Measuring the Informative and Persuasive Roles of...

MANAGEMENT SCIENCEArticles in Advance, pp. 1–14ISSN 0025-1909 (print) � ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.1110.1499

© 2012 INFORMS

Measuring the Informative and PersuasiveRoles of Detailing on Prescribing Decisions

Andrew T. ChingRotman School of Management, University of Toronto, Toronto, Ontario, Canada M5S 3E6,

[email protected]

Masakazu IshiharaStern School of Business, New York University, New York, New York 10012,

[email protected]

In the pharmaceutical industry, measuring the importance of informative and persuasive roles of detailing iscrucial for both drug manufacturers and policy makers. However, little progress has been made in disentan-

gling these two roles of detailing in empirical research. In this paper, we provide a new identification strategyto address this problem. Our key identification assumptions are that the informative component of detailing ischemical specific and the persuasive component is brand specific. Our strategy is to focus on markets wheresome drug manufacturers engage in a comarketing agreement, under which two or more companies marketthe same chemical using their own brand names. With our identification assumptions, the variation in the rel-ative market shares of these two brands, together with their brand specific detailing efforts, would allow usto measure the persuasive component of detailing. The variation in the market shares of chemicals, and thedetailing efforts summed across brands made of the same chemical, would allow us to measure the informativecomponent of detailing. Using the data for angiotensin-converting enzyme inhibitor with diuretic in Canada, wefind evidence that our identification strategy can help disentangle these two effects. Although both effects arestatistically significant, we find that the persuasive function of detailing plays a very minor role in determiningthe demand at the chemical level—the informative role of detailing is mainly responsible for the diffusion pat-terns of chemicals. In contrast, the persuasive role of detailing plays a crucial role in determining the demandfor brands that comarket the same chemical.

Key words : detailing; informative role; persuasive role; prescription drugs; decisions under uncertainty;diffusion

History : Received June 29, 2009; accepted November 7, 2011, by Pradeep Chintagunta, marketing. Publishedonline in Articles in Advance.

1. IntroductionIn the pharmaceutical industry, measuring the impor-tance of informative and persuasive roles of detailingis crucial for both drug manufacturers and policymakers. Understanding the relative importance ofthese two roles can help drug manufacturers allocateresources to detailing more efficiently. If the persua-sive role is important, firms can create artificial prod-uct differentiation by increasing their detailing efforts.On the contrary, if detailing is mainly informativeand its persuasive role is weak, the effectiveness ofdetailing will highly depend on the actual quality ofdrugs (i.e., side effects and efficacy profiles). Amongpolicy debates, many people believe that detailing ismainly persuasive and consumers will be better offif the industry reduces their detailing budget. Con-sequently, there are frequent calls for the industry torestrict detailing activities. However, if detailing ismainly informative in nature, putting restrictions on itmight slow down the adoption rate of new innovative

drugs. This in turn would hurt manufacturers’ profitsand their incentives to innovate, and lower consumerwelfare.

Despite its importance, little progress has beenmade in disentangling the informative and persua-sive roles of detailing. The main difficulty is thatboth effects would likely have positive impacts on thedemand for prescription drugs. If one only observessales and detailing efforts over time, it is hard to dis-entangle these two roles. In this paper, we providea new identification strategy to address this prob-lem. Our key identification assumptions are that theinformative component of detailing is chemical spe-cific and the persuasive component is mainly brandspecific. Our strategy is to focus on a market wheresome drug manufacturers engage in a comarketingagreement. Under such an agreement, two compa-nies market the same chemical using two differentbrand names. With our identification assumptions,the variation in the relative market share of these

1

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.

Published online ahead of print May 18, 2012

Ching and Ishihara: Measuring the Informative and Persuasive Roles of Detailing2 Management Science, Articles in Advance, pp. 1–14, © 2012 INFORMS

two brands, together with their brand-specific detail-ing efforts, would allow us to measure the persua-sive component of detailing. After controlling for thepersuasive effect, the variation in the market shareof chemicals, and the corresponding chemical-specificdetailing efforts summed across brands made of thesame chemical, would allow us to measure the infor-mative component of detailing. For instance, if detail-ing does not have any persuasive effect at all, ourassumptions would imply that the market shares fortwo brand-name drugs made of the same chemicalshould be roughly the same over time even if thedetailing efforts are very different across these twobrands (assuming the values of other marketing-mixvariables are about the same across brands).

More specifically, to model persuasive detailing,we follow the previous literature (e.g., Nerlove andArrow 1962) and allow a brand-specific persuasivedetailing goodwill stock to enter physicians’ utilityfunctions. To model informative detailing, we followChing and Ishihara (2010), which models informativedetailing as a means to build and maintain the mea-sure of physicians who know the most updated infor-mation about drugs.

Our identification strategy applies to both prod-uct level data and individual level data. As an appli-cation, we apply it to the product level data fromthe market of angiotensin-converting enzyme (ACE)inhibitor with diuretic (which is a subclass of hyper-tension drugs) in Canada. This market has threebrand-name drugs: Vaseretic, Zestoretic, and Prinzide.Zestoretic and Prinzide are made of the same chem-icals, but are comarketed by two different compa-nies. To investigate the validity of our identificationassumptions, we estimate two versions of the model:(i) two-chemical version that captures the comarket-ing environment and assumes that Zestoretic andPrinzide share one information set; (ii) three-chemicalversion that assumes that Zestoretic and Prinzidecould be made of different chemicals, and henceeach brand has its own information set, and bothinformative and persuasive effects of detailing arebrand specific. We find that the estimation results arecounterintuitive in the three-chemical version—thepersuasive effect of detailing is negative and insignif-icant. On the contrary, the estimation results from thetwo-chemical version are much more sensible—thepersuasive effect is positive and significant. This pro-vides support for our identification assumptions.

Based on the parameter estimates from the two-chemical version of the model, we find that the per-suasive function of detailing plays a very minor rolein determining the demand at the chemical level—theinformative function of detailing is mainly responsi-ble for the diffusion patterns of chemicals. In con-trast, the persuasive function of detailing plays a

crucial role in determining the demand for brandsthat comarket the same chemical.

The rest of this paper is organized as follows.Section 2 reviews the literature and discusses thebackground of the comarketing agreement. Section 3presents the demand model. Section 4 describes thedata. Section 5 discusses the results. Section 6 pro-vides the conclusion.

2. Literature Review andComarketing Agreement

2.1. Previous Literature on Persuasive DetailingLeffler (1981) argues that detailing plays both infor-mative and persuasive roles. He finds that new drugstend to receive more detailing than older drugs, andinterprets this as evidence that supports informativedetailing. He argues that physicians are relativelyunfamiliar with new drugs and hence if detailingprovides information about drug’s benefits and sideeffects, drug manufacturers would spend more detail-ing efforts for newer drugs. However, he also findsthat drug companies continue to spend significantamount of detailing efforts on old drugs and targetolder physicians. He interprets this as evidence for itspersuasive role, assuming that older physicians havealready known the older drugs’ efficacy and side-effect profiles.

