Brand Loyalty and Learning in Pharmaceutical Demand Chung-Ying Lee National Taiwan University August 15, 2015

Abstract Medicines for chronic conditions like high cholesterol, heart disease, and diabetes are repeatedly used for a long period of time. Consumer dynamics thus plays an important role in the demand for those drugs. The paper estimates a demand model with brand loyalty and learning using data from cholesterol lowering drug markets in the United States. The estimates suggest high switching costs and strong learning effects at the molecule level in the markets. Switching costs raise the predicted probability of choosing the same drugs in a row and learning largely increases patient stickiness to a molecule in the long run. I discuss implications of the estimated state-dependent effects for drug manufacturers, insurance companies, and healthcare policy makers.

1

Introduction

In recent years, branded drug manufacturers are facing generic competition with many of their blockbuster drugs. According to IMS Health, the patents of six of the ten best-selling prescription drugs in the US expired in 2011 and 2012. To retain revenue after patent expiration, branded drug manufacturers have employed several strategies, including pay-for-delay agreements (paying generic companies not to bring lowercost alternatives to market), presentation proliferation (selling a drug in new forms or dosages) and copay coupons (distributing a card directly to patients to lower their

1

out-of-pocket costs). The success of these strategies depends not only on how many patients drug manufacturers can attract today, but also on how many of the patients will stick to the brands tomorrow. Thus, patient stickiness plays an important role in these marketing strategies and understanding the sources of patient stickiness in pharmaceuticals can help manage the programs. This paper estimates the demand for pharmaceuticals by incorporating brand loyalty and learning to study patient stickiness. Using the micro-level databases for cholesterol-lowering drugs, I look at the switch patterns of patients with different lengths of treatment. The data shows inter-molecule switching probabilities are higher in early prescriptions and quickly decrease in a treatment, suggesting the existence of switching costs and learning about molecules. Switching costs make a patient loyal to a drug and lower the probability of choosing a different drug next time.1 On the other hand, patients learn the effects of drugs by taking it repeatedly and are more likely to stop experimenting once they find a drug that works well. I disentangle the two effects in a nested logit model and estimate the effects using patients’ prescription history. The estimates suggest high switching costs and strong learning effects at the molecule level in the US cholesterol lowering drug market. Intermolecule switching costs are larger than intra-molecule switching costs. The effect of learning about molecules gets larger as a patient takes a drug longer. In addition, switching costs largely raise the probability of choosing the same drug for new patients. Both learning and switching costs contribute to experienced patients’ stickiness in the long run. The paper adds to the literature on consumer dynamics in pharmaceuticals by exploiting the rich micro-level data and identifying effects of brand loyalty and learning. Coscelli (2000) uses a prescription level dataset from the Italian markets to study the relationship between probabilities of switching brands and patient and doctor attributes. He shows habit persistence of doctors and patients in prescription choices. Crawford and Shum (2005) use the same dataset to study inter-molecule choice by considering 1

In the paper, I use “switching cost” and “loyalty” interchangeably. Klemperer (1995) argues that brand loyalty can create switching costs. Studying the actual sources of switching costs in pharmaceuticals is beyond the scope of this paper.

2

consumer learning and patient heterogeneity. However, the lack of price variation in Italian markets makes it impossible to identify the switching cost in their structural analysis. My datasets from the US pharmaceutical markets can overcome this problem since drug manufacturers are allowed to compete in price in US. Moreover, my datasets have information on insurance plan design and this reveals variation in out-of-pocket costs for drugs with different status in an insurance plan. The price faced by a patient is her cost share rather than the full price. With the information on out-of-pocket costs, I can more adequately estimate the price elasticities. Dalen et al. (2011) and Lundin (2000) estimate binary choice models based on prescription level data with price information. They find price difference between branded drugs and generics an important factor in generic substitution. Both have limitations on the use of patients’ prescription history. Dalen et al. (2011) ignores past use and Lundin (2000) only considers switching cost by including the last prescription for each observation in the demand model. My paper takes into account both switching costs, which incur when patient chooses a drug different than the choice last time, and learning, which is revealed in the choice of the same drugs multiple times. Furthermore, I consider learning at the molecule level and the version (branded/generic) level. Incorporating learning about the generic version of a molecule in the demand model can help understand intra-molecule switching after patent expiration. Finally, the paper also contributes to the empirical study of consumer brand loyalty in the marketing literature. Past studies in this field (Krishnamurthi and Raj, 1991; Erdem, 1996; Allenby and Lenk, 1995; Keane, 1997; Dubé, Hitsch and Rossi, 2010; Osborne, 2011; Bronnenberg, Dubé and Gentzkow, 2012) focus on brand loyalty and state dependence in consumer goods. My paper looks at pharmaceuticals, which have more than 300 billion US dollar sales in a year. Consumer dynamics in this industry can be of interest to managers and policy makers. For example, drug manufacturers may want to know how large switching costs are for patients to try a drug when they are marketing a new product. Policy makers may be interested in the learning process in a treatment, which can provide directions to improve patient compliance or medication adherence.

3

The rest of the paper is organized as follows. Section 2 describes data and presents relevant statistics and trends. Section 3 develops models for demand. Section 4 discusses estimation strategies and results. Section 5 concludes.

