Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

The effect of television advertising on consumer price sensitivity : an investigation of frequently purchased… Kanetkar, Vinay 1989

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-UBC_1989_A1 K36.pdf [ 12.66MB ]
Metadata
JSON: 831-1.0098302.json
JSON-LD: 831-1.0098302-ld.json
RDF/XML (Pretty): 831-1.0098302-rdf.xml
RDF/JSON: 831-1.0098302-rdf.json
Turtle: 831-1.0098302-turtle.txt
N-Triples: 831-1.0098302-rdf-ntriples.txt
Original Record: 831-1.0098302-source.json
Full Text
831-1.0098302-fulltext.txt
Citation
831-1.0098302.ris

Full Text

T H E E F F E C T O F T E L E V I S I O N A D V E R T I S I N G O N C O N S U M E R P R I C E S E N S I T I V I T Y : A N I N V E S T I G A T I O N O F F R E Q U E N T L Y P U R C H A S E D P R O D U C T S By Vinay Kanetkar B. Arch. Indian Institute of Technology, Kharagpur, India, 1974 M. Arch. University of British Columbia, Canada, 1979 M. Sc. (Bus. Admin.) University of British Columbia, Canada, 1983 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY i n THE FACULTY OF GRADUATE STUDIES (FACULTY OF COMMERCE AND BUSINESS ADMINISTRATION) We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA January 1989 © Vinay Kanetkar, 1989 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for refer-ence and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Faculty of Commerce and Business Administration The University of British Columbia 2053 Main Mall, Vancouver, Canada V6T 1Y8 Date: A B S T R A C T .The central question studied is whether household price sensitivity increases or de-creases as the number of exposures to television advertising increases. As a means of answering this question, Salop's (1979) model of brand differentiation is generalized to incorporate advertising as a managerial decision variable. Salop's generalized model is linked to a random utility choice model to estimate the effects of advertising on consumer price sensitivity. The random utility choice models are then estimated using household level scanner panel data for two product categories (dry dog food and aluminum foil). The generalization of Salop's model shows that the direct effect of a firm's increased advertising is a lowering of price sensitivity. In addition, the model also shows that increased advertising may result in a higher or lower price sensitivity depending upon the advertising reaction of competitors. The empirical results for both product categories studied indicate that increased advertising is associated with higher price sensitivity. However, since the aluminum foil product category has only one major national advertiser, the results here cannot be fully explained by Salop's generalized model and thus show an important limitation of the model's applicability. Brand loyalty has proved to be the most important explanatory variable in random utility brand choice models for low cost brand identified consumer products (Guadagni and Little 1983). This dissertation develops a new measure of brand loyalty that is sensitive to the temporal pattern of household's previous brand purchases. For the product categories, studied, this new measure is superior to those proposed earlier. In addition, this work confirms the robust nature of Guadagni and Little's (1983) model of brand choice and shows that both television advertising and sales promotional variables have a significant impact on brand choice. ii Table of Contents Abstract ii Table of Contents iii List of Tables vi List of Figures viii Acknowledgement ix CHAPTER I. Effect of Marketing Mix Factors on Brand Choice . 1 1.1. Introduction 2 1.2. Effect of Brand loyalty, Advertising and Price on Brand Sales 5 1.3. Effect of Sales Promotion Variables on Brand Sales 8 1.4. A Review of Scanner Panel Studies 10 1.5. The Dissertation Objectives 14 1.6. The Dissertation Organization 17 CHAPTER II. The Effect of Advertising on Price Sensitivity 20 2.1. Conceptual and Inter-industry Comparison Literature 22 2.2. Recent Theoretical Literature 26 2.3. A Review of Recent Empirical Literature 36 2.4. Summary and Conclusions . 41 CHAPTER III. A Model of Advertising with Loyal and Non-loyal Consumers . .43 3.1 Introduction . 44 3.2. Product Market Operation 45 3.3. Model Assumptions and Overview 47 3.3.1 Microecohomic Assumptions 48 3.3.2 Consumer and Manufacturer Assumptions 50 3.3.3 Modelling Advertising Decisions 53 3.3.4 The Model Overview . . 55 3.4. The Model 56 111 3.4.1 List of Model Assumptions 56 3.4.2 The Optimal Price Derivation . . 58 3.4.3 Effect of Advertising on Price Sensitivity 63 3.5. Summary and Conclusions 68 CHAPTER IV. Analytical Methods, Hypotheses, and Data Sources 71 4.1. Introduction 72 4.2. Statistical Modelling Framework 74 4.2.1. Stochastic Process Choice Model . . .75 4.2.2. Random Utility Choice Model 76 4.2.3. Derivation of the Statistical Model 79 4.3. Variable Description and Model Hypotheses 84 4.3.1. Dependent Variables 85 4.3.2. Independent Variables 86 4.3.3. Model Hypotheses 92 .4.4. Demographic Comparisons . 95 4.5. Comparisons of Household and Store Level Sales Estimates . 97 4.6. Data Management Issues . 99 CHAPTER V. Model Estimates and Interpretations 102 5.1. Introduction : 103 5.2. Approaches to Compare Models 104 5.3. Brand Choice Models for Dry Dog Food 106 5.3.1. Alternative Model Specifications: Summary 107 5.3.2. Alternative Model Specifications: Estimation Results . . 108 5.3.3. Model Cross-validations 123 5.3.4. Price Sensitivity Interpretation 125 5.3.5. Model Interpretations 130 5.4. Brand Choice Models for Aluminum Foil 134 5.4.1. Alternative Model Specifications 134 5.4.2. Model Interpretations 138 5.5. Model Comparison across Product Categories 139 iv CHAPTER VI. Contributions, Implications, Limitations, and Future Work . . . 142 6.1. Contributions 143 6.1.1. Theoretical Contribution 144 6.1.2. Empirical Contribution 145 6.2. Managerial Implications • • • 149 6.2.1. Implications from Theoretical Model 149 6.2.2. Implications from Empirical Work 150 6.3. Limitations 153 6.4. Future Work . 156 6.5. In Conclusion 158 References Cited 192 Appendix A: Model Comparison by three brand loyalty groups 202 List of Tables 2.1. A Summary of Studies from Economics 159 2.2. A Summary of Studies from Marketing 160 2.3. Price Elasticity at Different Advertising Levels 161 3.1. Consumer Segments and their Responses to Selected Marketing Mix Variables 162 4.1. Purchase Share of Major Brands in Dry Dog Food Category . . . . . . . . 163 4.2. Incidence of Sales Promotion Activities by Product 164 4.3. Expected Signs for Independent Variables in Brand Choice Model 165 4.4. Demographic Comparisons - Family Size 166 4.5. Demographic Comparisons - Family Income 166 4.6. Demographic Comparisons - House Type 167 4.7. Demographic Comparisons - Home Ownership 167 4.8. Demographic Comparisons - Education 167 4.9. Demographic Comparisons - Occupation 168 4.10. Demographic Comparisons - Race 168 4.11. Household Wrap Brand Sales Comparisons 169 4.12. Dog Food Brand Sales Comparisons . 170 5.1. A Basic Sequence of Alternative Model Specifications for Dry Dog Food Brand Choice Models . 171 5.2. Effect of Alternative Measures and Alternative Forms of Advertising on Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates 172 5.3. Effect of Alternative Specifications with Disaggregated Form of Advertising on Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates 173 5.4. Effect of Alternative Measures of Brand Loyalty on Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates 174 5.5. Effect of Smoothing Constant for Guadagni and Little's Measure of Brand loyalty on Dry Dog Food Brand Choice Maximum Likelihood Parameter Estimates 175 v i 5.6. Effect of Alternative Specification for Guadagni and Little's Measure of Brand Loyalty on Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates . 176 5.7. Effect of Alternative Specification for Guadagni and Little's Measure of Brand Loyalty and Disaggregated Form of Advertising for Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates . 177 5.8. Effect of Household Income on Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates 178 5.9. Effect of Alternative Samples on Dry Dog Food Disaggregated Form of Advertising Brand Choice Models Maximum Likelihood Parameter Estimates 179 5.10. Effect of Alternative Samples on Dry Dog Food Aggregated Form of Advertising Brand Choice Models Maximum Likelihood Parameter Estimates . 180 5:11. Estimates of Time Trend for the Actual Brand Prices 181 5.12. Intercorrelations between the Residual of Actual Brand Prices and Mean Brand Prices . . . '. 182 5.13. Intercorrelations between Brand Advertising Exposures and Mean Number of Exposures per Purchase 183 5.14. Estimates of Marginal Change in the Regular Price as Result of the Television Advertising 184 5.15. Weighted Aggregate Elasticities using the Total Sample for 11 Brands of Dry Dog Food . . . . . . . . . . . . . 185 5.16. Aggregate Elasticities Evaluated at the Sample Means for 11 Brands of Dry Dog Food . . . . . 186 5.17. A Basic Sequence of Specifications for Aluminum Foil Brand Choice Models Maximum Likelihood Parameter Estimates 187 5.18. Weighted Aggregate Elasticities for Three Brands of Aluminum Foil . . . . 188 A.l. Effect of Alternative Specification for High Loyalty Group for Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates 209 A.2. Effect of Alternative Specification for Medium Loyalty Group for Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates 210 A.3. Effect of Alternative Specification for Low Loyalty Group for Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates 211 vn A.4. Comparison of Specifications for Three Loyalty Groups and Disaggregated Form of Advertising for Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates . . . . . . . . ^ 212 A.5. Comparison of Specifications for Three Loyalty Groups and Aggregated Form of Advertising for Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates 213 A.6. Elasticity Estimates for Selected Brand Choice Variables by the Brand Loyalty Groups . 213 L i s t of Figures. 2.1. Effect of Advertising Brand Price Sensitivity 189 3.1. Representation of Spatial Competition in One Dimension 190 4.1. Data Management Flow-chart for Dry Dog Food . . 191 A.l. Cumulative Distribution of Maximum Loyalty 215 V l l l A C K N O W L E D G E M E N T S . I wish to express my gratitude to Prof. Charles Weinberg and Prof. Doyle Weiss for their support throughout the doctoral program. As my dissertation advisors, they deserve the highest credits for moving me ahead and providing helpful criticisms at various stages. I also wish to thank Prof. Tae Oum for his continuous support and for encouraging me to think like an economist. I would like to acknowledge with gratitude the support received for this research from the effective frequency committee of the Advertising Research Foundation. In particular, I wish to thank Hilda Stewart, Larry Stoddard and Mike Naples of Advertising Research Foundation who helped in giving me access to the database used in this study. In addition, I wish to thank Ellen Walczek of Information Resources Inc. for providing details about data coding. I sincerely appreciate the efforts of Swee Hoon Ang and Gordon Brown in improving the English of this document, for it is in English where I need the most help. There are no words to express my deepest gratitude to my mother and my brothers who provided emotional support for my education in Canada. I owe more to them than to anyone else. Last but not the least, as a close friend and now as a life partner, my wife Raminder deserves the most credit for giving me enough strength and emotional support in writing and finishing this work. ix CHAPTER I Effect of Marketing Mix Factors on Brand Choice. 1.1. Introduction 2 1.2. Effect of Brand loyalty, Advertising and Price on Brand Sales 5 1.3. Effect of Sales Promotion Variables on Brand Sales 8 1.4. A Review of Scanner Panel Studies 10 1.5. The Dissertation Objectives 14 1.6. The Dissertation Organization . . 17 1 1.1. INTRODUCTION This dissertation is concerned with the effect of television advertising on consumer price sensitivity. As a means of understanding this effect, Salop's (1979) model of brand differentiation is generalized to incorporate advertising as a managerial decision variable. Salop's generalized model is linked to a random utility choice model to estimate the effects of advertising on consumer price sensitivity. A number of nested random utility choice models are then estimated using household level scanner panel data for two product categories (dry dog food and aluminum foil). The generalization of Salop's model predicts that the direct effect of a firm's increased advertising is a lowering of price sensitivity. In addition, his model also shows that increased advertising may result in a higher or lower price sensitivity depending upon the advertising reaction of competitors. The empirical results for one product category (dry dog food) confirms the prediction for Salop's generalized model that increased advertising results in a higher price sensitivity. The empirical results for the product category of aluminum foil, however, cannot be fully explained by Salop's generalized model and show an important limitation of the model's applicability. The empirical part of the work also demonstrates that the effects of sales promotional variables and advertising variables on brand choices vary across product categories. In addition, this work confirms the robust nature of Guadagni and Little's (1983) model and provides a significant improvement on their results by means of a different measure for the' brand loyalty variable. The background literature for this research is briefly reviewed below. For several decades marketing researchers have been attempting to develop a better 2 understanding of the determinants of sales and/or market share of branded and frequently purchased consumer goods. Such knowledge is sought not only for its own intrinsic value, but also to aid managers in making better marketing mix decisions. Much of the early theoretical and empirical work associated with this effort focused on the relationship between advertising and sales. Indeed, so many advertising - sales studies have been published that researchers are able to do "meta-analyses" on such specialized issues as the saturation effect of advertising (Aaker and Carman 1982), the effect of changing market environments on sales-response (Wildt and Winer 1983), and the effect of the data collection interval on the sales response coefficients of certain models (Clarke 1976, Weinberg and Weiss 1982, and Assmus, Farley and Lehmann 1984). Further, a review of this literature shows that there has been almost continuous controversy on (1) the appropriate form of the sales-advertising response function (Houston and Weiss 1974, Brodie and Kluyver 1984, Ghosh, Neslin and Shoemaker 1984), (2) the simultaneity of the advertising allocation decisions and the related sales response (Comanor and Wilson 1979, McAuliffe 1987), and (3) the proper period for data collection and analysis (Clarke 1976). Indeed, the lack of definitive progress on these issues with traditional econometric models is perhaps the principal reason why researchers have developed the so-called decision calculus models (Little, 1979). These models subjectively establish parameter values by using managerial judgments rather than by only estimating them from market data. Nevertheless, both in the decision calculus and econometric traditions, the response of sales to advertising has remained a critical concern. While many researchers have focused on the advertising - sales relationship, few have been fully satisfied with the degree of understanding achieved. One frequently offered explanation for this lack of progress has been the omission from the models of 3 appropriate competitive factors, particularly competitive price and promotional variables. These factors were often omitted because the associated data were expensive, difficult to collect, or otherwise not available to researchers. Consequently, the recent availability of scanner panel data, which in large part provides such information - both at the store and individual level - has been greatly welcomed. However, with this greater availability of competitive price and in-store sales promotional data, researchers have mainly focused on these variables. The early empirical work, for example by Guadagni and Little (1983), concentrated on the impact of such variables as promotional activity and brand loyalty on brand choice at the household level. More recents articles, e.g. Winer and Moore 1987, and Tellis 1988, have incorporated television advertising as an explanatory variable in their models. Our particular concern in this research is to distinguish between the main (or direct) effects of television advertising and the effects of advertising operating through the price variable. Furthermore, (national) television advertising programs and local promotional efforts (e.g., features, displays) may effect the price response of consumers differently. As a result, the major objective of this dissertation is to critically study and isolate the effects of television advertising on household price sensitivity. To accomplish this objective, a theoretical economic model is developed using Salop's (1979) model of brand differentiation to obtain market level demand. In addition, a closely related statistical model is set up to test the effects of television advertising on household price sensitivity. Finally, several variations of the statistical model are estimated using household level scanner panel data which includes television advertising exposure data. This is done for two product categories (dry dog food and aluminum foil). The purpose of this chapter is to set forth the objectives of this dissertation. To do this in a realistic manner, a review of the empirical research concerned with the effect of marketing mix elements on brand sales is critically examined. More specifically, in section 1.2., the empirical literature concerning the effect of brand loyalty, advertising and price on brand ;sales and/or market share is reviewed and commented on. The empirical literature concerning the effect of price deals and coupons on brand choices is reviewed in section 1.3 and in section 1.4, scanner panel studies concerned with household brand choices are reviewed. Finally, this chapter concludes with a discussion and overview of the dissertation's objectives and organization. 