`Prescriptions
`Author(s): Natalie Mizik and Robert Jacobson
`Source: Management Science, Vol. 50, No. 12 (Dec., 2004), pp. 1704-1715
`Published by: INFORMS
`Stable URL: http://www.jstor.org/stable/30048061 .
`Accessed: 20/01/2015 13:37
`
`Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
`http://www.jstor.org/page/info/about/policies/terms.jsp
`
` .
`JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
`content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
`of scholarship. For more information about JSTOR, please contact support@jstor.org.
`
` .
`
`INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.
`
`http://www.jstor.org
`
`This content downloaded from 208.85.77.1 on Tue, 20 Jan 2015 13:37:52 PM
`All use subject to JSTOR Terms and Conditions
`
`Page 1 of 13
`
`SENJU EXHIBIT 2207
`LUPIN v. SENJU
`IPR2015-01097
`
`
`
`MANAGEMENT
`SCIENCE
`Vol. 50, No. 12, December 2004, pp. 1704-1715
`ISSN 0025-1909 1 EISSN 1526-5501 104150121 1704
`
`infimeR
`DOI 10.1287/mnsc.1040.0281
`f 2004 INFORMS
`
`Are
`
`Marks"?
`Physicians
`Quantifying the
`"Easy
`and
`Effects
`of Detailing
`Sampling
`New
`Prescriptions
`Natalie Mizik
`Graduate School of Business, Columbia University, 3022 Broadway, New York, New York 10027-6902, nm2079@columbia.edu
`Robert Jacobson
`Department of Marketing, School of Business, University of Washington, Seattle, Washington 98195-3200,
`yusho@u.washington.edu
`
`on
`
`public attention and considerable controversy surround pharmaceutical marketing practices and their
`Much
`impact on physicians. However, views on the matter have largely been shaped by anecdotal evidence or
`results from analyses with insufficient controls. Making use of a dynamic fixed-effects distributed lag regres-
`sion model, we empirically assess the role that two central components of pharmaceutical marketing practices
`(namely, detailing and sampling) have on physician prescribing behavior. Key differentiating features of our
`model include its ability to (i) capture persistence in the prescribing process and decompose it into own-growth
`and competitive-stealing effects, (ii) estimate an unrestricted decay structure of the promotional effects over
`time, and (iii) control for physician-specific effects that, if not taken into account, induce biased coefficient esti-
`mates of detailing and sampling effects. Based on pooled time series cross-sectional data involving three drugs,
`24 monthly observations, and 74,075 individual physicians (more than 2 million observations in total), we find
`that detailing and free drug samples have positive and statistically significant effects on the number of new
`prescriptions issued by a physician. However, we find that the magnitudes of the effects are modest.
`Key words: pharmaceutical marketing; salesforce effectiveness
`History: Accepted by Linda Green, public sector applications; received March 8, 2004. This paper was with the
`authors 1 month for 1 revision.
`
`Introduction
`industry's promotional tactics lead to an increase in
`appropriate versus inappropriate use of drugs in a
`As the cost of prescription drugs continues to escalate,
`cost-effective manner.
`increased public attention is being focused on the
`Concern that pharmaceutical marketing practices
`marketing practices of the pharmaceutical firms as
`have exacerbated increases in public health costs has
`one source of the problem. Direct-to-physician activ-
`prompted government actions at the federal and state
`ities account for the bulk of U.S. pharmaceutical
`levels. For example, in 2002 the federal government
`firm promotional spending. IMS Health (2003) esti-
`issued a warning to the drug industry to curtail some
`mates that over $5.8 billion was spent in 2002 on
`of their marketing practices (Washington Post 2002).
`detailing, i.e., pharmaceutical sales representatives
`H.R. 2356, which calls for ongoing annual funding of
`(PSRs) visiting physicians to promote their firm's
`$75 million to conduct comparative cost-effectiveness
`drugs. In addition, the retail value of the free drug
`drug studies, was introduced in Congress in June
`samples distributed during these visits is estimated
`2003. A primary intent of this legislation is to pro-
`at $11.5 billion.
