Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues and an empirical illustration

Embed Size (px)

Citation preview

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    1/18

    Using cointegration analysis for modeling marketing interactions in

    dynamic environments: methodological issues and an empirical illustration

    Rajdeep Grewala,*, Jeffrey A. Millsb, Raj Mehtab, Sudesh Mujumdarb

    aDepartment of Marketing, Washington State University, Pullman, WA, 99164-4730 USAb

    University of Cincinnati, Cincinnati, OH, USA

    Received 6 June 1999; accepted 2 February 2000

    Abstract

    The authors argue that cointegration analysis is an intriguing development for analyzing marketing interactions in dynamic environments.

    Methodologically, the use of cointegration analysis requires statistical tests to determine whether this technique is appropriate for the system

    under investigation and, if it is appropriate, other statistical tests are needed to interpret the results. The authors collate a set of statistical tests

    and techniques to advance a comprehensive methodological framework that utilizes cointegration analysis to examine marketing interactions

    in dynamic environments. The framework is useful for analyzing marketing parameter functions with time-varying coefficients to investigate

    the relationship between market performance (e.g., sales, market share), marketing effort (e.g., advertising, sales promotion), and

    environmental conditions (e.g., market growth, inflation). The authors illustrate the utility of the framework for the famous case of Lydia

    Pinkham Medicine Company (LPMC). D 2000 Elsevier Science Inc. All rights reserved.

    Keywords: Cointegration analysis; Marketing interactions; Dynamic environments; Lydia Pinkham Medicine Company

    1. Introduction

    At the nucleus of marketing research and theorizing, lie

    marketing interactions. Marketing interaction mechanisms

    determine the relationship between marketing performance

    (e.g., sales, market share), marketing effort (e.g., advertis-

    ing, personal selling), and environmental conditions (e.g.,

    growth rate, competitive activities). Typically, researchers

    use market response models to investigate marketing inter-

    actions in order to examine the behavior of markets and

    predict the impact of marketing actions (Hanssens et al.,

    1990; Leone, 1995). Given the importance of marketinginteractions, scholars have proposed various methodological

    frameworks to model these interactions (cf., Wildt and

    Winer, 1983; Gatignon and Hanssens, 1987). Recent meth-

    odological advances in econometrics concerning cointegra-

    tion analysis provide a new technique to analyze these

    interactions. In this paper, we utilize recent advances in

    econometrics concerning cointegration analysis to illustrate

    a framework for analyzing marketing interactions.

    Since the path breaking paper by Granger (1981) and the

    subsequent conceptual and methodological developments

    by Engle and Granger (1987), cointegration analysis has

    become an integral part of non-stationary time series ana-

    lysis. Murray (1994) provided an intuitive explanation of

    cointegration. Murray (1994) uses the analogy of a drunkard

    walking her dog to explain the notion of cointegration. The

    drunk and her dog wander aimlessly, but make sure that they

    have an eye on each other and do not separate by more than

    a certain distance. Thus, even though both of them do not

    know where they are going, they do know that they are

    going together. In a way, the drunk and her dog arecointegrated. Formally speaking, two or more non-station-

    ary variables, which are integrated of the same order, are

    cointegrated if there exists a linear combination of these

    variables that is stationary. Specifically, cointegration ana-

    lysis involves time series data and multi-equation time series

    models, allowing for systematic and random parameter

    variation, with two or more variables.

    Marketing researchers have used multi-equation time

    series models to investigate various phenomena. For exam-

    ple, such models have been used to study the interaction

    between the structure of marketing function (brand vs.

    category management) and competition (cf., Zenor, 1994;

    * Corresponding author. Tel.: +1-509-335-5848; fax: +1-509-335-

    3865.

    E-mail address: [email protected] (R. Grewal).

    0148-2963/01/$ see front matterD 2000 Elsevier Science Inc. All rights reserved.

    PII: S 0 1 4 8 - 2 9 6 3 ( 9 9 ) 0 0 0 5 4 - 5

    Journal of Business Research 51 (2001) 127 144

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    2/18

    Curry et al., 1995); advertising and price sensitivity (cf.,

    Eskin and Baron, 1977; Krishnamurthi and Raj, 1985);

    advertising, temperature, price, and consumer expenditure

    (Franses, 1991); advertising, price sensitivity, and competi-

    tive reaction (Gatignon, 1984); advertising and product

    quality (Kuehn, 1962); advertising and product availability

    (Kuehn, 1962; Parsons, 1974); advertising competition(Erickson, 1995); advertising expenditure and advertising

    medium (Prasad and Ring, 1976); advertising and prior

    sales person contact (Swinyard and Ray, 1977); advertising

    and personal selling (Carroll et al., 1985); competitive

    behavior (Hanssens, 1980b); sales force effectiveness and

    environmental hostility (Gatignon and Hanssens, 1987);

    integrated marketing communications (cf., Beard 1996;

    Hutton, 1996); persistence modeling (Dekimpe and Hans-

    sens, 1995a,b); and consumer confidence (Kumar et al.,

    1995) among others.

    In most cases, conventional multi-equation time series

    analysis involves the use of Vector Autoregressive (VAR)models with two or more stationary variables (cf. Hamil-

    ton, 1994; Enders, 1995). Typically, one differences non-

    stationary difference variables to make them stationary

    and then uses them in a VAR model to investigate

    underlying data generation mechanisms (cf., Curry et al.,

    1995; Dekimpe and Hanssens, 1995b). Differencing non-

    stationary variables results in loss of information (cf.,

    Enders, 1995). Cointegration analysis provides a metho-

    dology for analyzing non-stationary variables, without

    making them stationary, thereby preventing loss of infor-

    mation due to differencing.

    Examples of marketing systems with non-stationary

    variables, which are related to each other and, thus, would

    benefit from cointegration analysis are plentiful. For in-

    stance, in a typical diffusion of innovation setting, where a

    new product is replacing an existing product, the sales of

    these two products, promotion and advertising spending,

    along with sales of competing products, are likely to move

    together and thereby be cointegrated. In addition, cointegra-

    tion analysis is a useful tool to examine sales force effec-

    tiveness (cf., Gatignon and Hanssens, 1987) and in

    understanding the implications of various pricing decisions

    and strategies on marketing performance (cf., Curry, 1993).

    These explications for application of cointegration analysis

    in marketing are by no means exhaustive and are meant asmere illustrations of the usefulness of cointegration analysis

    in investigating marketing interactions.

    Marketing researchers are just beginning to use coin-

    tegration analysis to study marketing interactions. Specifi-

    cally, a couple of studies (Baghestani, 1991; Zanias, 1994)

    examine the advertising sales relationship and Franses

    (1994) has studied the sales of new products. These studies

    and our illustrations demonstrate the utility of cointegration

    analysis; however, the intricate nature of theoretical re-

    search on cointegration limits its use. Our primary objec-

    tive is to summarize theoretical cointegration literature to

    facilitate its use by marketing scholars. Utilizing cointegra-

    tion analysis requires that all data series under investigation

    to be integrated of the same order, which implies that one

    has to perform statistical tests on the data series under

    investigation to make sure that the system under investiga-

    tion is suitable for cointegration analysis. In addition,

    drawing conclusions from the estimation results of coin-

    tegration analysis requires more statistical tests. The mainobjective of this article is to demonstrate a comprehensive

    methodological framework for analyzing multi-equation

    time series data using cointegration analysis. Such a frame-

    work is of considerable interest to both marketing scientists

    and marketing managers, as better understanding of mar-

    keting interactions is of interest to both parties. Both are

    interested in marketing interactions because they want to

    know what drives marketing performance. Our framework

    provides both parties with tools and a systematic method to

    study these interactions. Further, a comprehensive and

    consistent framework makes it easy to identify unifying

    principles that aid in empirical generalization and advance-ment of marketing science (cf., Bass, 1993, 1995; Bass and

    Wind, 1995). Finally, such a framework would be useful

    for pedagogic exposition.

    To achieve our objectives, we survey recent develop-

    ments in the econometrics and time series literature to

    collate a set of statistical tests and estimation techniques,

    which are useful in exploration of marketing interactions.1

    Based on our literature review, we illustrate the usefulness

    of cointegration analysis in marketing and provide the

    rationale for expecting specific type of behavior from

    various marketing variables. Furthermore, we demonstrate

    the proposed framework to model marketing interactions

    for the famous case of Lydia Pinkham Medicine Com-

    pany (LPMC).

