Modeling interaction and dynamic effects of pharmaceutical promotional expenditures

(1)

Modeling interaction and dynamic effects of

pharmaceutical promotional expenditures

by

Jan Albert Veldman

University of Groningen

Faculty of Economics and Business

Msc. Marketing Intelligence

July 2013

Email: janalbertveldman@gmail.com Phone number: -31(6) - 42 33 76 05

(2)

Management Summary

Marketing and the pharmaceutical industry has been a disputable combination. Between countries there are differences in the acceptance of the use of promotional instruments in the pharmaceutical industry. A reason for this is that marketing spending by pharmaceutical manufacturers has increased significantly the last years (Narayanan, Desiraju, & Chintagunta, 2004). Because of these high marketing spending politicians and academics have been debating the effects of the use of promotional instrument by the industry on social welfare for many years. For many instruments however the effects are still unclear.

In the marketing field accountability is becoming increasingly important (Rust et al., 2004). Showing that marketing is an investment instead of a cost has been the focus of many researchers. Accountability is important since it increases the influence of the marketing department in a firm (Verhoef & Leeflang, 2009). The present research contributes to the accountability research in the pharmaceutical industry by focusing on modeling interaction and dynamic effects of promotional expenditures in this industry. A marketing model is a representation of the most important elements of a perceived real-world system (Leeflang et al., 2000). The importance of such marketing models has increased over the last years (Wierenga, Bruggen, & Staelin, 1999). It helps managers interpret the growing amount of available data (Divakar, Ratchford, & Shankar, 2005).

(3)

For managers the importance of stock-variables implies that a long-term focus is important in the pharmaceutical industry. Promotional expenditures should be spread over several time periods to build a stock. For instance when physicians are visited by a sales representative they do not always have a patient that needs the promoted medicine. Consumers need time to influence their physician in prescribing them the medicine that they saw a add of. For these reasons it is important that managers focus on building a stock.

Between countries and politicians there are differences in the acceptance of direct-to-consumer advertising(DTCA). This research shows that DTCA can be an effective promotional instrument. Based on the relative importance it is shown that detailing is the most important direct instrument, but DTCA is the most important stock instrument. This research does not conclude on effects of promotional instruments on social welfare, it does conclude that there are significant effects of promotional instruments in the pharmaceutical industry. Outside of the scope of this research, this research also indicates that seasonal influences can be an important factor in explaining the variation is sales. Some medicine cure diseases that appear more during for instance the winter. Including a seasonal variable explains much. Based on the present research the following conclusions can be drawn:

- Detailing aimed at physicians is the most important direct variable in explaining the sales of a individual brand in the pharmaceutical industry

- Stock effects of promotions are more important than the direct effect in explaining the sales of a individual brand in the pharmaceutical industry.

- The stock of direct-to-consumer advertising is the most important overall variable in explaining the sales of a individual brand in the pharmaceutical industry.

(4)

Chapter 1: Introduction

A manager once said: “half the money that goes to marketing is wasted; the problem is I do not know which half.” This single statement is one of the main reasons for this thesis. This phrase tells lot; on the one side this manager knows he needs marketing to be successful, however on the other side he thinks marketing is a waste of money. The key here is the lack of accountability in the marketing expenditures. In other words marketing has difficulties in justifying its expenditures in terms of direct return on investment (ROI). If the manager would know which half of the marketing budget has a high ROI, and which half has a low ROI, he would surely cut the lower half; thereby making his investment more effective. Rust et al. (2004) confirm this, when they state: “having exhausted cost-saving opportunities in virtually every other function, marketing is next in the line of fire.” The marketing functions can only keep their budgets when they are able to justify their expenditures.

1.1 Pharmaceutical Industry

Accountability is especially interesting in the pharmaceutical industry. Besides the patients, the main stakeholders in the pharmaceutical industry are, manufacturers, physicians, governments, insurers, pharmacy benefit managers, wholesalers and pharmacies (retailers) (Manchanda, et al., 2005). The pharmaceutical industry is not a normal industry. For example innovation and patenting play a much larger role than in other industries. In addition, the margins manufacturers make are above average (Windmeijer et al., 2005). Another difference is that not the end user but mainly the physicians select a certain type and brand of medicine and not the patient. Physicians have a unique gate-keeping function (Campo et al., 2005). They prescribe the medicine; since special knowledge is required to know which product is the best product for the situation (Kremer et al., 2008). The medicine also is not directly paid by the patient; in Europe it is mostly paid by an insurance company (Windmeijer et al., 2005). From this we can see that the demand side of the medicine market differs significantly from a normal market. Hence, the actions by participants in this industry have a direct impact on the welfare of consumers and society (Manchanda, et al., 2005).

(7)

2 profitable industries since 1982 (Public Citizen 2001). The high profits in combination with the high marketing spending have made the combination of marketing and the pharmaceutical industry highly controversial.

1.2 Marketing and the pharmaceutical industry

The promotional expenditures in the pharmaceutical industry can be divided into two categories. There are (1) direct-to-physician (DTP) and (2) direct-to-consumer (DTCA) expenditures (Kremer et al., 2008). Examples of DTP are journal advertising, events and detailing. Especially detailing, visiting by sales representatives, is considered by researchers as an effective promotional instrument, since physicians appreciate the visits and find the information valuable (Campo et al., 2005; Narayanan, Manchanda, & Chintagunta, 2005). Examples of DTCA is advertising via mass media. In 1997 the US food and medicine administration changed its policy concering DTCA. They relaxed the restrictions on DTCA making the advertisments easier to broadcast due to the lower requirements, this lead to an explosive growth of the use of DTCA by the pharmaceutical industry in the US (Narayanan, Desiraju, & Chintagunta, 2004). Both DTP and DTCA can have two effects; (1) informative (indirect) or (2) persuasive (direct). During the product life cycle of a product this influence can change (Narayanan, Manchanda, & Chintagunta, 2005). The informative effect can be described as a representative of the industry explaining to a physician which new medicines are being developed and how they work. With the persuasive effect the goal of the representative is not just to inform the physician about the new product, but to persuade the physician in to prescribing that particular medicine.

(8)

3 know they are able to recover these cost faster (Kremer et al., 2008). This stimulates companies in permanently developing new medicines, and enable physicians to cure diseases that could not be cured before, the later improves again for the social welfare.