2

Data

The data are obtained from the MarketScan Commercial Claims and Encounters Database and the MarketScan Benefit Plan Design Database through National Bureau of Economic Research (NBER). The MarketScan Databases are constructed from privately insured paid medical and prescription drug claims. There are about 100 payers and more than 500 million claim records in the Databases. The Commercial Claims and Encounters Database provides the prescription-level data on date of service, drug characteristics, days of supply, full price, out-of-pocket cost, patient age, and patient gender. The “plan key” variable in the data can be used to link the Commercial Claims and Encounters Database to the Benefit Plan Design Database, which provides the health insurance information for the patient, including copayment and coinsurance for drugs in different tiers. My work focuses on the market of cholesterol-lowering drugs (statins) from 2005 to 2006. There are several reasons to analyze the market for the years. First, cholesterollowering drug is the largest therapeutic class in the US by spending in 2006 and had approximately 22 billion US dollar sales and 210 million prescriptions in the year.2 Cholesterol-lowering drugs can help lower rates of low-density lipoprotein (LDL) cholesterol in the blood and are usually taken repeatedly, or even for life, which gives the opportunity to observe the choices a patient made for several periods. This enables me to examine the dynamic aspect of pharmaceutical demand. Second, one of the bestselling drugs in this class, Zocor, lost its patent protection in mid 2006. With generic entry after patent expiration, I am able to see how patients respond to the availability of low-cost alternatives. Third, the drug manufacturer Pfizer employed aggressive strategies in this market, including pay-for-delay agreements and copay coupons, to 2

The Use of Medicines in the United States: Review of 2010, IMS Institute.

4

retain revenue from Lipitor after its patent expiration in 2011. Thus, understanding consumer switch between branded drugs and generics in this market can shed light on the possible outcomes of Pfizer’s strategies. I take three major steps to prepare the data to be used in this paper. First, I keep individuals who have at least one prescription recorded in both 2005 and 2006 to make sure the individuals in the sample are enrolled in both years. Second, I do not observe when a patient started to take the drugs. To avoid difficulty associated with left-censoring, I include only patients with their first prescriptions observed after June 30, 2005. By doing so, I assume that patients who haven’t took drugs for more than half a year are new in the market.3 Third, since I only observe the out-of-pocket cost for the drug purchased, I calculate the out-of-pocket costs for all available drugs faced by a patient by taking the average out-of-pocket cost for each drug across prescriptions belonging to the patient’s insurance plan. For some small insurance plans, there are no prescriptions for certain drugs and thus the average out-of-pocket costs for those drugs in the plan are not available. I drop patients in the plans with missing out-of-pocket costs. The filters together left me with 18,316 patients and 121,033 prescriptions. Table 1 presents the summary statistics. There are six molecules with a market share greater than 3% and branded simvastatin (Zocor) has generic equivalents which entered in mid 2006. The average full price for a 30-day supply for branded drugs falls between $78 and $118, while the average full price for generic lovastatin is only $21.4. The price for generic simvastatin is quite high since there was limited competition among generic manufacturers during the first few months of Zocor’s patent expiration.4 The out-of-pocket costs are the net payments from patients and they are about 22% to 3

In the sample, prescriptions with more than six months of supply are rare. Thus, based on the definition of a new patient, the cleaned data mostly covers individuals who entered this market for the first time ever and those who had not used drugs in this therapeutic class for a long time. The latter type of patients is assumed not to retain information on anything which happened before. Crawford and Shum (2005) use the same length to define a treatment initiation for ulcer. Brookhart et al. (2007) define a statin initiation as filling a prescription without having filled one in the past 12 months. Doing so for the study, however, would largely reduce the size of sample used for estimation. 4 US Food and Drug Administration (FDA) granted a 180-day exclusivity to Teva and Ranbaxy to sell generic simvastatins since they are the first challengers of Zocor’s primary US patent. Thus, during the 180 days after Zocor’s patent expiration, there were only three generic manufacturers: the two independent generic producers and Dr. Reddy’s Laboratories, which received a license from Merck to sell “authorized” generic simvastatin.

5

30% the full price of branded drugs. For the generics, the out-of-pocket costs are much lower since insurance companies usually ask for a lower cost share to induce patients to choose less expensive generic alternatives. I next discuss evidence of consumer dynamics in the data. I define days of treatment as the minimum of the days of supply for a prescription and the number of days between the prescription and the next one. Also, following Crawford and Shum (2005), I define a drug “spell” to be a sequence of one or more prescriptions to a single molecule or molecule-version. Table 2 shows that an average patient has 6.7 prescriptions and 10.7 months of treatment in the one year and a half window. The average number of spells for molecules is about 1.2 and for molecule-versions is about 1.3, implying that patients in the sample do not change drugs very frequently. The patterns of inter-molecule switching shown in Table 3 and Table 4 are consistent with this finding. The switch probabilities are high at the beginning of a treatment. Less than 3% of the patients switch to another molecule after the fifth prescription or 120 days of treatment. The high switch probabilities are probably resulted from experimentation. At the beginning of a treatment, patients are trying different drugs to find the best fit. Once they learn that a drug matches well, they would keep taking the drug for the rest of treatment. Finally, I look at intra-molecule switching by examining length of treatment with Zocor and probability of choosing generic simvastatin after Zocor’s patent expiration on June 23, 2006. Figure 1 shows that over 60% of patients who took Zocor before for various numbers of prescriptions choose generic simvastatin over the branded version when generic simvastatins are available. This implies that the low out-of-pocket costs for generic simvastatin make them very attractive even for experienced users. The nonmonotonic relationship between the number of prescriptions and probability of choosing generic simvastatin can be resulted from variations in length of supplies and copayments across prescriptions. In the data, I observe 30-day, 60-day, and 90-day supplies for most prescriptions. The variation in the days of supplies for a single prescription makes the number of prescription an imperfect measure of length of treatment. Figure 2 serves as a complement by showing that patients with a longer treatment with Zocor are more likely to stay with Zocor when there are cheaper generic equivalents in the market. The

6

conditional probability of choosing simvastatin declines from 81% for a patient with less than two months with Zocor to 67% for a patient who has taken Zocor for over five months. These results suggest that learning at the version level may take longer than learning at the molecule level. We cannot, however, draw any conclusions about learning without controlling for the drug costs.