    2. Methodological framework and conceptual

    underpinnings

    Marketing interactions, by their very definition, imply

    that interactions among several marketing effort variables,

    along with their interaction with environmental variables,

    determine marketing performance. Further, when firms take

    decisions concerning marketing effort, they may take mar-

    keting performance into consideration. In addition, environ-ment interacts with both performance and effort to further

    complicate matters. For example, the time of the year and

    advertising expenditure in the previous month together

    determine sales which in turn determines advertising ex-

    penditure this month which in turn influences sales. Multi-

    equation modeling helps in capturing this dynamic behavior

    1 We choose the statistical tests that in our opinion are most

    appropriate. We do not claim that these are the only or universally the

    best statistical tests for the purpose. Our objective is to provide and

    illustrate the steps of our framework and not to determine the goodness of

    one test vis-a-vis another.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144128

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    3/18

    in the market place. In Appendix A, we present a typical

    multi-equation model, which captures the dynamics of

    marketing interactions.

    To capture marketing interactions in a cointegration

    framework, we propose a nine-step framework to investi-

    gate the complex system represented in the two equations

    we present in Appendix A (Fig. 1). In the first four steps ofthe framework, i.e., unit root test, structural break test, unit

    roots with structural tests, and reconciling the results from

    the two unit root tests, we are concerned with determining

    the data generation process of each individual variables.

    Uncovering these aspects of the data generating mechan-

    isms, provides information whether the variables being

    studied are suitable for cointegration analysis or not. Sub-

    sequently, in the next two tests, i.e., cointegration test and

    estimation techniques, we use the results from the first four

    steps to model the interactions between environment, effort,

    and performance variables. Finally, in the final three steps,i.e., Granger causality, variance decomposition, and impulse

    response functions (IRFs), we use the inputs from the

    cointegration results to uncover interrelationships between

    the variables under investigation. In the remainder of this

    Fig. 1. Methodological framework.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144 129

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    4/18

    section, we enumerate on each of the nine steps in our

    framework and provide reasons for expecting certain beha-

    vior by marketing performance variables, marketing effort

    variables, and environmental variables.

    2.1. Unit roots2

    Dekimpe and Hanssens (1995a) operationalized the

    concept of stationary and evolving markets based on the

    unit root tests. The unit root tests examine each time series

    to determine whether the mean, variance, or autocorrela-

    tion of the underlying data generation process for each of

    these variables increases or decreases over time. A time

    series whose mean or variance or autocorrelation either

    vary over time or are not finite might be non-stationary

    and may have a unit root. Classical linear regression

    models requires data series under investigation to be

    stationary and if this assumption is violated it leads to

    the problem of spurious regression (Granger and Newbold,1974). Further, cointegration analysis requires all series

    under investigation to be non-stationary. Hence, it is

    important to identify, initially, the order of integration of

    the data generation process.

    One could hypothesize many marketing variables to be

    non-stationary based on their data generation process (De-

    kimpe and Hanssens, 1995b). For example, the vast litera-

    ture on diffusion of innovation suggests that the sales

    figures for a successful new product will grow during its

    initial years (cf., Mahajan et al., 1990). Further, one can

    expect price of some products to increase over time, perhaps

    due to inflation, and thereby be non-stationary. In addition,

    it is possible that price of some products decreases over time

    due to experience curve effects (Bass, 1995), thereby

    representing a non-stationary data generation process.

    2.2. Structural breaks

    The structural breaks represent a point or an interval in

    time, which denote modifications in the underlying data

    generation process. The modifying agent is usually an

    extraneous event. For example, structural breaks in sales

    might be due to interventions of federal regulatory agencies,

    as in the case of tobacco industry, where federal regulations

    on how and where to sell tobacco products are plentiful (cf.,Rogers, 1994; Economist, 1996; France, 1996). Other ex-

    amples of structural breaks include competitive new product

    introductions and new generation of products (cf., Norton

    and Bass, 1987, 1992; Mahajan et al., 1993).

    While analyzing 14 macroeconomic time series, Perron

    (1989) provided a startling finding that after correcting for

    structural breaks, like the exogenous oil price shock of

    1973, most of the macroeconomic series are either sta-

    tionary or trend-stationary. If a series is stationary or trend-

    stationary, cointegration analysis is not an option. Clearly,

    it is important to account for structural breaks when

    modeling economic time series to identify modifications

    in the data generation process. As Perron (1989) demon-strated, overlooking structural breaks might mislead con-

    clusions concerning the underlying data generation

    process, which may lead to model misspecification and

    wrong conclusions.

    There are two major issues concerning structural breaks.

    The first concerns the time when the break has its effect on

    the underlying data generation process: immediately after

    the event of interest or after a certain lag. Typically, either

    of the two cases is possible. If the event of interest is high-

    profile (e.g., oil shock of 1973), we might expect an

    immediate change. For low-profile interventions, like the

    actions of competitors or reprimand by federal agencies, thestructural change might be delayed, as the information

    needs time to diffuse through the social system (Mahajan

    et al., 1990).

    The second issue concerns the nature of the break. One

    can expect the mean of a series to change, or the slope of

    the data series to change, or changes in both mean and

    slope. An example of change in mean would be high-

    profile shocks, though this effect might be temporary.

    Interventions due to sales promotions or federal legisla-

    tion's fall in this domain. Changes in slope might be a

    result of federal regulations, competitive interventions, etc.,

    which need time to implement and diffuse through the

    social system. For instance, let us say that a federal agency

    issues a cease and desist order. The effect of this order

    could be gradual as information diffuses through the

    concerned social system (Mahajan et al., 1990). Finally, a

    successful new product introduction by the competitor can

    instantaneously reduce a firm's sales (change in mean) and,

    in addition, after the instantaneous effect, the influence of

    the new product may gradually erode more sales (change

    in slope).

    The time of structural change and the nature of the

    change are interesting in and of themselves. In addition,

    these univariate tests shed light on modifications in the data

    generation process for the time period under investigation.

    2.3. Unit roots after incorporating structural breaks

    Perron (1989) found most macroeconomic data series to

    be either stationary or trend-stationary after incorporating

    structural breaks. Traditional unit root tests (cf., Dickey

    and Fuller, 1981; Phillips and Perron, 1988) do not

    compensate for structural changes. Thus, it is possible

    that these traditional tests find unit roots in stationary

    process due to structural breaks. Hence, it becomes

    important to account for unit roots after incorporating

    structural breaks.

    2 In the remainder of Section 3, we elaborate on the rationale for

    expecting specific behavior form marketing variables. The statistical

    aspects of these tests and estimation techniques are discussed in Section

    4, where we illustrate the framework for the famous case of Lydia

    Pinkham Medicine.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144130

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    5/18

    2.4. Reconciling unit root tests before and after

    incorporating structural breaks

    The main reason to perform unit root tests after incor-

    porating structural breaks is to overrule the possibility that a

    structural break may be causing misperceptions concerning

    the stationarity of the variables under study. Further, thereexists a possibility that the unit root tests before and after

    incorporating structural breaks may not agree. If the results

    agree then we establish robustness of the findings. If the

    results do not agree, Perron (1989) suggests that one should

    proceed with and estimate models with the results from both

    unit root tests.

    So far, we have laid down the steps to investigate the

    underlying data generation process of each individual data

    series and have not examined the interaction between these

    data series. One can use the results from these steps to

    formulate an appropriate model for further investigation.

    Further, if we have two or more non-stationary time series,there is a possibility that these variables may be cointe-

    grated. In such a case, we must proceed with the coin-

    tegration tests, otherwise a VAR with stationary variables

    is appropriate.

    2.5. Cointegration

    In this step, we decide whether cointegration analysis is

    appropriate or not. If the variables under investigation are

    non-stationary and integrated of the same order, cointegra-

    tion analysis is mandatory. It is important to identify a

    cointegrating relationship between non-stationary variables

    because such a relationship implies an equilibrium between

    these variables and overlooking this equilibrium results in

    misspecifications in the error term (cf., Enders 1995). For

    example, we expect marketing effort to influence marketing

    performance, and for some products, we expect both types

    of variables to be non-stationary, e.g., in high-growth

    markets. Hence, we expect marketing effort and marketing

    performance to be cointegrated.

    2.6. Estimation

    We propose the use of standard VAR and Vector Error

    Correction Models (VECM) to estimate the relationshipbetween the variables under investigation. If some variables

    are non-stationary, but are not cointegrated, then one has to

    difference them in order to make them stationary. Subse-

    quently, we use these differenced transformed variables to

    estimate a VAR. Further, if one has cointegrated variables,

    one can estimate either a VAR in levels (i.e., with variables

    that have not been differenced), or VECM (cf., Toda and

    Yamada, 1996).