1.2. EFFECT of BRAND LOYALTY, ADVERTISING and PRICE on BRAND SALES. In this section, various attempts to generalize the sales-advertising or sales-price re-lationship are summarized. Much of the discussion is based on review articles by Clarke (1976), Assmus, Farley and Lehmann (1984), Simon and Arndt (1980), Aaker and Car-man (1982), and Tellis (1987). From these articles, it is generally concluded that (1) the short term elasticity of advertising is positive and statistically different from zero; (2) the impact of advertising exhibits diminishing returns to scale; (3) the long term impact of advertising, measured in terms of its mean carryover effect, is both positive and signifi-cantly different from zero; and (4) the effect of price on brand sales is negative and also significantly different from zero. In order to tie all these reviews together, a general model of the relationship is presented here. If 5# is sales of brand i in period t, An is advertising expenditure of brand i in period t, Pa is price of brand i in period t, and e,-t is a white noise disturbance in period t, then the model is Sit = a,• + A,-Stf_ i.-t-/?,- Aa+ rjiPit + eit (1-1) 5 This is one of several common functional forms used in the marketing literature to describe the sales relationship. In equation (1.1), parameter A, measures long term carryover effectf for brand i and the parameters /?,• and 77, indicate a short term effect of advertising and price respectively on sales. The purpose of this section is to summarize the various attempts of scholars to generalize from values of A, /?, and r? found in the empirical literature. Clarke's (1976) work sought to answer the question: how long does the cumulative effect of a "unit" of advertising persist? He analysed 69 econometric models published in 27 studies prior to 1974 and concluded that "••• the cumulative effect of advertising on sales lasts for only months rather than years". He reported a mean carryover effect (coefficient of the lagged dependent variable, A) for his 69 studies of 0.526. A similar attempt was made more recently by Assmus, Farley and Lehmann (1984). They analyzed 128 models reported in 28 different studies and concluded that the mean coefficient of the lagged dependent variable, the same measure of carryover effect used by Clarke, was 0.468. They also reported that the mean equilibrium short term elasticity of advertising was only 0.221. As a result of this earlier work, we conclude that the short term (current period) impact of advertising appears to be smaller and less important than the long term carryover effect of advertising. Simon and Arndt (1980) reviewed more than 100 published papers concerned with the form of the advertising response function. They concluded that the response function link-ing sales and advertising, measured either in terms of dollar volume or physical amounts, indicated diminishing returns to advertising efforts. Aaker and Carman (1982), with a somewhat different approach, reviewed 69 experimental studies involving an increase or a f This carry over coefficient is sometimes interpreted as a simple unexplained inertia effect ("brand loyalty") and is at other times interpreted as the long term effects of advertising efforts. 6 decrease in advertising levels. They concluded that a majority of the experiments involv-ing an increase k advertising did not show statistical evidence of an associated increase in sales. The authors explained that this result is due to a tendency by firms to "over-advertise." For the most part, the studies reviewed by Aaker and Carman measured only the short term effects of advertising. Their results are, therefore, likely consistent with the conclusions in the previous paragraph; that is current or short term effects of advertising are less important than the long term or carry over effects. Recently, Tellis (1987) reviewed 367 models in 40 studies prior to 1985 to estimate "average" brand demand price elasticity. One would expect that an increase in a brand's price will lead to a decrease in the brand's sales. Tellis found the mean price elasticity for the 367 models to be —1.76 with a standard deviation of 1.74. Thus, an average one percent increase in a brand's price is expected to result in 1.76 percent decrease in the brand's salesf. Again, there is a large body of empirical evidence concerning the effect of price on brand sales but less attention is given to the effect of advertising on a brand's price elasticity. In conclusion, there have been many studies of advertising's direct effect on sales, with the overall consensus that the response curve linking advertising and sales has the characteristic of diminishing returns to scale with a positive carryover effect. There is also a consensus that the effect of price on sales is negative. This research focuses on the effect of advertising on price sensitivity, for which there are few studies (see chapter II) and of these, there are conflicting results and only limited consensus. f The distribution of elasticities was such that less than 20% of them were greater than zero and the majority of these unusual estimates came from the study by Larnbin (1976). 7 1.8. EFFECT of SALES PROMOTIONAL VARIABLES on BRAND SALES The literature concerning sales promotional efforts and price deals follows a different research tradition from the ones reviewed above. Two of the important sales promotional activities in the low cost, frequently purchased, brand identified consumer product industry are dealing and couponing. Dealing involves a short-term (usually a week or less) price cut to the consumer (Blattberg, Eppen and Lieberman 1981). Couponing is also a price cut to the consumer but differs in that a consumer is required to present a coupon to receive a price cut (Narasimhan 1984). It is argued in this literature that dealing occurs either as a mechanism to enable price discrimination and/or as a method of accelerating consumer purchases and thereby shifting inventory holding costs to consumers. An overview of these alternative hypotheses is presented below and selected empirical research on couponing and dealing is reviewed. Economists like Salop and Stiglitz (1977, 1982), Varian (1980), and Shilony (1977) argue that retailers offer deals to discriminate between informed and uninformed con-sumers. Narasimhan (1984), and Bearden, Teel and Williams (1982) argue that couponing occurs to discriminate between price elastic (e.g. brand switchers) and price inelastic (e.g. brand loyal) consumers. On the other hand, Blattberg, Eppen and Lieberman (1981) and Eppen and Lieberman (1984) hold that dealing occurs as a means of sharing inventory holding costs between retailers and consumers. Further, Jeuland and Narasimhan (1985) and Neslinj Henderson and Quelch (1985) propose that consumers with higher inventory costs will buy goods at the regular prices while those with lower inventory costs will stock pile goods when prices are lower. Finally, Thaler (1986), starting from the perspective of behavioural decision theory, argues that responses to price decreases (e.g. promotional deals) differ from responses to price increases. Gurumurthy and Little (1986) and Winer 8 (1985, 1986) have provided empirical evidence to support this hypothesis. Selected parts of this empirical research on couponing and dealing are reviewed in more detail below. Bearden, Teel and Williams (1981) reviewed fifteen studies concerning the effects of cents-off coupons on sales. One of their conclusions was that consumer "deal proneness" was positively correlated to media exposure, and inversely related to brand loyalty. They reported that dealing can induce substantial brand switching, and coupled with "advertis-ing" (not classified by the authors as to whether the advertising was television advertising or feature advertising) can have a synergistic effect on total sales. Further, they reported that sales promotions are most effective for new product introductions or unfamiliar brands. Narasimhan (1984) proposes a price theoretic model to explain coupon usage. His model hypothesises that users of coupons are more price sensitive than nonusers of coupons and that the opportunity cost of time and other household resource variables are deter-minants of consumers' decisions about coupon use. Empirical evidence, however, pro-vided only limited support for this model; most of the estimated coefficients reported by Narasimhan were not significantly different from zero. Blattberg, Eppen and Lieberman (1981) posit that dealing occurs in retailing because retailers have higher inventory holding costs than consumers. Thus, the retailer is moti-vated to take a reduction in sales revenue if the consumer will incur some of the inventory costs. This implies that the consumer should exhibit stockpiling behaviour, i.e. buying higher quantities when the right brand is on deal. The study reported empirical evidence to support this hypothesis for brands of facial tissue, aluminum foil, liquid detergent, and waxed paper. In a related article, Neslin, Henderson and Quelch (1985) argue that price promotions induce 'forward buying' behaviour. Based on the argument that brand loy-alty induces different buying behaviour than otherwise, separate models were estimated for loyal and non-loyal consumers. The findings reported in this scanner panel study (for bathroom tissue and instant coffee) are somewhat disappointing. A large part of the varia-tion in household purchase quantity and interpurchase time appears not to be determined by brand promotions. While it is clear that sales promotional activities such as dealing and couponing affect sales (Guadagni and Little 1983, Tellis 1988), the mechanism by which these variables work is largely unknown. If dealing activities constitute informing consumers, and if all consumers are not equally receptive to dealing information, then dealing is a method of discriminating between the informed and the uninformed consumers.. However, it is not clear from the existing literature whether the effect of price deals is stronger than the effect of television advertising on consumer brand choices. These issues will be addressed in this research. 1.4. A REVIEW of SCANNER PANEL STUDIES The recent availability of scanner panel data provides purchase histories for large samples of households. The collection procedure directly records the purchasing behaviour of individual households at the level of an individual purchase. Such information can also be collected for extended periods and across stores to provide very comprehensive behavioural patterns. In this section, three out of four published studies (Guadagni and Little 1983, Neslin, Henderson and Quelch 1985, Tellis 1988 and Russell and Bolton 1988) that have used scanner panel data are reviewed. In addition to these published studies, there are several unpublished doctoral dissertations (Bayer 1985, Ortmeyer 1985, Landwehr 1986, and Gupta 1987) and unpublished manuscripts (Gurumurthy and Little 1986, and Hanssens 1987) that have examined brand choices and/or brand sales using scanner panel 10 data. Most of these studies used data for the product category coffee and extended the work initiated by Guadagni and Lit t le (1983). In general, the scanner panel studies have found that brand loyalty and size loyalty are the most important variables affecting household purchase decisions. In addition, as one would expect, all of the studies found that brand prices affect purchase decisions in a negative fashion. Studies for three different product categories are reviewed below. Guadagni and Lit t le (1983) report a model of household purchase choices for regular brands of coffee. The study was concerned wi th five brands in two sizes. The scanner panel which supplied data consisted of 2000 households, from which two random samples of TOO households were used for the analysis. One sample was used for calibrating the model and the other for validation purposes. The data spanned 78 weeks (Sept. 14, 1977 to M a r c h 12, 1980), and the data in each of the samples was broken into three time periods. The data from the first time period of 25 weeks was used to initialize the model parameters associated with brand and size loyalty in both samples. The second period, 32 weeks from the calibration sample, was used to estimate the model parameters. The last period of 21 weeks of data was used for testing model accuracy. In the validation sample, 25 weeks of data was used to initialize the model, and 53 weeks of data was used to assess the model's validity. The dependent variable (y^) for the study was a discrete variable and was set equal to one if the household in the sample purchased brand k; otherwise it was zero. The study used brand loyalty, size loyalty, presence of promotion, presence of promotional price cuts, product price at the time of purchase and brand and size effects as the explanatory vari-ables. The authors employed the mult inomial logit functional form to predict the puchase probabilities using the set of explanatory variables mentioned above. The functional form 11 of the multinomial ldgit model was exp(Et=i X*Pi) E^iEjiexp^-,-/?,-) where is the value of explanatory variable i for brand k, /?, is an unknown coefficient prob(yfc = 1) = ^ m „ x (1-2) for the explanatory variable i and is to be estimated. The number K is a set of brands under investigation, while m is the set of explanatory variables. Guadagni and Little reported that the multinomial logit functional form achieved a good fit when used with the regular coffee data. Further, brand and size loyalty (in part proxy variables, in our view, measuring the carryover effect of advertising) were extremely important factors in predicting the average household's choice probability. Consistent with our generalization in section (1.2), these factors showed a positive impact on the average brand choice probability. Brand price, presence of promotion, and presence of promotional price cuts were also important factors in customer choices. In addition, and as might be expected, the implied brand demand curve sloped downward. The study also reported that for the validation sample, the model prediction and the actual choices corresponded well. Parameter estimates for the validation sample, however, were not reported. Hence, it is not possible to draw conclusions with respect to the stability of parameter estimates reported for the model. Neslin, Henderson and Quelch(1985) also reported results from a study using scanner panel data. The study was concerned with consumer response to price promotions. More specifically, it was argued by the authors that price promotions induce forward buying' behaviour. The study used a sample of 2293 consumers over a 28-week period and investigated brand choices for bathroom tissue and instant coffee. The model used for the study was similar to the one summarized above except for its explanatory variables. The Neslin et al. study did not include carryover variables (brand and size loyalty), but 12 did include manufacturer and retailer newspaper advertising. The overall conclusion of the Neslin et al. study was "that a good deal of the variation in household purchase quantity and interpurchase time is not determined by promotions." Further, the reported r-square values were generally very small (less than 0.10), even though the study used a very large number of data points (about 60,000). In addition, only three out of eight (for bathroom tissue), and six out of thirteen (for instant coffee) estimated coefficients were significantly different from zero. As a result, it is difficult to interpret the impact of specific independent variables. Tellis (1988) published a study using scanner panel data for the toilet tissue product category. This study, like the study by McDonald (1971), is concerned with the effect of television advertising exposures on individual household brand choices in a competitive market setting. While McDonald's study focused on brand switching and did not control for other market factors such as price dealing, the study by Tellis was concerned with measuring the effect of advertising repetition on brand choices when the effect of advertising may be conditional on the strength of the brand loyalty variable. The data used by Tellis consisted of scanner panel records for household purchases of 12 key brands of toilet tissue over 52 weeks. The data also included weekly records of exposures to television advertisements. The exposures were determined by monitoring the household's television viewing behaviour and the airing of television commercials by use of computer controlled recording devices. The final estimation sample employed by Tellis consisted of 2634 purchases by 251 households. The Tellis study found, as did Guadagni and Little (1983), that brand loyalty is the strongest explanatory variable of brand choice. The next four important explanatory vari-ables were feature advertising, in-store display, price, and coupon in order of importance. 13 The television advertising variable had a positive (statistically significant) impact on brand choice, but the effect was small. Finally, the effect of advertising conditional on the brand loyalty variable was not significant at p < 0.05. As a result, the study does not offer sub-stantial insight into how advertising "works" on brand choices and it does not support the notion that the effects of advertising are conditional on brand loyalty. This dissertation will offer evidence that one of advertising's indirect effects is on short term price sensitivity. To conclude and summarize; the following three observations can be supported by the scanner panel studies reviewed above. First, if the brand loyalty variable is used in the brand choice models, then this is the strongest predictor of brand choices. This result is also consistent with the empirical literature reviewed in section 1.2. Second, feature advertising, in-store display, price promotional variables, and price also appear important in predicting brand choices but their order of importance varies from study to study. Third, the models have not utilized television advertising variables (except Tellis 1988) for predicting brand choices. 