Hurwitz and Caves (1988) find that pre-patent expi-ration cumulative detailing efforts slow down thedecline in post-patent expiry market shares of brand-name drugs. They interpret this as evidence for itspersuasive role. Rizzo (1999) also finds evidence thatdetailing lowers the price elasticity of demand andargues that it supports persuasive detailing. However,it should be pointed out that the results from Hurwitzand Caves (1988) and Rizzo (1999) are also consis-tent with informative detailing. As argued by Leffler(1981), informative detailing reduces the uncertaintyabout drug qualities, and hence could also achievesimilar empirical implications.

Narayanan et al. (2005) is the first paper that struc-turally estimates informative and persuasive roles ofdetailing in the pharmaceutical market by extend-ing the framework of Erdem and Keane (1996). Theiridentification argument builds on Leffler (1981). Morespecifically, they assume that drug companies knowthe true quality of their drugs when launching them,and informative detailing provides physicians withnoisy signals about the true quality. With this assump-tion, physicians will eventually learn the true qualityand hence detailing will not play any informative rolein the long run. As a result, the long-run correlationbetween sales and cumulative detailing efforts willidentify the persuasive role of detailing. The prod-uct diffusion paths then identify the informative role.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.

Ching and Ishihara: Measuring the Informative and Persuasive Roles of DetailingManagement Science, Articles in Advance, pp. 1–14, © 2012 INFORMS 3

It should be emphasized that in their framework,to separately identify the informative and persuasiveroles of detailing, it is crucial that (i) one assumesdetailing does not play any informative role in thelong run, and (ii) the data set needs to be long enoughso that it captures part of the product life cycle afterlearning is complete.1 In contrast, these features arenot necessary for our identification strategy.

Finally, Ackerberg (2001) argues that one can empir-ically distinguish informative and persuasive effectsof advertising by examining consumers’ purchasebehavior conditional on whether they have tried theproduct before. His insight is that advertisements thatgive consumers product information should primar-ily affect consumers who have never tried the brand,whereas persuasive advertisements should affect bothinexperienced and experienced consumers. His iden-tification argument requires one to observe indi-vidual level panel data, whereas our identificationstrategy applies even if one only observes productlevel panel data.

Compared with the previous studies, the main lim-itation of our identification strategy is that comar-keting agreement only happens in a relatively smallsubset of product categories. Therefore, one should becautious about how to generalize our results.

2.2. Comarketing AgreementComarketing in the pharmaceutical industry is a mar-keting practice where a company, in addition to itsown, uses another company’s sales force to pro-mote the same chemical and allow the partner com-pany to use a different brand name.2 According toCurrentPartnering (2009), the total number of comar-keting deals announced in the United States between2000 and 2008 is 208, and the yearly number hasremained at fairly steady levels. One reason whyan originator of the drug is willing to partner withanother company could be because it requires highfixed costs to build a sales force. The sales force inthe pharmaceutical industry requires extensive train-ing as they need to know the clinical trials resultsof the drug being promoted and their rivals’ drugs.Instead of paying such a high fixed cost, a companythat is short in their sales force of promoting a certaincategory of drugs might find it worthwhile to signa comarketing agreement with another company, andcharge its partner a royalty fee.

1 Byzalov and Shachar (2004), Mehta et al. (2008), and Narayananand Manchanda (2009) rely on similar identification arguments toestimate informative and persuasive advertising or detailing usingindividual level data.2 This definition of comarketing agreement is given by Current-Partnering (2009). There is another type of closely related marketingpractice where two or more firms market the same chemical underone brand name. CurrentPartnering calls this type of arrangementcopromotion agreement.

Comarketing agreements have also appeared in theautomobile industry (Sullivan 1998, Lado et al. 2003).Furthermore, for industrial products, it is commonthat different firms market identical products usingtheir own brand names (Saunders and Watt 1979,Bernitz 1981). In some countries such as Australia andJapan, firms also market generic drugs with brandnames (Birkett 2003, Iizuka 2011). Under these envi-ronments, we expect that our identification argumentscould also be applied.3

3. ModelWe modify the model proposed by Ching andIshihara (2010) to implement our new identificationstrategy. They model informative detailing as a meansto build and maintain the measure of physicians whoknow the most updated information about drugs, butignore persuasive detailing. Here, we model persua-sive detailing by including a detailing goodwill stockin the utility function for physicians.

The basic setup of the model is as follows. We con-sider a set of brand-name drugs, which treat thesame illness using similar chemical mechanisms. Letj = 11 0 0 0 1 J indexes brands, j = 0 denotes an outsidealternative, which represents other close substitutes.Some of the brands may be marketed under a comar-keting agreement and are made of the same chemical.Let k = 11 0 0 0 1K indexes for chemicals, where K ≤ J .We assume that each brand is made of one chemical.Let Ak be the set of brands that are made of chem-ical k. The characteristics of brand j ∈ Ak are givenby pj and qk, where pj is the price of brand j , andqk is the mean quality level of chemical k. Physiciansare imperfectly informed about the chemical’s meanquality level qk. Let I4t5 = 4I14t51 0 0 0 1 IK4t55 be a vec-tor of public information sets that describe the mostupdated belief about q = 4q11 0 0 0 1 qK5 at time t. Chingand Ishihara (2010) assume that I4t5 is updated by arepresentative opinion leader based on past patients’experiences. Let I k be the initial prior that physi-cians have when a drug made of chemical k is firstintroduced. For each chemical k, a physician eitherknows Ik4t5 or I k at time t. For simplicity, we assumethat physicians and the representative opinion leadershare the same initial prior belief. Let Mkt be the mea-sure of physicians who know Ik4t5; Mkt is modeled asa function of the cumulative detailing efforts at time t.

Our key identification assumptions are that (1) infor-mative detailing is chemical specific and (2) persuasive

3 However, the applicability of our identification strategy for indus-tries other than pharmaceutical may depend on the existence ofnonproduct factors that differentiate products under a comarketingagreement (e.g., after-sales services in automobile). If consumerscare about such nonproduct factors and advertising help consumerslearn about them over time, our identification strategy would notbe applicable unless researchers can control for them.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


detailing is brand specific. The first assumption impliesthat (a) Ik4t5 is updated based on past patients’ expe-riences for all drugs made of chemical k and (b) Mkt

depends on the sum of the cumulative detailing effortsforalldrugsmadeofchemicalk.Thesecondassumptionimplies that the persuasive detailing goodwill stock forbrand j only relies on the detailing efforts for brand j .

3.1. Updating of the Information SetA drug is an experienced good. Consumption of adrug provides information about its quality. Eachpatient i’s experience with the quality of a drug madeof chemical k at time t (q̃ikt) may differ from its meanquality level qk. As argued in Ching (2010a, b), thedifference between q̃ikt and qk could be due to theidiosyncratic differences of human bodies in reactingto drugs. An experience signal may be expressed as

q̃ikt = qk + �ikt1 (1)

where �ikt is the signal noise. We assume that �ikt

is independent and identically distributed (i.i.d.). Wefurther assume that �ikt is normally distributed withzero mean, and the representative opinion leader’sinitial prior on qk4I k) is also normally distributed:

�ikt ∼N401�2� 5 and qk � I k ∼N4q

k1�2

k50 (2)

The representative opinion leader updates the pub-lic information set at the end of each period usingthe experience signals that are revealed to the pub-lic. The updating is done in a Bayesian fashion.In each period, we assume that the experience signalsrevealed to the public is a random subsample of theentire set of experience signals.

