THE JOURNAL OF INDUSTRIAL ECONOMICS
`Volume XL VIII
`September 2000
`
`0022-1821
`No.3
`
`THE IMPORTANCE OF DOCTORS' AND PATIENTS'
`PREFERENCES IN THE PRESCRIPTION DECISION*
`
`ANDREA CosCELLit
`
`This paper studies the contribution of doctor and patient 'habit' to
`persistence in market shares in prescription drug markets. My unique
`panel dataset allows me to estimate the probability of switching brands
`as a function of patient and doctor attributes, with an emphasis on
`past prescribing behaviour so as to capture the degree of persistence. I
`find significant evidence of time~dependence in prescription choices
`for both doctors and patients, which seems to imply that in molecular
`submarkets in which brands are not allowed to compete on the basis of
`price, doctor and patient 'habit' at the micro~level can translate into
`sticky and persistent market shares at the aggregate level.
`
`I.
`
`INTRODUCTION
`
`IN THIS PAPER, I study the contribution of doctor and patient 'habit' to
`persistence in market shares among therapeutically equivalent pn~scription
`drugs. While, similar issues have arisen in the recent literature about the
`competition between generic and branded drugs, they are especially
`puzzling in the Italian pharmaceutical market. In Italy, regulatory fiat
`imposes uniform prices across all vendors of drugs which utilize the same
`active ingredient, thus eliminating price variation as an important avenue
`of differentiation among otherwise therapeutically equivalent drugs, which
`is true in drug markets with generic competitors. My unique panel dataset
`allows me to estimate the probability of switching brands as a function
`of patient and doctor attributes, with an emphasis on past prescribing
`behaviour so as to capture the degree of persistence.
`This analysis can shed. light on several aspects of market structure in
`the pha1maceutical industry. First, there is a growing body of literature
`
`*I acknowledge financial support from the European Commission through a TMR
`fellowship #ERBFMBICT972232. I would like to thank the Istituto Superiore dt Sanita' for
`the use of their data. This paper was previously circulated under the title 'Are Market Shares
`in Drug Markets Affected by Doctors' and Patients' Preferences for Brands?'. Seminar
`participants at Stanford University (OSB and Department of Economics), Royal Holloway
`and UCL have provided valuable comments. I would like to especially thank my principal
`advisor, Peter Reiss, who constantly helped me improve this paper, and the editor David
`Genesove for his useful comments. Mike Mazzeo, Fiona Scott Morton, Andrea Shepard,
`Matthew Shum and an anonymous referee have also provided many useful suggestions. All
`errors, however, are my own.
`i· Author's affiliation: National Economic Research Associates, J 5 Stratford Place, London
`WIN9AF, UK.
`email: Andrea. Cosceili@nera.com
`
`"' Blackwell Publishers Ltd. 2000, 108 Cowley Road, Q,f<)rd OX4 UF, UK, and JSO Main Street, Malden, MA 02148, USA.
`349
`
`SENJU EXHIBIT 2142
`INNOPHARMA v SENJU
`IPR2015-00903
`
`PAGE 1 OF 21
`
`

`
`ANDREA COSCELLI
`
`350
`that tries to explain observed market segmentation using data on national
`market shares. Empirical observations of market shares for trade-name
`and generic drugs in post-patent therapeutic categories in the US market
`usually indicate a degree of segmentation between branded drugs and their
`generic equivalents, arising from a finite cross-price elasticity between the
`two types of drugs (the cross-price elasticity between two homogeneous
`goods should be infinite). However, these studies ignore individual
`heterogeneity. The micro dataset at hand allows me both (i) to control for
`individual heterogeneity and (ii) to explore the degree of time-dependence
`in drug choices, both of which can be important in explaining the
`substantial and persistent differences in market shares among therapeuti(cid:173)
`cally equivalent drugs.