    However, before estimating a VECM, we have to deter-

    mine the cointegrating relationship that we can use as an

    independent variable in the VECM. This relationship is a

    linear combination of the cointegrating variable and is

    stationary. Various estimation procedures, such as Johan-

    sen's MLE, Box-Tiao, and OLS, are available to estimate

    the rank of the cointegrating vector (which equals the

    number of cointegrating vectors) and the cointegration

    relationships. Typically, Johansen's MLE (we use this

    method in our illustration) performs well with reasonable

    large sample sizes (cf., Johansen, 1988; Hargreaves 1992).Once we have obtained the VAR and/or VECM parameter

    estimates, we use these estimates to uncover the interactions

    among the variables in the system. Specifically, we use

    Granger causality, variance decomposition, and IRFs to

    study the dynamic system.

    2.7. Granger causality

    We expect marketing effort to determine marketing

    performance, in the words of time series literature, market-

    ing effort Granger causes marketing performance. Often the

    time paths of the two variables, marketing effort andmarketing performance, might show that the two variables

    move together, e.g., both increase and decrease together. A

    possible conclusion is that marketing effort is determining

    marketing performance. However, one can also argue that

    the firm is determining marketing effort based on marketing

    performance. After all, constant advertising to sales ratio

    strategies are quite common (Erickson, 1991). How do we

    determine whether effort is determining performance, or

    performance is determining effort, or both are determining

    each other? Granger causality can help determine this.

    2.8. Variance decomposition

    The decomposition of forecast error variance throws light

    on the effect size, i.e., how much of the forecast error

    variance of the focal variable is being explained by the

    variables of interests. For example, it helps to answer

    questions like how much of forecast error variance of sales

    is being explained by marketing effort and how much is

    being explained by environmental variables.

    2.9. Impulse response functions

    After giving a shock to a particular variable in a system,

    we use IRF to trace the time paths of all variables in thesystem. For instance, if we give a 10% shock to a firm's

    advertising (in other words, we increase/decrease the firm's

    advertising expenditure by 10%), we use IRFs to answer

    questions such as: Does the shock to advertising have a

    delayed effect on sales? How long does this effect last?

    What is the likely effect on the sales of competitor's

    product? What is the likely reaction of the competitor?

    To sum, the last three steps of the nine-step framework

    provide insights into the interactions between the variables

    under investigation. They aid in understanding the influ-

    ence of marketing effort and environment on marketing

    performance and also help to uncover any feedback from

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144 131

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    6/18

    performance to effort and/or environment. For the purpose

    of illustrating the framework, we examine the famous case

    of LPMC.

    3. Research setting: the LPMC

    We choose the LPMC to illustrate our methodological

    framework for two reasons. Besides easy access, we choose

    this database because two recent articles (Baghestani, 1991;

    Zanias, 1994) have examined this database using cointegra-

    tion analysis and we wanted to demonstrate that important

    underlying dynamics of the market process may be missed if

    one overlooks one or more of the steps of our framework

    (e.g., structural breaks).

    Palda (1964) provides a detailed review of the circum-

    stances that led to the disclosure of advertising and sales

    figures of LPMC. He also reviews the pertinent aspects of

    the history of LPMC. This section first traces the relevantevents that throw light on the advertising strategy of LPMC

    (drawing mainly from Palda, 1964).

    The primary product (nearly 99% of sales) of LPMC was

    a vegetable compound patented in 1873 alleged to cure a

    wide variety of ills related to ``women's weakness.'' The

    company relied solely on advertising as a means of promo-

    tion (Palda, 1964), changing the advertising copy only three

    times in the 54-year period. The aim of the advertising copy

    was to stimulate primary demand (Palda, 1964). Of the three

    advertising copy changes, two were due to orders issued by

    governmental regulatory agencies. The first of the two copy

    changes came about in November 1925 when the Food and

    Drug Administration (FDA) issued a cease and desist order.

    In 1938, the Federal Trade Commission had new objections

    to the then existing form of advertising of LPMC, which

    resulted in the second copy change in 1940. Winer (1979)

    succinctly summarized the advertising copy positioning

    strategy for LPMC as ``universal remedy'' for the period

    19071914, ``relief for menstrual problems'' for the period

    19151925, ``vegetable tonic'' for the period 19261940,

    and ``universal remedy'' again, for the period 19401960.

    In addition to these copy changes, LPMC followed an

    aggressive advertising strategy under Lydia Gove, who took

    over as director of the company in 1925. This streak of

    aggressive advertising (which started in 1926) reached its peak in 1934 with advertising to sales ratio of 85%. The

    then president of the company, Arthur Pinkham, took

    objection to the huge expenditure on advertising, resulting

    in a court case. This led to relatively lower levels of

    advertising from 1936. Schmalensee (1972) estimated that

    on average advertising was set at 64% of sales for the period

    19261936, whereas it was around 46% of sales for the

    other years.

    The vegetable compound did not have any close sub-

    stitute available for the time period (1907 1960) under

    investigation, thereby ruling out any competitive advertising

    effects (Palda, 1964). The price of the product, available in

    tonic and tablet forms, was fairly stable over this time

    period.3 Newspaper was the primary advertising medium

    used by the company and the media allocation remained

    fairly stable for the duration of the study.

    The primary role of advertising in the marketing strategy

    of LPMC, the lack of competitors, and the availability of

    detailed data result in a rather unique natural experiment forstudying the advertising sales relationship, with minimal

    variation in other variables (such as price, advertising

    medium, etc.). The uniqueness of the LPMC experience

    has led to an extensive literature analyzing the database.

    Beginning with Palda (1964), who estimated the Koyck-

    type distributed lag models using OLS, a host of researchers

    have applied increasingly sophisticated time series methods

    to study the LPMC data.4

    Recently, Baghestani (1991) uses cointegration analysis

    to investigate the advertisingsales relationship for LPMC.

    He found that the advertising expenditure and sales figures

    of LPMC are cointegrated in the order of one and, therefore,estimated an error correction model (ECM) to capture the

    short-run dynamics and long-run equilibrium conditions.

    Zanias (1994) replicated Baghestani's (1991) results and

    went on to show that forecasting with an ECM was more

    accurate in comparison to previous models. Further, Zanias

    found bi-directional Granger-causality between sales and

    advertising of LPMC.

    Despite their state-of-the-art application of (Baghestani,

    1991; Zanias, 1994) modern time series techniques, the

    results from these bivariate cointegration analyses could be

    misleading for the following two reasons. First, one cannot

    be sure that the results do not suffer from bias due to omitted

    variables, which could impact sales. In accordance with the

    law of demand, price is one such variable. In addition, the

    health of the economy is likely to influence demand and

    thereby sales. In order to remedy the omitted variable bias

    and to investigate the impact of price and the economic

    environment on sales, we include GDP to capture the level of

    economic activity and unemployment rate to utilize business

    cycles in addition to advertising expenditure and price.

    The second potential shortcoming, concerned with the

    political legal environment, is that of the changes in

    3 From 19151917, the price of the product in the liquid tonic form

    and the tablet form was US$7.28 and was increased to US$9 and then

    US$10 in 1918 and 1930, respectively. In May 1947, the price of the liquid

    form of the product was increased to US$11 and then to US$12 in January

    1948. The price for the tablet form of the product was increased to US$9.67

    in June 1948, to US$10.30 on March 1956 and finally to US$11 in

    November 1956 (see Palda, 1964, p. 39).4 These include Clarke and McCann (1973), Houston and Weiss

    (1975), Caines et al. (1977), Helmer and Johansson (1977), Kyle (1978),

    Weiss et al. (1978), Winer (1979), Hanssens (1980a), Mahajan et al., 1980,

    Erickson (1981), Bretschneider et al. (1982), Harsharanjeet et al. (1982),

    Heyse and Wei (1985), Magat et al. (1986). It is not our objective to survey

    the entire stream of research that this database has generated. Our analysis

    is based on two recent articles (Baghestani, 1991; Zanias, 1994), which

    utilize the techniques we explicate in this paper.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144132

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    7/18

    advertisement copy, which were required due to the regula-

    tions by federal agencies, namely FDA and FCC. Resolu-

    tions by such federal agencies are applied standards of what

    the concerned agency conceives to be of public interest and

    these resolutions reflect issues of not only politicallegal

    nature but also reflect culturalsocial values (Palamountain,

    1955). We expect these mandatory advertisement copychanges to influence the sales of LPMC and model these

    as constraints imposed by the legal environment. This

    surfaces in the statistical analysis in the form of structural

    breaks in the parametric model estimated. In light of this,

    we test for structural breaks and incorporate the detected

    breaks into the cointegration analysis.5

    4. Statistical analysis

    In this section, we use cointegration analysis to examine

    the impact of deflated GDP (RGDP), unemployment(UEMP), and deflated price (RPRICE) on both deflated

    advertising (RAD) and sales (RSALES). In addition, we

    investigate how advertising and sales influence each other in

    the presence of these three variables. In the case of LPMC,

    the price of the vegetable compound remained fairly stable

    for the period under investigation, but the price in real terms

    was changing. It is the price in real terms that truly reflects

    the cost of a product; therefore, we use real price as an

    explanatory variable.6 As the nominal price was fairly

    stable, one could view RPRICE as instrumenting inflation-

    ary pressures. To eliminate any spurious correlation due to

    inflationary effects between advertising and sales and to

    remain consistent across variables, we deflated both adver-

    tising expenditure and sales revenue. The consumer price

    index (base year 1967) was used to deflate advertising,

    sales, and price, and the GDP deflator (base year 1967) was

    used to deflate GDP.