According to the Bayesian rule (DeGroot 1970), theexpected quality is updated as follows:

E[

qk � I4t+ 15]

= E[

qk � I4t5]

+ �k4t5(

q̄kt −E6qk � I4t57)

1 (3)

where q̄kt is the sample mean of all the experience sig-nals that are revealed in period t; �k4t5 is a Kalmangain coefficient, which assigns the updating weight toq̄kt . Note that both �k4t5 and the perception variance,�2k 4t + 15, are functions of the variance of the signal

noise (�2� ), perceived variance (�2

k 4t5), the quantitiessold together with free samples at time t for all drugsmade of chemical k (nk

t ), and the proportion of expe-rience signals revealed to the public (�). They can beexpressed as

�k4t5 =�2k 4t5

�2k 4t5+ 4�2

�/4�nkt 55

and

�2k 4t + 15 =

141/�2

k 4t55+ 44�nkt 5/�

2� 50 (4)

3.2. Detailing and Measure ofWell-Informed Physicians

There is a continuum of physicians with measureone. They are heterogeneous in their information sets.A physician is either well informed or uninformedabout chemical k. A well-informed physician knowsthe current information set maintained by the repre-sentative opinion leader (Ik4t5). An uninformed physi-cian only knows the initial prior (I k). The number ofphysician types is then 2K .

The measure of well-informed physicians for chem-ical k at time t, Mkt , is a function of Mkt−1 and Dt =

4D1t1 0 0 0 1DJt5, where Djt is the detailing efforts forbrand j at time t. For simplicity, we assume that thisfunction only depends on Mkt−1 and Dk

t =∑

j∈AkDjt ,

i.e., Mkt = f 4Mkt−11Dkt 5. We capture the relationship

between Mkt and (Mkt−11Dkt ) by introducing an infor-

mative detailing goodwill stock, GIkt , which accumu-

lates as follows:

GIkt = 41 −�I 5G

Ikt−1 +Dk

t 1 (5)

where �I ∈ 60117 is the depreciation rate. We specifythe relationship between Mkt and GI

kt as

Mkt =exp4�0 +�1G

Ikt5

1 + exp4�0 +�1GIkt5

0 (6)

3.3. Prescribing DecisionsEach physician’s objective is to choose a drug so asto maximize the current period expected utility forhis or her patients conditional on his or her informa-tion set and other marketing variables such as persua-sive detailing and free samples. The demand systemis obtained by aggregating this discrete choice modelof an individual physician’s behavior.

The utility of patient i who consumes drug j madeof chemical k at time t is given by

uijt = �j − exp4−r q̃ikt5−�ppjt + �i1t + �ikt + eijt1 (7)

where �j is a brand-specific intercept; r is the coeffi-cient of absolute risk aversion; �p is the utility weightfor price; (�i1t + �ikt + eijt), which represents the dis-tribution of patient heterogeneity, is unobserved tothe econometrician but observed to the physicianswhen they make their prescribing decisions; and �iltcorresponds to the shock associated with the outsidealternative (l = 0) or inside alternatives (l = 1). Thissetup is equivalent to modeling physicians’ choiceas a three-stage nested process, where they choosebetween the inside goods and the outside good in thefirst stage (when �ilt is realized), then choose one ofthe chemicals in the second stage (when �ikt is real-ized), and then choose a brand in the third stage(when eijt is realized) if the chemical is comarketed bytwo or more firms. We assume that �ilt, �ikt, and eijt arei.i.d. extreme value distributed.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Note that q̃ikt is observed by physicians and patientsonly after patients have consumed the drug (butremains unobserved by the econometrician). Thus,physicians make their prescribing decisions based onthe expected utility of their patients. Let Ih4t5 denotephysician h’s information set at time t. Suppose thatdrug j is made of chemical k. If physician h is wellinformed about chemical k at time t, then Ihj 4t5= Ik4t5and his or her expected utility will be

E6uijt � Ih4t57= E6uijt � Ik4t57+�PG

Pjt +�SF Sjt

= �j − exp(

−rE6qk � I4t57+ 12 r

24�2k 4t5+�2

� 5)

−�ppjt

+�PGPjt +�SF Sjt + �i1t + �ikt + eijt1 (8)

where GPjt is a persuasive detailing goodwill stock for

drug j at time t with the depreciation rate �P , and �P

captures the effect of persuasive detailing; F Sjt is theamount of free samples given for drug j at time t, and�S captures the effect of free samples. If physician his uninformed about chemical k at time t, his or herexpected utility follows the same functional form asin Equation (8) except that Ihj 4t5 = I k. We emphasizethat (a) GP

jt is drug j specific rather than chemical kspecific, and (b) the depreciation rates for GI

kt and GPjt

are allowed to be different.In each period, physicians may also choose an out-

side alternative (i.e., other nonbioequivalent drugs).We assume the expected utility associated with theoutside alternative takes the following functionalform:

E6ui0t � Ih4t57= �0 +�tt + �i0t + �i0t + ei0t0 (9)

The time trend of the outside alternative allows themodel to explain why the total demand for insidegoods may increase or decrease over time. With theabove setup, we can construct the market shares in astandard way. The quantity demanded for drug j (njt)can then be expressed as

njt = Sizet · S(

j �Dt1 4E6qk � I4t571�k4t51Mkt−15Kk=13�d

)

+ �jt1 (10)

where Sizet is the size of the market at time t, S4j � ·5 isthe market share of drug j , �jt represents a measure-ment error, and �d is a set of demand side parameters.

3.4. IdentificationIt should be highlighted that if K = J (i.e., each brandcorresponds to a distinct chemical), the parametersof the informative and persuasive effects will mainlybe identified based on the functional form restric-tions. This is because the measure of well-informedphysicians (which is the main driver for the infor-mative effect), similar to the persuasive effect, is also

governed by a detailing goodwill stock. As a result,some empirical patterns (e.g., increasing sales trend)could be explained by either informative or persua-sive detailing. If the functional form assumptionsspecified here precisely capture the true nonlinearnature of these two effects, we can still separatelyidentify them and obtain consistent estimates in prin-ciples. Under this situation, however, the main sourceof data variation for identification is only the diffu-sion patterns of each brand.

If one has data from a market where some drugsuse a comarketing agreement (i.e., K < J ), one willbe able to use an additional source of data variationto help identify the persuasive effect: When two ormore companies use their own brand names to mar-ket the same chemical, our identification assumptionsimply that the variation of their relative market sharesand their relative detailing efforts would be mainlyresponsible for identifying the persuasive effect ofdetailing. For instance, if the persuasive effect is closeto zero, we expect to see that the relative marketshares should remain roughly the same across brandsthat are made of the same chemical even if their rel-ative detailing efforts vary significantly over time.On the contrary, if the data shows that their relativemarket shares are highly positively correlated withtheir relative detailing efforts, this tells us that the per-suasive effect is strong and positive. With this addi-tional source of data variation, we can control for thepersuasive effect and use the diffusion paths of eachdrug to identify the informative effect (i.e., the param-eters of learning process and initial prior beliefs).

