The Effects of Publicity on Demand: The case of anti-cholesterol drugs Andrew Ching * Rotman School of Management University of Toronto Robert Clark Institute of Applied Economics HEC Montreal Ignatius Horstmann Rotman School of Management University of Toronto Hyunwoo Lim Rotman School of Management University of Toronto November 16, 2010 * This is work-in-progress. Comments are welcome. Please direct all correspondence to: Andrew Ching, Rotman School of Management, University of Toronto, 105 St George Street, Toronto, ON, CANADA M5S 3E6. Phone: 416-946-0728. Fax: 416-978-5433. Email: [email protected].
48
Embed
The Effects of Publicity on Demand: The case of anti-cholesterol drugs
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
The Effects of Publicity on Demand:
The case of anti-cholesterol drugs
Andrew Ching∗
Rotman School of ManagementUniversity of Toronto
Robert ClarkInstitute of Applied Economics
HEC Montreal
Ignatius HorstmannRotman School of Management
University of Toronto
Hyunwoo LimRotman School of Management
University of Toronto
November 16, 2010
∗This is work-in-progress. Comments are welcome. Please direct all correspondence to: Andrew Ching, Rotman Schoolof Management, University of Toronto, 105 St George Street, Toronto, ON, CANADA M5S 3E6. Phone: 416-946-0728.Fax: 416-978-5433. Email: [email protected].
Abstract
This study investigates the effect of publicity on consumer demand. Although the marketing litera-
ture suggests that publicity can have an important influence on consumers’ choices, only a few studies
have attempted to quantify how media coverage can affect consumer demand. The main obstacle to
measuring the impact of publicity is that data on media coverage are difficult to interpret. To over-
come this obstacle, we develop an algorithm to collect information presented in news articles and to
map this information into a multi-dimensional attribute space. By applying our algorithm to news
articles covering statins (a class of anti-cholesterol drugs) accessible in Canada from 1993-2004, we
find evidence that media coverage plays a role in influencing physicians’ choice of statins. Our results
suggest that not all forms of publicity are equal. We find that comparison articles are much more
effective than non-comparison articles. Moreover, we find that the most effective type of publicity is
to provide information on the cholesterol-lowering ability of the drug, and the second most effective
type of publicity is about side-effects. Surprisingly, despite the fact that the ultimate goal of taking
anti-cholesterol medication is to reduce risks of heart disease, publicity related to this dimension has
very little impact on demand. We also find evidence that physicians/patients may generalize clinical
evidence on individual drug specific ability to reduce heart disease risks to the whole class of statins.
The results suggest that manufacturers of the drugs which are more potent in reducing cholesterol could
free-ride on the heart disease clinical evidence for other drugs. Our results could help firms decide which
types of post-marketing clinical trials to invest.
Publicity is an important marketing communication tool for firms. It could have a strong impact on pub-
lic awareness at a considerably lower cost than advertising, and function as a substitute/complement for
direct-to-consumer advertising (DTCA).1 Furthermore, consumers may give more credence to publicity
than to advertising, since publicity is generally conveyed to consumers in the form of news. Despite its
potential importance, previous research in marketing seldom investigated how publicity affects demand.
This is mainly because collecting publicity data and interpreting them is very challenging. This paper
contributes to the literature by (i) proposing a new method to interpret publicity data by mapping
the information in each news article (or broadcast) to a multi-dimensional attribute space, and (ii)
investigating how different types of publicity affect consumers’ demand, in particular, in the context of
prescription drugs.
More specifically, we study the impact of publicity on the demand for a class of prescription drugs
called statins, which are the most commonly used anti-cholesterol drugs. There are several reasons why
we study this market. First, there are large amounts of publicity surrounding this class of drugs, much
of it related to post-marketing clinical studies (Sillup and Porth, 2008). Second, it is plausible that
patients and physicians become informed about their efficacy and side-effects through the media that
they have access, e.g., newspapers, professional magazines, etc. Third, this is an important market for
public health and firms’ profitability. In the past decade, the most common cause of death is heart
disease.2 Medical research has consistently found a positive correlation between high cholesterol levels
and risks of coronary heart disease. Because statins are effective in lowering cholesterol levels, it has
been viewed as a way to reduce risks of heart disease and its demand has expanded very rapidly. From
the public policy standpoint, it is useful to know if media coverage would lead physicians to choose a
drug that has the strongest clinical evidence of reducing risks of heart disease at the time. From the1Principle of Marketing 7th edition, Chapter 15, Philip Kotler, Gary Amstrong and Peggy H. Cunningham2Public Health Agency of Canada ( http://www.phac-aspc.gc.ca/publicat/lcd-pcd97/table1-eng.php )
1
firms’ standpoint, understanding how the demand side perceives publicity could also help them design
their marketing campaign more efficiently.
Previous work has classified news stories as positive, negative, or neutral in a single dimension
(Berger et al., 2010; Chintagunta et al., 2009; Goldenberg et al., 2007; Huang and Chen, 2006). We argue
that such a single-dimensional classification could be misleading. For example, a news article might
report that an anti-cholesterol drug lowers cholesterol levels more effectively than do its competitors, but
that some patients experience serious side-effects. This article could be coded as positive, negative, or
neutral in a single-dimensional classification depending on the reader’s perspective. To avoid ambiguity
of context, we code the information of an article into a multi-dimensional attribute space. We consider
three attributes for our drugs: (i) their ability to lower cholesterol levels (short-term efficacy); (ii) their
ability to reduce the risks of heart disease (long-term efficacy); (iii) their side effects. Continuing with
the above example, our algorithm would classify the article as positive in the dimension of lowering
cholesterol levels and negative in the dimension of side effects. This multi-dimensional coding scheme
can reduce measurement errors made in single-dimensional coding schemes used in the previous studies.
More importantly, it allows us to measure how different types of information may affect consumer
demand. To the best of our knowledge, our research will be the first to use publicity data that is coded
in such a precise fashion.
We implement our methodology to code the publicity data that are accessible in Canada. In addition
to the publicity data, we have collected data on sales, detailing, journal advertising and clinical trial
outcomes from 1993 to 2004. One potential advantage of using the data from the Canadian market
is that direct-to-consumer advertising for prescription drugs is strictly regulated. While firms are
restricted in their use of DTCA, they may have a stronger incentive to communicate directly with
patients through publicity (e.g., by running more press releases). Therefore, the Canadian market may
provide a more interesting setting for studying the roles of publicity.