`vide objective scientific evidence to "reduce doctors'
`A detailing visit typically lasts two to five minutes
`reliance on marketing information from the pharma-
`during which time a PSR discusses one to three of
`ceutical industry" (Pear 2003). Given the fact that one
`the company's drugs. Information (and, at times, mis-
`of every five dollars spent on pharmaceutical drugs in
`information) about a drug's composition, therapeutic
`the United States is paid for by a state program, state
`value, proper dosage, and potential side effects is
`governments have also taken steps to counter PSR
`communicated (Zigler et al. 1995). Often, PSRs will
`influence. Most notably, several states have under-
`also dispense samples and possibly offer small gifts
`taken counterdetailing initiatives (Gold 2001). State
`to the physician. At issue is whether these interac-
`employees visit physicians in hopes of persuading
`tions with PSRs compromise physician integrity and
`them to switch from prescribing branded drugs to
`affect their prescribing behavior. More precisely, the
`key public policy issue is the extent to which the
`prescribing lower-cost generic drugs.
`1704
`
`This content downloaded from 208.85.77.1 on Tue, 20 Jan 2015 13:37:52 PM
`All use subject to JSTOR Terms and Conditions
`
`Page 2 of 13
`
`
`
`Mizik and Jacobson: Quantifying the Effects of Detailing and Sampling on New Prescriptions
`Management Science 50(12), pp. 1704-1715, v 2004 INFORMS
`
`1705
`
`Prescription drug expenditures are projected to
`remain the fastest-growing sector of health care
`expenditures. They are expected to account for 14.5%
`of $3.1 trillion health care expenditures by 2012
`(compared with approximately 10% in 2001). With
`recent legislation providing a Medicare drug benefit
`expected to cost the federal government $534 billion
`over the next decade, it is no wonder that the impact
`of pharmaceutical industry marketing practices is of
`keen interest to policymakers, the business commu-
`nity, and the general public.
`Two competing views have dominated discussion
`on the matter. The prevailing view contends that
`PSRs significantly influence physicians' prescribing
`behavior and that this influence has negative effect
`on patients' welfare, in that PSRs encourage physi-
`cians to prescribe more expensive branded drugs.
`Many public policy organizations and consumer
`advocacy groups adhere to this view (see, for exam-
`ple, www.nofreelunch.org). The prominent alternative
`view argues that PSRs do influence physicians' pre-
`scribing behavior, but that this influence is positive
`in that PSRs provide physicians with valuable infor-
`mation. As a result, physicians are better informed
`and make better choices for their patients. Pharma-
`ceutical companies and industry groups advocate this
`second view.
`Despite the substantial resources that pharmaceu-
`tical companies invest in promoting their products
`and the controversy associated with pharmaceutical
`marketing practices, surprisingly little is known about
`the magnitude of the impact that PSR visits and free
`drug samples have on physician prescribing behavior.
`Narayanan et al. (2003) report a pharmaceutical exec-
`utive as stating, "No one is really sure if sending the
`legions of reps to doctors' offices really works. Every-
`one is afraid to stop it, because they don't know what
`difference it's making" (p. 4).
`In point of fact, much of the evidence on PSR
`effectiveness is anecdotal. The empirical studies
`investigating the issue have been subject to data or
`methodological limitations that restricted their ability
`to control for potential biases and have come to
`contradictory conclusions regarding even the central
`issues: the effects of detailing on prescriptions (e.g.,
`Parsons and Abeele 1981 versus Gonul et al. 2001),
`of detailing on price elasticity (e.g., Rizzo 1999 versus
`Gonul et al. 2001), and even of price on sales (e.g.,
`Rizzo 1999 versus Gonul et al. 2001).
`We have obtained access to a unique database
`that allows us to undertake econometric analysis
`that overcomes a number of fundamental limitations
`existing in past research. In particular, making use
`of a dynamic fixed-effects distributed lag model that
`accounts for physician-specific effects likely to induce
`
`bias if left uncontrolled, we assess the effect of detail-
`ing and sampling on physician prescribing behavior.
`The large number of observations in the database (it
`involves a total of more than 2 million observations)
`allows us to accurately pinpoint the impact that inter-
`actions with PSRs have on the number of new pre-
`scriptions issued by physicians.
`We find that, although detailing and free drug sam-
`ples have a positive and statistically significant asso-
`ciation with the number of new prescriptions issued
`by a physician, the magnitudes of the effects are mod-
`est. As such, our results challenge the two domi-
`nant views and support the contention that, rather
`than being easy marks, physicians are tough sells.
`This realization is important because the public policy
`debate continues over how best to address the high
`cost of prescription drugs.