`Second, in recent years, we have witnessed a surge in direct advertising
`to consumers by pharmaceutical companies for prescription drugs sold in
`the US market. The amount spent on direct-to-consumer prescription drug
`advertising rose from US$35m in 1987 to US$357m in 1995, US$610m in
`1996, and over US$1 billion in 1997 (NERA (1999]). This spending choice
`reflects a widespread belief within the pharmaceutical industry that
`patients should have a role in the choice of prescription drugs. This paper
`directly studies the patient's role in pharmaceutical choice.
`Finally, the most important institutional features of the Italian market
`during the sample period such as the important role of licensed products,
`limited patient co-payments, and lack of direct financial incentives to
`doctors to prescribe cheaper drugs characterize almost every EU country
`(NERA [1999]).
`I use a new panel dataset provided by the Italian National Health
`Institute, which includes all the prescriptions in the anti-ulcer market from
`1990-1992 for a 10% random sample of the population of Rome aged
`15-85. This dataset allows researchers a glimpse into the dynamics of
`prescription behavior at the micro level which is not possible with the
`predominantly aggregate and/ or cross-sectional datasets which have been
`used in most studies of pharmaceutical markets to date.
`My main conclusions are as follows. I begin by testing the null hypo(cid:173)
`thesis of whether doctors and/ or patients are indifferent between different
`brands of the same molecule, as we would expect given their therapeutic
`equivalence. After rejecting the hypothesis, I attempt to isolate both the
`patient-level and the doctor-level factors which are responsible for product
`differentiation. I focus specifically on the degree of time-dependence in
`doctors' and patients' drug choices by testing whether the patients show
`state dependence in their purchasing patterns, and whether the doctors
`exhibit habit persistence. I find significant evidence of doctor and patient
`'habit', which imply that in molecular sub markets in which brands are not
`allowed to compete on the basis of price, habit persistence at the micro-
`
`©Blackwell Publishers Ltd. 2000.
`
`PAGE 2 OF 21
`
`

`
`PREFERENCES IN THE PRESCRIPTION DECISION
`
`351
`
`level can translate into sticky and persistent market shares at the aggregate
`level.
`The paper is organized as follows, In the next section I survey the
`previous empirical literature, In Section III, I describe the dataset used in
`the estimation, Section IV describes my empirical specification, while
`Section V reviews the results. A summary of the results make up the final
`section.
`
`II. DOCTORS' DEMAND
`
`While the present study focuses on doctors' demand for pharmaceutical
`products, most of the recent literature on pharmaceuticals (for example,
`Caves et at. [1991), Caves and Hurwitz [1988), Berndt eta!, [1997), Scott
`Morton [1997, 2000), and Scherer [1993)) has focused on supply-side issues
`(e.g,, entry, pricing, advertising, R&D races), In his comment on Caves
`[1991), Pakes [1991) argues that a panel following doctors'
`et at.
`prescriptions over time would be the only way to understand the major
`determinants of the demand for pharmaceuticals, The panel data I use
`allow me to separately identify doctor and patient effects,
`Much of the previous work on the demand for pharmaceuticals has used
`aggregate, market-share data, which are much better suited to measuring
`the degree of differentiation between various drugs rather than explain its
`causes, For example, Stern [1995] finds low substitutability between
`branded and generic drugs, while Ellison et at. [1997] find a high elasticity
`of substitution between generic and branded drugs,
`One recent microdata-based analysis of the demand for pharmaceuticals
`is that by Hellerstein [1998], She focuses on doctors' choices between
`branded and generic versions of drugs for which a patent has recently
`expired, Significantly, she finds some evidence of habit persistence in the
`prescription behavior of physicians, even after controlling for observable
`characteristics of physicians and patients. Unfortunately, her dataset does
`not allow her to test for patients' effects owing to data limitations, while
`her dataset allows for an analysis of financial incentives due to third-party
`payer variation. My dataset, on the other hand, has multiple observations
`for doctor-patient interactions, prescription of the same molecule by a
`single doctor to many patients, and prescriptions of the same drug by
`many doctors,
`Gorecki [1986, 1987] analyzes competition between patent holders
`and licensees in Canada: an institutional setting very similar to the
`Italian market He only observes aggregate data, but he is able to take
`advantage of the regulatory variation among Canadian provinces to
`identify competitive effects
`in his empirical analyses, Gorecki's
`conclusions in [1986] are consistent with my results: ',,, Since physicians
`still write, by and large, brand name prescriptions for the pioneering
`©Blackwell Publishers Ltd. 2000.