2
We incorporate our multi-dimensional measure of publicity into a demand system generated from
a standard discrete-choice demand model a la Berry (1994). Consumer utility functions over drugs can
be written as a function of product characteristics, including scientific evidence, publicity and other
marketing activities. Market shares are then determined as the aggregate outcome of consumer utility
maximizing decisions. Moreover, prices of prescription drugs are regulated in Canada. As a result,
prices change very infrequently over time. This institutional setting has also made it easier for us to
focus on studying how marketing communication activities would affect demand.
Our results suggest that not all forms of publicity are equal. We find that comparison articles
are much more effective than non-comparison articles. Moreover, we find that the most effective type
of publicity is to provide information on the cholesterol-lowering ability of the drug, and the second
most effective type of publicity is about side-effects. Surprisingly, despite the ultimate goal of taking
anti-cholesterol medication is to reduce heart disease risks, publicity related to this dimension has very
little impact on demand for a particular drug. We also find evidence that physicians/patients may
generalize clinical evidence on individual drug specific ability to reduce heart disease risks to the whole
class of statins. The results suggest that manufacturers of the drugs which are more potent in reducing
cholesterol could free-ride on the heart disease clinical evidence for other drugs. Our results could help
firms decide which types of post-marketing clinical trials to invest.
The rest of this paper is organized as follows. Section 2 reviews previous literature. Section 3
describes background information including the Canadian health care system, regulations on the phar-
maceutical industry, and the market for statins. Section 4 summarizes how we collect and code the
publicity data. Section 5 reports the estimation results. Section 6 is the conclusion.
2 Literature Review
Most of the previous empirical studies on publicity focus on studying the impact of critics and product
reviews on demand instead of media coverage (Basuroy et al., 2003; Berger et al., 2010; Chevalier and
3
Mayzlin, 2006; Godes and Mayzlin, 2004). The results of these studies appear to be ambiguous. While
some studies indicate that positive (negative) information increases (decreases) product evaluation
(Goldenberg et al., 2007; Huang and Chen, 2006), others find evidence that publicity (product reviews)
might increase sales of the product regardless of whether it is positive or negative (Berger et al., 2010).
Our research is related to the growing pharmaceutical marketing literature, which focuses on study-
ing the roles of detailing, journal advertising, DTCA and scientific evidence (Azoulay, 2002; Berndt
et al., 1997; Calfee et al., 2002; Chan et al., 2010; Ching and Ishihara, 2010b; Cockburn and Anis,
2001; Iizuka and Jin, 2005; Liaukonyte, 2009). This literature seldom investigates the roles of media
coverage. As far as we know, Chintagunta et al. (2009) is the only study that incorporate data on
news coverage when estimating the demand for pharmaceutical products. However, their focus is to
study learning through consumption experiences, and hence they simply incorporate publicity data as
a control variable in their analysis. They classify publicity solely based on the titles of articles, and
find puzzling results on the effect of their publicity variable. In particular, they find that news articles
have a positive influence on prescription choice, regardless how the title sounds (positive, neutral or
negative). They admit that this could be due to measurement error in their raw data or problems in
the data coding design. Our coding methodology could potentially address the shortcomings of the
literature mentioned above.
We should point out that our method is also related to the one proposed by Liaukonyte (2009), who
analyzes advertising content by classifying them as comparative and non-comparative advertising in
the over-the-counter analgesics market. She finds evidence that the impact of comparative advertising
on demand is more effective but shorter-lived than that of non-comparative advertising. However, for
each classification, she still follows the traditional method by coding each article as positive, negative
and neutral in a single dimension. Our multi-dimensional method of classification could provide richer
results than does her approach.
4
3 Background
3.1 Health Care System in Canada
The Canada Health Act adopted in 1984 entitles all legal Canadian residents to receive medically nec-
essary services without co-payment if the services are provided in hospital, or by practitioners.3 While
Canada’s public health system covers drugs administered in a hospital, it does not cover outpatient
prescription drugs costs, except for indigenous persons (i.e., First Nations and Inuit peoples), Canadian
Forces members, veterans, Royal Canadian Mounted Police members, and inmates of federal peniten-
tiaries. Moreover, provinces and territories administer publicly financed program to provide prescription
drug coverage concentrated on seniors, disabled citizens, and low-income persons with special needs.
In 2004, according to the Auditor General of Canada, about one million Canadians were eligible for
federal drug benefits and more than nine million were covered by provincial plans. About two-thirds
of Canadian residents are covered for prescription drugs by private insurance obtained through their
employer or purchased individually. Consequently, according to estimates for year 2000, 98% of the
Canadian population has some form of public or private sector drug plan coverage that provides a
degree of protection against severe drug expenditures (Fraser Group and Tristat Resources, 2002).
3.2 Regulation of Price and Direct-to-Consumer Advertising for Prescription Drugsin Canada
In Canada, prices of patented prescription drugs are strictly regulated. Health Canada introduced a
government agency, the Patented Medicine Prices Review Board (PMPRB) through amendments to the
Patent Act in 1987. This board regulates drugs that are still under patent and which have no generic
substitutes by establishing the maximum prices that can be charged in Canada for them (Anis and
Wen, 1998; Paris and Docteur, 2006). The PMPRB uses the term, “excessive” price, to characterize
either a high introductory price of a new medication, or a substantial increase in the price of an existing
medication. The essence of these regulations can be summarized as follows:3Health Canada http://www.hc-sc.gc.ca/hcs-sss/medi-assur/cha-lcs/index-eng.php
5
1. For a new strength or a new dosage form of an existing medicine, the price is considered excessive
if it does not possess a reasonable relationship to the average price of the existing medicine.
2. For a drug that represents a therapeutic breakthrough or provides a substantial improvement over
comparable existing medicines, the price is excessive if it exceeds the prices of all the comparable
products in the therapeutic class and the median of the prices in seven countries: France, Germany,
Italy, Sweden, Switzerland, the UK and the U.S.
3. For a drug that provides moderate, little or no therapeutic advantage over comparable medicines,
the price is judged excessive if it exceeds the price of comparable products in the Canadian market.