1.5. The DISSERTATION OBJECTIVES The major goal of the present research is to develop a better understanding of the effect of advertising on brand price sensitivity in low cost, frequently purchased, brand identified consumer products. This particular class of consumer products is studied for two reasons. First, large banks of data already exist for many of these goods. For example, Information Resources Inc. (IRI) has collected scanner panel data for over 100 products for several years. Similar databases are also held by BehaviorScan, Burke Marketing Research, Nielsen, and other commercial sources. In addition, the purchasing activity by a consumer of these goods requires only minimal human information processing, partially because the 14 cost to the consumer for making an incorrect decision is small. As a result, a simple economic model can be used to describe consumer behaviour for these products. As mentioned, the major goal of this dissertation is to improve our understanding of the effect of advertising on consumer price sensitivity. With respect to this goal, three questions concerning price sensitivity are to be explored by this disseration. These are as follow: (1) How does an increase in television advertising efforts of a brand affect consumer price sensitivity? (2) Since the brand loyalty variable is the best predictor of consumer brand choices, to what extent does an alternative definitions of the brand loyalty variable affect pre-dictive ability of the multinomial logit models and, most particularly, affect consumer price sensitivity? (3) Is the effect of sales promotional variables stronger than the effects of television advertising variables on household brand choices? All these questions are discussed in more detail below. Consumer Price Sensitivity: If a household is exposed to a higher number of television advertisements does the household become more or less price sensitive? To address this question theoretically, a model that allows for both the brand enhancing effects of adver-tising and competitive reactions is developed. More specifically, Salop's (1979) model of brand differentiation is generalized to incorporate advertising as a managerial decision vari-able. Under differing market conditions, this model predicts either increased or decreased price sensitivity. The dissertation investigates the applicability of this model to two mar-kets (dry dog food and aluminum foil) for which increased price sensitivity is reported. 15 Thus, the first question that will be answered is whether increased television advertising efforts result in reduced or increased brand price sensitivity. This question is both of a theoretical and empirical nature. Role of Brand Loyalty on Price Sensitivity: Since the brand loyalty variable is the best predictor of consumer brand choices, the operational definition of this variable may affect the way advertising and price interact. For example, Guadagni and Little (1983), as well as Tellis (1988) measured the brand loyalty concept by exponentially smoothing past brand choices. It is shown later that an alternative measure of the brand loyalty variable predicts consumer brand choices somewhat better than the measure used by either Guadagni and Little, or Tellis. In addition, this dissertation also reports on the influence of this definition of brand loyalty on estimates of the effect of advertising on consumer price sensitivity. Importance of Sales Promotion and Television Advertising: Past research has focused on the main effects of advertising, promotional efforts, and prices on brand sales (see sections 1.2 and 1.3 for review). In addition, past research has often not distinguished between the effects of sales promotional efforts (price deals, in-store display etc.) and those of television advertising. This is important because sales promotional efforts may serve to increase a brand's price sensitivity by attracting marginal customers. On the other hand, television advertising expenditures may promote brand loyalty among a brand's regular customers. The result of an increased proportion of brand loyal consumers may be to decrease customer price sensitivity for the brand. Thus, the third question (an empirical one) is whether the effect of sales promotional variables is stronger than the effect of television advertising variables on household brand choices. This is an important question to answer if advertising effects are to be better understood. The end product of this research is two-fold. First, a model that incorporates the var-16 ious impacts of price and television advertising on consumer brand choices is developed, thereby increasing our. theoretical understanding of the nature of consumer demand. Sec-ond, implications from several empirically estimated models provide pragmatic answers for marketing practitioners regarding the effect of advertising on consumer price sensitivity and confirms the theoretical model which links advertising with price sensitivity. 1.6. The DISSERTATION ORGANIZATION This dissertation is organized into five additional chapters. An overview of these chapters is provided below. In chapter II, several theoretical and empirical articles concerned with the effect of advertising on consumer price sensitivity are reviewed. Some studies argue that if a brand with higher advertising is perceived by consumers to be of higher quality, higher advertising should result in lower price sensitivity. On the other hand, other studies argue that advertising informs consumers of brand differences that are perceived to be important, and given that search cost of equivalent information is greater than that provided by advertising, increased advertising results in increased price sensitivity. This research is concerned with only the first argument and demonstrates that both responses are theoretically possible. The latter argument is mentioned and discussed partly for reasons of completeness and partly to provide a basis for developing the theoretical model that supports this dissertation. In chapter III, a theoretical model of consumer brand choice and manufacturer be-haviour, incorporating advertising and pricing decisions, is used to derive the effect of television advertising on brand sensitivity. More specifically, the behaviour of consumers 17 loyal to the target brandf, loyal to the competitive brand, and those who are brand switch-ers, is used to derive a demand function for the target brand. The derived demand function is such that the target brand's own price and the advertising of the competitive brand serve to reduce the target brand's demand. In addition, the target brand's demand is positively influenced by its own advertising and the price of the competing brand. This derived demand function is then used to obtain the profit maximizing price for the target brand. The effect of television advertising on the profit maximizing price is dependent on the proportion of consumers loyal to the target brand, and a reaction function of advertising intensity. These are conditions of brand differentiation and brand competition, and they help interpret observed statistical results about, the effect of television advertising on brand price sensitivity. In chapter IV, the statistical modelling framework necessary to test for, and estimate the strength of, the effect of television advertising on brand price sensitivity is described. In addition, the dependent and independent variables used in this study are described in detail. Finally, the IRI scanner panel database for the product categories of dry dog food and the aluminum foil are compared to census data to assess representativeness of our sample. In chapter V, the empirical results for the brand choice models are presented and interpreted. It is found that the effect of television advertising is to increase a brand's price sensitivity. This is found to be the case for both product categories. It is also found that modelling the television advertising variable in the disaggregate form (a variable for each brand advertised) offers a better explanatory model with more stable parameter f The target brand is. the nominal brand under study and, when appropriate, under active management by the firm and the brand chosen by a consumer after comparing it to the competitive brand. The optimal values of price and advertising are derived for this brand. 18 estimates than that of the aggregate form. The study also replicates (Guadagni and Little 1983 and Tellis 1988) findings for the variables of brand loyalty, television advertising, display, presence of a deal, and price. In chapter VI, the major contributions and their managerial implications are summa-rized and commented on. It is noted that the effect of television advertising is to increase a household's price sensitivity. It is also noted that the effect of the sales promotional variables is stronger than that of advertising variables for dry dog food. The reverse is observed for aluminum foil. The chapter concludes with a discussion of several areas that are thought to be fruitful for future investigations into these phenomena. 19 CHAPTER II The Effect of Advertising on Price Sensitivity. 2.1. Conceptual and Inter-industry Comparison Literature 22 2.2. Recent Theoretical Literature 26 2.3. A Review of Recent Empirical Literature 36 2.4. Summary and Conclusions 41 20 The purpose of this chapter is to provide a background for developing a theoretical basis for understanding the relationship between price and advertising. This relationship is somewhat difficult to describe directly, because the economics and marketing literature fre-quently includes discussion of other elements of the marketing mix to support propositions about price and advertising. Furthermore, the literature on the role of advertising, in some instances, is fraught not only with emotional overtones but also with conflicting modelling assumptions. The latter is, for example, evident in the debate between Dixit and Norman (1980) and Shapiro (1980) about the social welfare implications of advertising. - In addition, the roles of advertising have been researched for many years by marketers and economists with their necessarily different perspectives. For example, Marshall (1923) discussed the constructive and combative roles of advertisements. The main stream of economics, how-ever, has yet to develop tools to analyze the effects of advertising within the framework of either the standard Marshallian partial equilibrium analysis or the Arrow-Debreu general equilibrium analysis (Butters, 1976). Two major views are held about advertising's impact on prices. The first considers advertising to be persuasive, that is, advertising induces the consumer to buy a brand at a premium price. The other view considers advertising to be an information source that helps the consumer to make informed choices. Although these views and their implications in terms of both managerial and consumer behaviour are inconsistent, theoretical models to discriminate between these viewpoints are not found in the literature. Thus,, one of the tasks in this chapter is to provide a coherent picture of the literature, so that in the next chapter, we can propose a theoretical framework which will provide the basis for an empirical test for one of these two divergent views about advertising. This chapter comprises four sections. In section 2.1, conceptual arguments about 21 advertising's role in the market place, as provided by Marshall (1923) and Chamberlin (1962), are discussed. Also, some of the empirical studies involving inter-industry com-parison of product advertising and price are included. In section 2.2, several theoretical models that describe the effect of advertising on product competition are reviewed. A summary of empirical studies reported in the economics and the marketing literature con-cerning advertising's effect on product prices is presented in 2.3. A summary of the effects of advertising on consumer price sensitivity is included in 2.4. 2.1. CONCEPTUAL and INTER-INDUSTRY COMPARISON LITERATURE The purpose of this section is to integrate the conceptual arguments about the pro-cess underlying the effect of advertising on price sensitivity. To accomplish this task, the related work of Marshall (1923) and Chamberlin (1962) is reviewed. These conceptual studies suggest that advertising plays two roles: it informs consumers and it alters con-sumer preferences. In addition, the literature which compares the relationship between advertising intensity and its impact on product price and price sensitivity across indus-tries is reviewed (Borden 1942, Bain 1956, Comanor and Wilson 1974, Lambin 1976). A common conclusion from all these studies is that high levels of advertising tend to reduce consumer price sensitivity. Conceptual Studies Marshall (1923), one of the earliest scholars to write about advertising, considered the "constructive" use of advertising to be socially beneficial. He thought all measures designed to draw the attention of consumers to product offerings were constructive. Sales promotional activities such as samples and end-of-aisle displays are examples of construc-tive forms of advertisements. Marshall also identified two cases of "combative'' advertising 22 which he argued were socially wasteful. In the first case, he assumed that familiar ob-jects are preferred by consumers to unfamiliar ones. As a result, advertisements that are directed towards familiar products increase habitual purchase behaviour without adding utility (information) to the consumer. In the second case, Marshall argued that lavish advertising programs increase costs to a firm. As a result, the firm's profits are reduced if an increase in sales attributable to the advertising program does not fully cover the increased costs. Moreover, if the increase in sales is at the expense of a rival firm's sales (or profits) then advertising contributes to social waste. Marshall appeared to find ad-vertising aceptable when it provides information to consumers but considered it wasteful when advertising activities do not properly affect primary (industry level) demand. Chamberlin (1962) argued that advertising affects a product's demand by spreading information about the product and altering a consumer's wants, f Chamberlin recognized that these two effects are difficult to separate. He did, however, isolate conditions that affect the shape and the location of the demand curve. He argued that advertising is mainly concerned with price and availability of the product and this affects the shape of the demand curve and its elasticity. More specifically, advertising, by spreading price infor-mation, causes the demand curve to be more elastic. He argued further that the location of the demand curve (e.g. increasing quantity demanded at a given price) will be affected by advertising when advertising contains information about product benefits. Finally, if ad-vertising increases familiarity, and if familar products are preferred to unfamiliar products, advertising alters consumer's wants, which in turn alters the product demand. Chamberlin discussed this as a part of his theory of monopolistic competition; first published in 1933. However, the effects of advertising that he postulated have received only limited attention f According to Chamberlin consumer wants included consumer preferences, perception of quality of goods, and quantity of goods demanded 23 in the subsequent literature. ' In summary, both Marshall and Chamberlin argued that advertising directed towards providing product information may result in higher consumer price sensitivity. Chamberlin also argued that habitual purchasing behaviour induced by advertising activities may lead to a less elastic demand curve, i.e., consumers may become less price sensitive. Inter-Industry Comparisons Borden (1942), in a classic empirical study, asked one of the questions that relates to this research. He sought to find whether or not advertising causes demand for a product to be less elastic. Borden concluded that brand advertising promotes favourable consumer preferences, and that because of advertising consumers will repeat their purchase of a brand even though the brand is priced above a competitive brand. Therefore, advertising tends to make demand curves for a product more inelastic (less price sensitive). However, Borden also observed that, in the long run, some advertised brands were unable to maintain their premium price position. Bain (1956) studied industry entry conditions and their influence on business perfor-mance for 20 industries. One entry barrier identified by Bain and relevant to the present discussion concerned product differentiation. According to Bain, "• • • product differentia-tion is propogated by differences in the design or physical quality of competing products, by efforts of sellers to distinguish their products through packaging, branding, and the offering of auxiliary services to buyers, and by advertising and sales promotional efforts designed to win allegiance and custom of the potential buyer". Bain concluded that brand allegiance (loyalty), based mainly on prolonged advertising exposure and motives of con-spicuous consumption offered the predominant explanation for major entry barriers in the 24 cigarette, liquor, and "quality" fountain pens industries, and to a smaller extent in the soap industry. He argued that, as a result of the entry barrier, established firms were able to increase prices substantially above marginal costs. Comanor and Wilson (1974) proposed that there is an element of asymmetry between new and established firms in that established firms are experienced in catering to con-sumer needs. As a result, the marginal effects of advertising among firms depend on how consumers respond to the messages of different firms. If consumers can examine a new product and recognize its relative benefits, advertising is likely to have a small influence on consumer decisions. If, however, consumer ignorance is a common phenomenon, and the cost of obtaining comparative brand information is higher relative to potential consumer savings resulting from the information, then advertising may have a major influence con-sumer decisions. In this type of market, heavy advertising may create substantial barriers to the entry of new firms and established firms may charge higher prices than those in a competitively operating industry. Comanor and Wilson concluded, based on empirical studies for two consumer goods industries where advertising was above four percent of sales, that consumers on average paid prices more than 15 percent higher when compared to consumer goods industries where advertising levels were below four percent of sales. In a later article, Comanor and Wilson (1979) wrote that market power can be achieved due to advertising, and is shown by the willingness of consumers to purchase high priced, highly advertised products even when lower priced substitutes are available. They