    4.1. Unit root tests

    If a non-stationary time series yt can be made stationary

    after differencing itdtimes, then yt is said to be integrated of

    the orderd(denoted as yt$ I(d)). Tests suggested by Dickeyand Fuller (1981) and by Phillips and Perron (1988) are

    recommended to test for the order of integration of time

    series data.7 The augmented DickeyFuller (ADF) test, a

    generalized form of the Dickey Fuller test, is useful for

    testing for unit roots after incorporating appropriate lags.

    The following ADF equation is estimated:

    Dyt a0 yt1 ip

    i2

    biDyti1 4t 1

    where is the coefficient of interest. If we fail to reject H0:

    = 0, then the equation has a unit root, i.e., the underlying

    data generating process is non-stationary. However, it is

    possible that the equation has more than one root. Dickey

    and Pantula (1987) suggest that one could use the Dickey

    Fuller test recursively on successive differences of the

    concerned variable to detect multiple roots. While using the

    DickeyFuller tests, one must ensure that error terms are

    uncorrelated and have constant variances. Phillips and

    Perron (1988) developed a similar procedure to allow formilder assumptions about the distribution of the error

    terms. Note that the null hypothesis of non-stationarity

    forms the basis for both the Dickey Fuller test and the

    PhillipsPerron test.

    We utilize the Akaike Information Criterion (AIC) and

    Bayesian Information Criterion (BIC) as fit statistics for

    determining appropriate lag lengths. For RSALES and

    UEMP, both AIC and BIC gave two lags as appropriate.

    For the other three variables, there was no agreement

    between the two criteria. As the goal is to find proper

    relationships between variables, we took a conservative

    perspective and used the maximum of the appropriate lag

    length indicated by the two criteria.8 Hence, we use laglengths of three, four, and four for RAD, RPRICE, and

    RGDP, respectively.

    We present the results of the unit root tests in Table 1.

    The results show that the five variables are all integrated of

    order one, i.e., they are I(1), processes and, therefore, the

    system seems appropriate for cointegration analysis.9 How-

    ever, Perron (1989) found 14 macroeconomic time series to

    be either stationary or trend-stationary after correcting for

    structural breaks. In line with the evidence presented by

    Perron (1989), we went about testing for structural breaks in

    our five time series.

    4.2. Structural break tests

    For the LPMC, two possible events, besides the Great

    Depression, are suggestive of structural breaks. The first

    5 Note that we recognize that there is no way to be certain that one does

    not have an omitted variable bias. However, when theory suggests that

    specific variables are important (e.g., environment in the case of LPMC),

    one should attempt to, at least, control for them. In the case of LPMC,

    literature on marketing interactions suggests that we need to account for the

    environment (cf., Wildt and Winer, 1983; Gatignon and Hanssens, 1987).

    Based on this literature, we investigate the advertisingsales relationship

    for LPMC and control for the environmental effects.

    7 See Hamilton (1994) and Enders (1995) for thorough expositions.8 We estimated the concerned models with the lag lengths suggested by

    both AIC and BIC and got results similar to those from the conservative

    perspective.9 Nominal values of advertising (AD) and sales (SALES) were also

    tested for unit roots. Like Baghestani (1991) and Zanias (1994), these two

    series were found to be I(1).

    6 The vegetable compound was available in two forms, namely tablet

    and tonic. The price for both these forms was similar for most of the time

    period under investigation (see Footnote 3). In our analysis, similar results

    were obtained for both prices. For parsimony, we report results only for the

    price of tonic.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144 133

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    8/18

    is the two advertising copy changes initiated by the

    intervention of federal agencies. The first of these two

    advertising copy changes came in 1925 due to a ceaseand desist from the FDA. The second copy change was

    in 1940, when the FCC had objections to the existing

    advertising copy of LPMC. The second potential struc-

    tural break may result due to the streak of aggressive

    advertising strategy followed from 1926 to 1936 by Lydia

    Gove. We incorporate these structural breaks in our

    analysis as suggestive of the interventions by the legal

    environment. The years when LPMC had to change its

    advertising copy due to federal regulations can be used as

    suggestive of structural breaks in the advertising and

    sales series.

    Alternatively, some scholars suggest that one should let

    the data determine the time of structural change (cf.,

    Hansen, 1992). We followed Hansen's (1992) recommenda-

    tions and found structural breaks in the advertising and sales

    agreed with the dates suggested by the FDA and FCC

    interventions. These findings support the conjectures con-

    cerning the importance of the legal environment.

    We used Hendry's (1989) version of the Chow test,

    which relies on recursive updating of the residual sum of

    squares to test for structural breaks. The test was per-

    formed recursively for break in all time periods. Fig. 2

    shows the plot of t-values for this recursive test. As is

    evident from the figure, there were two breaks in both

    advertising and sales. Advertising had structural breaks in1925 and 1934, while sales had structural breaks in 1925

    and 1938. These dates agree with the advertising copy

    changes due to federal regulation and Lydia Gove's ag-

    gressive advertising strategy.

    We also performed the recursive structural break test on

    the other three variables. RPRICE had one structural break

    in 1933, RGDP had two structural breaks in 1931 and 1938,

    and UEMP had one structural break in 1930. As expected,

    the Great Depression seems to influence the breaks in these

    three variables. Subsequently, we used these structural

    breaks in Perron's (1989) test for unit roots in the presence

    of structural change.

    4.3. Unit root with structural breaks tests

    The Perron (1989) test for unit roots in the presence of a

    structural break in period t incorporates structural change in

    the period t = t + 1 and tests the following three null

    hypotheses against the appropriate alternatives. The first

    null hypothesis is of a one-time jump in the level of a unitroot process, and has the alternate of a one-time change in

    the intercept of a trend-stationary process.

    H1 : yt a0 yt1 m1DP 4t 2

    A1 : yt a0 a2t m2DL 4t

    where DP represents a pulse dummy variable. DP = 1 ift= t

    + 1; DP = 0 t T t + 1. DL represents the level dummyvariable and DL = 1, when t > t.

    The second null hypothesis is of a permanent change in

    the magnitude of the drift term vs. the alternate hypothesis

    of a change in the slope of the trend.

    H2 : yt a0 yt1 m2DL 4t 3

    A2 : yt a0 a2t m3DT 4t

    where DT represents a trend dummy. DT = t t, if t> t; DT= 0 if t t.

    The third null hypothesis involves a change in both

    the level and drift of a unit root process, and has the

    alternate of a permanent change in level and slope of a

    trend-stationary process.

    H3 : yt a0 yt1 m1DP m2DL 4t 4

    A3 : yt a0 a2t m2

    DL m3

    DT 4t

    Perron (1989) provides t-statistics for testing each of

    the above three hypotheses. These test statistics vary with

    the ratio of time until the structural break to the total time

    period under investigation. We conducted unit root tests

    for the three hypotheses for each of the five variables in

    our study (see Table 2 for results). RAD contained a unit

    root, with its first difference, i.e., D1RAD, being trend-

    stationary with one time change in the intercept.10

    RSALES also contained a unit root with its first differ-

    ence, i.e., D1RSALES, being trend-stationary with one

    time change in the intercept. RPRICE contained a unit

    root with its first difference, i.e., D1RPRICE, being trend-

    stationary with permanent change in both the slope and

    the intercept. Whereas, both RGDP and UEMP were trend-

    stationary. RGDP was trend-stationary with permanent

    change in both slope and intercept, whereas UEMP was

    trend-stationary with one time change in the intercept.

    10 Following standard convention in time series literature, we denote

    the difference variables as Dn[Variable Name]. Thus, the first difference(i.e., n = 1) of RSALES would be D1RSALES and the second difference

    would be D2RSALES.

    Table 1

    ADF and PP tests for unit roots

    An intercept term was included in all the tests.

    RAD RSALES RPRICE RGDP UEMP

    ADFa 8.52 6.09 7.86 0.83 12.03ADF (trend)b 15.48 14.59 13.55 6.85 11.89

    ADF (differenced)

    a

    707.75 170.81 26.11 248.95 269.55PPa 8.37 4.64 3.54 0.47 9.32PP (trend)b 11.36 7.61 5.94 7.34 9.25PP (differenced)a 36.15 32.53 35.99 36.15 26.91

    a Critical values for ADF and PP at 5% level of significance is 15.7for OLS autoregressive coefficient (Hamilton, 1994).

    b Critical values for ADF and PP at 5% level of significance is 22.4for OLS autoregressive coefficient (Hamilton, 1994).

