnson CL, Bachorik PS, Gordon DJ, et al. (1993), “Prevalence of high blood cholesterol among US adults. An update based on guidelines from the second report of the National Cholesterol Education Program Adult Treatment Panel,” Journal of the American Medical Association, vol. 269, pp. 3009-3014. Tesler, Lester et al. (1975) "The Theory of Supply with Applications to the Ethical Pharmaceutical Industry," Journal of Law and Economics, vol. 18, pp. 449. Vernon, John (1981), "Concentration, Promotion, and Market Share Stability in the Pharmaceutical Industry," Journal of Industrial Economics, vol. 19(3), pp. 246. Wilkes, Michael, Robert Bell, and Richard Kravitz (2000), "Direct-to-Consumer Prescription Drug Advertising: Trends, Impact and Implications," Health Affairs, vol. 19(2), pp. 110128. Wosinska, Marta (2001) “Direct-to-Consumer Advertising and Therapy Compliance,” mimeo.
28
Non-statin
Statin
Table 3: Choice set description 3/96 market share and Rx count *
10/99 market share and Rx count *
0% 0
2.3% 283
Added 4/96
8.5% 298
4.6% 578
Lipitor
Added 4/97
0% 0
53.3% 6665
Mevacor
Removed 3/96
27.8% 977
2.6% 327
Pravachol
OFF
13.9% 488
9.2% 1152
Zocor 5mg
ON in 1996 OFF in 1/97 - 4/98 ON in 4/98
0.4% 120
0.6% 69
Zocor 10mg+
ON
25.3% 888
17.9% 2235
Generic
ON
14.2% 500
5.2% 649
Brand-name
ON
4.4% 152
3.0% 372
Brand-name
OFF
2.7% 88
1.4% 179
Choice
Formulary status
Baycol
Added 10/99
Lescol
* Rx (prescription) count includes refills ** only 1/97 – 12/99 included in the calculation
Figure 2A: Number of Rx and DTC Advertising: Full Sample 14000
30000
25000
10000 20000 8000 15000 6000
Advertising dollars
Number of presctiptions (including refills)
12000
10000 4000
5000
2000
0
0 5 7 9 11 1 3 5 7 9 11 1 3 5 7 9 11 1 3 5 7 9 1996 1996 1996 1996 1997 1997 1997 1997 1997 1997 1998 1998 1998 1998 1998 1998 1999 1999 1999 1999 1999 All Cholesterol Rx
DTC spending
Figure 2B: Number of Rx and DTC Advertising: Sub-sample 1800
30000
1600
1400
1200
20000
1000 15000 800
600
10000
400 5000 200
0
0 5 7 9 11 1 3 5 7 9 11 1 3 5 7 9 11 1 3 5 7 9 1996 1996 1996 1996 1997 1997 1997 1997 1997 1997 1998 1998 1998 1998 1998 1998 1999 1999 1999 1999 1999 All Cholesterol Rx
DTC spending
Advertising dollars
Number of presctiptions (including refills)
25000
Figure 3: Cholesterol drug DTC spending 30,000.00 Lipitor introduced
Baycol introduced
Advertising expenditure (in $000s)
25,000.00
20,000.00
15,000.00
10,000.00
5,000.00
0.00 1 1996
4 1996
7 1996
10 1996
1 1997
4 1997
7 1997 Lescol
10 1997 Lipitor
1 1998
4 1998
Pravachol
7 1998 Zocor
10 1998
1 1999
4 1999
7 1999
10 1999
Figure 4A: Blue Shield market shares: Full Sample 100%
Contribution to total market
80%
60%
40%
20%