We should reemphasize that the traditional identi-fication argument also relies on two sources of datavariation to disentangle informative and persuasiveeffects. It requires one (i) to have a sufficient numberof observations of sales and detailing efforts in thelong run in order to identify the persuasive effect, and(ii) to use the diffusion patterns of brands to identifythe informative effect after controlling for the persua-sive effect. Nevertheless, such a long panel may notbe readily available. Under this situation, having datafrom markets with comarketing arrangement will beparticularly helpful in identifying these two effects.

We should note that like all structural estimationresearch, our results still need to rely on functionalform assumptions (Keane 2010). But with the extrasource of data variation provided by the comarketingenvironment, we should be able to identify informa-tive and persuasive effects more accurately comparedwith the traditional approach. It is also importantto recognize that our specification ignores one pos-sible function of detailing—it might increase physi-cians’ awareness of some brands sharing the samechemicals. If this function is important, our estimateswould suffer misspecification bias. In particular, the

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


importance of the persuasive effect would be over-estimated.4 Admittedly, with only product level data,it is very difficult to empirically tease out the impor-tance of this function from other potential structuralexplanations of slow diffusion. Nevertheless, we feelthat simply informing physicians that two brandsare made of the same chemical is relatively easyto accomplish. We also find that such informationis not hard for physicians to obtain. For instance,Physicians’ Desk Reference (a standard reference aboutdrugs) shows that Prinzide is another brand nameof Zestoretic right under its heading, and vice versa.Many popular online resources also provide thisinformation in the first few lines of their search results(e.g., http://www.MedicineNet.com). So relative tothe information about the chemical (its efficacy andside-effect profile), it seems much more likely thatphysicians know which brands are made of the samechemical. These institutional details suggest that theawareness issue should be of second-order impor-tance in our context. We therefore decide not to modelthis alternative explanation and leave it for futureresearch. Nevertheless, in §5, we will compare ourproposed framework with an alternative specification,which assumes the demand side does not know thatthere are two brands that comarket the same chemicalunder different brand names throughout the analysis.This allows us to shed some light on the validity ofour identification assumptions.

4. Data DescriptionWe apply our identification strategy to the market ofACE inhibitor with diuretic in Canada. This class ofcombination drugs is for treating hypertension. Datacome from IMS Canada. The revenue data is drawnfrom their Canadian Drugstore and Hospital Audit(D&H), the number of prescriptions is drawn fromtheir Canadian Compuscript Audit (CCA), the detail-ing minutes and free sample data are drawn fromtheir Canadian Promotion Audit (CPA). AlthoughD&H does not include purchases made by govern-ment, mail order pharmacies, nursing homes, or clin-ics, it covers more than 95% of the total sales.

The data set contains monthly data from March1993 to February 1999. There are three drugs inthe market—Vaseretic, Zestoretic, and Prinzide. Allof them are present throughout the sample period.Treating product/quarter as one observation, the totalsample size is 216. Vaseretic is marketed by Merck, itsgeneric ingredients are enalapril and hydrochloroth-iazide. It was approved by Health Canada in Septem-ber 1990. Zestoretic is marketed by AstraZeneca, itsgeneric ingredients are lisinopril and hydrochloroth-iazide. It was approved in October 1992. Interestingly,

4 The direction of bias for the informative effect is ambiguous.

Merck is the originator of lisinopril, and it signeda comarketing agreement with AstraZeneca. Merckalso markets lisinopril hydrochlorothiazide under thebrand name Prinzide. In other words, Zestoreticand Prinzide are made of exactly the same chemi-cals. Because Vaseretic was launched earlier, we con-sider Vaseretic as the incumbent, and Zestoretic andPrinzide as new entrants. The potential market size isdefined as the total number of prescriptions for drugsthat belong to ACE inhibitor, Thiazide Diuretic, andACE inhibitor with diuretic. It increases from 655,000to 860,000 during the sample period.

Table 1 shows the summary statistics. Figure 1shows the detailing minutes for the three drugs overtime. One common feature is that they all havehigh fluctuation. The detailing minutes for Vasereticand Zestoretic are roughly the same for the first30 months, but for the later period, Zestoretic on aver-age details more than Vaseretic. In general, Prinzidedetails much less than Zestoretic. Figure 2 shows thenumber of prescriptions dispensed in this market. Thesales for all three brands continue to increase evennear the end of our sample period. Being the firstin this market, Vaseretic controlled more than 80% ofthe sales at the beginning of the sample; Zestoretic’s

Table 1 Summary Statistics

Brand Mean SD Max Min

Number of prescriptionsVaseretic 4100706 67608 51446 21429Zestoretic 6138808 4190003 161330 322Prinzide 1181408 1116809 41447 131

Detailing minutesVaseretic 1103208 68901 31240 97Zestoretic 1162504 82806 41203 93Prinzide 51206 65007 31566 0

Free sample (no. of prescriptions)a

Vaseretic 7108 5208 29008 0Zestoretic 15205 10001 54504 0Prinzide 2008 2400 8301 0

PriceVaseretic 4005 8076 6902 2405Zestoretic 3403 8065 6105 1507Prinzide 3807 1506 8705 1602

aThe original data on free samples are measured in sample extended units:the number of packages multiplied by the number of pills per package.To incorporate the effect of free samples on the information updating processas part of consumption experience signals, we need to convert the sampleextended units into the number of prescriptions. Following Tu et al. (2005),we assume that one prescription lasts for 100 days. Based on the dailydosages of Vaseretic and Zestoretic/Prinzide, we set the daily consumptionto be 2.25 units for Vaseretic and two units for Zestoretic/Prinzide. The dailyconsumption times 100 would give us the amount of the sample extendedunits per prescription. It turns out that free samples represent less than 12%of the total number of prescriptions and hence it has negligible impacts onconsumer learning.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Figure 1 Detail Minutes vs. Time

0

500

1,000

1,500

2,000

2,500

3,000

3,500

4,000

4,500

0 10 20 30 40 50 60 70

Det

ailin

g m

inut

es

Time (1 = March 1993; 72 = February 1999)

Vaseretic

Zestoretic

Prinzide

Figure 2 Predicted and Actual Demand

0

2,000

4,000

6,000

8,000

10,000

12,000

14,000

16,000

18,000

0 10 20 30 40 50 60 70

Num

ber

of p

resc

ript

ions


Vaseretic (data)

Zestoretic (data)

Prinzide (data)

Vaseretic (predicted)

Zestoretic (predicted)

Prinzide (predicted)

Zestoretic

Vaseretic

Prinzide

share was only about 10%; Prinzide’s share was evensmaller (about 5%). It takes Zestoretic more than twoyears before it overtakes Vaseretic’s sales. However,Prinzide’s sales remain below Zestoretic through-out the period, even though Prinzide and Zestoreticare made of the same chemicals. The differences inthe number of prescriptions and detailing efforts forZestoretic and Prinzide indicate that the persuasiverole of detailing is likely important.