The problem with marketing in the pharmaceutical industry is that many of these the effects are unclear. In this respect Windmeijer et al. (2006) state that the use of DTP lowers the price elastisity, while Wieringa & Leeflang (2013) research showed that the model used by Windmeijer et al. (2006) had been based on wrong assumptions. Other authors such as Rosenthal et al., 2003 report strongly positive effects of DTCA advertsing, while others reveal negative elasticities (Ling, Berndt, & Kyle, 2002). Wosinska (2002) and Donohue & Berndt (2004) found that DTCA-advertising does effect the choice probability, but that the impact of DTP is significantly higher. From the above we can conclude that further research into the effects of the use of promotional instruments in the pharmaceutical industry would be interesting from several points of view; (1) politicians for having a base for decision making, (2) researchers for knowledge development and (3) managers for knowing how effective there investments are, ROI.

(9)

4 This thesis builds on the research of Wieringa & Leeflang (2013) and looks at the dynamic and interaction effects of marketing in the pharmaceutical industry. In other words the duration of an effect of marketing, and wheter the use of one instrument increase the effect of the other instruments. Thereby this thesis wants to contribute to the research on the effects of promotional instruments in the pharmaceutical industry.

1.3 Interaction and dynamic effects of promotional activities

As stated before many effects of the use of promotional instruments in the pharmaceutical industry are not known or are disputed. Two of these effects are: the interaction effects between instruments and the dynamic effects of an instrument. An interaction effect occurs when the use of a marketing instrument makes the other marketing instrument more effective. The more the first instrument is used the stronger the effect of the second instrument will be. Kremer et al. (2008) listed a number of interactions between promotional instruments and for instance the type of disease, or the region. Other authors like Campo et al. (2005) suggest a few interactions; for instance between price sensitivity and promotional instruments like events. Narayanan, Desiraju, & Chintagunta, (2004) made empircal generalizations about the interactions between several promotional instruments. They found a positive interaction between DTCA en detailing, and a negative interaction between detailing and price.

Dynamic effects can be captured by time-series data. A dynamic effect is the duration of the effect in time, for example if you launch a campaign in week 1 how long will it take till the effect of this campaign is measured. These effects can be only direct, short-term, or long-term effects. Narayanan, Machanda & Chintagunta (2005) already showed that there are differences in the dynamic effects of the several promotional instruments in the pharmaceutical industry.

1.4 Research question

As stated above the effects of the several promotional instruments are still unclear. This leads to problems for managers and policymakers, since, they lack a base for their decisions concuring the use of promotional instruments in the pharmaceutical industry. To focus further research this on effect the following research question is stated.

What is the relevance of interaction and dynamic effects of promotional expenditures in the pharmaceutical industry?

(10)

5 Are there significant interaction effects to be modeled between various marketing instruments in the pharmaceutical industry?

Are there significant dynamic effects to be modeled in the use of marketing instruments in the pharmaceutical industry?

The focus of this research is the US pharmaceutical industry; this has been chosen because of the available data and because of the controversy and the debates concerning the use of marketing in this industry. With this dataset a distinction can be made between DTP and DTCA promotional instruments.

1.5 Contribution to research

In marketing literature the effect of promotional instruments in the pharmaceutical industry has been studied thoroughly. This is mainly due to the characteristics of the industry and the debates around the use of marketing. In a meta-analysis Kremer et al. (2008) gave an overview of what effects of promotional activities by pharmaceutical companies is known and what is not yet known. The main conclusion from their analysis of studies states that the effectiveness of the separate promotional instruments is still unclear, and requires further research. What is known is that DTP has a positive effect on the sales of a brand, but this effect is modest. The pharmaceutical industry spends huge amounts of money on DTCA, however there are difficulties in justifying the large DTCA spending (Narayanan, Desiraju, & Chintagunta, 2004).

(11)

6

1.6 Structure of thesis

(12)

7

Chapter 2: Literature review

As mentioned in the introduction in this section related literature is reviewed. The literature review starts with a focus on accountability of marketing and how to measure and model accountability. After this the focus shifts to the promotional instruments used in the pharmaceutical industry. At the end of the literature review two effects; interaction and dynamic effects will be explained as they are the focus of this research.

2.1 Accountability

Accountability and the effect of the marketing on an organization did receive a lot of attention in the literature. In an article about the influence of marketing departments within a firm Verhoef & Leeflang (2009) state that accountability is one of the main drivers of the influence of a marketing department in a firm. An important development they see is that marketing is perceived as a cost and not as an investment. Accountability can turn this around since it shows the effects of marketing on firm performance. The importance of accountability in the field of marketing has been shown by many authors (Lehmann, 2004; Rust et al, 2004; O' Sullivan & Abela, 2007). Another reason why de influence of marketing in the boardrooms has decilined is the focus on shareholder value of top management, marketing however has a more long-term perspective (Doyle, 2000).

Accountability can be defined as the ability to justify expenditures in terms of return on investment (Verhoef & Leeflang, 2009). Rust et al. (2004) define accountability as the ability to measure the productivity of marketing. In their research accountability is about clarifying the ways in which marketing activities build shareholder value. The focus here is on marketing expenditures and how they influence marketplace performance. According to Rust et al. (2004) a company should have a business model where they track marketing expenditures and see how these expenditures influence what customers know, believe, feel and how they behave. It is important here to not only measure the effectiveness of marketing in financial metrics but also on nonfinancial metrics.

Ten years before the Verhoef & Leeflang research, Moorman & Rust (1999) already saw that a market orientation has positive impact on the firm’s performance. They define market orientation as ‘the organization wide generation, dissemination and responsiveness to marketing intelligence.’ It involves multiple departments sharing information about customers

Influence of the marketing

department/function Market Orientation Firm performance Accountability

(13)

8 and participating in activities to meet customer needs (Narver & Slater, 1990). Moorman & Rust (1999) concluded that the effectiveness of this orientation depends on the presence of marketing function, or department. The authors make a distinction between marketing function as a function and marketing as a set of values and processes. Within their research they show that to be effective in having a market orientation, the influence of marketing as a function is important. Figure 1 summarizes the steps from accountability to firm performance (O' Sullivan & Abela, 2007). Rust et al. (2004) summarize it as follows: “The effective dissemination of new methods of assessing marketing productivity to the business community will be a major step toward raising marketing’s vitality in the firm and, more important, toward raising the performance of the firm itself.” Accountability will help marketing to be a real function and not just a vision for the future. Moorman & Rust (1999) argue that it is important to have marketing as a function in the firm. In combination with the market orientation, marketing will have a positive impact firm performance.