`
`PAGE 3 OF 21
`
`

`
`352
`
`ANDREA COSCELLI
`
`brand, unless an element of price competition is introduced at the level of
`the pharmacist the pioneering brand will continue to dominate the market
`[ ... ]. Hence it is the combination of attempting to nullify quality
`differences between
`the pioneering and
`late entrant brands and
`introduction of price competition that results in the late entrants capturing
`market share'.
`Pharmaceutical markets are subdivided into therapeutic classes. Follow(cid:173)
`ing most of the recent economic literature on pharmaceuticals (e.g., Stern
`[1995]), I regard a therapeutic class as having several sub-markets. I define
`a therapeutic market as a 4-digit A TC code (for example, A02B contains
`all the anti-ulcer drugs), and a sub-market as a specified molecule (for
`example, ranitidine). The ATC code is an international classification
`scheme which classifies drugs by target part of the anatomy, mechanism of
`action, and chemical and therapeutic characteristics. This is a natural
`definition of demand because a 4-digit A TC code includes all the molecules
`which can theoretically be prescribed for a certain diagnosis. The
`molecules themselves differ according to side effects, interactions with
`other drugs, specific indications and prices. In the markets I study, a
`physician typically decides the appropriate molecule for the diagnosis and
`then she decides which trade-name 1 version of the molecule to prescribe to
`the patient.
`My work focuses on a particular therapeutic market: anti-ulcer drugs
`(A02B). I analyze this market because it accounts for a considerable
`proportion of world-wide expenditure on pharmaceuticals (around 5%,
`IMS International [1996)). Ulcers also required repeated treatment in the
`early 1990s/ a key feature of my analysis. I analyze six molecule
`submarkets (famotidine, ranitidine, nizatidine, roxatidine, omeprazole and
`misoprostole), which represent more than 90% of the prescriptions during
`the sample period (1990-1992). I restrict my sample to these six molecules
`because the other molecules represent more 'mature' and smaller sub(cid:173)
`markets, where some of the prices for identical brands differ. 3 In each sub(cid:173)
`market there is a patent-holder and licensees marketing the molecule.
`Much of the literature on trade-name drugs versus generics is concerned
`with
`the dimensions (among others,
`'perceived quality' or name
`recognition) according to which these products 'differ'. In my analysis I
`focus on competing drugs based on the same active ingredient and
`
`l All the drugs sold under a license or a patent in the Italian market have a trade name.
`2 It has recently been found that approximately 80% of peptic ulcers can be cured by
`eradicating Helicobapter Pylori, a bacterium responsible for the recurrence of ulcers, by using
`a combination of antibiotics and anti-ulcer drugs (Graham [1993]).
`3 Producers of older molecules had their prices equalized only upon applying for a price
`revision, which happened much later. Moreover, some of the producers in these excluded sub(cid:173)
`markets are very small firms for whom the assumption of identical quality might not hold.
`© BlackweH Publishers Ltd. 2000.
`
`PAGE 4 OF 21
`
`

`
`PREFERENCES IN THE PRESCRIPTION DECISION
`
`353
`
`marketed by important producers entering the market at the same
`time.4
`
`Ill. THE DATA
`
`The main dataset (provided by the Istituto Superiore della Santta ') records,
`for a 10% sample of the population of the Metropolitan Area of Rome
`aged 15-85, all the prescriptions in the anti-ulcer (A02B) drug market
`during the period 1990-1992. The sample is stratified according to age and
`gender; so that the results are representative of the Rome population.