PMPRB may use the median of the prices in seven countries as a reference when it is impossible
or inappropriate to identify comparable drugs in Canada.
4. The change in price of existing medications is to be considered excessive if the price increase
exceeds the increase in the general consumer price index.
DTCA is also regulated under Canada’s Food and Drugs Act and Regulations. Health Canada cur-
rently allows two forms of advertising: (i) reminder advertisements, which include only the brand name
and no health claims or hints about the product’s use and (ii) disease-oriented or help-seeking advertise-
ments, which do not mention a specific brand but discuss a condition and suggest that consumers ask
their doctor about an unspecified treatment (Gardner et al., 2003). Because of the extreme regulation,
pharmaceutical industry in Canada merely spent CAD 22 million in 2006 on DTCA (Mintzes et al.,
2009) compared to USD 4.2 billion in 2005 in the US, where DTCA restriction was relaxed in 1997.
3.3 The Market for anti-cholesterol drugs – Statins
There are two main types of cholesterols: LDL (the “bad” cholesterol) and HDL (the “good” choles-
terol). High amounts of the bad LDL will deposit cholesterol on the artery walls forming plaques.
More and more plaques will narrow the arteries lumen and may eventually block blood flow, and lead
6
to a heart attack. On the other hand, the good HDL takes excess cholesterol away and carries it
back to the liver to be excreted. It can also remove some of the cholesterol already attached to the
artery walls. Statins are a class of most commonly used anti-cholesterol drugs. They lower the level of
LDL by blocking HMG-CoA reductase, which is the enzyme that synthesizes cholesterol in the liver.
Clinical research results have suggested that statins can also raise HDL by 5 to 10 percents. However
some medical researchers argue that such modest increase may not have a significant positive effect in
reducing the risks of heart disease.4 Table 1 contains a brief synthesis of the main descriptive statistics
for the seven statins. In 1978, scientists at Merck identified the first statin, lovastatin, which hit the
market in 1987 under the brand name Mevacor after the long process of clinical trials needed for FDA
approval. To market statins effectively, Merck needed to convince patients/physicians that increase in
cholesterol levels is associated with cardiovascular diseases, and statins are safe and effective in reducing
LDL. Their marketing efforts, together with the government public awareness campaign, have made
the public to become more familiar with the potential risk of high cholesterol levels.5
In 1991, a rival drugmaker Bristol Myers Squibb introduced the second statin, Pravachol (pravas-
tatin). To compete against this new statin and to convince physicians of the long-term benefits of
statins, Merck introduced a new statin, Zocor(simvastatin) and sponsored a clinical study called 4S.
From the results of 4S published in 1994, researchers announced the findings that Zocor not only lowers
patients’ cholesterol levels but also reduces their risks of heart disease. With the help of the successful
clinical results, Zocor quickly became the number one seller in this market.
In 1985, researchers in Warner-Lambert succeeded to synthesize a different statin, Lipitor (atorvas-
tatin). To catch up the existing statins, Warner-Lambert signed an agreement to co-promote Lipitor
with Pfizer, the fifth largest drugmaker. In 1996, Pfizer announced results from a head-to-head clinical
trial called CURVES. The results showed that Liptior was more effective in lowering cholesterol lev-4THE MEDICAL NEWS (http://cme.medscape.com/viewarticle/479499 5)5CNN Money (http://money.cnn.com/magazines/fortune/fortune archive/2003/01/20/335643/)
7
els than the four older statins: Merck’s Zocor and Mevacor, Novartis’s Lescol, and BMS’s Pravachol.
However, this study does not establish any direct evidence on whether Lipitor (atorvastatin) can lower
the risks of heart disease. When Lipitor hit the market in 1997, Pfizer focused on communicating its
superior efficacy in lowering cholesterol levels with physicians. Interestingly, even though Lipitor did
not have a clinical trial that provided direct evidence that it can reduce risks of heart disease until
2001, its market had expanded steadily and rapidly since its introduction.6 In 2002, Lipitor achieved
estimated sales of $ 7.4 billion and became the best-selling product in the prescription drug market.
Not all statins were as successful as Lipitor and Zocor. In 1998, Bayer introduced the sixth statin,
Baycol (cerivastatin) and a clinical study sponsored by Bayer showed Baycol’s impressive effectiveness
in lowering cholesteorol levels.7 However, Public Citizen, an influential Washington-based consumer
advocacy group, issued a petition to FDA for the removal of Baycol claiming that Baycol had fatal side-
effects including Rhabdomyolysis, the rapid breakdown of skeletal muscle, and resultant renal failure.
In the U.S., 52 deaths were reported in patients using Baycol and so Bayer “voluntarily” withdrew
Baycol from the market worldwide in 2001.
In 2003, AstraZeneca released Crestor (rosuvastatin), a seventh statin, and claimed that it would
lower cholesterol levels far more effectively than other statins including Lipitor. An AstraZeneca spon-
sored head-to-head clinical study called STELLAR, provides clinical evidence that Crestor not only
lowers LDL (bad cholesterol) levels but also increases HDL (good cholesterol) levels significantly more
than does Liptior. However, in less than a year since the drug has first been marketed, there are
some cases of adverse reactions (more precisely, acute renal failure), which prompted Public Citizen
to petition for the removal of Crestor from the market. The FDA issued several warnings concerning
Crestor related side-effects, but it eventually declined the petition and concluded that it is not riskier6A positive correlation between high cholesterol levels and risks of coronary heart disease has been found in medical
research. Nevertheless, a drug that can lower cholesterol levels effectively does not necessarily mean that it can reduce therisks of heart disease. For instance, a recent clinical trial shows that a new anti-cholesterol combination drug, Vytorin,does not reduce the risks of heart disease even though it is very effective in lowering cholesterol levels (Park, 2008).
7http://www.pslgroup.com/dg/5b8a2.htm
8
than other statins on the market. Although safety issues for Crestor have been raised, Crestor has
experienced strong growth in the pharmaceutical market. We next turn to describe our data.
4 Data
The analysis integrates three different data sources: (i) publicity data covering statins derived from
Factiva, (ii) product level monthly prescription volume, detailing and journal advertising data for the
Canadian statin market from IMS Canada, and (iii) landmark clinical trial data for statins.