0% 5 7 9 11 1 3 5 7 9 11 1 3 5 7 9 11 1 3 5 7 9 1996 1996 1996 1996 1997 1997 1997 1997 1997 1997 1998 1998 1998 1998 1998 1998 1999 1999 1999 1999 1999
Baycol
Lescol
Lipitor
Mevacor
Pravachol
Zocor
Non-statin
Figure 4B: Blue Shield market shares: Sub-sample 100%
Contribution to total market
80%
60%
40%
20%
0% 5 7 9 11 1 3 5 7 9 11 1 3 5 7 9 11 1 3 5 7 9 1996 1996 1996 1996 1997 1997 1997 1997 1997 1997 1998 1998 1998 1998 1998 1998 1999 1999 1999 1999 1999
Baycol
Lescol
Lipitor
Mevacor
Pravachol
Zocor
Non-statin
Table 4A: Occurrence of switching in the full sample (38,558 patients) History Lipitor Zocor Pravachol Lescol fibrite Mevacor bile acid Baycol Zocor5 Niacin Zocor Lipitor Pravachol Lipitor Mevacor Lipitor Lescol Lipitor Pravachol Zocor fibrite Lipitor Mevacor Zocor Lescol Zocor Zocor5 Zocor Lipitor Zocor Lipitor fibrate Pravachol Lescol Zocor Pravachol bile acid Lipitor Lipitor Pravachol Mevacor Zocor All other combinations
Number of patients 11089 5852 3619 2141 2100 1980 924 511 271 262 total = 28,749 1091 868 579 499 448 429 348 150 130 114 104 99 93 84 83 Lipitor 81 4098 total = 9,609
Table 4B: Occurrence of switching in the sub-sample (4728 patients) History Lipitor Pravachol Zocor fibrate Mevacor Lescol bile acid Baycol Zocor5 Niacin Pravachol Lipitor Zocor Lipitor Mevacor Lipitor fibrate Lipitor Pravachol Zocor Lescol Lipitor Mevacor Zocor All other combinations
Number of patients 1202 821 600 365 282 235 184 98 42 35 total= 3,864 113 59 49 46 44 35 28 490 total = 864
Table 5: Mixed logit estimates for prescription drug choice Dependent variable: chosen alternative Sample: 5/96-9/99, 4728 patients, 11520 observations Standard errors are below the estimates 300 random draws are used for mixed logit simulations
variable last used same brand
mean
patient's copayment
mean St dev
off formulary constant (OFF) mean DIRECT-TO-CONSUMER ADVERTISING current direct-to-consumer mean advertising (DTC) DTC * off formulary (OFF) mean new user * DTC
mean
new user * DTC * OFF
mean
Reg 1 3.072 *** 0.014 -0.005 0.003 0.040 *** 0.005 -1.157 *** 0.137
Reg 2 2.228 *** 0.019 -0.009 ** 0.004 0.013 0.010 -1.919 *** 0.194
Reg 3 2.249 *** 0.019 -0.008 ** 0.004 0.013 0.010 -1.883 *** 0.194
Reg 4 2.248 *** 0.020 -0.008 ** 0.004 0.014 0.009 -1.902 *** 0.193
Reg 5 2.226 0.020 -0.010 0.004 0.017 0.009 -2.642 0.342
0.027 *** 0.007 -0.017 * 0.010
0.039 *** 0.009 -0.023 * 0.013
0.023 ** 0.011 -0.024 * 0.013 0.032 *** 0.010
0.046 *** 0.012 -0.073 *** 0.017 -0.015 0.014 0.093 *** 0.019
0.030 ** 0.012 -0.076 *** 0.017 -0.016 0.014 0.098 *** 0.019
PHYSICIAN-DIRECTED PROMOTIONAL VARIABLES current detailing (DOC) mean DOC * OFF
mean
lag free sampling (FREE)
mean
FREE * OFF
mean
*** ** * ***
0.185 *** 0.050 0.199 *** 0.064