The lack of accountability has undermined the influence of the marketing department in a firm, in other words the marketing department has not been able enough to show its contributions (Rust et al., 2004; Lehmann, 2004). The pressure to demonstrate the contribution to the firm performance is growing. In a response marketers are investing in the development of performance measurement abilities (Doyle, 2000). When a marketing department is able to measure its performance their influence within the firm increases, and the overall firm performance increases (O' Sullivan & Abela, 2007; Rust et al., 2004; Verhoef & Leeflang, 2009; Verhoef et al., 2011).

(14)

9 improved by adopting a more fact-based culture, in other words marketing decisions should be based on facts and financial plans (Verhoef et al., 2011). Hereby it is important that the focus is on overarching market asset, and that the right metrics are chosen.

2.2 Measuring performance measurement

(15)

10 As many academics already address, measuring the effect and productivity of marketing activities is challenging. Rust et al. (2004) mention three challenges: (1) relating marketing activities to long-term effects (Dekimpe & Hanssens, 1995). (2) The separation of individual marketing activities form other actions. In other words the measurement of a single marketing asset. (3) The use of purely financial methods is inadequate for justifying marketing investments, nonfinancial metrics are needed. Transforming and implementing nonfinancial metrics, such as brand equity, into currency outputs, dollars, is difficult (Morgan, Clark, & Gooner, 2002). Market models are tools which can help in the interpretation and combination of several metrics.

Through the years many more metrics were designed, the problem nowadays in the accountability of marketing however is not the number of metrics available, but understanding the metrics. According to Clark (1999) the marketing field currently has enough measures. What the field needs is an understanding of the interrelationships among the measures. In other words what the effect of one measurement is on the other. An example of the interrelations is the link between brand equity, customer satisfaction and for instance the market orientation of a firm.

2.3 Marketing Models

A method in making marketing more accountable is the use of marketing models. Marketing models can assist marketing managers in their decisions. The definition of a model is: ‘a representation of the most important elements of a perceived real-world system’ (Leeflang et al., 2000). Models can predict sales, and they can provide insights about structural relations that are not available from casual observations. Models can provide insights in how a market reacts on marketing activities such as advertising, pricing and promotions hereby making these activities more accountable. Based on this information marketing managers can adjust their actions (Leeflang et al., 2000). Models explicate the relations, both the process of model building and the model that ultimately results can improve inputs for marketing decisions. According to Wierenga, Bruggen & Staelin (1999); the importance of marketing models has grown over the years; this is mainly due to technological advances.

(16)

11 past 25 years. Not only the quantity of models produced increased, also the quality of the models improved (Lilien, 2011).

A successful model can increase firm profit and measure performance. Wierenga, Bruggen & Staeling (2001) present five factors that determine the success of a marketing management support system. These five factors are summarized in Figure 2. First the demand side of the model. The demand side of a model consists of three elements; the decision problem, the environment, and the decision maker. The second factor that determines the success is the

supply side of the model, the model for instance has to be functional. Thirdly the match between demand and supply of the model should be good. The fourth factor is design characteristics of the model, for instance the model, or at least the result should be presentable so that managers understand the out comings. According to Leeflang et al. 2000 there are five criteria for making a good marketing decision model. The model needs to be (1) simple, (2) robust, (3) adaptive, (4) complete and (5) evolutionary. These five design criteria for a good model are elaborated int the next paragraph. The fifth factor determining the success of a model is the characteristics of the implementation process.

2.3.1 Design Criteria for a marketing model

The five design characteristics/criteria for estimating a good model have been described by Leeflang et al. (2000). The first is simple; this means that a model should not be too complicated. Managers who do not have an analytical background should be able to ‘read’ the model. This criterion can be achieved by keeping the number of variables low, using relative variable and constraining parameter values. The second criterion is evolutionary. This for instance means that model can be adjusted or expanded when this is necessary. The third

Demand: Decisions Problems Decision Environment Decision Maker Supply: Functionality Types of models: (Data or knowledge driven)

Design Characteristics: Simple Robust Adaptive Complete Evolutionary Match Implementation Support of management Communication Training Success of a model

(17)

12 criterion is that the model should be complete; this means that the model should contain all the important and relevant element. Only then can the model be a good representation of reality. Adaptive is the fourth criteria, markets change and market behavior are dynamic, this means that a model should adapt to changes. Model building is not a one-time thing; a model needs to be adapted continuously. The last criterion according to Leeflang et al. (2000) is that a model should be robust. This implies that it is difficult for a user of a model to obtain bad answers.

2.3.2 Classification of marketing models

In their book ‘Building Models for Marketing Decisions’ Leeflang et al. (2000) use five dimensions to classify models; (1) intended use, (2) level of demand, (3) amount of behavioral detail, (4) time series versus causal econometric models and (5) models for single products versus models for multiple products.

2.3.2.1 Intended use

The intended use is why a firm or manager wants to use a model. Different purposes lead to different models. Three types of intended use can be distinguished. (1) Descriptive, used to describe decision- or other processes. This type of models lead to a better understanding of decision processes, and can be used to investigate the possibility of automation. Another type is (2) a predictive model. As the name already suggest this type of model can be used to forecast or predict the future. Demand models are used to estimate the level of demand based on performance variable, like marketing decisions. The last category is (3) normative models, these models can be used to produce a recommended course of action. An example of a normative model is a media allocation model.

2.3.2.2 Level of demand

The level of demand relates to the output the model should give, this can be distinguished in individual and aggregate demand. Examples of aggregate demand models are: (1) industry sales or product class sales, (2) brand sales, (3) market share. The same distinction can be made for individual demand. Then the brand sales will be the brand sales for the household.

2.3.2.3 Amount of behavioral detail

(18)

13 some behavioral detail is explicitly shown. The last class consist of models with a (3) substantial amount of behavioral detail. The more behavioral detail a model includes the more equations and parameters the model has, this leads to greater complexity.

2.3.2.4 Times series versus causal econometric models

Time series models capture behavior over time. This type of models describe the variation of a variable over time. Interactions between marketing instruments and this variation indicate how effective this instrument has been, at least in the short-term. With time series models it is possible to include lag, past, and lead, future, effects. With time series models it is for instance possible to investigate how consumers react to a price cut. Causal relationships can be shown.

2.3.2.5 Models for single versus models for multiple products

Models can also be used when there is more than one product to be considered. For instance when a company has a product line, the total performance of this line could be more important than the performance of a single product in the line. Examples of multiproduct models are product line interdependencies and resource allocation decisions.