Patients with a larger copayment difference between generic and

branded simvastatins can be more likely to switch to the generic version even if they may have taken Zocor for a long time and learned that the branded version works well. To disentangle the price effect from the learning effect, we need a model that can handle both at the same time.

3

Model

Demand for pharmaceuticals in the US markets is very complex and involves two major players: patients and doctors. The agency problems between them are interesting and important to consider. However, without detailed doctor-level data it is very difficult to separate doctor’s and patient’s roles in the demand model. Following Ching (2010), Crawford and Shum (2005), and Coscelli and Shum (2004), I assume that doctors and patients are a single decision-maker and that they choose drugs to maximize the patients’ utility. Consider a patient i who goes to a doctor in period t to seek treatment for high cholesterol. After diagnosing the patient and observing her copayments for each drug as well as her prescription history, the doctor selects drug j of molecule g from the choice set Jt for the patient or the outside option of non-medical treatment. A drug here is a combination of a molecule and a version (branded or generic), e.g. branded atorvastatin, branded simvastatin, generic simvastatin, etc. The utility patient i derives

7

from drug j in period t is Uijt = ξj + αpcijt + β1 I (mijt−1 = 1) + β2 I (dijt−1 = 1) m d +Xijt β3 + Xijt β4 + ζig + (1 − σ)ijt

(1)

= uijt + ζig + (1 − σ)ijt , where ξj is the base utility from drug j (drug fixed effect), pcijt the out-of-pocket cost for drug j paid by patient i in period t, mijt a dummy variable equal to one if the doctor chooses the molecule of drug j for patient i in period t, dijt a dummy variable equal m a vector of dummy to one if the doctor chooses drug j for patient i in period t, Xijt d a vector of dummy variables for variables for past use of drug j’s molecule, and Xijt

past use of drug j. The variable I (mijt−1 = 1) is an indicator which equals one if the doctor chose the same molecule of drug j for the patient in the last period, and zero otherwise. The coefficient β1 thus accounts for the dynamic behaviors of switching costs or experimenting. A positive β1 induces an inter-molecule switching cost while a negative β1 implies that doctors like to experiment with various molecules to find the best fit for the patient. That is, if β1 > 0, patient i’s utility is greater if she takes the same molecule in a row. If β1 < 0, the doctor prefers to seek diversity and try something different than her last drug choice for the patient. Similarly, I (dijt−1 = 1) is an indicator which equals one if the doctor chose the same drug (molecule-version) for the patient in the last period, and zero otherwise. If I (mijt−1 = 1) = 1 but I (dijt−1 = 1) 6= 1, the doctor chooses the same molecule this period as the molecule for the last period, but she chooses a different version of the molecule. This will happen when a doctor selected a branded drug last time but chooses its generic equivalent this time, or vice versa. If I (mijt−1 = 1) = 1 and I (dijt−1 = 1) = 1, the doctor chooses the same molecule and the same version for patient i in a row. If β1 + β2 > 0, the patient has inter- and intra-molecule switching cost and she has greater utility choosing the same molecule and version successively. If β1 + β2 > 0 and β2 < 0, the patient prefers to use the same molecule but try a different

8

version. m captures the effect of learning about the molecule of a drug. By including Xijt

dummy variables for different lengths of using a molecule, measured by number of prescriptions or treatment days, we will be able to see whether a patient that receives multiple prescriptions for the same molecule or several days of treatment with the molecule has learned that the molecule is effective in lowering her cholesterol level. 

For example, a positive coefficient on the dummy I 2 ≤



Pt−1

k=1 mijk < 4 implies that

the patient has tried the molecule two or three times and really likes it. In addition, 

comparing the coefficients on I 2 ≤



Pt−1



k=1 mijk < 4 and I 4 ≤



Pt−1

k=1 mijk < 6 will

reveal the difference in the degree of learning between a patient with 2 to 3 past prescriptions and a patient with 4 to 5 past prescriptions for the molecule. d captures the effect of learning about the version of a On the other hand, Xijt d includes dummy variables for different lengths of using a drug (moleculedrug. Xijt



version). If the coefficient on I 2 ≤ 

I 2≤

Pt−1

k=1 dijk

Pt−1

k=1 mijk



< 4 is positive and the coefficient on



< 4 is positive, a patient with 2 to 3 prescriptions for the drug likes 

the molecule as well as the version. However, If the coefficient on I 2 ≤ 

is positive and the coefficient on I 2 ≤

Pt−1

k=1 mijk < 4





Pt−1

k=1 dijk < 4 is negative, the patient with

2 to 3 prescriptions for the drug likes the molecule but prefers a different version. Moreover, we can study how lengths of treatment affect preference for a molecule and version by comparing the coefficients on the dummies for different lengths of treatment with a drug. To allow patient or doctor tastes to be correlated across drugs with the same molecules, I assume a nested logit model with molecules as groups. Under this assumption, ijt and (ζig + (1 − σ)ijt ) are both extreme value random variables, and σ is the nesting parameter with a value between 0 and 1. The within molecule correlation gets larger with σ. According to Berry (1994), the probability of patient i choosing drug j in period t is uijt