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144134

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    9/18

    Fig. 2. Structural breaks in advertising and sales.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144 135

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    10/18

    To summarize, after incorporating structural changes,

    there were three I(1) variables, namely RAD, RSALES,

    and RPRICE and two trend-stationary variables, RGDP and

    UEMP. Note that the finding concerning the two macro-

    economic variables is consistent with that of Perron (1989).

    4.4. Reconciling unit root tests before and after incorporat-

    ing structural breaks

    Our unit root tests showed all variables to be I(1)

    processes, whereas after compensating for structural breaks

    we find the three firm level variables, viz., advertising,

    sales, and price, to be I(1) processes, but the two macro-

    economic variables, GDP and unemployment to be trend-

    stationary. To investigate all possible scenarios, we decided

    to test model with five I(1) variables and a model with three

    I(1) variables.

    4.5. Cointegration tests

    In general terms, if there exists a stationary linear

    combination of two or more variables, all of which are

    integrated of the same order, say d, i.e., they are all I(d),

    such that this linear combination is integrated of orderI(dc), where c ! 1, then these variables are said to becointegrated. Cointegrated variables share a common sto-

    chastic trend (Stock and Watson, 1988).11 Engle and Gran-

    ger (1987) proposed a straightforward methodology to test

    for cointegration. Let Xt be a vector of n variables all

    integrated of order d, i.e., I(d). Estimate a vector A of size

    n such that

    AHXt et 5

    If et is integrated of order d c, where c ! 1, then thevariables in the vector Xt are said to be cointegrated of

    order c.

    Baghestani (1991) and Zanias (1994) used the Engle and

    Granger (1987) procedure to test for cointegration between

    advertising and sales of LPMC, which are both I(1). Further,

    they found the two variables to be cointegrated. Enders

    (1995), among others, points out that the inherent weakness

    of the Engle and Granger methodology is that it relies on a

    two-step estimation procedure; as a result, the inferences for

    the second step depend on which error term from the first

    step is used in the second step. Thus, it is possible that

    depending on the choice of the error term one could either

    find the variables to be either cointegrated or not cointe-

    grated. Enders (1995) recommends using Johansen's (1988)methodology, which relies on the relationship between the

    rank of a matrix and its characteristic roots.

    Johansen's method is a multivariate generalization of the

    DickeyFuller unit root test. Eq. (6) depicts this general-

    ization.

    DXt A1 IXt1 4t 6

    where Xt and 4t are the (n 1) vectors of variables anderrors, respectively, DXt represents Xt in first difference, AIis a (n n) matrix of parameters, and I is an (n n)identity matrix. The tests entail estimating the rank of (A1

    I

    ), which equals the number of cointegrating vectors. In

    Table 2

    Unit root test with structural breaks

    Variable Year of break L* t-statistic hypothesis 1 t-statistic hypothesis 2 t-statistic hypothesis 3

    RAD 1925 0.35 1.79 2.64 3.17RAD 1934 0.52 1.80 2.59 3.61D1RAD 1925 0.35 3.61 3.61 3.90

    D1RAD 1934 0.52 3.98a

    3.62 3.93D2RAD 1925 0.35 5.67b 5.50b 5.69b

    D2RAD 1934 0.52 5.53b 5.50b 5.45b

    RSALES 1925 0.35 1.50 3.72 3.11RSALES 1938 0.59 2.26 3.17 3.09D1RSALES 1925 0.35 4.18c 3.79 4.32a

    D1RSALES 1938 0.59 3.94a 3.83 3.89D2RSALES 1925 0.35 5.03b 4.92c 4.98b

    D2RSALES 1938 0.59 4.76b 4.89b 4.70c

    RPRICE 1933 0.50 0.11 2.83 0.03D1RPRICE 1933 0.50 3.34 3.30 4.39a

    RGDP 1931 0.46 2.22 3.61 4.24a

    RGDP 1938 0.59 2.11 3.71 3.91D1RGDP 1931 0.46 4.48b 4.58b 4.46a

    D1RGDP 1938 0.59 4.95b 4.55c 4.90b

    UEMP 1930 0.44 4.23c

    2.89 4.59c

    * L is computed as the number of years till present (i.e., test date) divided by the total number of years.a Significant at 1% level.

    b Significant at 2.5% level.c Significant at 5% level.

    11 See Hamilton (1994) and Enders (1995) for a thorough exposition of

    issues relating to cointegration.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144136

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    11/18

    practice, maximum likelihood estimation is used to obtain

    these cointegrating vectors and the ltrace and lmax statistics

    are used to test for the number of characteristics roots

    different from unity, which gives us the number of

    cointegrating vectors (Johansen and Juselius, 1990). These

    tests rely on the ordering of the characteristic roots (li) and

    Eq. (7) provides the estimates of these statistics for rcharacteristic roots.

    ltracer Tn

    ir1

    ln1 li

    lmaxrY r 1 Tln1 lr1 7

    where li is the estimated value of the characteristic root

    and T is the number of observations. The ltrace test has the

    null hypothesis of the number of distinct cointegration

    vectors being less than or equal to r against a general

    alternative. The lmax statistic tests the null hypothesis that

    there are r distinct cointegrating vectors against thealternative that there are r + 1 cointegrating vectors.

    Johansen and Juselius (1990) have provided the critical

    values for these statistics.

    We estimated the ltrace and the lmax statistics for two

    possible scenarios. First, after incorporating structural

    breaks, we found the three micro time series (advertising,

    sales, and price) to be I(1) and, therefore, we used these

    three series as the non-stationary series. We present these

    results in the top half of Table 3.12 As is evident from these

    results, there is a possibility of one cointegrating vector, i.e.,

    there exists one linear combination of these three variables

    which is stationary.13 Second, tests on the data generation

    process for the five variable system14 indicate the possibi-

    lity of three cointegrating vectors (see the bottom half of

    Table 3).15

    To incorporate these findings of structural change and thecointegration between advertising, sales, and price, we

    tested various specifications for the five variable system in

    a VAR framework with error correction components

    (VECM). Now we discuss these models.

    4.6. VAR and ECM

    VAR analysis is a symmetric simultaneous equation

    system. In general, a VAR system can be written as:

    Xt z i m

    i 1

    iXti Jt 8

    where Xt is an n-vector of variables, Z is an n-vector of

    constants, i is an n n matrix of coefficients, Jt is an n-vector of error terms, and m is the appropriate lag length. If

    any of the variables are non-stationary, then it is possible to

    difference or de-trend these variables before estimation to

    make them stationary. If two or more variables are

    cointegrated, then we include an error correction term (the

    linear combination of the cointegrated variables which is

    stationary) in the structural VAR analysis as an independent

    variable, which gives us the appropriate VECM. A VECM

    can then be written as

    DXt z im

    i1

    iDXti Jt 9

    where t1 is a vector of error correction components (i.e.,

    the cointegrating vectors) and DXt is a vector of first

    differences of the variables under investigation.

    There has been a debate as to which of the above

    specifications is appropriate. Until recently, the Johansen

    approach was popular and was used to determine cointe-

    grating relationship between variables under investigation.

    Subsequently, one would rely on these tests to estimate a

    VECM. However, recent works by Toda (1995), Toda and

    12 The cointegration vector obtained by maximum likelihood estima-

    tion is (see Johansen, 1988 for details): RAD0.473 RSALES37.062 RPRICE.