`
`PSR Influence on Physicians
`Most discussions of PSRs have focused on the factors
`facilitating their influence. Unquestionably, PSRs pro-
`vide physicians with information about new drugs,
`new indications, dosages, and interactions for existing
`medicines. Azoulay (2002) finds evidence that detail-
`ing diffuses product information. Avorn et al. (1982)
`report that 20% of surveyed physicians view informa-
`tion provided by PSRs as "very important" in influ-
`encing their prescribing behavior. Furthermore, PSRs
`are trained in persuading physicians. Detailing takes
`the form of presenting facts and, as has been doc-
`umented (Zigler et al. 1995), misrepresenting facts
`about the drug in an effective manner. Finally, mere
`exposure or salience effects might lead to a temporary
`increase in prescribing following a PSR visit. Numer-
`ous studies have reported high physician responsive-
`ness to PSR activity attributed it to PSR persuasive-
`ness (Avorn et al. 1982, Powers 1998).
`Less attention has been paid to the factors limiting
`PSR effectiveness. The key consideration here is that
`PSRs are not the only or even the primary source
`of information about drugs for physicians. Scientific
`papers, advice from colleagues, and a physician's own
`training and experience also influence prescribing
`practices. Indeed, most physicians view these other
`influences as far more important than that of PSRs
`(Peay and Peay 1990).
`PSR influence is limited by the fact that many
`physicians have skeptical or negative attitudes toward
`PSRs (Lichstein et al. 1992, McKinley et al. 1990).
`Attribution theory suggests
`that with
`low source
`credibility, which is determined by factors such as
`a source's trustworthiness and expertise (Dholakia
`and Sternthal 1977), arguments in a message will
`be discounted (Eagley and Chaiken 1975). Physicians
`recognize that PSRs are neither experts nor com-
`
`This content downloaded from 208.85.77.1 on Tue, 20 Jan 2015 13:37:52 PM
`All use subject to JSTOR Terms and Conditions
`
`Page 3 of 13
`
`
`
`1706
`
`Mizik and Jacobson: Quantifying the Effects of Detailing and Sampling on New Prescriptions
`Management Science 50(12), pp. 1704-1715, r2004 INFORMS
`
`pletely trustworthy. They realize that information pre-
`sented is biased toward the promoted drug and is
`unlikely to be objective or even accurate (Connelly
`et al. 1990). Thus, physicians will discount informa-
`tion received from a PSR.
`Some additional characteristics of physicians would
`seem to make them particularly tough sells. Friestad
`and Wright's (1994) persuasion knowledge model
`suggests that targets of persuasion use their knowl-
`edge about the persuasion agent and can effectively
`cope with and even achieve their own goals during a
`persuasion attempt, e.g., obtaining free drug samples
`that can be later distributed to patients. Campbell and
`Kirmani's (2000) tests of the persuasion knowledge
`model reveal that busy targets with accessible agent
`motivation (a profile that would fit most physicians)
`are particularly effective in resisting persuasion.
`When cast within the workings of other sources
`of influence, we would expect the ability of PSRs to
`influence physician behavior to be relatively small. As
`such, we hypothesize a relatively small effect of PSR
`activity on physician prescribing behavior.
`
`Previous Empirical Research
`The various studies assessing the effect of PSR activ-
`ity on physician prescribing behavior have generated
`conflicting results. Indeed, on some of the most cen-
`tral issues-ranging from the effects of detailing on
`prescriptions, of detailing on price elasticity, and even
`of price on sales-studies have come to diametrically
`opposite conclusions. Data and methodological limi-
`tations, however, raise concerns about the inferences
`drawn from these analyses.
`A few quasi-experimental studies of the issue orig-
`inate in the medical community. These studies com-
`pare physicians who did not see PSRs or were visited
`less frequently by PSRs to physicians who saw PSRs
`or were visited more frequently by PSRs (Chren and
`Landefeld 1994, Powers 1998). The limitation of these
`studies is that they are not randomized: PSRs do
`not determine which physicians to visit on a random
`basis. Rather, PSRs tend to see physicians who are
`more likely to utilize the drug or who prescribe in
`higher volume. This consideration invalidates these
`attempts to assess the effect of PSRs independent of
`controls accounting for motivation influencing PSR
`behavior.
`The ability to potentially control for other influ-
`ences is an advantage of regression-based analysis.