`This patient-level dataset contains over 310,000 observations. An obser(cid:173)
`vation records the identity of the prescribing doctor, the identity of the
`patient, the year and month, and the particular presentation form of the
`drug prescribed (for example, I package of ZANT AC 20 tablets, 150 mg
`each). An observation indicates exactly the drug bought by the patient,
`because the records are collected from pharmacies. In the patient-level
`dataset there are more than 3,400 doctors prescribing at least once to one
`of the in-sample patients. A supplementary dataset from the same source
`records all the prescriptions that 350 of these doctors wrote for any of
`their patients during the same period. The supplementary doctor-based
`dataset contains over 710,000 prescriptions and each observation records
`exactly the same information as the patient-level dataset. The final dataset
`used in my estimations has more than 75,000 observations; it retains all
`the observations in the patient-level dataset for the patients who received
`at least one prescription from one of the 350 doctors whose entire
`prescription history is known.
`
`Italian Market Three important characteristics of the Italian pharma(cid:173)
`ceutical industry are: (i) there is no price and third-party payer variation,
`(ii) the over-the-counter (OTC) market was tiny in the period of interest,
`and direct advertising to patients for prescription drugs had not yet
`started, 5 and (iii) during the sample period, the pharmacist had no power
`to substitute generics for trade-name drugs, as he does in many American
`states. 6
`
`Doctors' Prescribing Behavior Doctors heavily prescribe across brands:
`
`4 By doing this I believe I have effectively controlled for all 'objective' dimensions of
`differentiation between drugs; therefore I can proceed with my tests of doctor or patient
`indifference fairly confident that I have controlled for a large share of drug-specific
`heterogeneity.
`5 It is currently illegal throughout the EU and will remain so for several years even though
`the subject is now occasionally raised.
`6 Hellerstein's dataset, therefore, potentially contains a large amount of measurement error
`in the prescription variable for states where substitution with generics is mandatory.
`© Blackwell Publishers Ltd. 2000.
`
`PAGE 5 OF 21
`
`

`
`354
`
`ANDREA COSCELLI
`
`for example, 98.6% of the in-sample doctors prescribed each of the three
`available brands of omeprazole at least once during the sample period.
`Moreover, doctors usually prescribe multiple brands of each molecule in a
`given month. Thus doctors do not specialize in particular drugs over time
`or at a particular point in time.
`Patients are not limited to a single brand either. At a given point, 40%
`of them will have had experienced a shift to a new brand of the same
`molecule. Different doctors treat the same patient differently. Although
`the proportion of switches in
`the overall sample is 4.5%, among
`observations where patients change doctors, the proportion rises to 9%.
`This latter incidence, however, is much lower than the 48% switch rate that
`we would expect were the new doctor not to take into account the patient's
`history of prescriptions, and to prescribe to the patient according to the
`same proportion used for her other patients.
`Most strikingly, of those patients who were switched by the new doctor
`and then returned to their original doctor, 50% (44/88) are switched again
`when they go back to their usual doctor, and almost all of those (93.2%
`or 41 out of 44) go back to the treatment they received in the previous
`period. These patterns demonstrate clearly that the probability of receiving
`a new treatment is significantly influenced by the doctor's identity, and
`that doctors differ in their choice among therapeutically equivalent drugs
`for the same patient. Next, I present a formal econometric model which
`accommodates all these aspects of behaviour.
`
`IV. EMPIRICAL ANALYSIS
`
`IV(i). Theory
`
`I analyze the problem facing a doctor trying to choose among brands of
`a certain molecule. The entire analysis is conditional on the choice of the
`molecule, which is driven by a more complicated set of factors (indications,
`patient's general health, side effects, price, etc.); most of these factors are
`unobservable to the econometrician. By restricting myself to the analysis
`of 'homogeneous' goods, I can hold price and quality constant and focus
`on other factors driving the choice of competing brands. 7
`The central focus of the study is the question of what brand of a given
`molecule doctor i prescribes to a patient j given this patient's past
`prescription experience with this molecule. In particular, doctor i must
`decide whether to prescribe the brand the patient received in the previous
`period (hereafter o for 'old' brand) or a new brand, which might generate
`
`7 I do not analyze whether the doctor prescribed a different presentation form, because all
`the producers sell the same presentation forms, and there was therefore no need to change
`vendor if the doctor or the patient wanted a different presentation form.