2.4 Interaction effect

When using multiple marketing instruments, interaction effects can occur. An interaction effect occurs when the effect of one variable on the dependent variable depends on the level another variable. It represents the process that drives a response parameter. For instance the effect of a TV-campaign on sales can increase when simultaneously it is supported by a radio-campaign. In this example a customer can be remembered via a radio commercial to the TV-campaign, by means of recognizing a similar tune in both commercials. This recognition increases the marketing effectiveness. This effect is called the interaction effect, and is shown in Figure 3.

In the literature there is empirical support for the presence of such interactions in marketing (Gatignon & Hanssens, 1987). Interaction effects can be positive; synergy, and negative; jamming (Kaul & Wittink, 1996). Examples of positive interactions in the marketing are that

(19)

14 advertising effectiveness increases with product quality (Keuhn, 1962) and with prior salesperson contact (Swinyard & Ray, 1977). Interactions can also be negative as Gatignon & Hanssens (1987) showed in their summary of interactions, with the negative interaction between advertising and sales force. Jamming, a negative interaction, might occur when one instrument is making the other less effective. This might happen when there is too much overlap between the instruments. An example of a known negative interaction is the interaction between detailing and DTCA (Narayanan, Desiraju, & Chintagunta, 2004). Awareness of interactions and what type of interactions there are; help managers to decide on how to design their marketing mix. They can use this information to decide on the elements of the mix and the extend of resources that should be devoted and too which elements they should limit their resources.

2.5 Dynamic effect

Marketing instruments often do not reach their full effects in a single period. This means that the effect of an advertising campaign possibly does not end when the campaign is over (Leeflang et al., 2000). The effect or part of this effect will be perceptible for some time periods. When an effect of a variable is measured during a longer time period this is called the lead or the lag effect. When sales in period t are affected by advertising in period t-1, t-2, this is called a lag effect. When current sales are affected by future variables, for example an anticipated promotion, this effect is called a lead effect (Leeflang et al., 2000).

(20)

15 In the pharmaceutical industry dynamic effects are normally not measured over one time period. According to Neslin (2001) half of the ROI of a marketing investment is not measured in the first month. This implies that the effect of promotional instruments in the pharmaceutical industry can last for several months. Another reason for the existence of lag effects in the pharmaceutical industry is that a physician can be reached by marketing who do not necessarily have a patient who need the promoted medicine immediately. The same holds for consumers, they also need processing time, they for instance need to make an appointment with a physician, obtain a prescription and so forth (Osinga, Leeflang, & Wieringa, 2010). In Section 3 the use of a variable measuring the expenditures up to six months is described to accommodate the lag effects.

2.6 Promotional instruments in the pharmaceutical industry

The decision process in the pharmaceutical industry has traditionally focused on the physicians. Due to this focus physicians have been the main target of the pharmaceutical industries marketing effort; this is called direct-to-physician (DTP) marketing. In DTP three instruments are used; journal advertising, detailing and physician meetings (Kremer et al., 2008). The physicians decide what medicine they prescribe to the patient. In this focus the patients were passive participants in the decision process (Manchanda, et al., 2005). However the last years the role of the patient in the decision process has changed. This is due to the increased use of direct-to-consumer (DTCA) marketing. Patients have gained more and more knowledge about the medicine. This has led to consumers being increasingly inclined to assert their perspectives in the process (Campo et al., 2008).

(21)

16

2.6.1 Detailing

Detailing can be described as a personal visit of a sales representative to the physician (Kremer et al., 2008). During this visit the representative informs the physician about the newest medicines his or her pharmaceutical company developed. The purpose of the meeting is that the physician chooses the brand of the representative. In this sense detailing is persuasive, since the representative is trying to persuade the physician to prescribe its product. Detailing is the most effective promotional instrument used in the pharmaceutical industry (Campo et al., 2005). Especially in the introduction phase of a new medicine, detailing is shown to be effective. It is also shown that detailing is 26 up to 43 times as informative as a single feedback from a patient in terms of learning about the efficacy of the medicine. (Narayanan, Manchanda, & Chintagunta, 2005). Campo et al. 2005 studied the effect of detailing on the prescriptions rates of the medicine after a visit of a representative. Physicians state that the persuasive effect is present but short lived. They acknowledge that a visit of a representative leads to an increase of medicine prescription, however according to the physicians this effect is only short-term.

Physicians only have limited time, this is why they appreaciate the visit of a sales representative. For the physicians detailing is a quick and valuable source of information (Campo et al. 2005). It is expected that due to the limited time of physicians the duration of the effect of detailing is longer than one month. Physicians do not daily have time for a represenative visit. Campo et al. 2005 showed that a too high frequency of representative visit works counter productive. Based on this one or more lags in the effect of detailing is expected.

Based on this research a positive effect of detailing expenditures on sales is expected, leading to the following hypothesis:

H1: Detailing expenditures are positively related to sales

H2: Detailing expenditures of previous periods have a positive influence on sales

2.6.2 Journal advertising

(22)

17 better for new medicines. Persuasive ads work better with established low-risk medicines. Narayanan, Machanda & Chintagunta (2005) call this difference indirect (informative) and direct (persuasive) effects of marketing. As stated before physicians’ time is limited, a short remembering to the visit of the representative they just had will expect to have a positive interaction effect (Narayanan, Desiraju, & Chintagunta, 2004). A lag effect is expected based on the fact that a journal is read longer than the month it is published in. Thus the following hypothesis are postulated:

H3: Journal advertising expenditures is positively related to sales.

H4: The effects of detailing expenditures on sales are positively influenced by journal

advertising expenditures.

H5: Journal advertising expenditures of previous periods have a positive influence on sales.

2.6.3 Physician meetings

Besides detailing, physician meetings are considered by physicians as an important source of information (Narayanan, Manchanda, & Chintagunta, 2005). Physician meetings are meetings, or conferences organized by the pharmaceutical industry for physicians. The goal of these events is informative and persuasive. Informative because one aim is to inform physicians about the latest developments in new medicines. Persuasive because the organizing manufacturer hopes that the physicians will prescribe their medicine. These meetings do have a positive effect on sales, but not as large as detailing (Kremer et al. 2008). An interaction effect between physician meeting and detailing can be expected according to Neslin (2001). Thus the following hypothesis are postulated:

H6: Physician meeting expenditures are positively related to sales

H7: The effects of detailing expenditures on sales are positively influenced by physician

meetings expenditures

H8: Physician meeting expenditures of previous periods have a positive influence on sales

2.6.4 Direct to consumer advertising

(23)

18 focus towards DTCA (Campo et al., 2005). However the effect of DTCA has been disputed. According to Narayanan et al. (2004) the effect of DTCA is small and positive. Studies by Wosinska (2002) and Donohue & Berndt (2004) find that DTCA-advertising does effect the choice probability, however the effect is small. Manchanda et al. (2005) state that the effect of DTCA advertising can be found in category expansion, not in the increase of sales of a single brand. Because of the DTCA, patients can be encouraged to seek treatment, or increase the purchase frequency (Wosinska, 2002). Non-brand specific ads have more effect than brand specific ads.