4.1 Publicity Data
To investigate the impact of publicity on consumer demand, we collect publicity data covering statins
from 1986 to 2004 from Factiva. Factiva is a division of Dow Jones & Company that provides access
to more than 25,000 authoritative sources such as newspapers, journals, magazines, news and radio
transcripts, etc. We search for articles that contains a molecule name (i.e., atorvastatin, cerivastatin,
fluvastatin, lovastatin, pravastatin, rosuvastatin and pravastatin), and/or a brand name (i.e., Lipitor,
Baycol, Lescol, Mevacor, Pravachol, Zocor and Pravachol) of statins. We collect 37,115 articles in
the period of year 1986 to 2004 from Factiva. Table 2 presents the summary of sources in the our
database and the number of articles covering statins for each category. Then, we restrict the sample
to articles from Canadian Accessible Sources, to which Canadian physicians or patients may have
access. These sources include trade/professional magazines, academic medical journals, online sources,
Canadian television news, Canadian newspapers, and Canadian magazines, as well as U.S. television
news from four major television networks (ABC, NBC, CBS and FOX) and CNN, the eight biggest U.S.
newspapers with more than 500,000 daily circulations and the 25 top selling U.S. magazines. We assume
that the public might not have direct access to the press release from news agencies. Therefore, we leave
out articles from news agency such as Agence France-Presse, The Associated Press and Reuters. As a
result, we include 7,002 articles in the analysis. For each article, we extracted data such as headline,
9
source, contents and publication date. We detail our data collection approach in the appendix.
Table 3 presents the number of articles and their sources covering statins by year. The number of
articles covering statins in the Factiva database varies, mostly increases, in each year. For example, in
our Canadian sample, the number of articles (no. of sources) is 4 (3) in year 1986, and it increases to
2,688 (168) in year 2004. The number of sources covering statins also increases. There are two possible
reasons for the increase in the number of sources over time. First, it could be that the actual sources
have covered statins more than previous years and so Factiva has added more articles to its database
and it leads to increased number of sources. Second, Factiva might include more sources in its online
database. If the latter explains the increase in the number of sources across years, we need to weight
the each article by the number of sources available in each year. Therefore, we collect information on
when each source was included and excluded in Factiva database, and obtain the number of Canadian
sources available in the Factiva database over time. We find that the trend of the total number of
sources is similar to that for We weigh each article by dividing the number of sources available when
the article was published.
To assess the impact of publicity on consumer demand, it is essential to construct appropriate
measures of publicity. Since we often encounter articles which contain more than one message. Since
an article can be positive in one dimension but negative in another for a drug, simply coding its overall
tone as positive, neutral, or negative is likely to lead to ambiguity or measurement error. Therefore, we
interpret each article along multiple dimensions: (i) the ability of the drug to lower cholesterol levels
(short-term efficacy), (ii) the ability of the drug to reduce heart disease risks (long-term efficacy), and
(iii) the drug’s side effects. For each dimension, we label it as comparison if the article compares at
least two statins, and non-comparison otherwise.
For non-comparison articles, we score each article using a three-step Likert scale (+1, 0, -1) to
assess the positive, neutral or negative tone of the article, we assign “+1” (“-1”) if the article favors
10
(does not favor) the drug, and “0” if the article shows a neutral attitude toward the focal drug for each
dimension. For example, “Lipitor reduces cholesterol levels fast,” is counted as “+1” in the dimension
of short-term efficacy, “Baycol can cause a fatal damage to patients’ kidney” is considered to be “-1”
in the dimension of side effects, and “Pravachol was well-tolerated” is considered to be “+1” in the
dimension of side effects.
The coding of comparison articles is somewhat more complicated. In order to capture all of the
information contained in the message we score a comparison article according to the sum of its absolute
and relative attitudes towards each drug cited in the article. For the absolute score, we assign each
drug “+1”, “0”, or “-1” if the article shows a positive, neutral, or negative attitude towards the focal
drug, respectively, as we do for non-comparison articles. For the relative score, we assign the focal drug
the number of drugs that are reported inferior to the focal drug in the article. To illustrate how our
algorithm works, we provide the following examples:
1. When an article reports that both Lipitor and Zocor lowered cholesterol levels but Lipitor did
so more quickly than Zocor. For the absolute score, since the article states that both drugs are
effective in lowering cholesterol levels (short-term efficacy), both of them are assigned “+1”. For
the relative score, since the article reports that Lipitor is more effective than Zocor, Lipitor is
assigned “+1”. Therefore, we code Lipitor as “+2” = 1 (absolute) + 1 (relative), and Zocor as
“+1”= 1 (absolute) + 0 (relative) in the dimension of short-term efficacy.
2. When an article reports that Lipitor, Zocor and Pravachol were well-tolerated among patients
in a clinical trial. Patients with Pravachol showed no significant side effects and less patients
experienced side effects with Lipitor than with Zocor, we code Pravachol as ”+3” = 1 (absolute)
+ 2 (relative), Lipitor as “+2” = 1 (absolute) + 1 (relative), and Zocor as “+1” = 1 (absolute)
+ 0 (relative) in the dimension of side effects.
Table 4 presents a descriptive summary of publicity variables. While LC (Lowering Cholesterol
11
levels) dimension articles are often comparison, RH (Reducing risks of Heart disease) articles are mostly
non-comparison articles.
4.2 Prescription Volume, Detailing and Journal Advertising Data
The product level data obtained from the market research firm, IMS Canada, consist of monthly
observations of prescription volumes, detailing costs, and journal advertising pages for each statin
across Canada from March 1993 to December 2004. The market is defined as the national market for
month t, where t = 1, · · · , 130. The observation is defined as a molecule-month combination.
In figure 1, we plot the monthly prescription size for the four leading statin drugs (Lipitor, Zocor,
Pravachol and Crestor) in Canada. The prescription size for Lipitor reached almost one million by
2001 while the earlier arrivals, Zocor and Pravachol, had 300,000 and 150,000 monthly prescriptions,
respectively. Figure 2 shows the evolution of log of market shares for leading drugs relative to the
outside good.