argued that the information conveyed by advertising may create preferences for individual brands, and consequently, may result in a more inelastic demand curve. This work thus supports the view that advertising activities result in lower consumer price sensitivity. Lambin (1976) conducted an empirical study involving 107 European brands. The 25 data covered 16 product categories and seven European countries. Lambin argued for and tested a proposition that advertising intensity increases the capacity for the firm to charge higher prices. To test this proposition, he regressed estimated demand price elasticities of 22 brands with various measures of advertising intensities and found a significant negative correlation between price elasticity and advertising intensity as measured by brand advertising share. This result also supports the proposition that increased advertising activity will result in lower consumer price sensitivity. In summary, four inter-industry studies concerned with the effect of advertising on price sensitivity concluded that advertising decreases consumer price sensitivity, at least in the short run. 2.2. RECENT THEORETICAL LITERATURE The notion that advertising affects a brand's competitive position is commonly re-ported in the economics literature. In this section, several scholarly papers are reviewed that support this proposition. These papers view advertising as a tool which either informs consumers or affects consumer preferences. Some of the researchers explicitly assume that consumer preferences are affected by advertising because advertised brands are seen by consumers to be either higher in quality and/or more familiar than brands that are not advertised. To operationalize the competitive position of a brand, most authors have used a price level or, alternatively, some measure of price elasticity. For this discussion, the term price sensitivity will be used. Thus, if advertising resulted in a higher price level and/or resulted in a lower absolute price elasticity, then the effects of advertising imply a lower price sensitivity. The literature reviewed below is mainly concerned with the welfare implications of 26 advertising. The purpose in reviewing the literature, however, is to establish a basis for understanding the effect of advertising on consumer price sensitivity. Hence, in most cases, the entire contribution of each paper is not reviewed. Furthermore, while assumptions concerning the role of advertising are often explicit in most of the papers, assumptions concerning product variations and variations in consumer preferences are often implicit. Therefore, the discussion below may not indicate that the effect of advertising on con-sumer price sensitivity depends upon both product variations and variations in consumer preference. The literature review is organized in two parts. In the first part, several papers whose conclusions are that advertising results in lower price sensitivity are reviewed. In the second part, several papers are reviewed that draw the opposite conclusions. Literature Concluding that Advertising Results in Lower Price Sensitivity Several papers conclude that advertising serves to increase the price of the advertised brand. This implies that advertising decreases consumer price sensitivity. Initially, a study of a commodity product with identical consumer preferences (Dixit and Norman 1978) is reviewed. These authors argue that advertising increases the marginal utility of the commodity, thus lowering consumer price sensitivity. Following this, several papers are reviewed (Schmalensee 1974 and 1978; Kotowitz and Mathewson 1979; Boyer. Kihlstrom and Lafont 1984; Kihlstrom and Riorden 1984; and Milgrom and Roberts 1986) which assume that advertised brands are either of a better quality, or are more familiar to consumers, than non-advertised brands. . Dixit and Norman (1978) argue that if consumers are maximizing utility and firms are maximizing profit in a homogeneous product market that is at equilibrium without 27 advertising expenditure, then the presence of advertising causes the industry demand curve to shift to the right of the industry demand curve without advertising (see Figure 2.1). The demand shift occurs because the advertising efforts affect consumer taste in such a way that non-advertised products appear to be poor substitutes for advertised products. As a consequence, advertised products may have higher demand, or higher prices, or both, when compared to non-advertised products. One unique feature of Dixit and Norman's work is that the effect of advertising on price sensitivity was always negative even when market structure was varied from a profit maximizing monopoly to Cournot-Nash oligopoly, and to monopolistic competition. If all consumers responded to advertising in a similar manner, then Dixit and Norman's conclusions are valid. However, as pointed out by Shapiro (1980), these conclusions may not hold when consumers have varying price sensitivities. Furthermore, Dixit and Norman did not indicate the mechanism by which advertising changes consumer taste. Kotowitz and Mathewson (1979) assumed that advertising affects the consumer's per-ception of brand quality, since advertising informs consumers about brand attributes. The authors noted that "disseminative" and "persuasive" are two possible roles of advertising expenditures. They argued that disseminative advertising is truthful and easily verified whereas persuasive advertising may be either false or true and can only be verified by product usage. They further assumed a monopoly market structure and a rational but not fully informed consumer. Based on dynamic optimization behaviour of both the firm and consumers, the authors noted that in the short run it is possible for the firm to engage in persuasive advertising which may contain misleading information. Further, if product benefits are difficult to validate, and if advertising improves consumers' perception of a brand's quality, then advertising will tend to increase price. This implies that advertising 28 lowers consumer price sensitivity. ' Milgrom and Roberts (1986) proposed a model with price and advertising as decision variables that may signal product quality to customers. The model assumes that customers perceive high product quality by observing either a high price or a high level of advertising expenditure, or both. Product quality in the model, however, is not a choice variable. The authors argued that customers purchase a product during the initial period based on observed values of price and advertising. During this period, the perceived quality depends upon the extent to which a product is advertised and upon the product's price. A firm's sales of the product are assumed to increase with higher perceived quality. In subsequent periods, customers gain information about product quality by directly using the product, by communicating with other users, and by observing the firm's current period decisions about advertising level and price. Furthermore, in subsequent periods, the firm's sale of the product is assumed to increase as a result of an increase in the level of information about product quality. At equilibrium, the authors conclude that of all the strategies available to a firm producing a high quality product, the strategy of high price and a high level of advertising expenditure appeared to be the best option. For a firm producing a low quality product, a strategy of low price and low level of advertising is preferred. Based on this analysis, there may be a positive association between advertising and price, at least for a new product whose quality is generally not known. Kihlstrom and Riordan (1984) report a similar model of advertising effects in which firms do not choose prices. In other words, advertising alone is sufficient to influence a customer's perception of product quality. In this model, prices are determined by the resulting supply and demand conditions. However, like the Milgrom and Roberts model, when the system is in equilibrium, price and quality become positively correlated. 29 Thus, advertising through the influence of perceived quality may reduce a consumer's price sensitivity. Schmalensee (1974) formalized the view that advertising creates brand loyalty. He argued that advertising is more effective for established firms because consumers are more likely to be familiar with their products than the new products of firms entering a market. Therefore, if established and new entrant firms spend the same amount of money on advertising, then established firms are likely to have a greater proportion of brand loyal customers than new entrant firms. Brand loyal customers, at least in short run, may be willing to pay higher prices for the familiar products than for new products. In the long run, however, higher prices for established products will provide incentive for other firms to enter the market, and as a result, there will be increased competitive pressure on prices. Thus, in the long run, brand loyalty may not create any association between advertising and price. If familiarity results in brand loyalty, and no new brand enters the market, the model argues that there would be a positive association between advertising and price. In summary, several formal models that support (in at least some environments) the claim that advertising lowers consumer price sensitivity have been reviewed. A common argument in all the papers is that advertising affects consumer perception. If higher quality brands are priced higher and if consumers perceive highly advertised brands to be better quality brands, then, advertising results in lower price sensitivity. Similarly, if higher levels of advertising result in a higher proportion of consumers loyal to a brand, and the loyal consumers are willing to pay prices higher than the competitive brand price, advertising results in lower price sensitivity. One common weakness of all the models, however, is their implicit assumption that all consumers respond to advertising in a similar manner. This weakness is explored in the next chapter. 30 Literature Concluding that Advertising Results in Higher Price Sensitivity In this section the work of Nelson (1970, 1974, 1975 and 1978), Ehrlich and Fisher (1982), and Caves and Williamson (1985) is reviewed. These researchers argue that advertising is a source of product information for consumers and results in higher consumer price sensitivity. While most of the models reviewed in this section assume either monopoly or monopolistic competition as the market structure, Salop (1979) and Grossman and Shapiro (1984) provide a model for the oligopolistic market structure. The latter paper is concerned with advertising effects and argues that increased levels of advertising result in increased consumer price sensitivity. In a series of papers, Nelson argues that advertising provides information to consumers about product attributes such as the product's physical characteristics, prices, distribution outlet, etc. Nelson divided products into the categories of "search products" and "experi-ence products". With the former, he assumed that useful product attributes are evident to "the consumer by inspection, and there is little for the advertiser to gain by false or mis-leading advertising claims. Hence, advertising for these products tend to be informative. With experience products, however, important product attributes are difficult to verify except through actual use of the product. However, consumers through knowledge gained by repeat purchases, create incentive for the sellers to provide accurate information about the product. Thus, Nelson argued, for both product categories, sellers have an incentive to produce advertising that provides accurate information to consumers. This information, in turn, helps consumers make informed product purchase decisions. Nelson further reasoned that the elasticity of demand depends on the number of alternatives known or available to a consumer. If the consumer is aware of other brands or substitutes, demand will be more responsive to price changes. Since advertising is a low-cost method of making consumers 31 aware of substitute brands, advertising can increase the price sensitivity of demand and increase the competition in the market. Nelson's work is not based on a formal analysis and as a result we do not know the limitations of his reasoning. The proposal,. however, provides useful predictions for inter-industry comparisons. Ehrlich and Fisher (1982) extended Nelson's work on advertising as a source of infor-mation. They argued that consumers are not fully informed about the characteristics of different products or brands. In order to be informed, consumers either spend time search-ing for a product that matches their preferences or purchase a product and then learn about it. Thus, according to Ehrlich and Fisher, product price consists of the actual price plus the opportunity cost of time used to gather information about the product. This implies that consumers not only demand products, but also demand information about products. This, in turn, creates incentive for suppliers of products to supply information. Furthermore, if consumers perceive that it is cheaper to acquire information from adver-tising messages than other information sources, then they may be willing to pay more for the product to cover the cost of advertising. This implies that consumers may pay a higher price to acquire advertised products. The total cost of the product to the consumer, however, may fall since it includes both the search cost and purchase price. As a result, the demand curve will shift out to. the right in the presence of advertising because the total cost of consuming the product has declined. Ehrlich and Fisher thus use a concept of derived demand for advertising, and offer considerable flexibility in understanding the effect of advertising on product prices. More important though, this model incorporates the notion of opportunity cost of time and information which may not be directly measur-able, but which does affect consumer price sensitivity. Furthermore, the model is intended more for inter-industry comparison rather than for studying the interaction of advertising 32 on consumer price sensitivity. Salop (1979) and Grossman and Shapiro (1984) used a model structure similar to Ehrlich and Fisher's to study the consequence of product differentiation on a product market's equilibrium in terms of optimal brand prices, number of brands, and the level of advertising. Since Salop's work was the first published attempt to provide an economic structure to the problem of product differentiation, his paper is reviewed in detail by reporting his assumptions and main results. Following this, Grossman and Shipiro's papers are reviewed, but only the main features are discussed. Salop (1979) assumed an economy with only two industries. The industry under in-vestigation was monopolistically competitive with differentiated brands and decreasing average costs; the other industry was competitive producing a homogeneous commodity. Each consumer was assumed to buy either one unit or nonef of the differentiated commod-ity. The choice was made according to preferences, prices, and the distribution of brands in product space. Each brand was represented at an equal spatial distance from the other brands and each consumer was represented by the position of his or her "ideal" brand on the spatial map (see Figure 3.1). It was also assumed that consumer preferences were uni-formly distributed. The consumer's total cost of the purchase is the sum of a brand's price and a psychological cost; that is, a cost associated with the purchase of a brand that does not exactly match the most preferred set of product characteristics, i.e. the "ideal" brand. This latter cost varies across consumers while the brand's price is considered a constant in the market. In such a market, Salop argued that if a brand enjoyed a "monopoly" industry structure, then the brand is likely to capture all of the consumers in the brand's product f The model assumes that the consumer solves a two-stage maximization problem. In the first stage, the consumer decides whether to buy a differentiated product. In the second stage, the consumer compares alternative brands of the differentiated product for their relative costs. 33 space. If, on the other hand, a brand faced a "competitive" industry structure, then the set of consumers would be shared between competitive brands. From the above definitions of industry structures, Salop derived an implied demand function and concluded that the price sensitivity of the brand within a monopolistic industry structure is less than that of a brand within a competitive industry structure. Salop's model provides a useful frame-work for relating the industry structure to brand demand elasticity but it does not provide a basis for relating demand price elasticity to advertising. Salop's model is extended in chapter III to indicate how advertising can affect consumer price sensitivity. Grossman and Shapiro (1984) extended Salop's model by assuming that an advertise-ment in the differentiated product market provides full and truthful information about the brand it promotes. In addition, they assumed that consumers do not have alternative sources of brand or product information. Further, a consumer was assumed to remain unaware of the existence of a particular brand unless that consumer sees an ad for it. Re-ceiving information through an ad is assumed to be costless to the consumer. The authors argued that if a product class is homogeneous, then the minimum number of brands for which a consumer needs information is one. Any additional advertising about brands in such a product class results only in the redistribution of consumers among the competing brands. If, on the other hand, a product class is heterogeneous, then a utility maximizing consumer needs to be informed about all the brands in the product class. The authors note that advertising serves to inform consumers about the finer differences among com-peting brands. This role of advertising may result in making the product market more competitive by increasing the elasticity of demand for each brand. In a related paper, Caves and Williamson (1985) argued that when the cost of pro-ducing each product variety is fixed, firms that are not large enough may not be able to 34 supply the variety of brands to satisfy each buyer's particular need. Therefore only a small number of brands will be available in the market and they will be perceived by consumers as imperfect substitutes for one another. Thus, advertising serves to inform consumers about the finer differences among competing brands. If consumers are uninformed about brands in the market, then they are likely to choose a brand that incurs a higher psycho-logical cost than otherwise. This implies that higher advertising would result in a higher proportion of informed consumers and a lower psychological cost, on average. In summary, the papers reviewed in this section by Nelson, Ehriich and Fisher, Caves and Williamson, Grossman and Shapiro conclude that advertising results in higher price sensitivity. There are minor differences across papers to account for the hidden cost of product purchase. While Ehriich and Fisher term it search cost, Salop and others term it, more properly, psychological cost. In addition, if the cost of receiving information through advertising is lower than the search cost, it is possible to conclude that higher advertising results in higher price sensitivity. Furthermore, if advertising provides information that makes consumers aware of a brand or brands not previously known to them, all the above authors agree that one firm's advertising increases the consumer price sensitivity of its brand (although absolute sales may also be increased). In conclusion, if advertising affects consumer tastes, then it is likely'that this results in lower price sensitivity. On the other hand, if advertising serves to inform consumers about brands not known prior to the advertising, then advertising will increase price sensitivity. One common weakness of most of the models is their implicit assumption that all consumers respond to advertising in a similar manner. Furthermore, these models do not establish conditions that would create lower and higher price sensitivity as a result of advertising. One such model (an extension of Salop's work) is proposed in the next chapter to improve 35 upon these weaknesses. 