`©Blackwell Publishers Ltd. ;woo.
`
`PAGE 6 OF 21
`
`

`
`PREFERENCES IN THE PRESCRIPTION DECISION
`
`355
`a higher utility to the doctor (hereafter n for 'new' brand). More
`specifically, a choice problem where the decision maker decides either to
`stay on the diagonal of a transition probabilities model, or to move off(cid:173)
`diagonal to any other brand, is collapsed into a simpler problem where the
`decision-maker is confronted with the binary choice of either to stay on
`the diagonal or to move off-diagonal. Since the analysis is conditional on
`the choice of the molecule, 'new brand' does not include other molecules.
`
`IV(ii). Estimation Strategy
`
`The empirical model parameterizes the probability of switching brands as
`a function of patient and doctor attributes. I define as an 'old' brand, o, at
`time t, the brand consumed by the patient at time t- 1. This means that
`I define as a 'new' brand, any brand that differs from the 'old' one without
`taking into account whether the 'new' brand was previously prescribed.
`This means that my analysis focuses on 'first-order' state dependence, so
`that only one-period lags have an effect. Moreover, since only the year and
`month of the prescription are observed, there are 'ties' in the dataset. That
`is, there are patients who receive more than one prescription in a month.
`When the prescriptions are ranked in a chronological order, the pre(cid:173)
`scriptions where the patient, molecule and month are the same are
`randomly ordered. Finally, the prescriptions in the sample are written
`either for one package or for two packages of the same brand. The sample
`is evenly split between these two occurrences. Since I model a prescription
`episode, and not quantity, as my dependent variable, I do not distinguish
`between a prescription of two packages and a prescription of one
`package8
`I use probit specifications to test the null hypothesis of no doctor and/ or
`patient preferences. These models include doctors' fixed effects and
`patients' random effects to capture the unobserved (to the econometrician)
`component of doctors' and patients' preferences.
`The dataset suffers from the problem of initial conditions common to
`most dynamic panel data models. I observe a sample of doctors and
`patients for three years, but I do not have any information on their
`behavior before the sample period begins. Ideally, I would observe the
`doctors' behavior since
`they started practicing and
`the patients'
`prescriptions since they were first treated for ulcer problems. This is
`impossible; therefore following one of the suggestions put forth in the
`literature (Heckman [1981]), I assume that the prescription process starts
`anew if patients have not had a prescription for six months. Thus, I only
`
`&This implies a two~package prescription leads to the same degree of state-dependence as
`a one-package prescription.
`@ Blackwell Publishers Ltd. 2000.
`
`PAGE 7 OF 21
`
`

`
`ANDREA COSCELLI
`
`356
`use data on patients where I am able to observe whether for six months
`prior they did or did not receive ulcer medications. This basically means
`that when the process restarts, the decision makers do not retain the
`information on anything which happened before?
`
`IV(iii). Definition of the Variables
`
`Tables I and II list all the variables used in the estimations. SWITCH is
`the dependent variable and it takes value 1 if patient j receives a brand at
`timet different from what he received at time t- 1 (for the same molecule),
`it is 0 otherwise. We review those variables that are not self-explanatory.