DTCA can also have negative effect, since physicians have to spend time re-educating the patient (Campo et al., 2008). A 1997 survey, for example finds that 29% of consumers who saw a medicine ad talked to their physician about it (Wosinska, 2002). Overall physician’s attitude towards DTCA is negative. Physicians ultimately decide which product the patients will get, this could lead to a negative effect of DTCA. Physicians do not always value the patient involvement; they feel threatened in their expert position. They get annoyed with the patient’s interference especially in risky choice situations, or they simply ignore the patients’ opinion on which medicine to prescribe (Campo et al., 2005). There is dispute about the effect of DTCA, for this reason an effect is expected, however the sign is unknown.

Naryanan et al. (2004) showed a positive interaction between detailing and DTCA (Narayanan et al., 2004). This interaction only occurs when two conditions are met: firstly the DTCA should lead to patients asking for the brand, secondly the physician should not be predisposed against DTCA and willing to involve patients in the prescription decisions (Campo et al., 2005). From the used data it can not be seen if the conditions are met, for this reason an interaction is expected.Based on the previous the following hypothesis is drawn: H9: DTCA expenditures are related to the sales.

H10: The effects of detailing expenditures on sales are influenced by direct to consumer

expenditures

H11: DTCA expenditures of previous periods have an influence on sales. 2.7 Price in the pharmaceutical industry

(24)

19 has shown that the price knowledge of physicians is typically low (Campo et al., 2005). This leads to small price elasticity for the demand of medicines (Windmeijer et al., 2005). Another reason why price elasticity’s in the pharmaceutical industry are low is due to government regulations. In many countries the price of medicines are set by the government (Manchanda et al., 2005).

A new phenomenon in the pharmaceutical industry is co-payment. Co-payment is a form of consumer cost sharing; this implies that the end-user shares in the cost for the prescribed medicine. Because the end-user now has to pay part of the medicine himself they ask physicians to prescribe the cheapest medicine. This makes the end-user and physician more price sensitive (Manchanda et al., 2005). According to Manchanda et al. (2005) the role of price in the pharmaceutical industry is an open area for research. This however is not in the scope of the present research, since price is heavily regulated and physicians show little price knowledge; price as a variable is not included in the present research.

2.8 Conceptual model

(25)

20 + H3, H6 & H9 +/- +/- Sales Direct-to-consumer advertising Physicians meetings Journal advertising Previous (lag effect):

Direct-to-consumer advertising Physicians meetings Detailing Journal Advertising Detailing H2, H5, H8 & H11 H1 +/- H4, H7 & H10

(26)

21

Chapter 3: Research Design

In the present research the hypothesis stated in Section 2.6 is tested on the pharmaceutical industry data used by Wieringa & Leeflang (2013), with a model. Two types of effects are modeled; firstly the dynamic effects of all the instruments and secondly the interaction effects on detailing of: DTCA, physician meetings and journal advertising. As stated by Leeflang et al. (2000) models should be built in an evolutionary way. This implies that the model building process starts with a conceptual model (Figure 4) which is translated into an econometric model. Leeflang et al. (2000) in their book make a distinction between linear- and nonlinear models. A linear model is linear in the parameters, this implies that the original variables are linearly related or can be transformed to insure that the relation between the variables is linear. When this transformation is not possible in the model the model is nonlinear. In the present research a linear model is used. The method used for estimating the parameters is the ordinary least squares (OLS) regression. Using an OLS regression implies a number of assumptions that need to be satisfied; these are discussed in Section 3.4. The validation of the model is discussed in Section 3.5.

In an OLS regression the number of parameters that can be used is not infinite. In adding parameters the rule of thumb of 5 observations per parameter is taken into account. Based on the data set and a hold-out-sample for validation, the model may contain a maximum of 16 parameters. This is based on 96 observations per brand from which 16 observations are used for the hold-out-sample. The hold-out-sample implies that the model is estimated using the first 80 observations. The model predicts an outcome for all 96 observations, to check the predictive validity the last 16 observations are compared to the actual observations. Conclusions on pooling issues and the level of analysis from the Wieringa & Leeflang (2013) study are taken into account. In the next section the modeling of the interaction and dynamic effects will be described. Thereafter, the error term assumptions are discussed.

3.1 Data

(27)

22 brand. Due to time limitations, the estimation of 519 brands is out of reach. Every brand requires an own model estimation and own interpretation. For this reason a total of 10 brands has been selected to estimate the model. The selection has been done by looking at the largest brand brands. This implies that generic brands are excluded. For some categories, for instance medicines on flue, a seasonal fluctuation; seasonality is expected. Like Osinga, Leeflang & Wieringa (2010) the seasonality is inspected visually. In case of seasonality in sales, three quarterly dummies are added. In the pharmaceutical industry in some categories it can be logical that seasonal influences are present. For example during the winter more medicines to fight flue, colds and maybe depressions will be sold. However, for other categories, seasonality will not be logical. For example in the cancer category seasonal effects are not logical.

The dependent variable in the model is the brand sales in period t. In the dataset the sales are represented by the revenue of the brand in that specific month. The independent variables are the use of several promotional instruments and possible seasonality, all the data in the data set are monthly data.

3.2 Modeling Interaction Effects

In a linear model the interaction effect can be modeled as follows:

is a representation of the interaction effect between the first and the second parameter (Leeflang et al., 2000). The problem in modeling interaction effects is that an increase in predictor L variables contains 2L terms. This makes it necessary for a model builder to specify in advance which interactions to include. In this research three interaction effects are included; these are summarized in Table 1. Through the significance of the interaction parameter the presence of such an effect can be tested.

Hypothesis Interaction effect Expected sign

4 Detailing x Journal advertising +

7 Detailing x Physician Meetings +

10 Detailing x DTCA +/-

(28)

23

3.3 Modeling Dynamic Effects

There are several ways of modeling the duration of an effect. Clarke (1976) describes a few of these direct duration interval models. One of these is the direct lag model, as shown below. The lagged values of for instance advertising are included as independent variables:

where:

= sales at time t

= advertising expenditures at time t – j = disturbance term

A problem with using lag models is the loss of degrees of freedom; this is due to the increase in the number of parameters and from a decrease in the number of usable observations (Leeflang et al., 2000). Another problem that arrises is deteriming the rigth number of j, the number of lag periods included; the duration interval. This problem is called the truncation problem (Clarke, 1976; Leeflang et al., 2000).