2.3. A REVIEW of RECENT EMPIRICAL LITERATURE The purpose of this section is to review the empirical literature from economics and marketing that is concerned with the impact of advertising on product price sensitivity. In the preceding section, advertising was discussed as an information source that increases price sensitivity, and as a persuasive tool that affects tastes and in turn decreases price sensitivity. Such a distinction, however, is difficult to find in empirical work and one must be very cautious when interpreting empirical results in this respect. It is difficult to judge whether empirical results support one or the other or both of the theoretical propositions about the effect of advertising on price sensitivity. Review of Studies from Economics:. The results and methodology of seven studies (Schroeter, Smith and Cox 1987; Haas-Wilson 1986; Kwoka 1984; Feldman and Begun 1978 and 1980; Marvel 1979, Maurizi 1972; Cady 1976, and Benham and Benham 1976, and Benham 1972) that have appeared in the economic literature are summarized in Table 2.1. Most of these studies were intended to demonstrate the effect that advertising regulations have on consumer prices. In these studies, changes in advertising, if any, are not typically measured as a continuous variable and, as a result, one must examine these findings with considerable caution. However, a pattern across all seven studies does seem to show that restricting advertising efforts either increases product prices or, at best causes them to remain at the same level (The magnitude of the price increases found by the studies ranged between 3% and 33%). The effect reported by the seven studies is likely to be a result of local price adver-tising or local informational advertising, and should not be generalized to predict effects 36 from national media advertising programs that are generally thought to be designed for persuasive purposes. This is particularly true for the packaged goods industry. Two of these studies are reviewed below. Kwoka (1984) conducted a quasi-experiment to demonstrate the impact of advertising restrictions on prices charged to consumers by optometrists. He selected seven cities in the United States, four of which permitted some form of advertising while the remaining three restricted advertising by optometrists completely. Within each city, 21 optometrists were randomly chosen to be included in the study. To maintain experimental control over patient characteristics, seven patients with similar visual conditions were trained at two optometry schools with regard to an optometric examination. Each patient then went to three optometrists in each of the test cities for an eye examination. The following variables were measured: cost to consumer, time spent, and various other details about the tests and procedures employed during the examination. Kwoka concluded that restricting advertising increased the price of optometric service when the quality of service (measured in terms of the amount of time the examination consumed) was held constant. Further, he reported that optometrists who used only "in-store advertising" charged more for their services than those who advertised through multiple local media. Cady (1976) followed a more conventional approach to the problem by collecting information and constructing a regression model to explain the average consumer prices of ten commonly prescribed drugs. Cady's regression model included "advertising restriction" measured as a dummy variable. In addition, several demographic variables describing the region in which the store was located, and several store characteristic variables were also included as independent variables. His sample consisted of drug sales from 1848 drugstores across the U.S. The study reported that drugstores in the states which restricted consumer 37 advertising tended to have higher prices (4% to 9% higher) relative to states which had no restrictions. While there is, as yet, no conclusive explanation of these empirical findings, one may speculate that these effects may be due more to changes in the retail competitive market environment when advertising is allowed, than to any direct changes in consumer price sensitivity. While they represent important contributions, it is not clear from these studies how advertising efforts influence product price sensitivity. Review of Studies from Marketing Unlike the studies from economics, the ten studies reviewed from the marketing lit-erature (Krishnamurthi and Raj 1985, Wilkinson, Mason and Paksoy 1982, Farris and Reibstein 1980, Eskin and Baron 1977, Wittink 1977, Prasad and Ring 1976, Eskin 1975, Woodside and Waddle 1975, Curhan 1974, and Massy and Frank 1965) are based on mar-ket data for frequently purchased consumer products. Advertising levels were controlled experimentally in eight of the ten studies. Of these ten studies, seven measured television advertising while the remaining three measured in-store and newspaper advertising (see Table 2.2 and 2.3 for details). The seven studies involving television advertising studied the impact of variations in price and advertising levels on consumer price sensitivity. Four of them (Krishnamurthi and Raj, Prasad and Ring, Farris and Reibstein and Massy and Frank) reported that higher advertising levels resulted in a lower price sensitivity. The re-maining three studies, however, reported.the opposite effect; i.e. higher advertising levels resulted in a higher price sensitivity. All three studies reporting on in-store and newspa-per advertising concluded that higher levels of price advertising (either newspaper feature advertising or in-store display) resulted in higher consumer price sensitivity (Table 2.3). Two studies, one from each subset that involved television advertising, are summarized 38 below. • Krishnamurthi and Raj (1985) studied sales, prices, and advertising levels using AD-TEL split-cable TV data.f Their sample was demographically matched and consisted of 189 families in the experimental panel and 192 families in the control panel. The test brand was the market share leader. Both panels were initially exposed for 52 weeks to the same level of advertising efforts. For the next 24 weeks, the test panel was exposed to a higher level of advertising while the control panel received the same level of advertising as before. Each of the families in both groups were asked to maintain a weekly record of their pur-chases, including prices paid, deal or nondeal purchases, and the store used. The aggregate demand functions for the brand (aggregating across 76 weeks) was then estimated using the following independent variables: relative price, the multiplicative variable of the price times the dummy variable defined by whether or not a household belonged the experimen-tal panel, total grocery purchases (dollars), and total television viewing hours. The study reported that higher advertising levels resulted in a lower price sensitivity. Further, the study reported that higher advertising levels also resulted in higher levels of consumption, when other demand factors were held constant. More specifically, consumers in the high and low advertising conditions had price elasticities of ^0.79 and —1.33 respectively and these were significantly different at p < 0.05. Eskin and Baron (1977) studied the effect of price and advertising expenditures on new product sales from test market experiments. Four different products were studied and advertising expenditures and product prices were varied across the test cities. Two cities were assigned to each of the advertising conditions and store audits were used to measure f In A DTEL split-cable TV system, a sample of households is connected to cable system A and another to cable system B in a test city. ADTEL experiments often involve varying the number of advertising exposures across the cable systems. 39 each product's monthly unit sales for a six month period. Several other determinants of a store's potential sales were also collected. The study concluded that consumer price sensitivity for new brands increased with increased advertising expenditure, although not all product interactions were significant (p < 0.05). This is in direct contrast to the study by Krishnamurthi and Raj (1985), reviewed above. They reported a decrease in price sensitivity for the established market share brand leader when advertising levels were increased. In summary, we note that of the seven studies involving television advertising, four studies, using a variety of data sources for consumer goods, concluded that higher adver-tising levels result in lower consumer price sensitivity. The other three studies, however, reported the opposite effect. These conflicting empirical findings may be due to several factors. First, it is possible that higher advertising levels may result in a higher proportion of brand loyal consumers. Further, if brand loyal consumers are willing to pay a higher price for the same product than brand switchers are willing to pay, then one may find higher advertising levels resulting in lower price sensitivity. Second, if the competitive environment is dominated by a market share brand leader, then one may find that higher levels of advertising by the market leader may lower price sensitivity toward the leader. Third, if advertising is concerned mainly with informing consumers about the price and availability of newer brands, then advertising makes consumers aware of brands. This appears to result in increased consumer price sensitivity as proposed by Nelson (1970). These results, however, suggest the need for an improved understanding of the effect that advertising has on price sensitivity. To accomplish such a task, we propose a theoretical model in the next chapter. In subsequent chapters we will elaborate a methodology to test our theoretical model with data from a 40 consumer scanner panel. 2.4. SUMMARY and CONCLUSIONS In this chapter, several theoretical and empirical articles concerning the effect of advertising on consumer price sensitivity were reviewed. Most of the inter-industry studies reviewed appear to confirm the view that higher levels of advertising result in lower price sensitivity. , This view is further supported by several theoretical models. If a brand with higher advertising levels is perceived by consumers to be of higher quality, these models argue that higher advertising levels result in lower price sensitivity. Similarly, if higher advertising levels encourage higher familiarity levels, then lower price sensitivity is observed. Nelson's papers,, however, proposed the opposite theoretical point of view. Ehriich, Fisher and others further developed Nelson's point of view. This view argues that adver-tising may inform consumers about finer brand differences at a lower cost than the search costs incurred by consumers for product information. As a result, advertising may reduce the cost of product purchases to consumers. This perceived price reduction is considered to be an indication of increased price sensitivity that results from advertising efforts. Also reviewed was the empirical literature from economics and marketing. A review of papers concerned with the economics of advertising regulation suggested that restricting advertising efforts increases product prices. In other words, a higher level of advertising resulted in greater price sensitivity. A review of the literature from four marketing studies suggested that a higher level of advertising results in lower price sensitivity, while three other studies concluded the opposite effect. . In order to help improve our understanding of these seemingly inconsistent results, a 41 theoretical model based on some of the work reported in this chapter will be proposed. The model will identify the necessary conditions that cause advertising to increase or decrease consumer price sensitivity. Furthermore, the model will be specifically designed for the low cost, frequently purchased, and brand identified packaged goods industry, where television is the most important advertising medium. Following this, an empirical test of the model will be made using appropriate statistical procedures and scanner panel data. 42 CHAPTER III A Model of Advertising with Loyal and Non-loyal Consumers 3.1 Introduction 44 3.2. Product Market Operation 45 3.3. Model Assumptions and Overview 47 3.3.1 Microeconomic Assumptions 48 3.3.2 Consumer and Manufacturer Assumptions 50 3.3.3 Modelling Advertising Decisions 53 3.3.4 The Model Overview . 55 3.4. The Model 56 3.4.1 List of Model Assumptions 56 3.4.2 The Optimal Price Derivation 58 3.4.3 Effect of Advertising on Price 63 3.5. Summary and Conclusions 68 43 S.l, INTRODUCTION In chapter I, a review of several meta-analyses indicated that marketing mix variables such as price, advertising, deal amount, coupon amount, and in-store display have a major influence on brand choices. It was also noted that one of the "best predictors" of a brand's future sales was the brand's prior sales. Chapter II reviewed and discussed the theoretical and empirical literature concerning the effect of advertising on brand's price sensitivity. Two competing theoretical views concerning the impact of advertising on price sensitivity and their supporting literature were identified and discussed. The first argues that advertising, by differentiating among brands, induces a consumer to buy a brand at a premium price. This implies that higher levels of brand advertising will result generally, in lower price sensitivity. In the other view, advertising is seen to inform consumers about the availability and characteristics of alternative brands and thereby helps consumers make informed choices. This later view implies that higher levels of brand advertising will result in higher levels of price sensitivity. In this chapter, a theoretical model will be developed and presented so that the conditions under which one is expected to observe either increase or decrease in price sensitivity, as a result of increased advertising. The model reported below suggests that while higher levels of brand advertising may increase brand differentiation, competitive reaction can still result in higher levels of advertising causing higher price sensitivity. Several of the models reviewed in chapter II assumed that either advertising increases or decreases price sensitivity, but not both. The purpose of this chapter is to identify a set of conditions that are necessary for advertising to exert an influence on consumer price sensitivity and to show that both of the effects are possible. More specifically, a mathematical model will be developed to guide the empirical investigations discussed in 44 the next two chapters. The mathematical model developed here, like those described in chapter II, requires several assumptions. The set of predictions that result from the model depends upon the "reasonableness" of these assumptions. Hence, we will first justify the assumptions made in our model. We will then proceed to describe our model intuitively. This description is followed by a mathematical development of the model showing the effect of advertising on price sensitivity. The final section of this chapter is a summary of this chapter's main points and concludes with several comments about the later empirical work which is part of this thesis. S.S. THE PRODUCT MARKET OPERATION The purpose of this section is to provide a brief overview of the operation of the frequently purchased, brand identified, and low cost packaged goods market. This market consists of three groups of participants - consumers, retailers and manufacturers - although we will consider retailers as agents of manufacturers. In this section, the important characteristics of each of these participants are described, with particular attention to factors relating to advertising and price effects. In the analysis, decision making by retailers will not be considered since retailers may be expected to implement manufacturer brand policies (Watkins 1985). However, for completeness we will describe the role of retailers in the industry. Consumers of frequently purchased, brand identified, and low cost grocery products have long been thought to make purchases in a routine manner (Assael 1987, p. 92), with little effort, and selecting their choice from a small set of acceptable brands (Hoyer 1984). The purchase decision is seen to be mostly the result of habitual behaviour patterns rather than the outcome of a reasoned evaluation procedure involving substantial attribute 45 comparison across the competing brands (Kuehn 