    13 Though we find one cointegrating vector, it is not the same as that of

    Baghestani (1991) and Zanias (1994). First, with n variables, there is a

    possibility of finding n 1 cointegrating vectors. Thus, Baghestani (1991)and Zanias (1994) could have only gotten one cointegrating vector whereas

    we could potentially get two or four (depending on the model) cointegrating

    vectors. Further, the cointegration vector for Baghestani (1991) and Zanias

    (1994) had two terms (i.e., advertising and sales) whereas the cointegrating

    vector in our case has three (i.e., advertising, sales, and price) or five (i.e.,

    advertising sales, price, GDP, and unemployment) terms.14 There is general agreement that the unemployment series is

    stationary (cf., Perron, 1989). In the current data set, we found

    unemployment to be integrated of order one. We conjectured that this

    was due to the Great Depression years. Testing for structural breaks

    indicated that there was one structural break in 1930. After correcting for

    the break, as proposed by Perron (1989), unemployment was found to be

    trend-stationary. Nevertheless, for the time period under consideration

    unemployment is non-stationary.15 The cointegration vector obtained by maximum likelihood estima-

    tion is: RAD 0.386 RSALES 2.333 RPRICE + 1.478 RGDP 13.682

    UEMP.

    tI

    Table 3

    Cointegration tests: Johansen's methodology

    l Trace hypothesisal Trace

    statistic

    l Max

    hypothesisal Max

    statistic

    Three-variable system

    r 2, r > 2 0.98 r = 2, r = 3 0.98

    r 1, r > 1 9.89 r = 1, r = 2 8.91r = 0, r > 0 43.78b r = 0, r = 1 33.89b

    Five-variable system

    r 4, r > 4 0.72 r = 4, r = 5 0.72r 3, r > 3 7.45 r = 3, r = 4 6.73r 2, r > 2 15.98 r = 2, r = 3 8.53r 1, r > 1 32.59 r = 1, r = 2 16.61

    r 0, r > 0 75.88b r = 0, r = 1 43.29b

    a Null hypothesis is stated first, then after the ` comma'' alternate

    hypotheses is stated.b Significant at 1% level.c Significant at 5% level.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144 137

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    12/18

    Phillips (1993), Toda and Yamada (1996) and Phillips

    (1995) have shown that this approach may not be reliable,

    especially when less than 300 observations are available,

    which is true in our case. Toda (1995) demonstrates that,

    particularly with small data sets, inference concerning

    Granger causality can be more reliable if drawn from a

    VAR in levels form, because pre-test estimator biases are

    avoided. There may be some inefficiency due to the need

    to include enough lags to capture the non-stationary nature

    of the data, but this is likely to be small in comparison to

    the potential biases resulting from pre-testing (especially

    since unit root and cointegration tests have been shown to

    have low power for relatively small data sets).

    In accordance with the above mentioned literature, we

    adopt a pragmatic approach and consider inferences from

    both a VAR in levels and from a VECM specified accord-ing to the results of preliminary unit root and Johansen

    cointegration tests. The extent to which the results from

    these approaches coincide will provide some indication of

    the confidence with which we can draw conclusions from

    the data.

    Based on the above reasoning, we consider three

    different configurations. First, is a VAR in levels with

    the appropriate structural break for each of the five

    variable system. Second, is a VECM with the cointegrat-

    ing vector comprised of advertising, sales, and price, the

    three variable found to be I(1) after incorporating appro-

    priate structural breaks. Third, is a VECM with thecointegrating vector comprising the five variables. In the

    remainder of the article, we refer to these models as

    Models 1, 2, and 3, respectively.

    Before estimating the ECMs, we tested each model for

    the appropriate lag length (see Table 4 for results). As is

    evident from the table, a lag length of 1 or 2 was appropriate

    in most cases. In fact, we estimated the models with lag

    length of either 1 or 2 and obtained similar results. For

    parsimony, we report the results with lag length of 2.

    4.7. Granger causality tests

    The test for Granger causality in a VAR is to determine

    whether the lags of one variable enter into the equation of

    another variable. In the case of a VECM, where we have

    cointegrated variables, Granger causality requires the addi-

    tional condition that the speed of adjustment coefficient

    Table 4

    Lag length test

    Here, we report AIC followed by BIC.

    Model RAD RSALES RPRICE RGDP UEMP

    Model 1: VAR with structural breaks in levels

    Lag length 1 11.890 12.064 0.055 6.235 2.054

    12.714

    a

    12.888

    a

    0.694

    a

    6.909 2.504Lag length 2 11.593 11.906a 0.282a 5.932a 1.843a

    12.809 13.122 0.858 6.767a 2.451a

    Lag length 3 11.458a,b 12.041 0.184 5.949 1.84613.078 13.660 1.358 6.952 2.617

    Model 2: Three-variable cointegration

    Lag length 1 10.844 12.077 0.035 6.244 2.03211.743a 12.976a 0.789a 6.993 2.631a

    Lag length 2 10.581a 11.930a 0.312a 5.949a 1.87911.873 13.222 0.904 6.861a 2.638

    Lag length 3 10.590 12.058 0.199 5.971 1.855a,b

    12.286 13.754 1.420 7.051 2.781

    Model 3: Five-variable cointegration

    Lag length 1 10.126 11.994 0.033 6.247 2.08611.024a 12.893a 0.791a 6.996 2.686

    Lag length 2 9.962a 11.915a 0.271a 5.944a 1.87911.254 13.206 0.944 6.856a 2.639a

    Lag length 3 9.964 12.063 0.159 5.955 1.867a,b

    11.659 13.759 1.459 7.034 2.793

    a Optimal.b We checked for lag length of 4, but 3 was optimal.

    Table 5

    Granger causality

    Model RAD RSALES RPRICE RGDP UEMP

    A. Granger causality results: F-tests

    Model 1: VAR with structural breaks in levels

    RAD Granger caused 6.00

    a

    10.92

    a

    0.69 1.64 0.54RSALES Granger caused 2.12 7.46a 0.93 4.15b 2.34

    Model 2: Three-variable cointegration

    RAD Granger caused 20.01a 16.50a 12.43a 1.91 1.42

    RSALES Granger caused 0.41 0.76 0.05 2.72c 1.96

    Model 3: Five-variable cointegration

    RAD Granger Caused 18.41a 15.09a 11.21a 11.36a 13.11a

    RSALES Granger caused 1.19 1.57 0.81 2.72c 2.16

    B. Granger causality: speed of adjustment coefficients

    Model 2: Three-variable cointegration 1.287a 0.037 0.001 Model 3: Five-variable cointegration 1.365a 0.494 0.000 0.003 0.001

    a p < 0.01.b

    p < 0.05.c

    p < 0.10.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144138

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    13/18

    (beta coefficient of the cointegrating variable) to be different

    from zero. We used the likelihood ratio test to verify

    whether the lags of one variable enter into the equation of

    another. We present the results for the three models in Table

    5A and the results for the speed of adjustment coefficient for

    the two VECMs (last two models) in Table 5B.

    As is evident form Table 5A, for Model 1, the VAR inlevels, advertising is Granger caused by sales, whereas

    sales is Granger caused by GDP. The results for the two

    VECMs show that advertising is Granger caused by all the

    variables in the cointegrating equation, i.e., sales and price

    in Model 2 and sales, price, GDP, and unemployment in

    Model 3. In addition, for Models 2 and 3, the speed of

    adjustment coefficient is significant only for advertising. In

    order to understand the size of the impact of the macro-

    economic variables on sales and advertising we now turn

    to decomposition of forecast error variance and IRF.

    4.8. Variance decomposition and IRFs

    Decomposition of the variances of the forecast error is

    helpful in understanding the interrelationships amongst the

    variables in the system. The forecast error variance decom-

    position has information on the proportion of movement in a

    series due to innovations in the series itself and innovations

    in other series. IRFs demonstrate how one variable reacts to

    a shock in another variable. Plotting the IRFs is a practical

    way to visually represent the response in one series to a

    shock in another series.

    To compute variance decomposition and IRFs one must

    write the VAR process in its equivalent Vector Moving

    Average (VMA) form (Sims, 1980). That is, the VAR Eq.

    (8) can be written in its equivalent VMA form16

    Xt m I

    i0

    A4ti 10

    The mechanics behind variance decomposition is straight-

    forward. Taking the conditional expectation of Xt + 1 after

    updating it by one period in Eq. (10) gives

    EtXt1 a0 a1Xt 11

    where a0 and a1 are estimated coefficients. We can subtract

    the expected value from the actual value at period t + 1 toobtain a one-period-ahead forecast error. In a similar

    manner, we can compute forecast errors for n periods in

    the future. In the VMA form of the model, Eq. (10), the

    second term on the right hand side gives the n forecast error,

    i.e.,nIiH

    0i4t + n i. Putting restrictions on the VAR system

    decomposes the forecast error variance.

    Both variance decomposition and IRFs are sensitive to

    the ordering of variables in the VAR but the decomposition

    of forecast error variance converges over time to the

    unconditional variances. Table 6B displays the results from

    the Choleski decomposition of the 40th period forecast error

    variance for advertising and sales for the three models under

    consideration. For Model 1, VAR in levels, advertising, and

    salestaken togetheraccount for about 81% of forecast

    error variance in advertising, whereas unemployment ex-

    plains 11.9% of advertising forecast error variance. In sales,

    74.51% of forecast error variance is explained by sales

    itself, but the most important variable besides sales itself is

    GDP. In fact GDP explains over two times the forecast error

    variance explained by advertising, i.e., 12.47% vs. 6.19%.