Our identification strategy requires the informa-tive component of detailing to be chemical specific.One implication of this assumption is that when pos-itive information about a chemical becomes available,the sales of a brand should be positively influ-enced by the detailing efforts of the other brand

that comarkets the same chemical. We conduct asimple test for this implication. Following Chingand Ishihara (2010), we collect data on clinical trialsthat compare the efficacy of Zestoretic/Prinzide andVaseretic to create a measure of cumulative clin-ical outcome variable, and use it as a proxy forthe updated information sets.5 For each clinical trial,the relative outcome could either be (i) positivefor Vaseretic (i.e., negative for Zestoretic/Prinzide);(ii) positive for Zestoretic/Prinzide (i.e., negative forVaseretic); or (iii) no difference between Vaseretic and

5 The published clinical trial data is obtained from PubMed(http://www.pubmed.gov).

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Zestoretic/Prinzide. We then create a cumulative out-come variable for each chemical as follows. For eachclinical trial, we code its outcome as +1 (positive),−1 (negative), and 0 (no difference), and compute acumulative measure. We should also point out thatall the medical journal publications mention chemicalnames instead of brand names when they report clin-ical trial findings. Therefore, if sales representativesrefer physicians to some medical journal publicationsfor evidence to support their claims, physicians willsee the chemical name in those articles, suggestingthat the informative detailing should have a spillovereffect across brands that share the same chemical.

To test whether such a spillover effect exists, weregress the number of prescriptions on the inter-action between the cumulative clinical outcomes ofthe chemical and cumulative comarketing partner’sdetailing, controlling for other factors, and assumingthe coefficients are the same across all brands. Whencreating the cumulative detailing stock, we set thedepreciation rate to be 4.2% as in Berndt et al. (1997).6

We report two specifications here. Table 2 shows thespecifications and results. In both specifications, wefind that the interaction between the cumulative clin-ical outcomes and comarketing partner’s cumulativedetailing to be positive and statistically significant.This provides support for our hypothesis that thereis a spillover effect of informative detailing for drugsthat are comarketed.

In addition to the regressors used in specifica-tion (i), specification (ii) includes comarketing part-ner’s cumulative detailing stock. This variable allowsus to test whether there is any additional spillovereffect of detailing from a comarketing partner aftercontrolling for its interaction with the cumulativeclinical outcomes. According to the CI model, addi-tional informative spillover effect is possible becausedetailing could trigger or stimulate physicians to findout more information about the drug themselves bycontacting opinion leaders of the field (captured bythe representative opinion leader in the model). Thisallows them to learn about the patients’ consumptionexperiences revealed to the public, in addition to theinformation signals available from the published clin-ical trials. However, at the same time, we expect thatthe partner’s cumulative detailing stock also capturesthe brand-specific persuasive effect, which shouldhave a negative impact on the number of prescriptionof the focal drug. Overall, when combining these twocounteracting effects, we expect that the effect of thecomarketing partner’s cumulative detailing should besmaller than that of one’s own cumulative detailing

6 The depreciation rate of detailing stock estimated by Ching andIshihara (2010) is 4.5%, which is very close to what Berndt et al.(1997) find.

Table 2 Ordinary Least Squares Regression of the Number ofPrescriptions on the Interaction Between the CumulativeClinical Outcomes and Comarketing Partner’s CumulativeDetailing

SpecificationDependent variable:Number of prescriptionsjtVariable (i) (ii)

Pricejt −4056 −4068

470635 470415Cum_Clinicaljt − 5186201 −61212084

4370045 4371095Cum_Detjt 00183 00241

4000135 4000205Cum_Detjt ×Cum_Clinicaljt 00215 00215

4000125 4000125Cum_Det_Partnerjt 00062

4000175Cum_Det_Partnerjt ×Cum_Clinicaljt 00148 00144

4000085 4000085Constant −1112005 −3106004

4444035 4678045

Adjusted R-squared 00871 00878No. of observations 216 216

Notes. Standard errors are in parentheses. Estimates shown in bold are sig-nificant at the 5% level. Definition of variables are as follows:Cum_Detjt : cumulative detailing minutes for brand j at time t .

—We follow Berndt et al. (1997) and set the depreciation rate at 4.2%.Cum_Det_Partnerjt : cumulative detailing minutes for brand j ’s partner at

time t .—If j is Zestoretic (Prinzide), Cum_Det_Partnerjt = Cum_Det for

Prinzide (Zestoretic) at time t .—If j is Vaseretic, Cum_Det_Partnerjt = 0.

Pricejt : price of drug j at time t .Cum_Clinicaljt : cumulative outcomes of direct comparison clinical trials for

brand j ’s chemical at time t .—For Zestoretic (j) and Prinzide (l), Cum_Clinicaljt = Cum_Clinicallt .

(and its sign could be either positive or negative).This is consistent with our estimation results for spec-ification (ii)—we find that the comarketing partner’scumulative detailing stock is positive and significant,and is much smaller than own cumulative detailing stock(0.062 versus 0.241). The evidence provides supportfor using the Ching and Ishihara (2010) approachto model consumer learning and informativedetailing.

5. ResultsWe estimate the models using the method of sim-ulated maximum likelihood. We apply the pseudo-policy function approach proposed by Ching (2010b)to control for the potential endogeneity problem ofdetailing. In the appendix, we present the spec-ifications of the pseudo-detailing policy functions.They are similar to the one used in Ching and

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Table 3 Parameter Estimates

Three-chemical Two-chemical

Estimates SE Estimates SE

Learning parameters� 2� 00620 00296 00174 00025q

1−2006 5047 −2402 3096

q2

−1809 4012 −1406 4083q

3−1709 3086

� 2 00320 00142 00166 00025q1 1 1q2 6200 7027 3609 6029� 1/30,000 1/30,000

Preference parameters�0 0 0�1 −3030 00195 −3069 1080�2 −3078 00294 −4026 1020�3 −3091 00107 −4027 1037r 00034 00001 00031 00005�p 6.71E−05 1.01E−04 4.43E−05 1.26E−05�t −00011 00001 −00014 00004�P −4090E−06 4.72E−06 9.48E−07 9.26E−08�S −4.56E−08 1.34E−08 −5.59E−09 4.74E−09

Detailing stock parametersêP 00010 00005 00084 00015êI 00013 00001 00013 00003�0 −2022 00127 −1020 00381�1 4.01E−05 3.41E−06 3.07E−05 7.62E−06

Other parameters for error termsSD (�) 17508 9067 17008 3706SD (�) 1 1SD (�) 00264 00020 00116 00031SD (e) 00024 00004

Log likelihood −2149008 −2150000

Notes. Estimates shown in bold are significant at the 5% level. Brands (j): 1—Vaseretic (incumbent); 2—Zestoretic(entrant); 3—Prinzide (entrant). q1: quality for Vaseretic; q2: quality for Zestoretic and Prinzide (in the two-chemicalversion).

Ishihara (2010). We also use their procedure to handlethe initial conditions problem.7

5.1. Parameter EstimatesIn our specification, we treat Vaseretic, Zestoretic,and Prinzide as inside goods. We combine all otherdrugs that belong to ACE inhibitor with diuretic, ACEinhibitor, and Thiazide Diuretic as the outside good.Brand 1 is Vaseretic, brand 2 is Zestoretic, and brand 3is Prinzide. Let q1 be the quality for Vaseretic, andlet q2 be the quality for Zestoretic and Prinzide. Foridentification reasons, we need to normalize q1, thescaling parameters for the number of consumptionexperience signals, �, and the intercept term for theutility of the outside good, �0. We set q1 = 1, � =

1/301000, and �0 = 0.