1962 and Houston and Weiss 1976). This view is reinforced by Engel and Blackwell (1982) who report that consumers perceive low social or financial risk in the purchase of such products. In addition, Kuehn also concluded' that many consumers make purchase decisions for grocery products based exclusively on their experience with the brand unless situational factors intervene. As a result, for consumers who are making a low involvement or habitual purchasing decision, advertising by the purchased brand will have primarily a reinforcing effect on their current choices. Frequently purchased packaged goods are mostly sold in supermarkets. The retail price of a package is thought to be an important decision variable for some consumers (Assael 1987). It is also an important tactical variable used by manufacturers. Yet in many instances, the weekly or daily posting of consumer prices is controlled by retailers (Watkins 1985). It is also commonly understood in the industry that retailers prefer to carry a brand and size combination that is popular with consumers and therefore expect such brands to remain on the shelf for shorter durations (Mason and Mayer 1981, p. 606). The following discussion concerns brand management and advertising practices in the industry. Most of the frequently purchased packaged goods industries are oligopolies dominated by a small number of manufacturers (Jones 1986). It is commonly the case that the same manufacturer may have from five to ten different brands competing in the same product market with the marketing efforts for each of these brands managed independently of the other brands by a brand manager (Kotler and Turner 1979). While many of these brands are close substitutes for one another and other competing brands, advertising efforts are directed towards informing consumers that each brand is indeed unique. Manufacturers perceive that advertising directed toward this end is a relatively important sales generating activity (Jones 1986). This is further evidenced by the fact that many companies budget 46 from five to ten jps»rcent of annual sales on consumer focused advertising efforts (Comanor and Wilson 1974}, Furthermore, it is also common for a company to allocate from 80 to 90 percent of the total advertising budget to television advertising. Television advertising for these products tends to focus on one or two important product characteristics (Stewart and Furse 1986 ;and Assael 1987) and by and large appears to be aimed at existing brand users (who may, in fact, be users of more than one brand) to reinforce their perception of brand benefits. In summary, consumers of frequently purchased, brand identified, and low cost prod-ucts typically make purchases in a routine or habitual manner. Much of the industry's advertising effort is directed towards establishing a unique selling position for a brand and thereby reinforc'kig the habitual purchasing pattern of present brand customers. To model such an industry, it is necessary to simplify some of the components that are not directly relevant to understanding the effect of advertising on consumer price sensitivity. 8:3. MODEL ASSUMPTIONS AND OVERVIEW To model such an industry, the model developed by Salop (1979) provides an attractive framework. Salop's model is unique and valuable since it allows one to model consumer and firm behaviour simultaneously. X The model, however, does not allow variations in consumer responses to advertising, and it has thus been modified and extended in order to do that. Most important, the modified model (described below) offers a structure with which to detect the direction of advertising and price interactions. In other words, the model allows one to specify the circumstances under which advertising increases or decreases the level of brand price sensitivity. It is important to note that Salop's model, as $ As is common, we explicitly model only the consumer and the manufacturer. Although Robert Steiner (1973) and others consider retailers as separate agents, we do not pursue this approach here. 47 modified in this chapter, is also consistent with the standard microeconomic framework. Thus, for example, the model of perfect competition or the monopoly industry model are special cases of the model developed here. The remainder of this section is organized into four subsections. First the microeco-nomic assumptions that are used in our model will be described. Next we will describe and discuss the assumptions concerning the actions of consumers and manufacturers, and then we will proceed to discuss several assumptions about modelling advertising decisions. We will conclude this section with an overview of the proposed model. S.S.I. Microeconomic Assumptions Our modelling effort is concerned with the behaviour of companies and consumers. We assume that the companies are maximizing total profit from a particular brand and that consumers are minimizing the total cost of a particular purchase while achieving a given utility level|. In the model, a brand's total profit is calculated as the difference between total revenue and total costs of production and advertising expenses. The consumer's total cost of the purchase is the sum of a brand's price and a "psychological" cost. Psychological cost is the cost associated with the purchase of a brand that does not exactly match the consumer's most preferred set of product characteristics. Such a cost varies across consumers while the brand's price is considered a constant in the market. These assumptions are made because it is reasonable to believe that companies and consumers may be implicitly pursuing these objectives. Furthermore, these objectives are simpler to work-with than assumptions like maximizing sales or market share. Finally, as noted by Varian (1984) these assumptions may be reasonable first approximations to many other f In terms of microeconomic theory, consumers minimize total expenses and set their utility for a brand equal to a base reference price. In other words, we are using Hicksian expenditure function (Varian 1984). 48 assumptions such as maximizing sales or market share. ' In our model, a company is choosing a level of price and a level of advertising efforts while a consumer chooses a brand from a set of competing brands. When both consumers and managers have achieved optimal levels in their choice variables, it is customary to say that a state of economic equilibrium has been reached. This is the state at which no one can further improve their objective function through their behaviour, given certain limitations about the behaviour of others. Given that such an optimal equilibrium price exists, the question to be asked is how will it respond to changes in optimal advertising efforts. That is, what are the changes to the optimal price with respect to changes in the optimal level of advertising? The direction of these changes will show how price sensitivity varies with changes in advertising. We will be using Salop's model which is based on notions of spatial competition to demonstrate the effect of advertising on brand price. It is important to recognize, however, that the model is also related to the standard microeconomic model of an industry with perfect competition. For example, if consumers do not perceive some loss when they purchase a brand that does not exactly match their preferences, then the spatial competition model would behave like the model for perfect competition (Capozza and Van Order 1978). As a result, one theoretical test of the spatial competition model is to demonstrate that as the cost per unit loss of benefit approaches zero (i.e. consumers do not perceive product differentiation), price should approach marginal cost. In fact, this is found to be the case in the proposed model (see page 62). 49 S.S.2. Consumer and Manufacturer Assumptions Consider a market with three brands, A, B, and C and assume that they represent a segment of a larger perceptual map (see Figure 3.1). Further, consider each of these brands to be positioned along only one perceptual dimension such as meatiness for dog foods, strength for aluminum foils, or nutritional value for ready-to-eat cereals. Only one dimension is considered for three reasons. First, if there is more than one dimension of product quality, then a linear combination of multiple dimensions can always be con-structed to obtain single index variable of product quality (Arnold, Oum, Pazderka, and Snetsinger 1987; and Bresnahan 1981). Second, it is common to use only a few dimensions in theoretical spatial models. For example, the Defender model (Hauser and Shugan 1983), which is the basis for one of marketing's most popular spatial mapping methods, relies on a two dimensional preference representation. Third, as stated earlier, advertising efforts generally focus on one or two characteristics of the advertised product. It is also assumed that both brands A and C charge identical prices (p) and that brand B charges a different price, p. This assumption reduces the number of variables in the model, and implicitly means that the behaviour of all competitors to brand B is alike. Generalization to multiple brand prices poses neither a conceptual nor a mathematical problem, although it is cumbersome, f In the empirical work reported in the next two chapters, however, no restriction is made to just two brands. A test will be made to see if all competitors behave alike and if this is not the case, then the empirical model will be modified to reflect the behavior of multiple competitors. For ease of exposition, brand B is called the target brand and all others are referred to as the competing brands. t If there are n competitive brands in the market each with unique price, then target brand optimal price will depend upon (n— 1) competitive prices which is more cumbersome to accomodate within the model. 50 To be consistent with the assumption of equal prices for the competing brands, it is also assumed that the perceptual distances (see Figure 3.1) between brands A and B, and between B and C are identical, (say x units). Eaton and Lipsey (1975) argued that long run equilibrium in terms of the location of these brands implies equal distance between them. It is also assumed that consumers for a given purchase occasion must purchase exactly one unitf of one of the three brands. By assuming that each of the consumers will buy one of the three brands, the probabilistic component of their not purchasing anything at all is not considered. Furthermore, Greenhut, Norman and Hung (1987) report that this assumption will result in a linear (in price) aggregate demand function which is desirable for the purpose of this model. They also show that modelling the consumer's choice combinations of both brand and quantity will result in the aggregate demand function being non-linear in prices. Although such a topic may require further research, it is assumed in this model that consumption of a product category such as dog food remains fixed for all consumers. This implies that quantity demanded is fixed for all consumers.. The "location" of a consumer on the perceptual map is defined by the consumer's most preferred set of product characteristics. Thus, a consumer located at a distance z from the target brand prefers z more units of the product characteristics than what the target brand has and (x— z) units less than what competing brand C has. For example, Consumer Reports assigns a nutritional index (dietary fibre plus protein minus added sugar, sodium and fat) to each of several brands of ready-to-eat cereals. In this case, hypothetical brands t The model assumes that the consumer solves a two-stage utility maximization problem. In the first stage, the consumer decides whether to buy a differentiated product subject to a budget constraint. In the second stage, the consumer compares alternative brands for their relative costs. Since advertising is likely to influence the second stage, subsequent discussion is concerned with the brand comparison process. 51 A, B, and C have nutritional indicies of 50, 60, and 70 respectively, and the example consumer prefers an index of 63 (i.e. 2=3 see Figure 3.1). For simplicity, it is assumed there are N consumers in the market whose preferences are uniformly distributed along the perceptual map between points A and G. If it is assumed that consumer preferences are normally distributed, for example, then mathematical com-plexity would be added to the model (a complicated form of integration) without providing additional insights about the effect of advertising on price sensitivity. Thus, in this case, over the line segment of length x (say, half of the perceptual map) there will be y con-sumers. Since there are only three brands in the market, very few of the N consumers' will be able to buy the brand that exactly matches their most preferred product characteris-tics. This means that, in general, consumers will buy a brand that differs to some degree from their most preferred characteristics and incur some psychological cost. If z denotes the distance away from the target brand and a consumer prefers product characteristics of amount L, then the consumer faces a product benefit loss (or psychological cost) of z units (Salop 1979 and Grossman and Shapiro 1984). Further, if t is a per unit cost associated with the benefit loss, then zt will be a cost associated with the benefit loss. The consumer's total cost with respect to the purchase of the target brand will be the actual price (p) plus the benefit loss (zt). A theoretical concept of reference pricef is also used in the model. This is the price that a consumer expects to observe at point-of-purchase (Winer 1986). While the concept of reference price is mostly a theoretical one, there is considerable empirical support for this construct for a frequently purchased product like coffee (Gurumurthy and Little 1985 f Reference price is the price that a consumer expects to observe at point-of-purchase (Winer 1986). Reservation price, however, is the maximum acceptable price that a con-sumer expects to pay for a brand that exactly matches his or her preference (Scherer 1980). Both are different from the actual price (p) a consumer pays to buy a brand 52 and Winer 1986). As we will see below, use of this concept for a theoretical purpose does not require its measurement in the later empirical work. As a result, it is a very useful model building tool. S.S.3. Modelling Advertising Decisions It was noted in chapter II that if advertising is directed towards building favourable consumer preferences (Borden 1942, Bain 1956, Dixit and Norman 1978 and Schmalensee 1974) then, as a consequence of advertising, consumers may be willing to pay a premium over the competitive brand prices. Similarly, Comanor and Wilson (1974, 1979), Kotowitz and Mathewson (1979), and Milgrom and Roberts (1986) argue that if advertising improves the perception of brand quality, then as a result of advertising, consumers may be willing to pay a premium over competitive price for this reason as well. Both of these arguments imply that advertising may increase the consumers reference price for the advertised brand. If advertising intensity is at level zero, then the price a consumer expects to pay is termed a base reference price (v). If advertising intensity is at a level of a, the reference price will adjust to v+ h(a). According to the literature cited above, the function h(a) is a ndndecreasing function in a (i.e. || > 0). Furthermore, if the advertising intensity level is zero, the associated increase in the reference price must be zero (i.e. h(0) = 0). One way to approximate h(a) is to take a Taylor series approximation about the base level of zero advertising intensity. Thus, the first order approximation can be written as h(a) « h(0)+ §^( a ~ 0). Since, h(0) — 0, one can write the brand's adjusted reference price at advertising intensity of a to be v+ a|^. Substituting the marginal change (|^ ) into the expression of adjusted reference price as a positive constant (k), one can write the value of the adjusted reference price at an advertising intensity of a as v + ak. 53 To operationalize the measure of advertising intensity (a), the level of reach, r, and frequency, /, of an advertisement are defined. A fraction, r, of the target population is exposed to the advertising message on / separate occasions. One functional form, often used in the advertising industry for this measure of intensity, is reach times frequency. It is assumed, however, that a more general functional form links advertising intensity to reach and frequency. More specifically, assume a = ip(r,f). Furthermore, advertising intensity will be assumed to be increasing (or, at least, not decreasing) with increasing levels of reach and frequency (dtp/dr > 0 and dip/df > 0). To achieve higher levels of advertising intensity, however, rate of change of advertising intensity with respect to frequency and reach is negative (i.e. function ip is a concave with respect to r and /). Advertising is a cost to the firm. A reasonable cost function is A = g(r,f) advertising dollars per consumer. If there are N consumers in the market, then the total cost of advertising will be A X N. We assume that the function g(-, •) is increasing in r and / (dg/dr > 0 and dg/df > 0), so that rate of change cost to achieve higher values of r and / is positive (in terms of calculus, g(-,-) is a convex function). Three groups of consumers are distinguished based on their past purchase behaviour: a group that is loyal to the target brand, a group loyal to the competitive brand, and a group of brand switchers. A loyal consumer is the one who has made "several" prior purchases of a brand. Furthermore, a consumer can either be loyal to the target brand or to the competitive brand. Finally, consumers who are loyal to neither brand will be called brand switchers. Each of these groups responds differently to advertising for their own brand and competing brands. This differing responsiveness is operationalized in the model as follows. For consumers loyal to the target brand, the marginal change in the base reference price 54 resulting from the advertising efforts of the target brand and the competitive brand is denoted by ki and ki respectively. Similarly, for consumers loyal to the competitive brand, the marginal change in the base reference price resulting from the advertising efforts of the target brand and the competitive brand is denoted by kc and kc respectively. Finally, for brand switching consumers, the marginal change in the base reference price