In the previous section only one version of a lag model was represented. In the literature many more models are discussed, like the distributed lag model, partial adjustment model, etc. (Clarke, 1976). When j is the number of lag periods, the number of usable observations is T-j. The number of degrees of freedom, # available observations - # parameters, of the error term then is T-2-2j. With every additional lag period, one additional observation is lost, which means a stronger decrease in degrees of freedom. This strong decrease in degrees of freedom is due to the first observation that does not have a lag. The lag variable is created by shifting all the observations one period. The larger the j the smaller the number of degrees of freedom becomes (Leeflang et al., 2000)

(29)

24 time i, and is the proportion of the stock of a period that is carried over, similar to Rizzo (1999) the is set to 0.7 (Narayanan, Manchanda, & Chintagunta, 2005). A stock variable solves the degrees of freedom problem that arises when introducing multiple lag periods, the problem is solved since only one additional variable is added which reduces the degrees of freedom by 2.

In this research for four promotional instruments eight dynamic effects are included; four one period lag variables and four stock variables. These are summarized in Table 2. Through the significance of the dynamic parameter and the better fit of the model, the presence of such an effect can be tested.

Hypothesis Dynamic effect Expected sign

2 Detailing t-1 & Stockvariable +

5 Journal advertising t-1 & Stockvariable + 8 Physician meetings t-1 & Stockvariable + 11 Direct-to-consumer t-1 & Stockvariable +/- Table 2: Dynamic effects

3.4 Error term assumptions

When using an OLS regression a few error-term assumptions have to be tested. Violation of the assumptions can lead to wrong estimation of the parameters and a wrong estimation of the variance of the parameters. According to Leeflang et al. (2000) it is wrong to assume that the disturbance term satisfy the assumptions. The reality is never ideal, for instance all the relevant predictor variables are rarely available. For this reason the following assumptions are tested: 1. for all jt; 2. for all jt; 3. is normally distributed 4. Multicollinearity 3.4.1 Independence

(30)

25 that other variables capture the omitted variable’s effect leading to biased parameters. Reasons for the violation of this assumption can be a measurement error in one or more predictor variables. Another reason can be that there is still a pattern in the error-term which is not accommodated in the model, for instance seasonal influence.

The assumption of independence of the error-term is tested using the Durbin-Watson test. This test tests for the amount of information still remaining in the residual, for instance seasonal influence or a lag effect. A remedy for the detection of autocorrelation is a re-estimation by the generalized least square - (GLS) method, which is written as

(Leeflang et al., 2000).

3.4.2 Homoscedasticity

The second assumption; for all jt, is that the error term is homoscedastic, in other words that the error term has the same variance for all possible values of a predictor variable (Leeflang et al., 2000). The result of violating this assumption is a wrong estimate of the variance of effect, the remedy for when this assumption is violated is a re-estimation by GLS. Homoscedasticity is tested visually, via a scatterplot, when there is a constant relationship between the residuals and the predictors, the error term is homoscedastic.

3.4.3 Non-normality

The assumption of a normally distributed error term also has to be assessed. The disturbance term needs to be normally distributed for the standard test statistics for hypothesis testing and confidence intervals to be applicable. Normality is tested with the Kolmogorov-Smirnov test.

3.4.4 Multicollinearity

(31)

26

3.5 Validation

The success of a model depends heavily on the quality of the data. The quality of the data is determined by availability, quality, variability and quantity of the data (Leeflang et al., 2000). As discussed previously the model is estimated on brand level, this implies that the quantity of the data is not very high. The variability of the data could also be better, for example DTCA is not used without the support of other instruments. This can make the modeling of interaction effect difficult. Besides the variability and the quantity of the data, the data set still contains good data based on the availability and the quality. However due to the lower quantity and variability of the data the model can perform less well. The goal of the model-building process is to end with an acceptable, final model. The quality of the results of this final model is assessed during the validation process.

Based on the assumptions the model is re-estimated. The model is validated based on three validation methods (Leeflang et al., 2000): (1) face, (2) statistical and (3) predictive validity. The first validity face validity relates to the believability of the structure and outputs of a model. This is based on theoretical and common-sense expectations and broadly accepted results. For example price is expected to have a negative effect on sales and advertising is expected to have a positive effect.

The second type of validity is statistical validity. This validity can be addresses in several ways; for instance the error term assumptions as describe previously, the significance of the equation and parameters is also an important check. The fit of the model is tested using the R2 and adjusted R2. The fit of the models is also addressed using the AIC measure, with the AIC nested models can be compared. The third type of validation is the predictive validity, in other words the model should also hold outside the data used for estimation. This is done by using a hold-out sample. The test used is the mean absolute percentage error (MAPE), the formula for assessing MAPE is as follows.

(32)

27

3.6 Model specification

Based on the conceptual model an OLS regression model has been specified. The model contains 18 parameters, when seasonality is included. Without seasonality the model contains 15 parameters; both are in line with the rule of thumb of minimum of 5 observations per parameter.

(1)

= Sales revenue for brand b in month t

= Detailing expenditures for brand b in month t

=Journal advertising expenditures for brand b in month t =Physician meeting expenditures for brand b in month t =Direct-to-consumer advertising for brand b in month t

= Detailing expenditures for brand b in month t-1

=Journal advertising expenditures for brand b in month t-1 =Physician meeting expenditures for brand b in month t-1 =Direct-to-consumer advertising for brand b in month t-1 =Stock variable for detailing expenditures for brand b in month t

=Stock variable for journal advertising expenditures for brand b in month t =Stock variable for physician meeting expenditures for brand b in month t

=Stock variable for direct-to-consumer advertising expenditures for brand b in month t =Interaction effect between detailing expenditures and journal advertising for brand b in month t

= Interaction effect between detailing expenditures and physician meeting expenditures for brand b in month t

= Interaction effect between detailing expenditures and direct-to-consumer advertising for brand b in month t

= dummy proxies for seasonal effects: 1 if the observation is in month t and, 0 otherwise. Here the winter season is chosen as baseline (months 1-3). l = 1 is spring (months 4-6). l = 2 is summer (months 7-9). l =3 is autumn (months 10-12)

(33)

28

Chapter 4: Results

In this section the results of the model estimations are presented and discussed. First the selected brands are described. This is followed by the parameter estimates of the 10 models. These models are tested on the assumptions as described in Section 3.4. The multicollinearity problem is investigated via three options: (1) deleting variables, (2) stopping the stock variable at t-1, (3) summarizing the physician oriented promotional instruments. Based on these test the re-estimation has been described in Section 4.4. The re-estimated model is tested on three types of validly in Section 4.5.