`Past research has made use of different regression
`techniques to assess PSR influence. Unfortunately,
`it has been inadequate in controlling for physician-
`specific effects. Parsons and Abeele (1981) use data
`for 24 months and 14 territories to model the number
`of prescriptions sold in a given territory for a given
`month as a function of sales calls. Interestingly, the
`
`estimated main effect of detailing was negative. The
`most dominant explanatory factor in the model is
`sales lagged one period, which would reflect per-
`sistence in behaviors and carryover effects magnify-
`ing the influence of detailing. Alternatively, lagged
`sales could be reflective of territory-specific effects
`that are not modeled and, as such, could lead to
`biased estimates. Wotruba (1982), for example, raises
`this possibility of territory-specific effects to question
`the reported effects of detailing.
`Rizzo (1999) uses annual data for the period 1988-
`1993 and for 46 drugs to estimate a brand-level model
`linking prescriptions for a drug for a given year
`to pharmaceutical company marketing activities. He
`finds that price is negatively related to sales and that
`detailing is anticompetitive in that it decreases price
`sensitivity. Detailing is also found to have a direct
`positive effect on sales. Surprisingly, no consideration
`is given to the dynamic properties of sales. The classic
`spurious regression characteristics, i.e., very high R2
`in the presence of substantial unmodeled autocorre-
`lation, appear to be present. As such, questions exist
`about the validity of both the point estimates and
`standard errors reported in the analysis.
`Gonul et al. (2001) use data involving 1,785 patient
`visits to estimate a multinomial logit model assess-
`ing factors influencing physician prescribing behav-
`ior. Exactly opposite to the findings of Rizzo (1999),
`they report that price has a positive effect on prescrip-
`tion probabilities and that detailing increases price
`sensitivity. They find positive effects of detailing and
`sampling, but do not discuss the implications of their
`magnitudes. These magnitudes, calculated based on
`descriptive statistics, imply elasticities that are sur-
`prisingly large.' The elasticity estimates for the seven
`drugs studied, evaluated at the mean level of detail-
`ing and sampling, average 41% for detailing and 48%
`for sampling. Particularly for samples, which have a
`negligible marginal cost, their estimated coefficients
`imply enormous returns to enhanced PSR activity. In
`point of fact, these substantial effects could arise, not
`from the influence of PSR activity, but rather as an
`outgrowth of a joint correlation with an omitted factor
`from the model, e.g., larger practices prescribe more
`and receive more free samples.
`A concern, which Gonul et al. (2001) explicitly
`is over the role of physician-specific
`acknowledge,
`effects that can induce a bias in the estimated coeffi-
`cients. They state (p. 84),
`prescription behavior patterns might be strongly influ-
`enced by factors other than the explanatory vari-
`ables we include in our model. Examples are physi-
`
`1 The elasticity of prescription probability Pj to covariate Xjk in a
`conditional logit model is calculated as (dPj/Pj)/(aXjk/Xjk) = f3k * Xjk *
`(1- P).
`
`This content downloaded from 208.85.77.1 on Tue, 20 Jan 2015 13:37:52 PM
`All use subject to JSTOR Terms and Conditions
`
`Page 4 of 13
`
`
`
`Mizik and Jacobson: Quantifying the Effects of Detailing and Sampling on New Prescriptions
`Management Science 50(12), pp. 1704-1715, l2004 INFORMS
`
`1707
`
`to 11 years; annual sales range from under $0.5 billion
`to more than $1 billion; they come from different ther-
`apeutic areas. Although the effect of detailing can
`vary across drugs, analysis of these three drugs offers
`some generalizable insights, not only because they
`provide a cross-section of drugs in the marketplace,
`but because they represent more than 4 million PSR
`interactions with physicians.
`Model
`We employ the following dynamic fixed-effects dis-
`tributed lag regression model to assess the effect of
`detailing and sampling on new prescriptions:
`
`cians' unobservable personal characteristics.... Ignor-
`ing these factors might bias the coefficients of the
`included explanatory variables.
`The extent to which their estimates are biased by
`the failure to control for unobservable factors remains
`unanswered, but this is one consideration that might
`account for the large estimated effects.