Previous research has documented that marketing activities have an influence on the prescription
choices of physicians. To control for the impact of detailing and journal advertising, we incorporate
detailing spending and journal advertising pages for each drug. To convert nominal to real dollars of
detailing, we use Consumer Price Index from Statistics Canada. Our dataset does not include the period
when Mevacor, Pravachol or Zocor was introduced. In 1993, these three drugs might have accumulated
goodwill stocks of detailing and journal advertising efforts. To control the initial condition problem, we
follow Ching and Ishihara (2010b) and assume that before March 1993, these three drugs would have
spent the same monthly detailing and journal advertising efforts as the average of detailing and journal
advertising spending between Mar 1993 and Feb 1994. Figures 5 and 6 graph the monthly detailing and
journal advertising stocks for four leading drugs. The market entries of Lipitor and Crestor coincide
with large detailing and journal advertising efforts. Zocor and Pravachol stopped detailing and journal
advertising efforts when the generic products were introduced in the market.
12
The demand system used here will include an outside good (i.e., we allow patients with high LDL
to choose treatments other than statins, or no treatment at all). We therefore need to calibrate the
potential market size for statins, which includes high LDL patients who are on statins and other anti-
cholesterol drugs, and those who choose not to take any drugs. In order to estimate what percentages
of Canadians have a high cholesterol problem, we use data from the Canadian Heart Health Survey,
recorded between 1986 and 1992. The study suggests that 33% of the total Canadian population aged
16 to 65 and 85% of that over 65 have a high cholesterol (i.e., high LDL) problem. We multiply this
number by total Canadian population for each age group in a given month, as defined by the Statistics
Canada, and use the result as a proxy for the total number of potential patients for statins. In order
to convert total population with high cholesterol problem into the number of prescriptions, we assume
that each patient visits a physician and receives a prescription once per 90 days. Based on this measure
of potential market size, we calculate each brand’s and outside good’s market shares in each month.
For most product categories, price is an important factor that affects consumers’ purchase decision
on the product. However, we do not include prices in our demand system because the institutional
details of Canada suggest that it may not play an important role. Price regulation for prescription
drugs in Canada often leads to infrequent price changes. Table 6 shows the price which the province of
Ontario pays for Ontario Drug Benefit program beneficiary. The table suggests that prices for statin
have hardly changed over time, and this should alleviate the concern of price endogeneity (the data
suggests that price is hardly correlated with unobserved demand shock that varies over time), and
including brand fixed effects in our model should suffice in capturing the impact of prices. Moreover,
around 98% of Canadian residents possess a certain type of coverage for prescription drugs.
4.3 Landmark Clinical Trials
Azoulay (2002), Ching and Ishihara (2010a) and Cockburn and Anis (2001) find evidence that clinical
trials have a significant impact on physicians’ prescription decisions. Hence, to control for the impact
13
of clinical evidence, we collect landmark clinical trial data from the U.S. National Library of Medicine
(www.medscape.com) and include indicators for each one in our empirical analysis. Table 7 lists the
landmark trials we include in the analysis. We simply include the clinical trial dummies which become
one after the trial has been released. The dummies only apply to the focal drugs in the clinical trials.
For instance, the REVERSAL study published in March 2004 favors Lipitor over Pravachol. We create
a dummy variable for REVERSAL Lipitor equal to one when the observation for Lipitor is post-2004
and zero otherwise.
5 Econometric Model and Results
5.1 Model
We adopt the discrete-choice approach of Berry (1994) to model the demand for statins. This approach
offers several benefits: the derived demand equations are consistent with consumer theory and the
entry of new products can be easily incorporated. Consumer utility functions over drugs can be written
as a function of product characteristics, including scientific evidence, publicity and other marketing
activities. Market shares are then determined as the aggregate outcome of consumer utility maximizing
decisions. We define the utility of a patient i when she buys an anti-cholesterol drug j at time t as
follows:
Uijt = STK Xjt · β +17∑
k=1
Landmarkkjt · θk +7∑
k=1
γj + ξjt + εijt, (1)
where STK Xjt = {STK Publicityjt, STK Detailingjt, STK JournalADjt} represents a vector of
goodwill stock of publicity, detailing and journal advertising, respectively; Landmarkkjt denotes a
dummy variable for landmark trial k for drug j at time t (Landmarkkjt is equal to one if drug j receives
a landmark trial k favoring its efficacy before time t, and zero otherwise); ξjt may be interpreted as
the mean of the consumer valuation of an unobserved product characteristic; εijt is assumed to be i.i.d.
extreme value distributed with mean zero across consumers. We normalize the mean utility of the
outside good (j = 0) to be zero, and hence Ui0t = εi0t.
14
Detailing and journal advertising have been known to have long lasting impacts on demand (Azoulay,
2002; Berndt et al., 1997). Therefore, in explaining current period sales, stocks of cumulative detailing,
journal advertising, and publicity are more appropriate measures than flows. Following Azoulay (2002),
we allow for the possibility that the stock variables decay over time with a monthly carryover rates of
δ = .95. For example, the stock variable of detailing is defined as:
Since we assume that error terms εijt follow the extreme value distribution, the market share of
drug j at time t is given by the logit formula:
Sjt =exp(STK Xjt · β +
∑19k=1 Landmarkkjt · θk +
∑7k=1 γj + ξjt)
1 +∑7
j=1 exp(STK Xjt · β +∑19
k=1 Landmarkkjt · θk +∑7
k=1 γj + ξjt). (3)
It follows that,
log(Sjt
St) = STK Xjt · β +
19∑
k=1
Landmarkkjt · θk +7∑
k=1
γj + ξjt, (4)
where Sjt is the market share of drug j; ξjt are the unobserved characteristics of drug j at time t and
become the error terms. Detailed descriptions of included variables are listed in each table of results.
5.2 Endogeneity of Detailing and Journal Advertising
The previous literature has recognized that detailing and journal advertising could be endogenous. A
firm’s detailing and journal advertising effort might be correlated with shocks to its drug’s expected
quality over time, which are unobservable to researchers. If a manufacturing firm observes these shocks,
it is likely to choose a level of detailing and/or journal advertising to reflect the shock. Hence, if
researchers do not incorporate the shock into the model, the coefficients on the detailing and journal
advertising variables will be biased. We assume that clinical trial results and publicity are the major
sources of information about the quality of drugs to consumers and physicians. To control for the
unobserved shock to expected quality, we collect all the clinical trial data for statins between 1993
15
and 2004 and publicity data between 1986 and 2004, and incorporate them in the analysis. By doing
so, we attempt to control for the endogeneity problem of detailing and journal advertising caused by
unobserved shock to drugs’ expected quality.