`
`Patients' Variables
`Time-invariant #PRESCRIPTIONS distinguishes among patients ac(cid:173)
`cording to the seriousness of their ulcer problem (e.g., chronic versus
`
`TABLE I
`PATIENT-LEVEL VARIABLES USED IN THE EsTIMATION
`
`Dependent variable
`SWITCH
`
`Patient's variables
`GENDER
`AGE
`#PRESCRIPTIONS
`
`#DOCTORS
`
`#MOLECULES
`
`#SPELL-MOLECULE
`#PAST SWITCHES
`#MONTHS
`
`NEWDOCTOR-TEMP
`
`NEWDOCTOR-PERM
`
`NEWDOCTOR-RET
`
`Takes value l if the brand prescribed is different from the
`brand previously prescribed, 0 otherwise
`
`Takes value 1 if female and 2 if male
`Patient's age
`Total number of prescriptions that the patient receives in
`the sample
`Total number of different physicians who prescribed at
`least one drug to the patient
`Total ntimber of different molecules that the patient
`consumes in the sample
`Number of prescriptions of the molecule up to timet
`Number of (within~molecule) switches up to timet
`Actual number of months elapsed between the prescription
`at timet and the one at t- 1
`Dummy equal to 1 if the physician is a temporary one for
`the patient, 0 otherwise
`Dummy equal to l if the physician is a new permanent one
`for the patient, 0 otherwise
`Dummy equal to I if the patient returns to a previous
`physician, 0 otherwise
`
`9 Most of the models were rerun with a 3~month window instead. The results are not
`qualitatively different.
`© Blackwell Publishers Ltd. ;woo.
`
`PAGE 8 OF 21
`
`

`
`PREFERENCES IN THE PRESCRIPTION DECISION
`TABLE II
`DOCTOR-LEVEL VARIABLES USED IN THE ESTIMATION
`
`357
`
`Doctors' characteristics-anti-ulcer market
`Average monthly quantity prescribed by the doctor in the entire
`QUANTITY
`market in the previous six months
`Average monthly herfindahl index across brands in the entire
`market in the previous six months
`Average monthly herfindahl index at the molecule level in the
`entire market in the previous six months
`
`HERFBRAND
`
`HERFMOLE
`
`Doctors' characteristics-molecule-specific
`Share of prescriptions of the molecule by the doctor in the
`MOLESHARE
`previous month
`Weighted proportion of the prescriptions of the molecule written
`for the old brand in the 2 previous months (last month's share
`plus 0,9 of the previous month's share)
`
`% OLD BRAND
`
`occasional), while #MOLECULES differentiates patients according to
`their willingness to change treatment. #DOCTORS controls for patient(cid:173)
`specific preferences for changing doctor. Patients who change doctors
`more often probably gather more information on possible treatments;
`therefore it ought to be more difficult for a doctor to switch them; on the
`other hand, these patients might be more experimental.
`
`Time-varying There are three time-varying covariates for the patient,
`which are crucial to measure patient 'habit'.
`#SPELL-MOLECULE increases over time with any prescription of
`the molecule. #PAST SWITCHES increases only upon a within-molecule
`switch. These two ad-hoc variables proxy for the switching tendency of
`a patient. For example, a patient who receives the tenth prescription of
`the molecule ranitidine, has #SPELL-MOLECULE= 10,
`if #PAST
`SWITCHES= 0, this indicates that the patient has always been with the
`same brand. Finally, #MONTHS counts the actual number of months
`elapsed between prescription episodes. It cannot exceed six, given the
`definition of a treatment episode. Finally, there is a series of prescription(cid:173)
`specific dummies defining whether the patient is receiving the prescription
`from a substituting physician (NEWDOCTOR-TEMP), 10 has permanently
`moved to a new physician (NEWDOCTOR-PERM), or is returning to
`the usual physician (NEWDOCTOR-RET) after receiving a prescription
`from a substitute. These variables explore whether doctors have different
`preferences for vendors of a particular molecule. If prescriptions decisions
`
`10 Most of the temporary substitutions occur when the physician is on vacation, or the
`patient is out of town when he needs a prescription.
`©Blackwell Publishers Ltd. 2000.
`
`PAGE 9 OF 21
`
`

`
`358
`
`ANDREA COSCELLI
`
`are determined solely by the patient's condition, a change in the pre(cid:173)
`scribing doctor should not vary the patient's likelihood of being switched.
`
`I use doctor covariates which are time-varying. Most
`Doctors' Variables
`of them summarize the prescribing behavior of the physicians in the six
`months immediately prior to the prescription episode.