4.1 Brand selection and description

The selection of brands that are used for estimating the model has been done along a few criteria. The data set contains 127 generic and 393 brand brands. The generic brands are taken out. Since there is no longer a patent on these medicines and the same medicine is sold by many manufacturers, the relevance of the promotional instruments is much less. The second criteria is the size of the brand, the larger the brand the more promotional effort is expected. The third criteria is that there are enough observations per brand that can be used. This implies that a brand that is launched half way the observation period will not be used. This is for instance the case with Brand 246. The final selection the selection of the nine brands is shown in Table 3. The development of the revenues of these over time brands can be seen in Figure 5. From the graph it can be seen that seasonality can be expected with Brand 46. Except the Brands 502&94, all the selected brands show an increase in revenue over time. MedID MedName Category

code Launch Year Init Period Mean revenue Standard deviation 375 Priolosec 23400 1989 1 $177444.80 106220.70 392 Prozac 64340 1988 1 $155153.80 49768.07 512 Zoloft 64340 1992 1 $98240.10 40330.73 502 Zantac 23400 1992 1 $98019.75 65793.82 510 Zocor 32111 1992 1 $97279.68 57461.26 94 Claritin 14110 1993 4 $81749.44 54212.16 46 Augmentin 15600 1984 1 $78191.76 32898.68 347 Paxil 64340 1993 1 $75966.47 45752.03 368 Premanir Tabs 52112 1992 1 $75293.85 12646.67 378 Procardia 31700 1992 1 $74734.39 24278.20

(34)

29 In Table 4 the average amount of expenditures per promotional instrument is shown. From this table it can be concluded that except Brand 512, 502 & 46 all brands use all promotional instruments. From this graph it can be seen that all brands except brand 94 use DTL as their main promotional instrument. For the brands 375, 510 & 94 DTCA is an important promotional instrument. In Figure 6 the development of revenue and promotional expenditures is presented, here it can be seen that for brand 375 DTCA became increasingly more important, up to time period 50 DTCA was not used. It can also be visually seen that compared to other instruments DTCA shows higher peaks, these peaks can be observed with all brands that use DTCA. Detailing has been used throughout period 1 -96. The descriptions of the other brands are available on request.

Table 4: Average monthly promotional expenditures

0 1000 2000 3000 4000 5000 6000 7000 375 392 512 502 510 94 46 347 368 A ve rag e m o n th ly p ro m o tion al e xp e n d itu re s in $ Brand DTL JAD MTG DTC 0 50000 100000 150000 200000 250000 300000 350000 400000 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 M o n th ly R e ve n u e in $ Period 375 392 512 502 94 46 347 368

(35)

30 Figure 6: Development revenue and promotional expenditures of brand 375

4.2 Model assumptions

The assumptions are tested on three brands, namely brands 375, 392 & 512. Brand 512 has no DTCA expenditures, for this reason the DTCA’s variables are left out of these estimations. The test for testing the assumptions as described in Section 3.4.

4.2.1 Multicollinearity

The VIF’s and the number of significant variables of all the estimated models to solve the multicollinearity problem are shown in Table 5. An overview of the estimated models and included variables is given in Table 6. From the first estimation it can be concluded that multicollinearity is a problem. Multicollinearity can be detected using the variance inflation factor (VIF), the correlation between independent variables and logical arguments. In this model all three detections methods indicate the presence of multicollinearity. For instance in the model of Brand 392 ten variables have a VIF indicating multicollinearity. Logically seen, there are also a few indicators for multicollinearity, for instance the negative parameters of nearly all the promotional instruments. Only the stock variable has a positive and logical sign. The multicollinearity problem is accessed in four different ways: (1) deleting correlating variables, (2) stopping the stock variable at t-1 instead of t, (3) summarizing the physician oriented promotional instruments and (4) specifying a multiplicative model with interactions between all variables.

(36)

31 Model 1 2 3 4 5 6 7 8 9 10 11 12 Brand 375 R2 # VIF>10 # Significant AIC .962 .951 .939 .931 .965 .970 .925 .924 .903 .926 .817 .734 13 10 4 0 12 0 5 2 0 0 0 0 6 7 5 3 12 4 7 6 4 5 2 4 1981.80 1955.66 1971.03 1973.97 1890.30 1925.881 1966.55 1946.51 1967.60 -312.080 -279.411 1481.597 Brand 392 R2 # VIF>10 # Significant AIC .916 .905 .897 .887 .917 .894 .761 .750 .746 .770 Not .556 10 6 0 0 7 0 5 2 0 0 Sig. 0 5 5 4 4 3 6 4 5 6 3 4 1847.55 1873.50 1875.34 1876.58 1824.19 1833.873 1931.63 1912.32 1911.99 -224.707 1440.287 Brand 512 R2 # VIF>10 # Significant AIC .665 .615 .589 .556 .670 .614 .333 .308 .308 .423 .271 .250 7 5 0 0 5 0 2 0 0 3 0 0 5 4 3 2 6 2 3 2 2 2 2 2 1931.66 1961.42 1963.74 1962.03 1907.34 1912.255 1981.05 1961.04 1961.04 -329.217 -314.788 1357.543 Bold model indicates chosen model to accommodate for multicollinearity

Table 5 Multicollinearity diagnostics

Model Variables Model Variable

1 DTP, MTG, JAD, DTCA direct, lag, stock, interaction 7 PhysProm, DTCA direct, lag, stock, interaction 2 DTP, MTG, JAD, DTCA direct, stock, interaction 8 PhysProm, DTCA direct, lag, stock t-1, interaction 3 DTP, MTG, JAD, DTCA direct, stock 9 PhysProm, DTCA direct, lag, stock t-1

4 DTP, MTG, JAD, DTCA stock 10 Multiplicative model, DTP, MTG, JAD, DTCA, stock t-1 5

6

DTP, MTG, JAD, DTCA direct, lag, stock t-1, interaction DTP, MTG, JAD, DTCA direct, stock t-1

11 12

Multiplicative model, PhysProm, stock t-11 GLS Re-estimation of Model 6

(37)

32

Model 6 without lag and interaction with stock t-1

Brand 375 Brand 392 Brand 512 Brand 502 Brand 94 Brand 46 Brand 347 Brand 368 Brand 378