5.3 Is Publicity Exogenous?
Figures 3 and 4 present three major statins’ (Liptor, Pravachol and Zocor) stock of LC (Lowering
Cholesterol levels) comparison score and RH (Reducing risks of Heart disease) non-comparison score,
respectively. The graphs suggest that the LC comparison score and RH non-comparison score of Lipitor
is highly correlated. However, Pravachol and Zocor have show different patterns in LC comparison levels
and RH non-comparison levels. This will allow us to disentangle the effects of two different types of
publicity.
There may also be some concern that publicity itself is endogenous. This is particularly true for
articles that discuss sales of the drug. In other words, publicity might be correlated with unobserved
shocks affecting sales of the drug. To address this concern we do not include articles purely related to
sales of drugs (e.g., an article which reports that the sales of Lipitor have improved from the previous
year) since we expect that the more popular drugs would have more meida coverage which emphasizes
sales of the drug. The articles we include in our analysis refer to efficacy and side effects of statins and
this type of publicity is generally yielded by the release of clinical trial results. In some broad economic
sense, one can argue that these clinical trial results are correlated with unobserved shocks of sales over
time. However, considering the time needed for publication and uncertain outcomes, we consider these
clinical trial results and publicity as being uncorrelated with month-to-month changes in unobserved
shocks.
5.4 Results
Table 8 reports the estimation results of Equation 4 [only use the log regressions]. We take log of
all the stocks of marketing and publicity variables when running these regressions. The literature
16
has consistently found evidence that drugs are experience goods (Crawford and Shum, 2005; Ching,
2010a,b; Ferreyra and Kosenok, 2010), and detailing activities contain information which helps the
demand side to learn the true quality of the products (Ching and Ishihara, 2010b,a; Leffler, 1981;
Narayanan et al., 2005). When patients/physicians slowly learn the true quality of the products from
consumption experiences, it is plausible that the marginal return of marketing activities would diminish
over time. Our prior is that publicity serves a similar purpose, and hence it seems likely that its effects
would also diminish over time. By taking logs of these variables, we are able to capture this feature.
The first specification only includes stocks of detailing and journal advertising. It does not include
any publicity variables. This regression serves as a benchmark for comparison. Consistent with the
previous literature, we find that the effect of the detailing stock is positive and statistically significant.
The stock of journal advertising is much weaker - in fact, it is not significant.
The second specification includes the “traditional” type of publicity variables, where we do not
break down the information content into a multi-dimensional attribute space. Instead, we simply
have two variables: (i) comparison articles, and (ii) non-comparison articles. Under this traditional
scheme, we code one article as positive (+1) when the sum of the three-step Likert scales across three
dimensions is greater than zero, or negative (-1) when the sum is smaller than zero. These variables
mimic the previous approach which codes articles in a single dimension fashion. The results show that
the effect of the stock of comparison articles is positive and significant, while the effect of the stock
of non-comparison articles is insignificant. Interestingly, the effect of the stock of journal advertising
now becomes positive and significant. This may indicate that the first regression suffers from omitted
variable bias.
In the third specification, we consider publicity about a drug along three dimensions: (i) its ability
to lower cholesterol levels (LC); (ii) its ability to reduce the risks of heart disease (RH); (iii) its side-
effects (SE). We further distinguish whether the publicity is comparison or non-comparison, except
17
for RH messages. For RH messages, we only consider non-comparison articles because the number
of comparison articles is close to zero for RH. This is because there are very few (landmark) clinical
trials which conduct head-to-head comparisons in this dimension. Although the ultimate goal of taking
statins is to reduce risks of heart disease, it is possible that patients/physicians might mainly focus
on the ability of drugs to lower cholesterol levels because it is well-documented that LDL and heart
disease rates are positively correlated. It is important to empirically investigate whether this is the case
because of its managerial and public policy implications. Our results show that the effect of the stock
of LC comparison articles is positive and significant, while the effect of LC non-comparison articles is
not significant. Interestingly, the effect of the stock of RH is not significant. The results support our
hypothesis that patients/physicians mainly pay attention to the LC dimension. We also find that the
effect of the stock of SE comparison articles is positive and significant. However, we find that the effect
of SE non-comparison articles is negative and significant, which is counterintuitive.
Regarding LC comparison articles, we find that many articles are consistent in reporting the relative
strength of statins in lowering cholesterol levels. It is common for articles to compare Lipitor with
Mevacor, Lescol, Pravachol and Zocor - all of them are not as potent as Lipitor. Even though we have
given Lipitor an extra unit increase when coding such articles,8 it might not be enough to reflect the
additional benefits that Lipitor gets if we restrict the coefficient for LC comparison to be the same
across drugs. Moreover, Baycol and Crestor are more potent than Lipitor, and they are introduced
after Lipitor has been on the market for 1 year and 6 years, respectively. It turns out that most of
the comparison articles involving Baycol and Crestor are typically comparing it with Lipitor. After
reading these articles, it seems likely that at least some of the patients/physicians would infer that
Baycol or Crestor is better than other statins as well. Since our coding scheme does not reflect this,
the effect of LC comparison articles for Baycol and Crestor may be stronger than that for Lipitor.
Based on this intuition, we also allow LC comparison to have heterogeneous effects across drugs in the8When two drugs are compared, the more potent one receives +1 and the less potent one 0 for the measure of relativity.
18
fourth specification. The estimation results are consistent with our intuition. The point estimates of LC
comparison are all positive and highly significant. The coefficient of Crestor and Baycol are close to each
other, and higher than Lipitor. The differences are statistically significant. Moreover, the coefficient
of Lipitor is much higher than those of other statins. Interestingly, the sign of the SE non-comparison
variable now switches from negative to positive, and is statistically significant. This suggests that
its negative coefficient in specification 3 could be due to misspecification bias. Nevertheless, the RH
non-comparison variable remains insignificant.