`
`Constant across molecule markets Table II describes the variables in
`detail. QUANTITY distinguishes between heavy and light prescribers of
`anti-ulcer drugs, and proxies for the size of the doctor's patient base.
`HERFBRAND proxies for the doctors' preference to prescribe multiple
`brands, while HERFMOLE controls for the dispersion in prescribing
`behavior of the doctor due to choice of d(fferent molecules, which is
`dictated by heterogeneity in the patients' pool. Its coefficient does not have
`any specific economic meaning: it simply allows me to interpret the
`coefficient on HERFBRAND as representing doctors' preference for
`prescribing multiple brands, rather than their need to prescribe multiple
`drugs (molecules) due to therapeutic considerations. This distinction is
`based on the assumption that the doctor first chooses the molecule that
`suits the patient and then decides which brand to prescribe.
`
`Molecule-specific % OLD BRAND captures habit persistence by the
`doctor. MOLESHARE proxies for the importance that the molecule has
`for the doctor.
`
`Molecule markets The number of competing products, the time since
`the molecule entered the market, the size of the vendors, and the amount
`of advertising all differ across molecule markets. I use molecule dummies
`to proxy for the different competitive conditions in each molecule market.
`Finally, I include monthly dummies to control for exogenous changes
`(such as regulatory changes) that occur during the sample period.
`
`Summary Statistics Table III provides sample summary statistics. There
`are slightly more than 75, 000 observations in the original sample, an
`observation being a prescription. I exclude the following observations:
`(i) patients who received prescriptions from doctors whose entire pre(cid:173)
`scription history is not known, (ii) the first prescription of each molecule
`for each patient, so as to initialize the process, and (iii) all prescriptions in
`the first 6 months and those more than 6 months apart for the same patient
`because of the initial conditions problem discussed above. This leaves
`43,840 observations.
`There are more than 5, 000 different patients and 350 doctors in the
`sample. I observe a change of brand (SWITCH) in 4% of the observations.
`There are slightly more prescriptions written to females in the sample,
`© Blackwell Publishers Ltd. 2000.
`
`PAGE 10 OF 21
`
`

`
`PREFERENCES IN THE PRESCRIPTION DECISION
`TABLE III
`SUMMARY STATISTICS FOR THE SAMPLE
`
`359
`
`Dependent variable
`Switch
`
`Patient's variables
`Gender: 1 if female, 2 if male
`Age
`#In~sample prescriptions
`#Molecules prescribed in-sample
`#Different doctors prescribing
`#Prescriptions of the molecule
`#Past switches up to timet
`#Months between prescriptions
`New doctor-temp
`New doctor-perm
`New doctor-ret
`
`Mean
`
`Std. Dev. Min
`
`Max
`
`0.040
`
`O.I97
`
`0
`
`1.484
`62.73
`29.478
`1.87
`1.629
`I4.955
`0.396
`1.232
`O.Ql8
`0.02I
`0.024
`
`0.499
`I2.65
`I9.75
`1.09
`!.II
`I2.68
`1.205
`1.277
`O.I34
`O.I43
`O.I55
`
`I-F
`I5
`2
`I
`I
`2
`0
`0
`0
`0
`0
`
`2-M
`85
`I26
`8
`I4
`I06
`I8
`6
`I
`I
`I
`
`Doctors' characteristics-anti~ulcer market
`Average monthly quantity prescribed in-sample
`Herfindahl brand level
`Herfindahl molecule level
`
`876.I68
`0.243
`0.485
`
`297.568
`0.069
`O.II3
`
`I07.08
`0.09
`O.I92
`
`25I2.7I7
`0.68
`0.93
`
`Doctors' characteristics-molecule specific
`%Old brand prescribed
`Molecule share
`
`0.97
`0.528
`
`0.405
`0.256
`
`O.o25
`0.025
`
`1.9
`I
`
`Molecule sub-market
`ran/dine
`nizatidine
`roxatidine
`omeprazole
`
`0.742
`0.04
`0.008
`0.097
`
`0.438
`O.I99
`0.09I
`0.297
`
`0
`0
`0
`0
`
`while the average age is around 62. Turning to the time-varying covariates
`for the patients: the average prescription is written to a patient who is
`purchasing the particular molecule for the fourteenth time (#SPELL(cid:173)
`MOLECULE) and who has already switched brands 0.4 times (# PAST
`SWITCHES). The average time between two prescriptions (#MONTHS)
`is less than two months. Finally, in approximately 6% of the prescriptions,
`the prescribing doctor differs from the one who wrote the previous
`prescription. The number of prescriptions where the patient temporarily
`visits a doctor different from
`the one responsible for the previous
`prescription (NEWDOCTOR-TEMP) is about 1.8%; whereas 2% of the
`prescriptions are written by a new doctor, who then becomes the usual
`doctor for the patient (NEWDOCTOR-PERM). Finally, 2.4% of the
`prescriptions are written by the usual physician after the patient has been
`prescribed something by a substitute in the previous period.