Overall sig. .000 .000 .000 .000 .000 .000 .000 .000 .000 Intercept -3581.43 34959.49 -76738.22 200045.95 -14278.86 66490.24 15761.47 73544.85 54221.30 Detailing 5.41 1.88 2.39 7.35*** 2.38 -2.26 1.36 -4.10*** -1.87 Journal -10.40 20.73* 1.28 51.70* 1.09 41.41* -61.53*** -15.33 4.29 Meetings -13.26 6.56* 6.67 24.77*** .38 22.29*** 14.42*** 5.10 15.97 DTCA 2.63* 2.01** -2.80 2.41*** 396.40 1.88 1.08 DetailingStock-1 8.14*** 2.19*** 5.03*** .71 3.36*** .07 4.71*** 1.81 1.04 JournalStock-1 27.44*** 5.44 7.17 12.66 -12.33 -33.91*** -52.81*** -19.44*** 7.81 MeetingStock-1 -9.89 14.42*** 12.85*** 25.22*** 1.41 22.44*** 8.69** 11.55*** 22.713 DTCAStock-1 4.56*** 1.82*** 3.33*** 1.54*** 141.01 -.17 .000 Season -4278.17* R2 .940 .894 .614 .937 .818 .702 .841 .850 .673 Adjusted R2 .935 .884 .587 .931 .801 .670 .826 .836 .651 Durbin-Watson .456 .495 .205 .556 .280 .564 .265 1.266 .178 AIC 1925.881 1833.873 1912.255 1844.053 1903.310 1861.826 1857.968 1606.759 1807.57 * Sig at 10% ** Sig at 5% *** Sig at 1%, BOLD parameters indicate VIF >10

(38)

33 Mean-centering is not considered as a solution to the multicollinearity problem. In their work Echambadi & Hess (2007) prove that mean-centering is not the sollution in multiple regression models. They show that: ‘mean-centering neither changes the computational precision of parameters, the sampling accuracy of main effects, simple effects, interaction effects nor the R2.” The determinants of the cross product matrix X’X are identical for uncentered and mean-centered data. Hereby not solving the multicollinearity problem. Echambadi & Hess (2007) state that researchers should not use mean-centering as a sollution to multicollinearity problems. For this reason mean-centering is not considered.

In Table 7 the results of Model (6) are shown, the other estimates are available on request.

4.2.1.1 Deletion of variables

There are a few solutions to multicollinearity; the most logical solution in this case is to delete variables that capture the same effect. In this case the direct lag effect and the stock variable capture the same effect. The direct lag- effect has been deleted from the model, leading to Model (2), since the stock variable captures a longer period of time, which in the case of the pharmaceutical industry is more logical. The re-estimation result are shown in Table 5, these results still show VIF’s >10. This is due to the correlations between the direct and the interaction effects. Based on these high correlations the interaction effects were taken out, leading to Model (3). However in Model (3) there were still a number of parameters with a VIF >10, based on this the direct effects were taken out, leading to Model (4), which only has four stock variables as independent variables.

(4)

4.2.1.2 Stopping the stock variable at t-1

The focus of this research is modeling the interaction and dynamic effects. Model (4) only takes into account dynamic effects. To reduce the correlation between the stock variables and the other variables the stock variable is stopped at t-1 instead of t. This leads to Model (5). As can be seen from Table 5 stopping the stock variable at t-1 does not solve the multicollinearity problem, however it does lower the VIF’s.

(39)

34 Model (6) is estimated without the lag and the interaction effect since these groups of variables are the main cause of the multicollinearity. Model (6) shows no signs of multicollinearity, the VIF’s are <10, and there are a number of significant parameters.

(6)

4.2.1.3 Summarizing the physician oriented promotional variables

Table 5 shows that it is not possible to model the interaction effect with Model (1). This is partly due to the high correlations between the different promotional instruments. Multicollinearity is often caused by manager’s decisions. In the current dataset for example DTCA is always supported by another instrument. This decision led to multicollinearity making it impossible to investigate the effect of the individual instruments. To investigate the interaction between DTCA and all physicians oriented promotional instruments, the physicians oriented promotional instruments are summarized, as seen in the formula: . This leads to Model (7) where there is only one interaction. As can be seen from the Table 5 summarizing the physician oriented promotional instruments does not solve the multicollinearity problem. However the high VIF’s are only the lag and stock effects.

(7)

Model (8) is specified and estimated with the stock t-1 variable, this model reduces the multicollinearity problem, however the VIF’s are still too high, this originates from high correlations between the interaction variable and the direct variable.

(40)

35 (9)

4.2.1.4 Multiplicative model

Based on the three solutions described above, it can be concluded that including the interaction effect in the model in the pharmaceutical industry is not possible. With a multiplicative model it is possible to model interactions, however the disadvantage is that interaction will be between all variables, hence making it impossible to explain the individual interaction effect. A log transformation is required before estimating the multiplicative model. Because log(0) is not possible al the values of 0 are recoded into log(x+1). This transformation is done for Model 6 and Model 9.

(10)

(11)

The results presented in Table 5 indicate that the multiplicative model is not a real improvement. Accommodating for interactions does not really improve the number of significant parameters and it does not improve the R2 much.

Based on the high R2, the low VIF’s and number of significant parameters Model (6) is chosen as the best model. This implicates that due to multicollinearity problems it is not possible to model individual interaction effects between the promotional instruments, hereby rejecting hypothesis 4, 7 & 10. The rest of the assumptions are tested using Model (6).

4.2.2 Independence of the Error term

(41)

36

4.2.3 Homoscedasticity

Homoscedasticity is tested visually with a scatterplot between the residuals and the predicted variables, there is homoscedasticity is there is a consistent relationship, heteroskedasticity when the relationship is erratic. In this case there is homoscedasticity. Figure 7 shows the scatterplot for Brand 375. The other scatterplots are available on request.

Figure 7 Scatterplot residuals Brand 375

4.2.4 Non-normality

Based on the Kolmogorov-Smirnov p-value of 0.200 (> .05) for eight brands, it can be concluded that the residuals are normally distributed. For two Brands; 347 & 378 the residuals are not normally distributed. For these models the hypothesis that each unobservable error follow a normal distribution cannot be rejected, this implies that the standard test statistics for hypothesis testing and confidence intervals are not applicable for these models. Brands 347 & 378 will therefore not be included in the interpretation.

4.4 Re-Estimation

Modeling interaction and dynamic effects of pharmaceutical promotional expenditures