Pr (dijt = 1 | Xit ) =  P

k∈Ig e

e 1−σ uikt 1−σ

9

σ 

1+

P P g

k∈Ig e

uikt 1−σ

1−σ  ,

(2)

where Xit denotes the set of conditioning variables for individual i in period t and Ig the set of products in molecule g. The probability of choosing the outside good is 1

Pr (di0t = 1 | Xit ) = 1+

P P g

uikt

1−σ k∈Ig e

1−σ .

(3)

There are some limitations in the model that should be discussed before I show the estimation results. By making the assumption that doctors are perfect agents for their patients, I ignore a doctor’s learning from her other patients and other relevant doctor effects. Two institutional features in the US pharmaceutical markets can mitigate concerns about the information asymmetries arising from the agency issues. First, direct-to-consumer advertising in the US for prescription drugs makes patients more aware of existence of particular drugs and drug characteristics. The advertising also encourages patients to talk about drugs with their doctors. With the prevalence of advertising, patients become active in discussing prescription treatment and can influence doctors’ decision-making. A 2004 FDA survey reports that 31% of respondents asked their doctors about prescription treatment and 29% of them asked about a specific brand. Also, 13 to 22% of surveyed doctors felt some pressure to prescribe a requested drug.5 Second, patients in the US can make the version decision at the pharmacies and choose between branded and generic versions of a molecule unless their doctors write a prescription as “dispense as written.” According to Shrank et al. (2011), only 3.4% of prescriptions for cholesterol lowering drugs filled in 2008 and 2009 at CVS, the largest pharmacy chain in the US, were designated as dispense as written. This implies most high cholesterol patients in the US have the option of substituting versions of a prescribed molecule. Another limitation is the lack of patient conditions in the model. Patients with different cholesterol levels may have different switching costs or learning effects. A drug can reduce the cholesterol level more effectively for more severe patients and they can thus learn more quickly about the drugs they have tried. Without information on 5

Aikin, KJ.; Swasy, JL.; Braman, AC. U.S. Department of Health and Human Services; Food and Drug Administration. Patient and physician attitudes and behaviors associated with DTC promotion of prescription drugs - summary of FDA survey research results. 2004. Nov 19. 2004.

10

patient severity, the parameters in the model would only capture the average switching costs and learning effects. I cannot identify patient heterogeneity due to the data limitation, even though this would be an interesting feature to include. Given the limitations discussed above, the results presented below should be interpreted with the caveats in mind. I would like to allow for the relevant doctor effects and patient heterogeneity in pharmaceuticals as important topics for further research.

4

Estimation

With the specification of the likelihood of dijt given Xit , we can estimate the parameters using maximum likelihood (MLE). The log-likelihood can be written as  log (L(θ)) =

Jt Nt X T X X t=1 i=1 j=1

 I (dijt = 1) log  

 uijt 1−σ

P

k∈Ig

e

uikt 1−σ

 e  σ     , (4) 1−σ u P P ikt 1−σ 1+ g e k∈Ig

where θ is the vector of the unknown parameters. McFadden (1973) has shown that the loglikelihood function is globally concave and discussed the conditions for MLE to be consistent and asymptotically normal in this application. The out-of-pocket cost variable (pcijt ) is endogenous since drug with a better (unobserved) quality can give the patient greater utility and cost more. To alleviate the bias associated with this endogeneity problem, I make use of the panel structure of the data by including the drug fixed effect in estimation. Drug quality is not observable but is assumed to be fixed over time. Thus, adding drug fixed effect can help solve the problem from endogenous out-of-pocket costs. Identification of the other parameters relies on the variation in the corresponding variables. I present several sets of estimation results. I first focus on the subsample with the patients’ first five prescriptions since the data analysis shows more inter-molecule switches in the first five prescriptions or in the first 150 days of treatment. Iizuka (2012) uses the same panel length to study adoption of generic pharmaceuticals. Also, I separately estimate the model using number of prescriptions and days of treatment as the measure of length of treatment. To demonstrate the importance of including drug fixed effects, I show the results with and without drug fixed effects. Finally, as a robustness check, I discuss the estimation results based on the full sample at the end of the section. Table 5 summarizes the estimation results for the subsample. First note that the pseudo