7 Because our focus is on identifying the informative and persuasiveroles of detailing, instead of the relative importance of differentsources of information, we do not include the data on clinical trialsoutcomes when estimating the structural model here. Ching andIshihara (2010) show how to incorporate data on clinical trials inthe learning process.

We estimate two versions of the model: One makesuse of the comarketing identification argument andthe other does not. More specifically, in the versionthat uses the comarketing identification argument, weassume that the demand side knows Zestoretic andPrinzide are made of the same chemical, and thusthe information sets for the two brands are identicalfor all time periods. We refer to this version as thetwo-chemical version. In the version that does not usethe comarketing identification argument, we assumethe demand side does not know that Zestoretic andPrinzide are made of the same chemical, and thusthe updating of each brand’s information set is basedsolely on the past experience signals revealed fromthat brand. But we still maintain the assumption thatthe true mean qualities for Zestoretic and Prinzideare the same. We refer to this version as the three-chemical version.

Parameter estimates are reported in Table 3. Otherthan the persuasive effect of detailing (�P ), most ofthe parameter estimates appear to be qualitatively

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


similar for both versions. The time trend of the out-side good (�t) is negative and significant, indicatingthat the value of the outside good relative to insidegoods is declining over time. This is consistent withthe continuous expansion of the demand for Vaseretic,Zestoretic, and Prinzide. The parameter estimates forthe true mean quality and initial priors are all sta-tistically significant. The true mean quality of thechemical for Zestoretic and Prinzide (q2) is higherthan that of the chemical for Vaseretic (q1). The ini-tial prior mean qualities of both chemicals are lowerthan their true mean qualities. Most of the preferenceparameters are significant and have the right sign.Note that the price coefficient (�p) is significant inthe two-chemical version, but very small; moreover,it is insignificant in the three-chemical version. This isnot surprising because Canada provides prescriptiondrug coverage to patients who are 60 or older, andmost patients who have hypertension are the elderly.We find that the effects of free samples (�S) are neg-ative in both versions and significant in the three-chemical version. This suggests that some physiciansmay use free samples as a substitute of writing pre-scriptions. But we should note that the impact of freesamples is very small.

The parameter that appears to be estimated verydifferently across these two versions is the persua-sive effect of detailing (�P ). It is negative and insignif-icant in the three-chemical version, whereas it ispositive and significant in the two-chemical version.The result in the three-chemical version contradictsthe conventional beliefs that the persuasive effect ofdetailing is present (see, e.g., The Economist 2003).Why is the estimated persuasive effect negative andinsignificant in the three-chemical version? As dis-cussed earlier, the identification of informative andpersuasive effects in the three-chemical version ismainly achieved by the functional form assumption,and relies on one source of data variation—the dif-fusion patterns of brands. For this data set and thefunctional forms chosen here, the model can fit thediffusion patterns well without relying on the persua-sive effect.8

Recall that the two-chemical version makes useof an extra source of data variation to help iden-tify the model: The persuasive effect is identified bythe correlation between the relative market shares ofZestoretic and Prinzide and their relative cumulativedetailing efforts. After controlling for the persuasiveeffect, the informative effect is identified by the cor-relation between the relative market share of chem-

8 Note that this identification problem is not specific to the waywe model informative detailing. We obtain similar results whenwe follow Narayanan et al. (2005) and model informative detailingas noisy signals of the true quality. The results are available uponrequest.

Table 4 Goodness of Fit: Mean AbsolutePercentage Error


Vaseretic 00049 00068Zestoretic 00077 00083Prinzide 00234 00146

Total 00360 00297

icals and the chemical specific detailing efforts. Themore sensible estimate of the persuasive effect in thetwo-chemical version provides support for our comar-keting identification assumption, and demonstratesthat our estimation strategy uses the information inthe data more accurately when estimating these twoeffects of detailing.

To show the goodness of fit of the model, we simu-late 5,000 sequences of quantity demanded (expressedin terms of number of prescriptions) for Vaseretic,Zestoretic, and Prinzide, where each sequence cor-responds to a simulated sequence of 4E6q � I4t57s5Tt=1obtained from the demand model and Equation (3).We simulate the model from the inception dateof Vaseretic (the first ACE inhibitor with diuretic).Because we do not observe the data between theinception dates of the drugs and the first period inour data, we need to impute the missing values ofdetailing, free sample, price, and size of the market.We follow the procedure explained in §4.4 of Chingand Ishihara (2010).

To investigate the goodness of fit for the two- andthree-chemical versions, Table 4 reports the meanabsolute percentage error by comparing the simu-lated demand and actual demand for three brands.It shows that the two-chemical version provides bet-ter overall goodness of fit (0.297 versus 0.360). Thisprovides additional support for the two-chemical ver-sion, which assumes that the demand side knowswhich brands engage in a comarketing agreement.In Figure 2, we graphically show the goodness of fitfor the two-chemical version. It shows that the modelis able to fit the data quite well.

5.2. Quantifying the Importance of Informativeand Persuasive Detailing

In this subsection, we examine the economic impor-tance of informative and persuasive detailing. In par-ticular, we are interested in investigating how thedemand for each brand as well as the total marketdemand change when we eliminate (1) the informa-tive function of detailing and (2) the persuasive func-tion of detailing. We use the two-chemical version toconduct this simulation exercise.

We first examine the importance of informativedetailing. The left-hand side panel of Figure 3 shows

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Figure 3 Importance of Informative and Persuasive Detailing

No persuasive effectVaseretic

0 20 40 60 0 20 40 60

Zestoretic

0 20 40 600 20 40 60

Num

ber

of p

resc

ript

ions

02,0004,0006,0008,000

10,00012,00014,00016,00018,000

02,0004,0006,0008,000

10,00012,00014,00016,00018,000

Prinzide


0

1,000

2,000

3,000

4,000

5,000

6,000

0

1,000

2,000

3,000

4,000

5,000

6,000

7,000

8,000

No informative effect

BaseNo informative effect

0

1,000

2,000

3,000

4,000

5,000

6,000

7,000

8,000

0 20 40 60

BaseNo persuasive effect

0 20 40 600

1,000

2,000

3,000

4,000

5,000

6,000

7,000

the results. To simulate the demand without infor-mative detailing, we set �1 = 0. We simulate 5,000sequences of quantity demanded for Vaseretic,Zestoretic, and Prinzide with and without informa-tive detailing and compare their average predictedquantities taking the data on detailing, free sam-ples, and price as given. We see that the averagepredicted quantities decrease because of the elimina-tion of informative detailing except at the later peri-ods for Vaseretic. The main driving force behind thiscounterfactual exercise is that the measure of well-informed physicians effectively stays at a very lowlevel (determined by �0) over time. In the earlier peri-ods, Vaseretic, being the incumbent, is mainly com-peting with the outside alternative. As a result, thiscreates an immediate negative impact on its numberof prescriptions. Note that the time trend of the out-side alternative is negative. As a result, the demandfor the inside alternatives still increases over timein this counterfactual exercise. It turns out that thedemand for Vaseretic without the informative func-tion exceeds that under the base case in the later peri-

ods. This is because there are very few physicianswho know Zestoretic/Prinzide has higher qualitythan Vaseretic. On the contrary, the magnitudes of thenegative impact on Zestoretic and Prinzide increaseover time. In the base case, the predicted total numberof prescriptions for Zestoretic and Prinzide is roughly18,800 at the end of our sample period. After elimi-nating the informative function of detailing, their pre-dicted total number of prescriptions drops to 7,700.