resulting from the advertising efforts of the target brand and the competitive brand is denoted by ks and ks respectively. The specific measurement procedure used to operationalize the brand loyalty construct is discussed in the next chapter. , (. • • . • Three reasons suggest the rationale for the use of these groups in the proposed model. First, these groups are expected to respond to marketing mix variables differently (see summary in Table 3.1). For example, brand loyal consumers may use brand advertising to evaluate a brand after purchase, whereas brand switchers may use advertising prior to purchase to.compare competing brands. Second, it was noted in chapter I that one of the "best predictors" of a brand's future sales is the brand's prior sales. Brand loyalty provides a basis for modelling such a component. Third, the degree of brand familiarity and future choice probabilities can be expected to vary systematically over these groups. This is important because these groups can be expected to have different adjusted reference prices. S.S.4- Model Overview . Before developing the model in detail, an intuitive sketch of the model is given. Consider a market with two brands and three groups of consumers: a group loyal to the target brand, a group loyal to competitive brands and a group of brand switchers. Consumers incur two types of costs when purchasing a brand: the actual price paid and 55 the psychological cost associated with the product benefit loss when the purchased brand does not match the consumer's most preferred set of product characteristics. The consumer compares the total cost of purchase (i.e. the adjusted reference prices) for two brands $ and purchases the brand that offers, comparatively, the lowest cost. It is this comparison process that offers a basis for deriving the aggregate demand function for a brand. The model's demand function is such that the target brand's price and the competing brand's advertising decrease the target brand's demand. Furthermore, target brand's demand is positively affected by the target brand's advertising level and the competing brand's price. This demand function is used to derive a company's profit maximizing price. The effect of advertising on the profit maximizing price is dependent on several factors including the proportion of consumers loyal to the target brand, consumer responses to advertising, and the reaction of competition to changes in advertising levels. More specifically, it is concluded that while levels of brand advertising may increase product differentiation, competitive reaction can still cause the end result to be lower price. Sj. THE MODEL The description of the proposed model is organized into three subsections. In the first subsection, the assumptions associated with the model are listed. In the next subsection, these assumptions are used to derive the optimal price and its properties in relation to other models of competition. The final subsection describes and comments on the effect of advertising levels on price. 8.4-1- List of Model Assumptions Assumption 1: There are several brands of a single product that is differentiated along t Since all competitors behave indentically, we are in effect modelling a two brand market. 56 a single dimension of quality. The single dimension of quality may be a composite index of several physical brand attributes. Assumption 2: Consumers incur a cost when they are not able to purchase a brand which is an exact match between characteristics they most desire and the purchased brand's actual characteristics. The cost per unit of characteristics lost is t dollars. Assumption 8: Potential consumers are distributed uniformly along the same single dimension of quality as are brands. Assumption 4- All consumers buy one unit from the product class on each purchase occasion. Consumers act to minimize the total cost associated with a purchase while companies are maximizing profit. Assumption 5: Advertising cost per consumer is given by a cost function A = g(r,f), where r and / are the reach and frequency of the advertising program. It is assumed that g(-, •) is a convex function. Assumption 6: Advertising intensity (a) is determined by reach (r) and frequency (/). Thus, o = ip(r,f). It is also assumed that -ip{-, •) is a concave function. Furthermore, advertising intensity affects a consumer's base reference price. Thus, if a is the advertising intensity, k the marginal increase in reference price as a result of advertising, and v the base reference price when advertising intensity is zero, then v+ak is the adjusted reference price resulting from advertising efforts. Assumption 7: There are N consumers segmented into three groups. Group one is loyal to the target brand with NO consumers, group two is loyal to the competitive brand with NO consumers and group three, the brand switchers, has NO consumers (0 = 1 — 0 — 0). Consumer preference of these three groups are uniformly distributed. While 57 companies cannot price discriminate across these consumer groups, these consumer groups may respond differently to the advertising messages of each brand. As a result, these groups may have different adjusted reference prices. For consumers loyal to the target brand, the marginal changes in the base reference price due to the advertising efforts of the target brand and the competitive brand are denoted by and ki respectively. Similarly, for consumers loyal to the competitive brand, the marginal changes in the base reference price resulting from the advertising efforts of the target brand and the competitive brand are denoted by kc and kc respectively. Finally, for brand switching consumers, the marginal changes in the base reference price resulting from the advertising efforts of the target brand and the competitive brand are denoted by ks and ks respectively. Assumption 8: The production cost function has constant marginal and fixed costs. C — F + cq, where q is output demanded, C is total production cost, F is fixed cost, and c is marginal costf. 8.4-2. The Optimal Price Derivation A model of brand purchasing is now derived for the three groups of consumers men-tioned above and is used to make predictions about the effect of advertising on price sen-sitivity. First, a demand function is derived for the consumers loyal to the target brand, then a similar approach is used to derive demand functions for the other two groups. If v is the base reference price for a consumer loyal to the target brand and he or she is exposed to advertising at level a from the target brand, then at the next purchase occasion his or her adjusted reference price will be v + afy for the target brand. Similarly, f Marginal cost (c) may be a function of quality. For example, c= c[z) and marginal cost may be increasing in quality and it becomes increasingly expensive to achieve higher levels of quality. Since the focus of present work is on the optimal price and advertising decisions of the target brand, the marginal cost aspect of quality is not explicitly modelled. 58 if advertising is a for the competing brand, then the competing brand's adjusted reference price will be v + aki for that consumer. If a consumer loyal to the target brand is located at a distance z from the target brand and considers buying it, then that consumer will incur a loss of zt if he or she in fact does buy the brand. Similarly, if the consumer buys the nearest competing brand, then the consumer will incur a loss of | x — z\ t, where x is the distance between the target and the competing brands. If the consumer buys the target brand, then he must be at least indifferent between the brand that charges price p and a competing brand that charges price p. Thus, expense minimizing consumer behaviour implies that v + ak, - (p + tz) > v + ah) - [p + t(x - z)]. (3.1) Note that v + ak\ and v + all are reference prices of the target brand and the competitive brand respectively that a (target brand loyal) consumer expects to spend and the other two terms are expenses to the consumer. In addition, if "z is the value for which the above expression holds as an equality, then by re-arranging terms we obtain p + xt+ aki — ak{ ~ V z =  :— —. 2t Since consumer preferences are assumed to be uniformly distributed, all loyal consumers within the distance z of the target brand will buy the target brand. If N is the number of consumers located along the line segment A to C and competing brand prices are identical, then the target brand's aggregate demand will be 2Nz. Thus, if the proportion of consumers loyal to the target brand is denoted by 0, then total demand for consumers 59 loyal to the target brand (g^)t will be q 1 = 29Nz Simplifying, 2 m r , 7 i ~~^~\P + tx - p+ aki — ak\). NO q L = — {p + tx-p + ak - ak) (3.2) The demand function differs from that of Salop's in two important respects. First, the concept of brand loyalty is incorporated into the model, Salop's model does not consider this (it assumes $ = 1). Second, the effect of advertising intensity is incorporated through the notion of reference prices, Salop's model assumes ki = \ = 0. Note also that the demand function, (3.2). above implies, as one would expect, that demand is sloped downward with respect to a brand's own price (-^ - < 0) and is positively sloped to the competing brand's price (^- > 0). Furthermore, the intensity of a brand's own advertising has a positive influence on the target brand demand (-^ - > 0) while the influence from the competing brand's advertising is negative (-^ - < 0). It is possible to derive similar equations for those consumers loyal to the competitive brands. Thus, if the total demand for the target brand from the consumers loyal to the competitive brand is denoted by q G, the proportion of consumers loyal to the competitive brands is denoted by 9, while kc and kc denote marginal changes in the reference price due to advertising intensities, a and a respectively. Then a N9 . - . . , q = —£-{p + tx — p-\? akc - akc}. (3-3) Similarly, if the total demand for the target brand from the brand switching consumers is denoted by q s, the proportion of consumers that are brand switchers is denoted by 9 while t Note that the brand's aggregate demand is Nj\ dz = 2Nz. In this derivation N is a density of consumer preferences and z measures the area of the spatial map on either side of brand A. 60 ks and k3 denote marginal changes in the reference price due to advertising intensities, a and a. Then NO qS = —j-{p + tx - p + aks - ak3}. (3-4) We are now in a position to write the target brand's total demand (q). Thus, q=qL + qc + qs- (3.5) If c is the marginal cost of production and F is total fixed cost, then the target brand's total profit is 11 = pq- cq- N-g{r,f) - F. Substituting the value of q from expressions (3.5) and re-arranging terms, we have, TI = (p - c){qL + qC + qS) - N-g(r,f) - F. If the target brand is maximizing profit, then to satisfy the first order condition of maxi-mization with respect to price (p), advertising reach (r) and frequency (/), we must have dii' an an „ — = np = o, — = nr = o, and — = n, = o. op • or of Further, dq1 dqc dqs ATdg(r,f) nr = (P - c)(— + —.+ — - N— , and or or or or / KP R df df df ' df Using the first of these three equations, the optimal price (p) can be written as p = c {crL+qC+qS) d{qL+ qc+qs)/dp' Substituting values of partial derivatives, namely, dqL/dp, dqc/dp and dqs/dp in the above expression, substituting values of q1, qG and qs from expressions (3.2), (3.3) and (3.4) and 61 simplifying, we obtain, „ _ 1 - dp/dp p + xt + 6(aki - aki) + 9(akc - akc) + 9(ak3 - aks) p~ 2-dp/dPc+ ~~~ (2 - dp/dp)  (3- 6) In this expression, the term dp/dp indicates a reaction of the competitive brand to the target brand's price change and it is called a reaction function. For example, if two competitors react by changing prices in the same direction, dp/dp is a positive constant. On the other hand, if two competitors react and change their prices in opposite direction to each other, then dp/dp is a negative constant. Finally, if two competitors change their prices randomly (without consideration of competition), then dp/dp is considered to be zero. Furthermore, an intense competitive market implies that dp/dp is —1 and collusive market implies dp/dp is 1 (Capozza and Van Order 1978 and Bresnahan 1981). In the case of perfect competition, where there are a large number of competitors, Capozza and Van Order argued that the reaction function tends to zero (dp/dp = 0) with the cost of benefit loss zero (t = 0) as well. Further, since consumers in such a market treat each brand like a commodity, all consumers are brand switchers (9 = 0 and 9 = 0), and advertising does not alter reference price i.e. ki = ki = kc = kc — ks = kf = 0. Finally, since all competitors charge identical prices (i.e. p = p), substituting-these in (3.6) we find that the optimal price is equal to marginal cost. Thus, one of Capozza and Van Order's requirements for spatial competition is satisfied. When the industry structure is a monopoly, then there is only one brand in the market. In addition, since there is no competitor to react to, dp/dp is zero. Moreover, a is zero, the optimal price is p = p, and all consumers are loyal(0 = l ) . Substituting these values in (3.6) we obtain p = c+ xt+ aki. This means that if the industry structure is a monopoly, the optimal price will be some markup over marginal cost. Furthermore, the monopolist 62 would use advertising to raise a brand's optimal price above that of perfect competition. These two extreme cases show that our model can be reduced to both the monopoly or the perfect competition industry models. This provides additional assurance that the model is a reasonable one. A further examination of equation (3.6) reveals that the profit maximizing price is postively related to competitive brand prices (p), the marginal cost of production (c), the perceptual distance between two brands (x), and the trade-off cost associated with brand characteristics (t). Most critically for the present analysis, the effect of advertising intensity (a) on the optimal price (p) cannot be determined a priori. Several possibilities are explored below. 8.4.8. Effect of Advertising on Price To determine the direction of change in the optimal price as a function of a change in advertising intensity!, we totally differentiate the first order condition of maximization of price with respect to price and advertising intensity. This condition states that the marginal profit must equal zero, or f l p = 0. Denoting = npp and = jl p a, we may write nPprfp -1- npa(k = 0. Solving for dp we obtain dp da n pa PP Note that f In this discussion, a higher price implies lower price sensitivity and vice versa. 63 Furthermore, 2N Tin =—(dp/dp-1). To satisfy the second order condition of profit maximization, Ylpp must be less than zero. This is possible, if and only if, the value of the price reaction function (dp/dp) is less than one. This implies that at the equilibrium, the reaction of the competitive brand to the target brand's price change cannot be perfectly matched or overresponded to. This notion, however, is well recognized in the industrial economics literature by Bresnahan (1981) and used in a marketing application by Shugan and Jeuland (1983). One important implication of this analysis is that (1 - dp/dp) will always be positive. When dp/dp = 1, then n p p = 0 and the effect of advertising on price is infinite t. Although this intensive competitive behaviour is common in retail markets such as gasoline, the profit maximization condition excludes the possibility of such behaviour. Substituting values of ITpa and n p p in the expression for ^  we obtain, dp _ [9kt + 0kc + 9ks] - (da/da)[6kt + $kc + 0k3] la ~~ 2(1 - dp/dp) ^ Note that the second order condition for profit maximum discussed above implies that the right hand side denominator, 2(1 — dp/dp) will always be positive. This in turn implies that the sign of the right hand side expression depends upon the proportion of brand loyal consumer and the direction of reaction function for advertising intensity, (da/da). Thus, the change in the target brand price sensitivity as a result of a change in advertising intensity (our proxy for that is ^ ) , cannot be assigned a positive or negative sign without additional assumptions or information. Finally, the terms in the square parentheses in (3.7) indicate a weighted average (weighted by the proportion of brand loyal and brand t Since the effect of purchase price on the reference price is assumed to be linear with the coefficient of the price variable, —1, dp/dp = 1 will result in a degenerate solution. 64 switching consumers in the market) of the reference price change due to advertising. We will now discuss the necessary conditions for higher or lower brand prices as a result of increased advertising efforts. For clarity of exposition, let 4> be an index that measures the degree to which the target brand's advertising is imitated by the competitive brand. Formally, <f> is da/da. Thus, when <j> equals zero, it implies that the competitors do not react to the target brand's advertising action. Further, when (f> equals one, it implies that the competitive brand matches exactly the advertising actions of the target brand. Finally, when <p equals minus one, it implies that the competitive brand reacts exactly opposite to the target brand's advertising actions. The implications for each of these cases are now considered. First, consider the consequence of competitors not reacting to the target brand's advertising action. Formally, this is the case when <f> is zero. This requires setting da/da = 0 in expression (3.7) and results in dp = 6k{ + 9kc + 0ks da 2(1 - dp/dp). ' [ ' Note that all terms on the right hand side of (3.8) are positivef; hence we conclude that under these circumstances higher advertising by the target brand leads to higher prices. In other words, higher advertising levels without competitive reactions leads to lower consumer price sensitivity. This, of course, is an outcome of brand differentiation argued by Bain (1956), Comanor and Wilson (1974), and others. We will now investigate a case where the competitor mimics the target brand's action about advertising decision. More formally, we are considering the case when <j> is one. One implication of such competitive behaviour is that the advertising reaction of the competitor t The right hand side denominator 2(1 — dp/dp) will always be positive to satisfy the second order condition of profit maximum. 