11

R-squared values are between 0.67 and 0.72, suggesting that the model fit is good despite the limited patient information I can obtain from the data. All price (copay) coefficients are significant and precisely estimated. But the price coefficients are more negative in the models with drug fixed effects, which implies that the price coefficient is overestimated if we do not add the drug-specific intercepts. The estimated nesting parameters are about 0.5 and statistically significant under the fixed-effect models. This suggests that the correlation in the “tastes” for drugs with the same molecules is quite large and should not be ignored. The estimates for the coefficients on I(dijt−1 = 1) and I(mijt−1 = 1) are both positive and significant. This means that there is a positive cost if a patient switch to a different molecule or version. Other things being equal, patients prefer to take the same drug in this period as the drug they took in the last period. Furthermore, inter-molecule switching cost is much higher than intra-molecule switching cost since the sum of the two estimated coefficients is larger than the estimated coefficient on I(dijt−1 = 1). The dollar value of inter-molecule switching cost, calculated by dividing the sum of coefficients on I(dijt−1 = 1) and I(mijt−1 = 1) by the absolute value of price coefficient in Model (2), is $111.98 while the dollar value of intra-molecule switching cost is only $14.09. Results from Model (2) shows that the effects of learning about molecules are all positive and significant. Learning effects at the molecule level increases quickly in the first five prescriptions. P  P  t−1 t−1 The coefficient on I m = 4 more than doubles the coefficient on I m = 1 , ijk ijk k=1 k=1 implying that patients who have had four prescriptions for the same molecule derive much more utility than patients with only one prescription with the molecule. The finding is consistent with the patterns shown in Table 4. Estimated coefficients on the dummies for cumulative prescriptions with drugs are also positive and significant, which suggests positive learning effects at the version level. Results from Model (4) are qualitatively consistent with those of Model (2). Most coefficients are significant and have the expected signs. Switching costs and learning effects can be further illustrated in changes in the probability of choosing a drug for patients with different lengths of taking the drug. Table 6 presents the predicted probabilities of choosing the same drug in the first five prescriptions. The probabilities are calculated using the estimates from Model (2) in Table 5 . The first column in the table shows that the choice probabilities vary a lot for the first prescription. In my model for pharmaceutical demand, a patient’s choice for the first prescription is based on the out-ofpocket costs and the base utility for each drug. Thus, drugs with a high choice probability for the first prescription have either relatively high “quality” (Lipitor) or relatively low cost

12

(Simvastatin).6 For the second prescription, the probability of choosing the same drug increases dramatically for most drugs. The probability of choosing Lipitor jumps from 26.85% to 96.96% after the first use. The probability of choosing the less attractive drug Pravachol also largely increases from 3.99% to 78.32%. High switching costs thus make a patient quite likely to stay with a drug even when they use the drug just once. The probabilities of choosing the same drug for more experienced patients are affected by both learning and switching costs. Most of the predicted probabilities in columns two to five of Table 6 increase monotonically with the length of treatment but at a diminishing rate. It is interesting to note Zocor’s relatively low choice probabilities through the first five prescriptions. For a patient who has taken Zocor for four prescriptions, the probability of choosing Zocor again is only 33%, compared to more than 90% for most of the other drugs. This is because patients’ learning about Zocor benefits the generic competitor which shares the molecule but has a much lower copayment. Also, as discussed above, the intra-molecule switching cost is much smaller than the inter-molecule switching cost.7 The generic simvastatin is thus very attractive to the experienced Zocor users, which lowers the probability they stick with Zocor. Table 7 shows similar predicted choice probabilities with different days of treatment. As a robustness check, I estimate the model using the full sample and calculate the predicted choice probabilities. Table 8 summarizes the estimation results. The estimated price coefficients and nesting parameters are very close to the estimates based on the subsample with the first five prescriptions. Most of the other estimates are qualitatively similar. The only noticeable difference is the higher estimated intra-molecule switching costs in Table 8. This can be resulted from the loyal Zocor users who have taken Zocor for a long time. My data shows that more than 50% of patients who have taken Zocor for more than 300 days stay with Zocor even when there are cheaper generic equivalents available. Table 9 and Table 10 present the predicted choice probabilities. The patterns are also quite consistent with those shown in Table 6 and Table 7.

5

Conclusion

The paper estimates a nested logit demand model with brand loyalty and learning using data from cholesterol lowering drug markets in the US. The estimates suggest large inter-molecule switching costs and strong learning effects at both molecule level and version level in the 6

The estimated fixed effect (not shown) is the highest for Lipitor. Since Zocor and generic simvastatin are the only two drugs in the sample that share a molecule, the estimated intra-molecule switching costs are the costs of switching between Zocor and generic simvastatin. 7

13

demand for cholesterol lowering drugs. The large inter-molecule switching costs dramatically increase the probability of choosing the same drugs in a row for new patients. Experienced patients obtain more utility as they take the same molecule longer. This, together with the switching cost, raises the probability of staying with a molecule for experienced patients. On the other hand, the relatively low intra-molecule switching costs give generic drugs an advantage of competing with the branded counterparts, as the learning effect from the branded drugs can benefit generics with a low cost. These state-dependent effects in pharmaceutical demand documented in the study have several implications for drug manufacturers, insurance companies, and health care policy makers. First, when rolling out a new product in a market with existing competitors, drug manufacturers should lower the high inter-molecule switching costs to encourage doctors and patients to try it. They can provide free samples to reduce the financial costs or invest in advertising, such as sending representatives to doctors or direct-to-consumer advertising, to lower the information costs. Second, to encourage generic use and create savings for insurance plans, insurance companies can consider providing more incentive to those who have never tried the generics. Since it is easier for experienced users of a molecule to switch to its generic equivalents after patent expiration, insurance companies can focus more on patients who have never tried molecules with expiring patents by lowering or eliminating their copayment for their first prescription for the molecules. Finally, policy makers can consider making doctors and/or patients aware of the high switching costs so that they can be more careful in choosing drugs at the beginning of a treatment. The study leaves the analysis of doctor effects and patient heterogeneity as important and interesting future research opportunities. Detailed data on both sides are needed to estimate their separate learning effects and perform other relevant analysis. Understanding doctors’ and patients’ roles in pharmaceutical demand can definitely help make more precise policy recommendations.