`The unit of measurement for the quantity variable (QUANTITY) is
`the defined daily dose, which is an international measurement unit. For
`©Blackwell Publishers Ltd. 2000.
`
`PAGE 11 OF 21
`
`

`
`ANDREA COSCELL!
`
`360
`each molecule, the daily number of milligrams of the chemical compound
`that the average patient needs is defined. Therefore any package can be
`transformed into a number of days of therapy for the average patient.
`Finally, the dummy variables for the molecule markets indicate that 74%
`of the prescriptions in the sample were for ranitidine, followed by
`omeprazo/e with 9%.
`Table IV compares the market shares in the original sample to those in
`the estimation samples. First, column (I) shows that within-molecule
`market shares differ. Second, the main estimation sample is remarkably
`similar to the original sample, while the switching sample in column (3)
`
`TABLE IV
`MARKET SHARES IN THE SAMPLE
`
`Column (I) shows market shares in the overall sample; in column (2) market shares in the
`sample used in the estimation are reported. Column (3) is a subset of the sample used in
`column (2), conditional on the patients' being prescribed a new brand. Column (4) indicates
`the expected brand frequency in the switch sample based on the total in~sample market shares
`
`Overall
`sample
`
`Estimation
`sample
`
`Estimation
`sample and
`switch
`
`Conditional
`probabilities
`
`Famotidine
`FAMODIL
`GASTRIDIN
`MOTIAX
`
`Ranitidine
`RANIBEN
`RANIBLOC
`RANIDIL
`TRIGGER
`ULCEX
`ULKOBRJN
`ZANTAC
`
`Nizatidine
`CRONIZAT
`NIZAX
`ZANIZAL
`
`Roxatidine
`GASTRALGIN
`NeoH2
`RoxiT
`
`Omeprazoie
`LOSEC
`MEPRAL
`0MEPRAZEN
`#Observations
`
`0.467
`0.448
`0.083
`
`0.02
`0.021
`0.419
`0.021
`0.029
`0.002
`0.483
`
`0.359
`0.581
`0.059
`
`0.332
`0.199
`0.468
`
`0.304
`0.352
`0.343
`75,316
`
`0.458
`0.453
`0.088
`
`0.03
`0.025
`0.432
`0.024
`0.024
`0.003
`0.467
`
`0.371
`0.566
`0.062
`
`0.292
`0.161
`0.546
`
`0.343
`0.329
`0.328
`43,840
`
`0.415
`0.415
`0.169
`
`0.07
`0.062
`0.387
`0.032
`0.033
`0.005
`0.405
`
`0.437
`0.437
`0.125
`
`0.285
`0.428
`0.286
`
`0.303
`0.37
`0.326
`1,775
`
`0.437
`0.454
`0.108
`
`0.031
`0.026
`0.435
`0.017
`0.026
`0.007
`0.454
`
`0.453
`0.466
`0.08
`
`0.3
`0.272
`