In the medical literature, researchers believe that high cholesterol levels are an indicator for high
risks of heart disease. A drug that can lower cholesterol levels would likely reduce risks of heart disease.
However, without a clinical trial that establishes a direct causal effect of taking the drug on heart attack
incidence rates, its manufacturer is not allowed to directly tell doctors or advertise that the drug can
reduce the risk of heart attack. Consequently, some drug companies are willing to sponsor landmark
clinical trials to establish such evidence. Most of the RH articles report on clinical trial results. We
therefore find it surprising that RH non-comparison variable is not significant. By investigating the
medical continuing education literature, we find evidence that many physicians may not strictly follow
evidence-based prescribing practice (Mamdani et al., 2008). It appears that some physicians may
assume a drug’s ability to lower the LDL level will likely translate to reducing heart disease risks and
ignore the possibility that it may have other side-effects that work the other way. This could explain
why the RH non-comparison variable is insignificant after controlling for LC.
Another possibilty is that physicians and patients generalize the ability of one drug to reduce heart
disease risks to the whole statin class. In order to test this hypothesis, we estimate specification 5 where
we create RH sum variable by summing RH non-comparison across drugs, taking the log, and allowing
it to interact with the LC comparison variable. The interaction term is positive and significant, which
supports our hypothesis. In specification 6, we allow the effects of the interaction term to differ across
19
drugs. Because of the multi-collinearity issue, we do not allow the LC comparison variable to have
heterogeneous effects across drugs. These interaction terms are all positive and significant except for
Pravachol. Overall, the results from specification 6 are consistent with those of 4 and 5.
Previous studies that have investigated the impact of marketing-mix on sales in this market usually
do not use clinical trial outcomes or publicity data with the exception of Azoulay (2002) who focuses
on clinical trial data. But even Azoulay (2002) does not use publicity data. The missing clinical trials
and publicity information in these studies may lead to the potential endogeneity problem of detailing
or prices. The typical remedy that these studies use is the instrumental variable approach. In our case,
we believe that the endogeneity problem is less of a concern. Pharmaceutical prices are regulated in
Canada and cannot be freely changed by firms. Moreover, we have included more information than
what the previous studies did in our regressions. In particular, this is the first study that measures
publicity in a multi-dimensional way. We therefore should alleviate the endogeneity problem of detailing
due to omitted variables compared with previous studies.
We should also note that the correlation between journal advertising and detailing is quite high. The
correlations are 0.824 and 0.721 for their flow and stock variables, respectively. Such high correlations
suggest that it could be hard to separately measure their effects (Berndt et al., 1997). This could
explain why the point estimates of the stock of journal advertising vary across specifications. Since the
estimate of the stock of detailing seems to be more stable. As a robustness check, we have run another
set of regressions where we drop the stock of journal advertising. Most of the results reported above
are robust in this specification. The results of this set of regressions are reported in tables 12 and 13
in the appendix.
Venkataraman and Stremersch (2007) and Ching and Ishihara (2010a) find evidence that new in-
dications or clinical evidence could change the effectiveness of detailing. We therefore consider an
alternative specification where we allow the stock of detailing to interact with the clinical trial dum-
20
mies. This allows us to capture the idea that the effect of detailing could be heterogeneous across drugs.
We find that our results regarding publicity are largely robust in this specification as well. The results
of this set of regressions are reported in tables 14 and 15 in the appendix as well.9
5.5 Do firms view publicity as substitute/complement for detailing and journaladvertising?
As we mentioned in the introduction, publicity might serve as a substitute for detailing and/or journal
advertising. As a test of whether this is the case, we regress detailing/journal advertising on the
publicity variables. The results are reported in tables 9 and 10. Under specification 2 in table 9 for the
detailing regression, we find that LC comparison and RH non-comparison variables are both positive
and significant. In table 10 for the journal advertising regression, we find that the LC comparison
variable is also positive and significant. If firms treat publicity as a substitute for detailing/journal
advertising, we would expect that the coefficients for these variables should be negative. The positive
coefficient could be due to the fact that most of the media coverage reports some new clinical findings.
As argued in the previous literature (Azoulay, 2002; Ching and Ishihara, 2010b), the effectiveness of
detailing/journal advertising may increases with positive clinical outcomes, and as a result, firms may
also do more detailing. This may lead to the positive coefficients in these regressions.
We therefore run two more regressions for detailing/journal advertising response functions by con-
trolling for the impact of landmark clinical trials. In specification 3 in tables 9 and 10, we simply
include the clinical trial dummies which become one after the results from the trial have been released,
but the dummy only applies to the focal drug. However, a clinical trial about drug A (say Lipitor)
could affect the marketing activities for other drugs as well. Conceivably, if drug A receives good news,
the marginal return of detailing for the rival drugs might decrease as it now becomes harder to convince
doctors to switch away from drug A. We therefore allow the clinical trial dummies to affect drugs other9We have also considered to separate SE comparison and non-comparison, and RH non-comparison variables by drug.
However, the number of observation becomes quite small and the individual effects of these variables cannot be reliablyestimated.
21
than the ones being investigated in the trial. The results are reported in specification 4 in tables 9
and 10. In both specifications 3 and 4, we see that the publicity variables are no longer significant.
The results confirm our conjecture above. Moreover, they also indicate that firms’ detailing or journal
advertising decisions may not be influenced by publicity in general.10 The coefficients for clinical trial
dummies for both detailing and journal advertising estimating functions are reported in table 16 in the
appendix.
6 Conclusion
This research attempts to study the role of publicity on the choice of prescription drugs. In the study
we employ rich data sets to reduce a bias caused by omitted variables affecting the prescription choices.
The result presented here demonstrates that product market in the statin class is shaped by firms’
detailing, journal advertising, publicity and scientific evidence.
By allowing the information that is provided in the text of publicity the model is able to show
that different types of publicity have different impact on the choice of prescription drugs. The results
also raise the interesting question of the extent to which publicity is a strategic instrument available to
pharmaceutical companies. We will investigate this further in the future.
10It is important to note that our analysis is at the product level. It is possible that firms’ decisions to detail mightchange at the physician level in response to publicity Chen and Tan (2010).
22
References
Anis, Aslam H., Quan Wen. 1998. Price Regulation of Pharmaceuticals in Canada. Journal of HealthEconomics 17(1) 21–38.