`
`Empirical Analysis
`A key benefit of utilizing pooled time series cross-
`sectional (panel) data is the ability to test for and
`control the effect of unobserved fixed factors. These
`unobserved factors, if left uncontrolled, can induce
`bias in the coefficient estimates of the explanatory fac-
`tors included in the model. Past research has either
`not used panel data or not made full use of the ben-
`efits of panel data analysis. We make use of pooled
`time series cross-section observations (24 months of
`observations across 74,075 physicians) and panel data
`statistical methods (i.e., a dynamic fixed-effects dis-
`tributed lag regression model) to assess the effect
`of detailing and sampling on physician prescribing
`behavior.
`
`Data
`Access to the data was gained from a U.S. pharmaceu-
`tical manufacturer with the only condition of ensur-
`ing the anonymity of the firm and the drugs in the
`study. Two different sets of data were merged to form
`the database. One data set pertains to the number
`of new prescriptions for the studied drugs and their
`competitors issued by physicians during a month.
`The new prescription measure reflects both new and
`repeat usage, but does not reflect refills accompany-
`ing the prescriptions. These data cover a 24-month
`period for three widely prescribed drugs. The second
`data set pertains to detailing and sampling activity
`by PSRs for the same three drugs. The two data sets
`were merged into one database containing prescrib-
`ing and promotional activity information by month
`and physician.
`To reduce the possible influence of extreme val-
`ues (outliers) that would arise from, for example,
`data entry errors and the common practice of one
`physician signing for all samples that later get dis-
`tributed to a group of physicians attending a confer-
`ence, we excluded the top 0.5% of observations for the
`number of details, samples, and new prescriptions.
`We later undertook sensitivity analysis on alternative
`definitions of outliers (e.g., 0%, 1%, 5%) and found
`results in close correspondence across these alterna-
`tive samples.
`Table 1 presents basic background information and
`descriptive statistics for the drugs included in our
`study. The drugs differ on a variety of dimensions:
`They have been on the market from less than 1 year
`
`Prescribeit
`6
`=- a + E
`j=0
`
`6
`
`6
`j * Detailsit-j + E yj * Samplesit-j
`j=0
`6
`+ E Aj * Competitort- + E
`i=1
`
`
`
`* Prescribeit-
`
`=O
`T
`
`+
`
`11
`
`, * Time(r) + i, Ks Specialty(s) * Trendt
`
`s=1
`
`+=1
`+ Eit'
`(1)
`where Prescribeit, Detailsit, Samplesit, and Competitorit
`are, respectively, the number of new prescriptions
`issued, the number of PSR visits, the number of free
`drug samples received, and the number of new pre-
`scriptions issued for competitive drugs by physician i
`at time period t. Time(7) is an indicator function that
`takes on the value 0 prior to the time period r and 1
`from the time period r on, Specialty(s) is an indi-
`cator function that takes on the value 1 when the
`specialty area of the physician is s, 0 otherwise (i.e.,
`separate dummy variables for each of the 11 spe-
`cialty areas), and Trend is the observation number
`for a given month and year. Because it includes both
`current-term and lagged variables in the model, Equa-
`tion (1) allows for a wide range of possible effects and
`influences, e.g., serial correlation (current-effects) and
`state-dependent (persisting) dynamic relationships.