References Allenby, Greg M, and Peter J Lenk. 1995. “Reassessing brand loyalty, price sensitivity, and merchandising effects on consumer brand choice.” Journal of Business & Economic Statistics, 13(3): 281–289. Berry, Steven T. 1994. “Estimating discrete-choice models of product differentiation.” The RAND Journal of Economics, 242–262.

14

Bronnenberg, Bart J, Jean-Pierre H Dubé, and Matthew Gentzkow. 2012. “The Evolution of Brand Preferences: Evidence from Consumer Migration.” American Economic Review, 102(6): 2472–2508. Brookhart, M Alan, Amanda R Patrick, Sebastian Schneeweiss, Jerry Avorn, Colin Dormuth, William Shrank, Boris LG van Wijk, Suzanne M Cadarette, Claire F Canning, and Daniel H Solomon. 2007. “Physician follow-up and provider continuity are associated with long-term medication adherence: a study of the dynamics of statin use.” Archives of Internal Medicine, 167(8): 847–852. Ching, Andrew T. 2010. “Consumer learning and heterogeneity: Dynamics of demand for prescription drugs after patent expiration.” International Journal of Industrial Organization, 28(6): 619–638. Coscelli, Andrea. 2000. “The importance of doctors’ and patients’ preferences in the prescription decision.” The Journal of Industrial Economics, 48(3): 349–369. Coscelli, Andrea, and Matthew Shum. 2004. “An empirical model of learning and patient spillovers in new drug entry.” Journal of Econometrics, 122(2): 213–246. Crawford, Gregory S, and Matthew Shum. 2005. “Uncertainty and learning in pharmaceutical demand.” Econometrica, 73(4): 1137–1173. Dalen, Dag Morten, Kari Furu, Marilena Locatelli, and Steinar Strøm. 2011. “Generic substitution: micro evidence from register data in Norway.” The European Journal of Health Economics, 12(1): 49–59. Dubé, Jean-Pierre, Günter J Hitsch, and Peter E Rossi. 2010. “State dependence and alternative explanations for consumer inertia.” The RAND Journal of Economics, 41(3): 417– 445. Erdem, Tülin. 1996. “A dynamic analysis of market structure based on panel data.” Marketing Science, 15(4): 359–378. Iizuka, Toshiaki. 2012. “Physician agency and adoption of generic pharmaceuticals.” American Economic Review, 6(120): 2826–58. Keane, Michael P. 1997. “Modeling heterogeneity and state dependence in consumer choice behavior.” Journal of Business & Economic Statistics, 15(3): 310–327.

15

Klemperer, Paul. 1995. “Competition when consumers have switching costs: An overview with applications to industrial organization, macroeconomics, and international trade.” The Review of Economic Studies, 62(4): 515–539. Krishnamurthi, Lakshman, and Sethuraman P Raj. 1991. “An empirical analysis of the relationship between brand loyalty and consumer price elasticity.” Marketing Science, 10(2): 172–183. Lundin, Douglas. 2000. “Moral hazard in physician prescription behavior.” Journal of Health Economics, 19(5): 639–662. McFadden, Daniel. 1973. “Conditional logit analysis of qualitative choice behavior.” Osborne, Matthew. 2011. “Consumer learning, switching costs, and heterogeneity: A structural examination.” Quantitative Marketing and Economics, 9(1): 25–70. Shrank, William H, Joshua N Liberman, Michael A Fischer, Jerry Avorn, Elaine Kilabuk, Andrew Chang, Aaron S Kesselheim, Troyen A Brennan, and Niteesh K Choudhry. 2011. “The consequences of requesting ‘dispense as written’.” The American journal of medicine, 124(4): 309–317. Wooldridge, Jeffrey M. 2010. Econometric analysis of cross section and panel data. MIT press.

16

Table 1: Summary Statistics Product

Molecule

Branded/ generic

Lipitor atorvastatin B Vytorin ezetimibe/simvastatin B Crestor rosuvastatin B Zocor simvastatin B Lovastatin lovastatin G Pravachol pravastatin B Simvastatin simvastatin G 5 others B Note: The level of observations is prescription.

Date of entry Jan Jul Aug Jan Feb Nov Jun < Jul

97 04 03 92 02 91 06 02

In-sample market share

Ave full price for 30 days

44.34 17.83 10.60 9.78 8.23 3.56 3.00 2.67

84.9 78.6 79.7 115.3 21.4 117.7 93.6 80.7

Ave copay for 30 days 18.6 18.4 19.7 30.0 6.3 21.5 9.1 21.3

Table 2: Overall Statistics Variable

Obs

Mean

Number of Rx 18,316 6.69 Days of treatment 18,316 321.43 Number of spells (molecule) 18,316 1.19 Number of spells (molecule-version) 18,316 1.25 Age 18,316 52.27 Female 18,316 0.46 Note: The level of observations is patient.