    As far as the three-variable cointegration model is con-

    cerned (Model 2), advertising mainly explains itself

    (64.99%) with GDP and unemployment together explaining

    nearly 20% of forecast error variance in advertising. For

    sales, besides sales itself, price (24.4%) seems to be the

    most important variable. Here again, GDP and unemploy-

    ment, taken together, explain more forecast error variance in

    sales than advertising, i.e., 20.15% vs. 16.06%. For the five-

    variable cointegration model, the results are similar to the

    three variable cointegration model. Advertising (78.91%)

    explains the bulk of its own forecast error variance, but the

    second variable is GDP (13.99%). For sales, besides salesitself, we again find price (25.69%) to be the most important

    variable. Again, the superiority of GDP over advertising in

    explaining the forecast error variance in sales is demon-

    strated (15.29% vs. 7.56%). We now discuss the IRFs.

    The elements in the matrix A1i in Eq. (10) are called

    impact multipliers. The impact multipliers, taken together,

    form the IRF. We plotted the IRFs with upper and lower

    90% confidence bounds obtained by Monte Carlo integra-

    tion estimates of standard errors (see Doan, 1992 for

    details). The IRFs were consistent across models. In Fig.

    3, we present the IRFs of interestthe response of adver-

    tising and sales to 10% shock. As is evident from these

    16 Again, see Hamilton (1994) and Enders (1995) for details.

    Table 6

    Forecast error variance decompositionresults after 40 periods

    All figures in this table are percentages.

    Model RAD RSALES RPRICE RGDP UEMP

    Model 1: VAR with structural breaks in levels

    RAD variance

    decomposition

    35.24 46.71 4.34 1.80 11.90

    RSALES variance

    decomposition

    6.19 74.51 3.63 12.47 3.20

    Model 2: Three-variable cointegration

    RAD variance

    decomposition

    64.99 6.21 8.63 9.89 10.28

    RSALES variance

    decomposition

    16.06 39.39 24.40 13.39 6.76

    Model 3: Five-variable cointegration

    RAD variance

    decomposition

    78.91 0.51 4.22 13.99 2.36

    RSALES variance

    decomposition

    7.56 46.03 25.69 15.29 5.42

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144 139

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    14/18

    IRFs, a 10% shock to advertising results in sales instanta-

    neously rising by about 9% with the effect dying out in

    about two periods. A 10% shock to sales has a similar

    impact on advertising.

    4.9. Discussion

    As in previous research (Baghestani, 1991; Zanias,

    1994), we found advertising and sales to be integrated

    Fig. 3. Impulse response functions of interest. (a) Model 2: shock to advertising response sales lag. (b) Model 2: shock sales response advertising lag. (c) Model

    3: shock advertising response sales lag. (d) Model 3: shock sales response advertising lag.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144140

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    15/18

    of order one. In addition, we find price, GDP, and

    unemployment also to be integrated of order one. Further,

    we find two structural breaks in both advertising and sales.

    The structural breaks in advertising are in 1925 and 1934.

    The first structural break coincides with the first federal

    regulation against LPMC. The second break seems to be a

    result of the depression and it appears that the impact of

    depression took some years to set in. The two breaks in

    sales were in 1925 and 1938. The break in 1925 coincided

    with the first federal reprimand, whereas the second break

    is at the end of the aggressive advertising streak by Lydia

    Gove. We do not observe any impact of the second federal

    intervention in 1940, perhaps, because the product had

    already acquired a negative reputation. The one break in

    Fig. 4. Time plots of advertising and sales.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144 141

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    16/18

    price is in 1930 and, as expected, coincides with the

    great depression.

    After incorporating structural breaks, the unit root tests

    suggest that the order of integration of advertising, sales,

    and price is the same. Even though the order of integration

    of advertising and sales remain the same after incorporat-

    ing structural breaks, the breaks alter the data generation process for the two series. This is evident from the time

    plots of advertising and sales (Fig. 4) and is confirmed by

    the Perron's (1989) test.

    We identify one cointegrating vector, but this vector is

    comprised of advertising, sales, and price and not advertis-

    ing and sales as in previous research (Baghestani, 1991;

    Zanias, 1994). Further, unlike the previous two research

    endeavors, we find that advertising does not Granger cause

    sales. It seems, at least in the case of LPMC, that the LPMC

    executives determined the advertising levels by relying on

    previous year's sales. However, advertising did not signifi-

    cantly influence sales. Perhaps, this insight explains thesecond structural break in advertising. As the advertising

    levels were based on sales, the advertising expenditure was

    decreased when the depression had a significant influence

    on sales. The variance decomposition results confirm the

    weak effect of advertising on sales, as the environmental

    variables explain more forecast error variance than advertis-

    ing. The IRFs show that advertising does have a short-term

    effect on sales, but the effect of sales on advertising is much

    stronger. This leads more support to the thesis that advertis-

    ing levels were determined based on sales.

    5. Conclusion

    Modeling of marketing interactions is important for

    both marketing researchers and marketing practitioners.

    With the growth in availability of single source data (cf.,

    Curry, 1993) time series modeling is becoming more

    important for both academicians and practitioners. We

    borrow from the recent literature in time series on

    multi-equation modeling to collate a set of econometric

    tests and estimation techniques necessary for the use of

    cointegration analysis. Cointegration analysis will aid inthe analysis of dynamic marketing interaction models and

    help in uncovering the underlying dynamic process. The

    framework provides guidelines as to the steps necessary

    for the use of cointegration analysis.

    Appendix A. Multi-equation model

    where Aijk(L) is the polynomial in the lag operator L. We can write Eq. (A) as:

    PE1 A APP AE1 E1 AE2 E2 AEE 4

    where PE1 is the (p + e1) 1 vector of p performance variables and e1 endogenous effort variables; Pis the p 1 vector of performance variables; E1 is the e1 1 vector of endogenous effort variables; E2 is the e2 1 vector of exogenous effortvariables; Eis the e 1 vector of environmental variables; and 4is the (p + e1) 1 vector of error terms. Further, A is the (p +e1) 1 vector of constants; Ap is the (p + e1) p matrix of coefficients; AE

    1is the (p + e1) e1 matrix of coefficients; AE2 is

    the (p + e1) e2 matrix of coefficients; AE is the (p + e1) ematrix of coefficients.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144142

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    17/18

    We illustrate the cointegration analysis for the famous

    case of the LPMC. We show that recent research (Baghes-

    tani, 1991; Zanias, 1994) had overlooked certain important

    aspects of the analysis (e.g., structural break tests), which

    resulted to their concluding bidirectional Granger causality,

    whereas we found that advertising does not Granger cause

    sales. In addition, our analysis uncovers the incidence andnature of extraneous environmental interventions. Future

    research should use cointegration analysis to study the

    advertising sales relationship in a competitive setting. Ques-

    tions like: (1) which firm's (market leader or follower)

    advertising spending follows the other; (2) what determines

    the followers' advertising spendingfirms own sales or the

    market leaders advertising spending, etc., can be easily

    addressed by using cointegration analysis. In addition,

    cointegration analysis can also be used to study other

    dynamic situations like CEO compensation and the share

    price of the firm's etc. Further, marketing practitioners will

    find the framework handy, which is likely to in empiricalgeneralizations and advancement of marketing science.

    Acknowledgments

    A part of this paper was presented at the Marketing

    Science Institute conference at Berkeley in April 1997.

    The authors appreciate the helpful comments of Martin

    S. Levy and Ravi Dharwadkar on the earlier versions of

    this manuscript.

    References

    Baghestani H. Cointegration analysis of advertising sales relationship. J Ind

    Econ 1991;39:671 81.

    Bass FM. The future of research in marketing: marketing science. J Mark

    Res (February) 1993;30:16.

    Bass FM. Empirical generalization and marketing science: a personal view.

    Mark Sci 1995;14(3):G619.

    Bass FM, Wind J. Introduction to special issue: empirical generalizations in

    marketing. Mark Sci 1995;14(3):G15.

    Beard F. Integrated marketing communication: new role expectations and

    performance issues in the clientad agency relationship. J Bus Res

    (November) 1996;37:20715.

    Bretschneider SI, Carbone R, Longini RL. An adaptive multivariate ap-

    proach to time series forecasting. Decis Sci 1982;13:66880.Caines PE, Sethi SP, Brotherton TW. Impulse, response identification and

    causality detection for Lydia Pinkham data. Ann Econ Soc Meas

    1977;6:14763.

    Carroll VP, Rao AG, Lee HL, Shapiro A, Bayus BL. The navy enlistment

    experiment. Mark Sci (Fall) 1985;4:35274.

    Clarke DG, McCann JM. Measuring the cumulative effect of advertising: a

    reappraisal. Combined Proceedings. American Marketing Association,

    1973. pp. 1359.

    Curry DJ. The New Marketing Research Systems. NY: Wiley, 1993.