`A key characteristic of Equation (1) is that it allows
`for a physician-specific effect, i.e., the intercept ai
`is allowed to vary by physician. This consideration
`acknowledges that physician behavior patterns are
`influenced by unobserved or unobservable factors,
`e.g., physician characteristics. To the extent that these
`unobserved factors are correlated with detailing and
`sampling, analysis not controlling for their effects will
`result in biased estimated effects for the marketing
`phenomena. Although a Hausman (1978) specification
`test can empirically assess the role played by fixed
`effects, we have a priori reason to believe that these
`unobserved factors will in fact be correlated with mar-
`keting activity. For instance, larger practices will gen-
`erate more prescriptions and will also attract more
`
`This content downloaded from 208.85.77.1 on Tue, 20 Jan 2015 13:37:52 PM
`All use subject to JSTOR Terms and Conditions
`
`Page 5 of 13
`
`
`
`1708
`
`Table 1 Drug Profiles
`
`Mizik and Jacobson: Quantifying the Effects of Detailing and Sampling on New Prescriptions
`Management Science 50(12), pp. 1704-1715, w2004 INFORMS
`
`Drug A
`0.5 to 1 billion
`3 years
`
`12
`
`1.73
`(1.75)
`4.34
`(9.76)
`13.18
`(14.81)
`2-3
`
`Drug B
`over 1 billion
`11 years
`
`Drug C
`under 0.5 billion
`6 months
`
`18
`
`1.98
`(1.70)
`7.79
`(13.71)
`8.82
`(10.43)
`2-3
`
`11
`
`1.73
`(1.44)
`4.02
`(7.73)
`2.27
`(3.58)
`2-3
`
`Sales range (US$)
`Time on the market at the
`beginning of the study period
`Estimated number of competitors
`in the respective therapeutic area
`Mean number of details per
`physician per month
`Mean number of free drug samples
`per physician per month
`Mean number of monthly new
`prescriptions per physician
`Average number of refills following
`one new prescription
`Recommended duration of therapy 3 months, and as maintenance with 3 months and up to 9 months Not yet established
`periodic patient reexamination
`Average
`
`Above average
`
`Average
`
`42.91
`(43.63)
`
`48.80
`(53.73)
`
`22.46
`(19.03)
`
`6
`
`6
`
`6
`
`Cost of therapy relative to other
`drugs in the therapeutic area
`Mean number of monthly new
`in
`prescriptions per physician
`the respective therapeutic area
`Therapeutic area is
`Relatively new
`Well established
`Well established
`Specialty area of top prescribers
`Primary carett
`Primary careft
`Psychiatry
`Number of physicians
`10,516
`30,005
`55,896
`Number of data points
`252,384
`720,120
`1,341,504
`Note. Standard errors are in parentheses. The sum of the number of physicians
`in each of the three data sets is greater than the total
`in the study (74,075) because some physicians are in the upper 60 prescribing percentile
`number of physicians
`for more than one of
`the drugs in our study.
`ttThe primary care specialty area includes family practice, general practice, internal medicine, and osteopathy.
`detailing. As such, a spurious positive correlation, i.e.,
`involve both a direct effect on prescriptions and an
`indirect effect that arises through persistence in physi-
`unrelated to any potential effects of detailing, will
`cian behavior. The total effect of detailing and sam-
`exist between detailing and prescriptions due to a
`pling can be calculated as
`joint correlation with practice size.
`Because the effects of PSR activity are unlikely to
`be limited to the month when the visit occurred, we
`allow for current and lagged effects for both detailing
`and sampling. The number of observations available
`in our data sample allows us to directly estimate sep-
`arate effects for lagged terms. As such, we do not
`need to impose a specific decay pattern as a neces-
`sity for preserving degrees of freedom. We do, how-
`ever, need to specify the length of time that a PSR
`visit might influence physician behavior. We select a
`six-months lag length under the view that the effect
`of a visit will dissipate substantially over this period,
`but we also test for longer-term lagged effects. Still,
`we expect the effects of detailing to be largest in the
`months directly following the visit. The cumulative
`direct effect of detailing and sampling can be obtained
`and
`simply by summing the coefficients, i.e., E61=0j
`Lagged values for new prescriptions (Prescribeit_)
`capture autocorrelation in the series that arises
`through inertia and persistence in physician behavior.
`These autoregressive effects play a key role in that
`they magnify the effects of detailing and sampling.
`That is, the total effects of detailing and sampling
`
`6
`
`1-yyj
`j=1
`
`,
`
`/Pj/ 1-Bb-[ j and y/
`
`j=l
`
`j=0
`
`j=O
`respectively.
`We include both lagged own prescriptions and
`lagged competitors' prescriptions, thus the model
`separates lagged total demand dynamics into two
`key components: competitive substitution and own
`demand growth. Lagged own prescriptions will have
`a positive effect on prescriptions. Lagged competitors'
`prescriptions will have a negative impact on prescrip-
`tions because they capture the substitution effects
`that physicians make between competing drugs. The
`current competitive prescriptions, however, will cap-
`ture two different phenomena. They will reflect not
`only substitution effects, but also changes in total
`demand due to market expansion or contraction. In
`this regard, current-term competitive effects will act
`as a proxy for two different phenomena with oppo-
`site effects, i.e., negative substitution effects and posi-
`tive market demand effects. As such, the current-term
`coefficient (AO) will depend on the relative magnitude
`of the two conflicting effects and, therefore, the sign
`of the effect cannot be postulated a priori.