17

Std. Dev. 3.90 136.94 0.49 0.54 8.20 0.50

Min

Max

1 8 1 1 0 0

22 838 13 13 64 1

Table 3: Inter-molecule Switch by Prescriptions Rx No. 2 3 4 5 6 7 8 9 10 >10

# patients

# patients with a molecule switch

17737 16377 14400 12073 9578 7194 5653 4634 3864 11769

1045 710 507 363 247 171 91 67 66 128

Percentage 5.9% 4.3% 3.5% 3.0% 2.6% 2.4% 1.6% 1.4% 1.7% 1.1%

Table 4: Inter-molecule Switch by Days of Treatment Days of Treatment

# patients

30 - 59 60 - 89 90 - 119 120 - 149 150 - 179 180 - 209 210 - 239 240 - 269 270 - 299 >= 300

18290 17665 16952 16040 15193 13836 12938 11825 10437 10152

# patients with a molecule switch 741 476 626 274 236 263 138 145 132 262

18

Percentage 4.1% 2.7% 3.7% 1.7% 1.6% 1.9% 1.1% 1.2% 1.3% 2.6%

Table 5: Nested Logit Estimates (First Five Prescriptions) (1) Variable

Est.

pc Nesting parameter (σ) I(dijt−1 = 1) I(mijt−1 = 1)

(2) S.E.

−0.0137 0.8449 0.2707 2.7205

(3)

Est.

S.E.

0.0008 −0.0299 0.0129 0.5497 0.0504 0.4213 0.0963 2.9268

1.7060

0.1050

0.5193

0.1896

I

2.0884

0.1031

0.8364

0.1834

I I I I I I

d =4 k=1 ijk

2.3288

0.1056

1.0331

0.1820

2.5430

0.1142

1.2704

0.1897

−0.1115

0.0572

0.5975

0.1796

−0.1194

0.0517

0.6679

0.1717

−0.1466

0.0502

0.6767

0.1693

−0.0942

0.0508

0.7591

0.1740

(4) S.E.

0.0011 −0.0138 0.0382 0.8494 0.1470 0.2235 0.1580 3.3924

Lengths of treatment
in prescriptions Pt−1 I mijk = 1 k=1


Pt−1 mijk = 2 k=1
Pt−1 mijk = 3 k=1
Pt−1 m ijk = 4 k=1
Pt−1 d =1 k=1 ijk
Pt−1 d = 2 ijk k=1
Pt−1 d =3 k=1 ijk
Pt−1

Est.

Est.

S.E.

0.0008 −0.0300 0.0127 0.5663 0.0439 0.2742 0.0757 3.4022

0.0011 0.0306 0.1092 0.1232

0.1604

Lengths of treatment in days
Pt−1 I 25 ≤ D × mijk < 75 k=1 ik

1.2119

0.0834

0.0763

I 75 ≤

1.4085

0.0851

0.1913

0.1616

I

1.6697

0.1206

0.6151

0.1987

I I I I


Pt−1 Dik × mijk < 125
Pk=1 t−1 125 ≤ Dik × mijk < 175 k=1
Pt−1 175 ≤ D × mijk k=1 ik
Pt−1 25 ≤ D × dijk < 75 k=1 ik
Pt−1 75 ≤ Dik × dijk < 125
Pk=1 t−1 125 ≤ D × dijk < 175 k=1 ik
Pt−1

I 175 ≤

D × dijk k=1 ik

Fixed Effect

No

Yes

1.6457

0.0916

0.6402

0.1520

−0.0222

0.0478

0.9426

0.1484

−0.0160

0.0470

1.0183

0.1494

−0.1003

0.0583

0.7499

0.1721

−0.1009

0.0426

0.7930

0.1401

No

Yes

# obs 78,314 78,314 78,314 78,314 Log-likelihood value -50349.7 -44042.9 -50457.4 -44114.3 Pseudo R-squared 0.6721 0.7132 0.6714 0.7127 Note: Homoskedasticity-robust standard errors are reported. Dik is days of treatment for patient i in period k.

19

Table 6: Predicted Probabilities of Choosing the Same Drug (First Five Prescriptions) Cumulative Rx with product Product Lipitor Vytorin Crestor Zocor Lovastatin Pravachol Simvastatin Other

0

1

2

3

4

26.85% 12.80% 8.38% 1.42% 5.73% 3.99% 38.21% 2.61%

96.96% 92.73% 88.83% 25.21% 84.09% 78.32% 97.87% 69.99%

97.92% 94.95% 92.14% 28.56% 88.62% 84.19% 98.48% 77.46%

98.30% 95.85% 93.50% 29.11% 90.53% 86.73% 98.71% 80.84%

98.76% 96.95% 95.19% 33.17% 92.94% 90.00% 99.02% 85.31%

Note: Choice probabilities are evaluated at mean out-ofpocket costs.

Table 7: Predicted Probabilities of Choosing the Same Drug (First Five Prescriptions) Cumulative days with product Product Lipitor Vytorin Crestor Zocor Lovastatin Pravachol Simvastatin Other

0

25-74

75-124

125-174

≥175

26.31% 12.49% 8.17% 1.13% 5.61% 3.88% 39.84% 2.55%

97.52% 94.02% 90.74% 30.63% 86.74% 81.64% 98.52% 74.26%

97.94% 95.01% 92.24% 35.50% 88.80% 84.35% 98.78% 77.77%

98.16% 95.53% 93.01% 26.47% 89.89% 85.80% 98.82% 79.68%

98.29% 95.84% 93.49% 27.33% 90.55% 86.70% 98.90% 80.87%

Note: Choice probabilities are evaluated at mean out-ofpocket costs.

20

Table 8: Nested Logit Estimates (Full Sample) (1) Variable

Est.

pc Nesting parameter (σ) I(dijt−1 = 1) I(mijt−1 = 1)

(2) S.E.

−0.0141 0.8288 0.3387 4.1305

(3)

Est.

S.E.