Reg 6 2.253 *** 0.018 -0.007 * 0.004 0.005 0.012 -2.745 *** 0.351 0.029 ** 0.013 -0.052 *** 0.019 -0.016 0.018 0.041 0.026 0.152 0.050 0.187 0.065 0.038 0.017 0.171 0.034
***
-1.584 0.828 3.070 0.510 -1.422 0.231 1.568 0.158 1.012 0.257 0.939 0.141 1.548 0.368 1.754 0.151 1.647 0.310 1.996 0.154 -0.357 0.183 1.597 0.130 -6.962 0.969 3.816 0.476 set to 0 set to 1 -1.602 0.239 1.602 0.186 0.731 0.448 1.386 0.191
*
*** ** ***
ALTERNATIVE SPECIFIC CONSTANTS mean Baycol constant
0.643 *** 0.162
St dev mean Lescol constant
-0.308 *** 0.060
St dev mean Lipitor constant
1.538 *** 0.055
St dev mean Mevacor constant
0.980 *** 0.146
St dev mean Pravachol constant
1.889 *** 0.147
St dev mean Zocor 10mg+ constant
0.528 *** 0.059
St dev mean Zocor 5mg constant
-1.465 *** 0.106
St dev generic non-statin constant brand non-statin constant
mean St dev mean
baseline -0.597 *** 0.065
St dev brand nonformulary non-statin
mean St dev
0.034 0.160
-1.224 0.793 3.090 0.508 -1.057 0.188 1.632 0.157 1.901 0.069 1.001 0.134 1.076 0.252 1.697 0.147 2.434 0.218 1.901 0.146 0.159 0.115 1.667 0.129 -7.033 0.983 4.168 0.485 set to 0 set to 1 -1.470 0.235 1.637 0.186 0.026 0.325 1.549 0.180
Log-likelihood -14,115 -13,807 *** significant at 1%, ** significant at 5%, * significant at 10%
*** *** *** *** *** *** *** *** ***
*** *** ***
*** ***
***
-1.255 0.793 3.081 0.507 -1.019 0.185 1.591 0.156 1.888 0.069 0.972 0.135 1.082 0.250 1.647 0.145 2.395 0.218 1.874 0.146 0.150 0.115 1.656 0.129 -7.066 0.996 4.173 0.493 set to 0 set to 1 -1.445 0.234 1.612 0.186 0.011 0.325 1.528 0.181
-13,802
*** *** *** *** *** *** *** *** ***
*** *** ***
*** ***
***
-1.122 0.776 3.003 0.500 -1.031 0.187 1.605 0.157 1.903 0.070 1.007 0.135 1.082 0.251 1.667 0.146 2.376 0.219 1.915 0.149 0.186 0.114 1.632 0.128 -6.961 0.995 4.135 0.493 set to 0 set to 1 -1.454 0.236 1.622 0.187 0.025 0.325 1.534 0.180
-13,790
*** *** *** *** *** *** *** *** ***
*** *** ***
*** ***
***
-1.512 0.827 3.093 0.512 -1.494 0.237 1.658 0.159 1.033 0.257 1.076 0.133 1.542 0.367 1.799 0.153 1.896 0.305 1.895 0.152 -0.322 0.184 1.613 0.131 -7.491 1.003 4.140 0.486 set to 0 set to 1 -1.544 0.236 1.624 0.185 0.753 0.433 1.333 0.183
-13,753
* *** *** *** *** *** *** *** *** *** * *** *** ***
*** *** * ***
-13,735
*** *** *** *** *** *** *** *** *** ** *** *** ***
*** ***
***
Table 6: Promotional expenditures by brand: 1/1996 – 12/1999
Brand
Mean direct-toconsumer advertising
Mean free sampling
Mean detailing
(in $1000)
(in 1000s of pills)*
(in $1000)
0
1,746 (24 months)
2,540 (25 months)
1,505 (48 months)
2,301 (48 months)
5,870 (36 months)
4,896 (36 months)
708 (48 months)
434 (48 months)
5,959 (27 months)
3,159 (48 months)
3,421 (48 months)
3,473 (47 months)
4,234 (48 months)
3,225 (48 months)
0
0