`
`This content downloaded from 208.85.77.1 on Tue, 20 Jan 2015 13:37:52 PM
`All use subject to JSTOR Terms and Conditions
`
`Page 6 of 13
`
`
`
`Mizik and Jacobson: Quantifying the Effects of Detailing and Sampling on New Prescriptions
`Management Science 50(12), pp. 1704-1715, @2004 INFORMS
`
`1709
`
`The other variables in Equation (1) are time-period-
`specific indicators and specialty-specific trends. The
`time-period-specific indicators (the coefficients 8,)
`allow for the fact that the number of prescriptions
`can shift across time periods. These intercepts capture
`not only seasonal effects but all brandwide influences
`that shift prescribing behavior across all physicians
`(e.g., price changes, changes in the set of alternative
`medications available, changes in advertising cam-
`paigns, etc.). That is, the inclusion of the time-specific
`indicator variables will capture all effects common
`across physicians, which would include the diffu-
`sion pattern for the drug, research reports in scien-
`tific journals, any negative or positive publicity for
`the brand or its competitors, etc. The 11 specialty-
`specific trends K, capture influences that shift pre-
`scribing behavior across all physicians in a particular
`specialty. After we take first differences, the time-
`period-specific indicators and the specialty-specific
`trends are transformed into time-period-specific inter-
`cepts and specialty-specific intercepts.
`To remove the influence of physician specific effects
`(i.e., ai), we take first differences of Equation (1) to
`obtain
`
`6
`
`data for antiulcer drugs.2 We found that, after making
`use of competitive sales as a proxy variable, monthly
`changes in brand detailing exhibited little correlation
`(0.01) with monthly changes in competitor detailing.
`Given the absence of correlation at the brand level, the
`correlation at the physician level can also be expected
`to be similarly small and, indeed, even smaller. As
`such, the magnitude of this bias in the estimated effect
`of detailing caused by the exclusion of competitive
`detailing from the analysis can be expected to be min-
`imal, or even completely absent.
`We estimate Equation (2) using instrumental vari-
`able estimation as ordinary least squares will generate
`biased estimates of the coefficients for APrescribeit-1
`and ACompetitorit. By construction, APrescribeit_1 will
`be correlated with the differenced error term qit
`and, just as substitution effects cause competitor pre-
`scriptions to influence own prescriptions, own pre-
`scriptions will influence the amount of competitor
`prescriptions. Following Anderson and Hsiao (1982),
`we use lagged values of the levels of the series
`(values at time period t - 2 and earlier) to gener-
`ate instrumental variable estimates for APrescribeit-1
`and ACompetitorit. This procedure generates consis-
`tent (i.e., asymptotically unbiased) estimates of the
`parameters and their standard errors.3
`Results
`For each of the three drugs in our study, we estimated
`the Equation (2) regression model. Table 2 reports
`the estimated coefficients. Figures 1 and 2 graphically
`depict the estimated direct effects of detailing and
`sampling, respectively.
`Persistence in Prescribing Behavior. For each of
`the three drugs in the study we observe signifi-
`cant persistence in physicians' prescribing behavior.
`Although the first-order autocorrelation is the most
`substantial for all three drugs, significant higher-order
`effects are present as well. For Drug A the estimated
`coefficients for months 1 through 6 of 0.357, 0.205,
`
`2 We thank Ernst R. Berndt for granting us access to these data. See
`Berndt et al. (2003) for a complete description of the data. Consis-
`tent with lack of correlation between firm and competitor detailing,
`we also found little difference in the estimated effect of detailing
`on sales generated from a model that included competitive detail-
`ing versus a model that excluded competitive detailing (i.e., the
`difference was less than 3%).
`3The consistency of the lagged-variable-based instrument hinges
`on the assumption that the disturbances are serially uncorrelated;
`see, for example, Arellano and Bond (1991) for a discussion. We
`included higher-order lagged endogenous variables in the model
`to ensure this would be the case and we find no evidence to the
`contrary. For example, for Drug A the second-order autocorrela-
`tion of the residual is -0.0029, with a t-stat of -0.96; the 3rd order
`autocorrelation is 0.00035, with a t-statistic of 0.12. Indeed, our esti-
`mated residual correlation matrix is in near exact correspondence
`with that postulated by a fixed effects, autoregressive model for
`each of our three drugs (Arellano 2003).
`
`=
`
`+
`
`6
`
`j=1
`
`* APrescribeit-j
`
`it. (2)
`
`APrescribe