0.0008 −0.0286 0.0115 0.5591 0.0259 0.8506 0.0361 3.5959

0.7785

0.0455

0.4463

0.0701

I 4≤

1.1652

0.0556

0.7223

0.0767

I I I I I I I

I 10 ≤

k=1

dijk

1.5774

0.0786

1.1092

0.0986

1.7692

0.1042

1.2178

0.1219

2.0101

0.1009

1.3905

0.1163

−0.1448

0.0233

0.0761

0.0596

−0.1177

0.0222

0.2189

0.0602

−0.1816

0.0262

0.1128

0.0678

−0.1022

0.0264

0.3027

0.0732

−0.0110

0.0231

0.5449

0.0743

(4) S.E.

0.0014 −0.0143 0.0443 0.8124 0.0913 0.3545 0.0855 4.1540

Lengths of treatment in
prescriptions Pt−1 I 2≤ mijk < 4 k=1


Pt−1 mijk < 6 k=1
Pt−1 6≤ mijk < 8 k=1
Pt−1 8≤ m ijk < 10 k=1
Pt−1 10 ≤ mijk k=1
Pt−1 2≤ d ijk < 4 k=1
Pt−1 4≤ d <6 k=1 ijk
Pt−1 6≤ d <8 k=1 ijk
Pt−1 8≤ dijk < 10 k=1
Pt−1

Est.

Est.

S.E.

0.0008 −0.0283 0.0128 0.5312 0.0291 0.8440 0.0403 3.5987

0.0011 0.0315 0.0696 0.0711

Lengths of treatment in days
Pt−1 I 60 ≤ D × mijk < 120 k=1 ik

0.5262

0.0507

0.1557

0.0834

I 120 ≤

1.1429

0.0628

0.7720

0.0920

I

1.2391

0.0674

0.9090

0.0908

1.4444

0.0782

1.0664

0.0986

1.5417

0.0756

0.9904

0.0943

−0.0471

0.0284

0.2673

0.0732

I I I I I I


Pt−1 Dik × mijk < 180
Pk=1 t−1 180 ≤ Dik × mijk < 240
Pk=1 t−1 240 ≤ Dik × mijk < 300
Pk=1 t−1 300 ≤ D × mijk k=1 ik
Pt−1 60 ≤ D ik × dijk < 120
Pk=1 t−1 120 ≤ D × dijk < 180 k=1 ik
Pt−1 180 ≤ Dik × dijk < 240
Pk=1 t−1 240 ≤ D ik × dijk < 300 k=1
Pt−1

I 300 ≤

D × dijk k=1 ik

Fixed Effect

No

Yes

−0.1593

0.0306

0.1465

0.0759

−0.2147

0.0295

0.0737

0.0707

−0.1574

0.0274

0.2009

0.0699

−0.0621

0.0238

0.4483

0.0700

No

Yes

# obs 121,033 121,033 121,033 121,033 Log-likelihood value -56610.4 -50194.3 -56756.7 -50297.2 Pseudo R-squared 0.7632 0.7900 0.7626 0.7896 Note: Homoskedasticity-robust standard errors are reported. Dik is days of treatment for patient i in period k.

21

Table 9: Predicted Probabilities of Choosing the Same Drug (Full Sample) Cumulative Rx with product Product Lipitor Vytorin Crestor Zocor Lovastatin Pravachol Simvastatin Other

0

1

2-3

4-5

6-7

8-9

≥ 10

28.82% 14.71% 9.65% 2.06% 6.49% 4.32% 30.73% 3.23%

97.19% 93.64% 90.11% 30.08% 85.55% 79.40% 96.66% 73.99%

98.47% 96.48% 94.43% 34.45% 91.68% 87.76% 97.89% 84.11%

98.53% 96.62% 94.66% 41.89% 92.01% 88.23% 98.17% 84.69%

98.76% 97.14% 95.46% 36.49% 93.17% 89.88% 98.20% 86.77%

99.30% 98.37% 97.40% 47.14% 96.05% 94.05% 98.92% 92.11%

99.51% 98.85% 98.15% 60.66% 97.18% 95.74% 99.30% 94.31%

Note: Choice probabilities are evaluated at mean out-of-pocket costs.

Table 10: Predicted Probabilities of Choosing the Same Drug (Full Sample) Cumulative days with product Product Lipitor Vytorin Crestor Zocor Lovastatin Pravachol Simvastatin Other

0

1-59

60-119

120-179

180-239

240-299

≥ 300

28.66% 14.67% 9.59% 2.37% 6.51% 4.30% 30.87% 3.01%

97.16% 93.60% 90.02% 30.26% 85.55% 79.26% 96.40% 72.54%

98.59% 96.76% 94.86% 43.87% 92.37% 88.66% 98.12% 84.38%

97.88% 95.19% 92.43% 37.37% 88.90% 83.80% 97.33% 78.14%

98.76% 97.15% 95.46% 34.46% 93.24% 89.90% 97.91% 86.02%

99.04% 97.80% 96.48% 40.87% 94.73% 92.06% 98.39% 88.91%

99.36% 98.52% 97.62% 53.97% 96.42% 94.56% 98.98% 92.32%

Note: Choice probabilities are evaluated at mean out-of-pocket costs.

22

Figure 1: Choice Between Zocor and Simvastatin by Number of Prescriptions with Zocor

23

Figure 2: Choice Between Zocor and Simvastatin by Days of Treatment with Zocor

24