We next consider the importance of persuasivedetailing. The simulation exercise is done in a similarfashion as above. To simulate the demand with-out persuasive detailing, we set �P = 0. The right-hand side panel of Figure 3 displays the results.It shows that the decrease in demand for Vasereticis almost zero, and many physicians switch fromZestoretic to Prinzide, causing the demand forZestoretic to decrease and the demand for Prinzide toincrease. Interestingly, the total combined demand forZestoretic and Prinzide has hardly changed.

Overall, our results show that even though the per-suasive effect is statistically significant, and plays an

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


important role in determining the relative demand ofbrands that share the same chemical, it plays a veryminor role for the demand at the chemical level—thediffusion patterns of chemicals are mainly explainedby the informative effect. The reasons why the per-suasive effect has very little influence on the demandat the chemical level is because its impact on physi-cians’ expected utility is small relative to the impactdue to the informative effect. However, conditioningon choosing a chemical being comarketed by morethan one brand name, physicians’ decisions on whichbrand to prescribe no longer depend on the infor-mative role of detailing. Thus, the persuasive effecthas a much stronger influence on the comarketingbrands. It is important to note that our models do notassume the diffusion patterns of chemicals are mainlydriven by the informative role of detailing a priori.For instance, the data could reveal that the persuasiveeffect is so strong that it explains not only the brandswitching behavior for comarketing brands, but alsomost of the diffusion patterns at the chemical level.

6. ConclusionIn this paper, we propose a new identification strategyfor measuring the informative and persuasive rolesof detailing. Our identification argument makes useof time series properties of sales and detailing effortsfor markets where some brands are marketed undera comarketing agreement. Using the data on ACEinhibitor with diuretic in Canada, we find that boththe informative and persuasive roles of detailing arestatistically significant in this market. By simulatingour model, we show that the persuasive role is mainlyresponsible for brand switching for brands that sharethe same chemicals. However, the persuasive effectplays a very small role in explaining the diffusion pat-terns at the chemical level—the informative role ofdetailing is mainly responsible for this.

Our results could have important implications forboth policy makers and drug manufacturers. Oneimplication is that if we follow some policy advo-cates’ suggestions and limit the amount of detailingdone by drug manufacturers, this may slow downthe rate of learning for physicians significantly. As aresult, physicians may make less informed decisionsfor their patients. Another implication for drug man-ufacturers is that there is an informational externalityproblem for companies that engage in a comarketingagreement. This suggests that when they structure thedetails of a comarketing agreement, it is important totake this externality into account. Our proposed iden-tification strategy potentially allows drug companiesto quantify the values of the externality.

Finally, we note two limitations of our study. First,our results only rely on one subclass of drugs. In the

future, it would be important to examine whetherthe quantitative results obtained here are robust byapplying our identification strategy to more classes ofdrugs. Second, the choice of comarketing agreement isendogenous. It is possible that the firm that decides tolicense the drug (i) may be constrained by the numberof sales persons employed, or (ii) has a much weakersales force in marketing the therapeutic class to whichthe drug belongs. The former reason should not posea problem in affecting the parameter estimates, butthe latter one could because our econometric speci-fication essentially assumes away the potential het-erogeneity in the efficiency of sales force. However,in our application, Zestoretic and Prinzide are mar-keted by AstraZeneca and Merck, respectively, andboth drug companies are very well established in theindustry. We feel that their sales force training shouldbe fairly similar and hence the heterogeneity of theirsales force quality may not be a serious concern.Investigating how companies choose their partners tocomarket products and its implications on our identi-fication argument will also be an important topic forfuture research.

AcknowledgmentsThe authors thank Pradeep Chintagunta, the associate edi-tor, and three anonymous referees of Management Science fortheir constructive comments. The authors also thank AbeDunn; Jean Eid; Avi Goldfarb; Nitin Mehta; and seminarparticipants at the University of Toronto, New York Univer-sity, Bureau of Economic Analysis, Conference on HealthEconomics and the Pharmaceutical Industry in Toulouse,2008 International Industrial Organization Conference, 2008North American Econometric Society Summer Meeting,2009 American Economic Association Annual Meeting, 2009Canadian Economic Association Annual Conference, 2009Marketing Science Conference, 2010 Invitational ChoiceSymposium, and 2010 Canadian Health Economics StudyGroup Meeting for their helpful feedback. All remainingerrors are those of the authors. The authors are gratefulto CurrentPartnering for sharing their report on comarket-ing agreement. The authors acknowledge the financial sup-port provided by the Michael Lee-Chin Family Institute forCorporate Citizenship at Rotman School of Management.Andrew Ching also acknowledges the financial supportfrom the Social Sciences and Humanities Research Councilof Canada.

Appendix. Controlling for EndogeneityProblem of DetailingTo control for the potential endogeneity problem of detail-ing, we apply the pseudo-policy function approach pro-posed by Ching (2010b). To use his method, we approximatemanufacturers’ detailing policy functions, and jointly esti-mate them together with the demand model. In our model,the state variables consist of 4E6qk � I4t571�2

k 4t51Mkt−15 ∀k.In addition, we include an instrumental variable in thepseudo-detailing policy functions. We follow Ching and

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Table A.1 Parameter Estimates for Pseudo-Detailing Policy Functions


Estimates SE Estimates SE

�10 6002 2020 6050 1041�11 3052 1000 2054 00986�12 −00430 2009 −4001 4010�13 −8096 1000 −1309 3057�14 3030 5700 5044 3200�15 00113 00712 −00006 00141�20 7021 00969 7024 00308�21 14047 3209 −1306 9032�22 −22600 9601 14605 6054�23 −8304 12004 1809 1604�24 64700 30108 −38308 1502�25 00027 00111 00027 00041�30 1209 2016 1003 1008�31 14406 10308 16404 5091�32 9803 12508 −15702 7054�33 −1173209 1126008 −29304 1006�34 11103 39504 42003 1707�35 −00244 00712 −00297 00112s.d.(�) 1058 00097 1052 00025

Notes. Estimates shown in bold are significant at the 5% level. Brands (j):1—Vaseretic (incumbent); 2—Zestoretic (entrant); 3—Prinzide (entrant).

Ishihara (2010) and define Fjt as the total detailing min-utes for the set of drugs in the cardiovascular category,which are produced by the manufacturer of brand j , but notexplicit substitutes for ACE inhibitors with diuretics. Thespecifications of the pseudo-policy functions are discussedbelow. In Table A.1, we report the estimates for the pseudo-detailing policy functions that correspond to the demand-side estimates in Table 3.