Azoulay, Pierre. 2002. Do Pharmaceutical Sales Respond to Scientific Evidence? Journal of Economics& Management Strategy 11(4) 551–594.
Basuroy, Suman, Subimal Chatterjee, S. Abraham RavidSource. 2003. How Critical Are Critical Re-views? The Box Office Effects of Film Critics, Star Power, and Budgets. Journal of MarketingResearch 67(4) 103–117.
Berger, Jonah, Alan T. Sorensen, Scott J. Rasmussen. 2010. Positive Effects of Negative Publicity:When Negative Reviews Increase Sales. Marketing Science 29(5) 815–827.
Berndt, Ernst R., Linda T. Bui, David H. Reiley, Glen L. Urban. 1997. The Roles of Marketing,Product Quality and Price Competition in the Growth and Composition of the U.S. Anti-Ulcer DrugIndustry. The Economics of New Goods, NBER Studies in Income and Wealth 58 551–594.
Berry, Steven T. 1994. Estimating Discrete-Choice Models of Product Differentiation. RAND Journalof Economics 25(2) 242–262.
Calfee, John E, Clifford Winston, Randolph Stempski. 2002. Direct-to-Consumer Advertising and theDemand for Cholesterol-Reducing Drugs. Journal of Law and Economics 45 673–690.
Chan, Tat, Charkravarthi Narasimhan, Ying Xie. 2010. An Empirical Model of Consumer Learningand Evaluation on Multiple Attributes. Working Paper, Washington University in Saint Louis.
Chen, Yutian, Wei Tan. 2010. Strategic Advertising: Evidence in the Pharmaceutical Industry. WorkingPaper, California State University, Long Beach.
Chevalier, Judith A., Dina Mayzlin. 2006. The Effect of Word of Mouth on Sales: Online Book Reviews.Journal of Marketing Research 43(3) 345–354.
Ching, Andrew T. 2010a. A Dynamic Oligopoly Structural Model for the Prescription Drug Marketafter Patent Expiration. International Economic Review 51(4) 1175–1207.
Ching, Andrew T. 2010b. Consumer Learning and Heterogeneity: Dynamics of Demand for PrescriptionDrugs after Patent Expiration. International Journal of Industrial Organization 28(6) 619–638.
Ching, Andrew T., Masakazu Ishihara. 2010a. Measuring the Informative and Persuasive Roles ofDetailing on Prescribing Decisions Working Paper, University of Toronto.
Ching, Andrew T., Masakazu Ishihara. 2010b. The Effects of Detailing on Prescribing Decisions underQuality Uncertainty. Quantitative Marketing and Economics 8(2) 123–165.
Chintagunta, Pradeep K., Renna Jiang, Ginger Z. Jin. 2009. Information,Learning, and Drug Diffusion:the Case of Cox-2 Inhibitors. Quantitative Marketing and Economics 7(4) 399–443.
Cockburn, Iain M., Aslam H. Anis. 2001. Hedonic Analysis of Arthritis Drugs. Medical Care Outputand Productivity . National Bureau of Economic Research, Inc, 439–462.
Crawford, Gregory S., Matthew Shum. 2005. Uncertainty and Learning in Pharmaceutical Demand.Econometrica 73(4) 1137–1173.
23
Ferreyra, Maria Marta, Grigory Kosenok. 2010. Learning about New Products: An Empirical Studyof Physicians’ Behavior. Economic Inquiry (in press) .
Fraser Group and Tristat Resources. 2002. Drug Expense Coverage in the Canadian Population. Pro-tection from Severe Drug Expenses .
Gardner, David M., Barbara Mintzes, Aleck Ostry. 2003. Direct-to-Consumer Prescription Drug Adver-tising in Canada: Permission by Default? Canadian Medical Association Journal 169(5) 425–427.
Godes, David, Dina Mayzlin. 2004. Using Online Conversations to Study Word-of-Mouth Communica-tion. Marketing Science 23(4) 545–560.
Goldenberg, Jacob, Barak Libai, Sarit Moldovan, Eitan Muller. 2007. The NPV of Bad News. Inter-national Journal of Research in Marketing 24 186–200.
Huang, Jen-Hung, Yi-Fen Chen. 2006. Herding in Online Product Choice. Psychology and Marketing23(5) 413–428.
Iizuka, Toshiaki, Ginger Zhe Jin. 2005. The Effect of Prescription Drug Advertising on Doctor Visits.Journal of Economics & Management Strategy 14(3) 701–727.
Leffler, Keith B. 1981. Persuasion or Information? The Economics of Prescription Drug Advertising.Journal of Law and Economics 24(1) 45–74.
Liaukonyte, Jura. 2009. Is Comparative Advertising an Active Ingredient in the Market for Pain Relief?An Empirical Model of Advertising Content. Working Paper, Cornell University.
Mamdani, Muhammad, Andrew Ching, Brian Golden, Magda Melo, Ulrich Menzefricke. 2008. Chal-lenges to Evidence-Based Prescribing in Clinical Practice. The Annnals of Pharmacotherapy 42(5)704–707.
Mintzes, Barbara, Steve Morgan, James M. Wright. 2009. Twelve Years’ Experience with Direct-to-Consumer Advertising of Prescription Drugs in Canada: A Cautonary Tale. PLoS ONE 4(5).
Narayanan, Sridhar, Puneet Manchanda, Pradeep K. Chintagunta. 2005. Temporal Differences in theRole of Marketing Communication in New Product Categories. Journal of Marketing Research 42(3)278–290.
Paris, Valerie, Elizabeth Docteur. 2006. Pharmaceutical Pricing and Reimbursement Policies in Canada.OECD Health Working Papers: No. 24.
Park, Alice. 2008. Is vytorin a failure? Time http://www.time.com/time/health/article/0,8599,1703827,00.html.
Sillup, George P., Stephen J. Porth. 2008. Ethical issues in the pharmaceutical industry: an analysis ofUS newspapers. International Journal of Pharmaceutical and Healthcare Marketing 2(3) 163 – 180.
Venkataraman, Sriram, Stefan Stremersch. 2007. The Debate on Influencing Doctors’ Decisions: AreDrug Characteristics the Missing Link? Management Science 53(11) 1688–1701.