    Curry SD, Mathur SK, Whiteman CH. BVAR as a category management

    tool: an illustration and comparison with alternate techniques. J Fore-

    casting 1995;14(3):181 99.

    Dekimpe MG, Hanssens DM. The persistence of marketing effect on sales.

    Mark Sci 1995;14:(1):121.

    Dekimpe MG, Hanssens DM. Empirical generalizations about market evo-

    lution and stationary. Mark Sci 1995;14(3):G109 21.

    Dickey D, Fuller WA. Likelihood ratio statistics for autoregressive time

    series with a unit root. Econometrica 1981;49:105772.

    Dickey D, Pantula S. Determining the order of differencing in autoregres-

    sive processes. J Bus Econ Stat 1987;15:45561.

    Doan TA. RATS User's Manual, Version 4. Evanton, IL: Estima, 1992.

    Economist. The Cigarette Wars: Stop Smoking. May 11, 2123, 1996.Enders W. Applied Time Series. New York: Wiley, 1995.

    Engle RF, Granger CWJ. Cointegration and error correction: representation,

    estimation and testing. Econometrica 1987;55:25176.

    Erickson GM. Using ridge regression to estimate directly lagged effects in

    marketing. J Am Stat Assoc 1981;76:766 73.

    Erickson GM. Dynamic Models of Advertising Competition. Boston:

    Kluwer Academic Publishing, 1991.

    Erickson GM. Advertising strategies in a dynamic oligopoly. J Mark Res

    1995;32:2337.

    Eskin GJ, Baron PH. Effect of price and advertising in test market experi-

    ments. J Mark Res (November) 1977;14:49908.

    France M. Commentary: blowing smoke in the tobacco wars. Bus Week

    1996;3479:37.

    Franses PH. Primary demand for beer in Netherlands: an application of

    ARMAX model specification. J Mark Res 1991;28:2405.Franses PH. Modeling new product sales: an application of cointegration

    analysis. Int J Res Mark (December) 1994;11:49102.

    Gatignon H. Competition as a moderator of effect of advertising on sales. J

    Mark Res (December) 1984;21:38798.

    Gatignon H, Hanssens DM. Modeling marketing interactions with applica-

    tion to salesforce effectiveness. J Mark Res (August) 1987;26:24757.

    Granger CWJ. Some properties of time series data and their use in econo-

    metric model specification. J Econometrics 1981;16:12130.

    Granger CWJ, Newbold P. Spurious regression in econometrics. J Econo-

    metrics 1974;2:11120.

    Hamilton JD. Time Series Analysis. New Jersey: Princeton Univ.

    Press, 1994.

    Hansen BE. Testing for parameter instability in regressions with I(1)

    processes. J Bus Econ Stat 1992;10:321 35.

    Hanssens DM. Bivariate time series analysis of relationship between ad-vertising and sales. Appl Econ 1980a;12:32940.

    Hanssens DM. Market response, competitive behavior and time series ana-

    lysis. J Mark Res 1980b;17:47085.

    Hanssens D, Parsons LJ, Schultz RL. Market Response Models: Econo-

    metric and Time Series Analysis. Boston: Kluwer Academic Publish-

    ing, 1990.

    Hargreaves CP. Macroeconomic Modeling of the Long Run. Aldershot,

    UK: Edward Elgar Publishing, 1992.

    Harsharanjeet JS, Ephraim SF, Hrishikesh VD. Measuring dynamic

    marketing mix interactions using translog functions. J Bus

    1982;55:40115.

    Helmer RM, Johansson JK. An exposition of the box-Jenkins transfer

    function with an application to the advertising sales relationship. J

    Mark Res 1977;14:227 39.

    Hendry DF. PC-GIVE: An Interactive Modeling System. Oxford: Univ. ofOxford Press, 1989.

    Heyse JF, Wei WWS. Modeling the advertising sales relationship. J Fore-

    casting 1985;4:165 81.

    Houston FS, Weiss DL. Cumulative advertising effects: the role of serial

    correlation. Decis Sci 1975;6:47181.

    Hutton JG. Integrated marketing communication and the evolution of mar-

    keting thought. J Bus Res (November) 1996;37:15562.

    Johansen S. Statistical analysis of cointegration vectors. J Econ Dyn Con-

    trol 1988;12:231 54.

    Johansen S, Juselius K. Maximum likelihood estimation and inferences on

    cointegration with applications to the demand for money. Oxford Bull

    Econ Stat 1990;52:169 09.

    Krishnamurthi L, Raj SP. The effect of advertising on consumer price

    sensitivity. J Mark Res (May) 1985;22:11929.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144 143

  • 8/9/2019 Using cointegration analysis for modeling marketing interactions in dynamic environments: methodological issues a

    18/18

    Kuehn AA. How advertising performance depends on other marketing

    factors. J Advertising Res (March) 1962;2:210.

    Kumar V, Leone RP, Gaskins JN. Aggregate and disaggregate sector fore-

    casting consumer confidence measures. Int J Forecasting (September)

    1995;11:36177.

    Kyle PW. Lydia Pinkham revisited: a box-Jenkins approach. J Advertising

    Res 1978;18:319.

    Leone RP. Generalizing what is known about temporal aggregation andadvertising carryover. Mark Sci 1995;14(3):G141 50.

    Magat WA, McCann JM, Morey RC. When does lag structure really matter

    in optimizing advertising expenditures? Manage Sci 1986;32:18293.

    Mahajan V, Bretschneider S, Bradford J. The feedback approach to

    modeling structural shifts in market response. J Mark (Winter)

    1980;44:7180.

    Mahajan V, Muller E, Bass FM. New product diffusion models in marketing:

    a review and direction for research. J Mark (January) 1990;54:1 26.

    Mahajan V, Sharma S, Buzzell RD. Assessing the impact of competitive

    entry on market expansion and incumbent sales. J Mark (July)

    1993;57:3952.

    Murray MP. A drunk and her dog: an illustration of cointegration and error

    correction. Am Stat (February) 1994;48:379.

    NortonJA,BassFM.Adiffusiontheorymodelofadoptionandsubstitutionfor

    high technology products. Manage Sci (September) 1987;33:1069 86.Norton JA, Bass FM. Evolution of technological generations: the law of

    capture. Sloan Manage (Winter) 1992;33(2):66 77.

    Palamountain JC. The Politics of Distribution. NY: Greenwood Press, 1955.

    Palda KS. The Measurement of Cumulative Advertising Effects. NJ: Pre-

    ntice-Hall, 1964.

    Parsons LJ. An econometric analysis of advertising, retail availability and

    sales of a new brand. Manage Sci (February) 1974;20:938 47.

    Perron P. The great crash, the oil price shock and the unit root hypothesis.

    Econometrica 1989;57:301 20.

    Phillips PCB. Fully modified least squares and vector autoregression.

    Econometrics 1995;63:1023 78.

    Phillips PCB, Perron P. Testing for a unit root in time series regression.

    Biometrica 1988;75:33546.

    Prasad VK, Ring LW. Measuring sales effects of some marketing mix vari-

    ables and their interactions. J Mark Res (November) 1976;13:391 6.

    Rogers A. Will tobacco stocks catch fire. Fortune 1994;129(9):12.

    Schmalensee R. The Economics of Advertising. Amsterdam: Northern Hol-land Publishing, 1972.

    Sims C. Macroeconomics and reality. Econometrica 1980;48:1 49.

    Stock J, Watson M. Testing for common trends. J Am Stat Assoc

    1988;83:1097 107.

    Swinyard RW, Ray ML. Advertisingselling interactions: an attribution

    theory experiment. J Mark Res (November) 1977;14:50916.

    Toda HY. Finite sample performance of likelihood ratio tests for coin-

    tegrating ranks in vector autoregressions. Econometric Theory

    1995;11:1015 32.

    Toda HY, Phillips PCB. Vector autoregressions and causality. Econometrica

    1993;61:136793.

    Toda H, Yamada H, Inference in possibly integrated vector autoregressive

    models: finite sample evidence. ISER Discussion Paper, Osaka Univer-

    sity, Osaka, Japan, 1996.

    Weiss DL, Houston FS, Windal P. The periodic pain of Lydia E. Pinkham.J Bus 1978;51:91101.

    Wildt AR, Winer RS. Modeling and estimating in changing environments.

    J Bus 1983;56(3):36588.

    Winer RS. An analysis of time-varying effects of advertising: the case of

    Lydia Pinkham. J Bus 1979;52:56376.

    Zanias GP. The long run, causality and forecasting in the advertising sales

    relationship. J Forecasting 1994;13:60110.

    Zenor MJ. The profit benefits of category management. J Mark Res (May)

    1994;31:20213.

    R. Grewal et al. / Journal of Business Research 51 (2001) 127144144