In the two-chemical version, we assume the researcherknows that Zestoretic and Prinzide are made of the samechemical. Let us index Vaseretic, Zestoretic, and Prinzide asbrand 1, 2, and 3, respectively. Also, let us index the chem-ical for Vaseretic as chemical 1 (k = 1), and the chemical forZestoretic and Prinzide as chemical 2 (k = 2). Let us definethe following variables:

ãuqkk′t = E6u

qkt � I4t57−E6u

qk′t � I4t571

E6uqkt � I4t57 = −exp

(

−rE6qk � I4t57+ 12 r

24�2k 4t5+�2

� 5)

0

Note that E6uqkt � I4t57 is part of the expected utility that

depends on E6qk � I4t57 and �2k 4t5; ãu

qkk′t is the difference

between this partial expected utility from choosing chemicalk and k′.

Let 4 · 5 be an indicator function. We specify the pseudo-detailing policy function for Vaseretic (j = 1) as

logD1t = �10 +4�11 +�13 ·M2t5·41−M1t5·�ãuq12t�· 4ãu

q12t>05

+ 4�12 +�14 ·M2t5 ·M1t · �ãuq12t� · 4ãu

q12t < 05

+�15 · F1t + �1t1

where �jt is a prediction error. The pseudo-detailing policyfunctions for Zestoretic and Prinzide (j = 213) are speci-fied as

logDjt = �j0 +4�j1 +�j3 ·M1t5·41−M2t5·�ãuq21t�· 4ãu

q21t>05

+ 4�j2 +�j4 ·M1t5 ·M2t · �ãuq21t� · 4ãu

q21t < 05

+�j5 · Fjt + �jt 0

In the three-chemical version, we assume the researcherdoes not know that Zestoretic and Prinzide are made ofthe same chemical. Let us index Vaseretic, Zestoretic, andPrinzide as brand 1, 2, and 3, respectively. Also, let usindex the chemicals for Vaseretic, Zestoretic, and Prinzideas chemical 1, 2, and 3, respectively. For j = 11213, define

k∗

j = arg maxk∈8112139\8j9

E6uqkt � I4t570

We specify the pseudo-detailing policy function forbrand j as

logDjt = �j0 +4�j1 +�j3 ·Mk∗j t5·41−Mjt5·�ãu

q

jk∗j t�· 4ãu

q

jk∗j t>05

+ 4�j2 +�j4 ·Mk∗j t5 ·Mjt · �ãu

q

jk∗j t� · 4ãu

q

jk∗j t< 05

+�j5 · Fjt + �jt 0

In this specification, each firm’s detailing policy functiondepends on the strongest competitor’s state. This simplify-ing approach offers the advantage that we do not need toincrease the number of parameters. Yet the above specifi-cation still captures that firms care about the states of allthe rivals in the sense that they need to find out who theirstrongest competitor is.

ReferencesAckerberg, D. A. 2001. Empirically distinguishing informative and

prestige effects of advertising. RAND J. Econom. 32(2) 100–118.Berndt, E., L. T. Bui, D. H. Reiley, G. L. Urban. 1997. The role of mar-

keting, product quality, and price competition in the growthand composition of the U.S. anti-ulcer drug industry. The Eco-nomics of New Goods, NBER Studies in Income and Wealth, Vol. 58.University of Chicago Press, Chicago, 277–322.

Bernitz, U. 1981. Brand differentiation between identical products:An analysis from a consumer law viewpoint. J. Consumer Policy5(1–2) 21–38.

Birkett, D. J. 2003. Generics—equal or not? Australian Prescriber26(4) 85–87.

Byzalov, D., R. Shachar. 2004. The risk reduction role of advertising.Quant. Marketing Econom. 2(4) 283–320.

Ching, A. T. 2010a. A dynamic oligopoly structural model forthe prescription drug market after patent expiration. Internat.Econom. Rev. 51(4) 1175–1207.

Ching, A. T. 2010b. Consumer learning and heterogeneity: Dynam-ics of demand for prescription drugs after patent expiration.Internat. J. Indust. Organ. 28(6) 619–638.

Ching, A. T., M. Ishihara. 2010. The effects of detailing on pre-scribing decisions under quality uncertainty. Quant. MarketingEconom. 8(2) 123–165.

CurrentPartnering. 2009. Co-promotion and Co-marketing Agreementsin Pharma, Biotech, and Diagnostics. CurrentPartnering, a divi-sion of Wildwood Ventures Limited, York, UK.

DeGroot, M. H. 1970. Optimal Statistical Decisions. McGraw-Hill,New York.

Economist, The. 2003. Pushing pills: Marketing drugs to doctorsis turning into a tricky business. (February 13), http://www.economist.com/node/1580138.

Erdem, T., M. P. Keane. 1996. Decision making under uncertainty:Capturing dynamic brand choice processes in turbulent con-sumer goods markets. Marketing Sci. 15(1) 1–20.

Hurwitz, M. A., R. E. Caves. 1988. Persuasion or information? Pro-motion and the shares of brand name and generic pharmaceu-ticals. J. Law Econom. 31(2) 299–320.

Iizuka, T. 2011. Physician agency and adoption of generic pharma-ceuticals. Working paper, University of Tokyo, Tokyo.

Keane, M. P. 2010. Structural vs. atheoretic approaches to econo-metrics. J. Econometrics 156(1) 3–20.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Lado, N., O. Licandro, F. Perez. 2003. How brand names affectthe price setting of car makers producing twin cars? Workingpaper, European University Institute, Florence, Italy.

Leffler, K. 1981. Persuasion or information? The economics of pre-scription drug advertising. J. Law Econom. 24(1) 45–74.

Mehta, N., X. (Jack) Chen, O. Narasimhan. 2008. Informing, trans-forming, and persuading: Disentangling the multiple effectsof advertising on brand choice decisions. Marketing Sci. 27(3)334–355.

Narayanan, S., P. Manchanda. 2009. Heterogeneous learning andthe targeting of marketing communication for new products.Marketing Sci. 28(3) 424–441.

Narayanan, S., P. Manchanda, P. K. Chintagunta. 2005. Temporaldifferences in the role of marketing communication in newproduct categories. J. Marketing Res. 42(3) 278–290.

Nerlove, M., K. J. Arrow. 1962. Optimal advertising policy underdynamic conditions. Economica 29(114) 129–142.

Rizzo, J. 1999. Advertising and competition in the ethical pharma-ceutical industry: The case of antihypertensive drugs. J. LawEconom. 42(1) 89–116.

Saunders, J. A., F. A. W. Watt. 1979. Do brand names differentiateidentical industrial products. Indust. Marketing Management 8(2)114–123.

Sullivan, M. W. 1998. How brand names affect the demand for twinautomobiles. J. Marketing Res. 35(2) 154–165.

Tu, K., N. R. C. Campbell, M. Duong-Hua, F. A. McAlister. 2005.Hypertension management in the elderly has improved:Ontario prescribing trend, 1994 to 2002. Hypertension 45(6)1113–1118.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.

Measuring the Informative and Persuasive Roles of...

Documents

Transcript of Measuring the Informative and Persuasive Roles of...