65 is unitary (i.e. da/da = 1). Substituting da/da = 1 into (3.7) we obtain, dp _ [9ki + 9kc + 9k3\ - \0k~i + 9kc + 9k3] Ta ~~ 2(1 - dp/dp) (3.9) It is still not possible to determine whether increased advertising levels will result in lower or higher price. There are three possibilities. Thus, In words, if the weighted average change in the reference price caused by target brand advertising is more than the weighted average of the reference price change resulting from the competitive brand advertising, then increased advertising results in higher price. On the other hand, if the weighted average change in the reference price resulting from the target brand advertising is less than the weighted average change in the reference price for the competitive brand advertising, then increased advertising results in lower price. This condition implies that consumer price sensitivity for the target brand will be higher if (1) the competitive brand imitates the advertising action of the target brand, and (2) the weighted average change in reference price of the competitive brand is more than the same change for the target brand. The next possibility to be discussed involves the case of the competitor reacting to the advertising decisions of the target brand with an opposite action. More formally, we are considering the case when <j> is minus one (i.e. da/da = — 1), Substituting da/da — —1 into (3.7) we obtain The right hand side of (3.10) is positive. This implies that the price of the target brand will go up, if the competitive brand reduces its advertising activity in response to increased advertising of the target brand. In other words, if the competitive brands react to the dp _ [9ki + 9kc + 9k3\ + [$k, + 9kc + 9k3] ~da ~~ 2(1 - dp/dp) (3.10) 66 advertising decisions of the target brand with an opposite action, then higher level of advertising by itifae target brand results in lower price sensitivity. It is sometimes argued that brand loyal consumers ignore competitive brand advertis-ing (Assael 1987) and that brand switchers ignore all advertising. In this case, it follows that advertising •should be directed towards brand loyal consumers so that these consumers become less price sensitive (presumably sales promotion would be used to attract brand switchers). We can investigate the validity of such an argument in the context of our model. This case implies that the marginal change in reference price for the target brand's loyal consumers due to competitive advertising will be zero. For consumers loyal to the target brand, competitive brand price is p while Vis is the marginal response of these con-sumers to competitive advertising. On the other hand, consumers loyal to the competitive brand ignore the target brand's advertising. This implies that kj = kc = 0. Furthermore, since brand switchers are assumed to ignore all forms of advertising, then ks = ks — 0. Substituting these values into (3.7) we obtain dp _ 9kt - 9kc(da/da) , . da ~ 2(1 -dp/dp) ' 1 J It is interesting to note that even when these restrictions are imposed we cannot determine whether higher advertising levels will lead to lower or higher price. This implies knowing the proportion of consumers loyal to brand (9) by itself, is not sufficient to predict the influence of advertising on consumer price sensitivity. One must know in addition, the reaction of competition to advertising. Expression (3.11) can be further analyzed to derive the implication of increased advertising intensity on brand price by using the approach adopted in expressions (3.8) to (3.10). For example, consider the case where the competitor mimics the target brand's advertising decision, i.e. da/da = 1. Substituting da/da = 1 67 into (3.11) we obtain dp 6ki — 9kc (3.12) da 2(1-dp/dp)' It is still not possible to determine whether increased advertising levels will result in lower or higher price. There are three possibilities. These are In words, consider two multiplicative variables. One for the target brand involving multi-plication of the proportion of consumers loyal to the target brand (6) and reference price change due to its increased advertising for these consumers (ki). The other variable for the competitive brand involving multiplication of the proportion of consumers loyal to the the value of the multiplicative variable for the target brand is more than the multiplicative variable for the competitive brand, then increased advertising results in higher price. On the other hand, if the value of the multiplicative variable for the target brand is less than the value of the multiplicative variable for the competitive brand, then increased adver-tising results in lower price. Thus, in deciding the directionality of the effect of changes in advertising on brand price, one needs to know the direction of the competitive reaction to the target brand's advertising and the magnitude of marginal changes in the reference prices for the both competitive and target brands. 8.5. SUMMARY and CONCLUSIONS In this chapter, by taking into consideration the behaviour of three groups of con-sumers, that is, consumers loyal to the target brand, consumers loyal to the competitive brand and brand switchers, a demand function for the target brand is derived. The derived demand function is such that the target brand's price and the advertising activities of the competitive brand (9) and and reference price change for this group of consumers (kc). If 68 competing brand have a negative influence on the target brand demand. Furthermore, the target brand demand is positively influenced by the target brand's advertising and the price of the competing brand. This demand function is then used to obtain the profit maximizing price of the target brand. It is found that the effect of advertising intensity on the profit maximizing price is dependent on the proportion of brand loyal consumers and the competing brand's reaction function for advertising intensity. More specifically, if two brands make advertising decisions independently of each other, then higher advertising levels will result in lower price sensitivity and higher prices. If, on the other hand, the competitive brands follow the advertising actions of the target brand and the weighted average change in reference price resulting from the target brand's advertising is less than the weighted average change in reference price caused by the competitive brand's advertising, then higher advertising levels will result in lower prices. It is also concluded that advertising directed towards brand loyal consumers may or may not result in lower price sensitivity. Thus, to decide the directionality of the effect of changes in advertising on consumer price sensitivity, the direction of competitive reaction to advertising must initially be determined. For an oligopolistic competitive market, the model described here provides a useful procedure to interpret whether an increase in advertising will result in an increase or a decrease in consumer price sensitivity. The procedure, however, requires an analysis and evaluation of the three conditions (listed below) with respect to reference price, advertising intensity, and brand loyalty measures. These conditions are listed below and are closely associated with discussion of brand loyalty, and brand competition. The procedure is outlined below. (1) Determine whether pricing and advertising activities of competing brands covary, and 69 if so, in what direction. (2) Determine the proportion of brand switchers in the population by defining the appro-priate measures of brand loyalty. (3) Determine the size of changes in reference prices attributable to advertising, and the weighted average (weighted by a measure of brand loyalty) change in reference price due to advertising. With this procedure, it is possible to predict whether or not there will be a higher or a lower consumer price sensitivity. In the following chapter, attention is given to understanding the database relevant to the testing of the propositions concerning consumer price sensitivity and this procedure is applied in chapter V (see section 5.3.4). 70 CHAPTER IV Analytical Methods, Hypotheses, and Data Sources. 4.1. Introduction : 72 4.2. Statistical Modelling Framework 74 4.2.1. Stochastic Process Choice Model 75 4.2.2. Random Utility Choice Model 76 4.2.3. Derivation of the Statistical Model . . . • 79 4.3. Variable Description and Model Hypotheses 84 4.3.1. Dependent Variables 84 4.3.2. Independent Variables 86 4.3.3. Model Hypotheses 92 4.4. Demographic Comparisons 95 4.5. Comparisons of Household and Store Level Sales Estimates 97 4.6. Data Management Issues 99 71 4.1. INTRODUCTION In chapter III, household level brand choice behaviour is used to develop a theoret-ical model. Development and analysis of this model shows how the effect of television advertising affects household price sensitivity. Specifically, increased advertising directly, results in a lower price sensitivity. On the other hand, if there is the imitative competitive reaction of advertising, then an increase in television advertising can increase household price sensitivity. The purpose of this chapter is to describe the statistical modelling framework that will be used to test whether the effect of television advertising is to increase or decrease household price sensitivity. The procedures used to derive the necessary variables for statistical modelling are also described in this chapter. Additionally, in this chapter an assessment is made to determine whether the sample data used in the statistical modelling is representative of a meaningful population. Before the chapter summary is presented, a brief description of the database used in this dissertation is provided. One database particularly suited to study the effect of television advertising on house-hold price sensitivity is a sample collected by Information Resources Inc. (IR1) for the city of Eau Claire, Wisconsin. The database is made available for this research by the Advertis-ing Research Foundation. This database was collected by IRI's single-source data system in which a sample of household television sets were monitored for advertising exposure. For the same time period, the corresponding households' purchases were recorded through a supermarket based scanner system. Additional details of the data collection procedure is reported in Eskin (1985) and Clarke (1987). The entire database contains five broad product categories: household wrap, dog food, detergent, spaghetti sauce, and cookies. For each product category, purchasing history from January 24, 1983 to September 17, 72 1984 (88 weeks) was recorded for more than 2,000 households. Television advertising ex-posure, however, was only recorded for a subset of about 700 housholds. In addition, total store sales data were collected for each of 88 weeks for all the brands in the five product categories. In section 4.2, the brand choice modelling framework is reviewed, with emphasis on McFadden's (1973) conditional logit model. In this discussion, it is noted that among the popular discrete choice models, the logit model is a better alternative for the purpose of this dissertation than the probit model. In addition, the logit model is also better suited than market share based models (Oum 1979, Carpenter and Lehmann 1985 and Russell and Bolton 1987) because they require either temporal aggregation, or household level aggregation or both, and hence, may lead to loss of information. In section 4.3, the dependent and independent variables used in the logit choice model are discussed. In particular, the procedures used to derive brand prices, television advertising measures, brand loyalty indicators, and sales promotional variables for brands purchased and competitive brands are discussed. This section concludes with a summary of expected algebraic signs (positive or negative) for all coefficients in the brand choice model. In section 4.4, the sample data are compared to the census data for the congressional district in which Eau Claire, Wisconsin is located. The comparisons suggest that the IRI sample when compared to census estimates contains fewer households with one member and more households with three and four members. The proportion of sample house-holds earning more than $25,000 per year is also substantially overrepresented in the IRI sample, as is the number of households living in a single family detached owner occupied homes. The sample also overrepresents the number of households headed by members with 73 managerial and professional occupations. Finally, the sample household heads have more formal education than those in the census data. These differences between the I R I sam-ple and the census data may also reflect characteristics of a typical shopper at suburban supermarkets. Al though it is evident that the I R I sample is overrepresenting the "upscale" household, it is sti l l possible that the I R I sample households may have purchased the same brands that the total populat ion purchased. In section 4.5, brand purchases for the two products analyzed (dog food and household wrap) are compared to total brand sales obtained from store level sales data. The comparison suggests that estimates from the two sources are very close. T h i s is an assurance that the IRI sample household purchases are reasonable in representing the total sales for most brands in these product categories, at least among the stores monitored. In section 4.6, two sample composition issues are discussed. The first issue concerns a household making multiple purchases on the same purchase occasion. The other issue is concerned wi th households making infrequent purchases. These issues are discussed with reference to a large size scanner panel database. 4.2. STATISTICAL MODELLING FRAMEWORK The purpose of this section is to describe the statistical modelling framework that w i l l be used to test whether an increase in television advertising intensity leads to an increase, or a decrease, in household price sensitivity. To this end, the choice of modelling framework is described by examining two alternative modelling approaches. T h e n , the selected approach is described in detail. For testing hypotheses concerning the effect of television advertising on household price 74 sensitivity, several alternative statistical modelling frameworks are examined according to three criteria. First, the statistical model should be consistent with the economic model reported in chapter III. Second, the statistical model must be able to handle several brand level variables such as price, advertising exposure, and deal amount, in addition to several household level variables such as household income and household size. Finally, a practical consideration requires that the statistical model be easy to estimate using a typical mainframe computer. Two commonly used approaches for modelling consumer choice behaviour are stochastic process choice models and random utility choice models. Both are reviewed below although only one, the random utility choice model, fits the requirements. 4-2.1. Stochastic Process Choice Model The main idea behind the stochastic brand choice process models (Massy, Montgomery and Morrison 1970) is to relate successive brand choices of households. These choices are then modelled according to their order, heterogeneity, and stationarity. If the successive brand choices are independent, the brand choice model is a zero order model. If the successive brand choices are dependent, the brand choice model is a higher order model. Similarly, if the model parameter(s) vary across households, then the brand choice is heterogeneous. Finally, if the model parameter(s) remain the same over time, the brand choice model is stationary. A number of stochastic choice models reported in the literature are based completely on a matrix consisting of successive choice behaviour aggregated over a sample of households and/or over time. These models seldom include managerially controlled explanatory variables such as brand prices, television advertising, etc. Several attempts to include 75 explanatory variables have proved algebraically complex and computationally difficult (Telser 1962, Kuehn and Rohloff 1967, Lilien 1974, Blattberg and Jeuland 1981 and Jones and Zufryden 1982). Another difficulty with these models is that a number of different model specifications may be supported equally well by a single set of data (Givon and Horsky 1979). While these are important weaknesses for the purpose of this dissertation, the main idea that two successive brand choices may be related is important, and is also well known in the literature. This idea is pursued in more detail in a future section where the independent variables are discussed. 4-2.2. Random Utility Choice Model The random utility choice models (McFadden 1973) are concerned with explaining a household's brand choices given information (a vector of attributes) about the chosen and competitive brands, and about the household. One specific functional form of the random utility choice model that has received considerable attention in the marketing literature is the multinomial conditional logit choice model (Gensch and Recker 1979; Jones and Zufryden 1980; Arnold, Oum and Tigert 1983; Guadagni and Little 1983; Neslin, Henderson and Quelch 1985; and Tellis 1988) and the purpose of this subsection is to describe the main features of this model. In addition, several advantages of using the logit choice model to test the effect of television advertising exposures on household price sensitivity are summarized. Before a full description of the model is given, an intuitive description of the model is given below. Consider a household facing a choice among several brands at a given purchase occa-sion. Information on brand characteristics such as price, deal amount, television advertising exposure, and household characteristics such as income, and family size, for each of these 76 brands is available. Using this information, one would like to predict the household's brand choice. McFadden (1973), faced with this situation, decomposed the household's utility into two components: one determined by brand characteristics and the other a stochastic or random component. The utility of a brand to the household was assumed to be a lin-ear function of brand characteristics, and it was assumed that households maximize their utility at each purchase occasion. The random component of utility was assumed to be distributed independent of brand choice and identical across households according to the type I extreme-value distribution (Johnson and Kotz 1970). (This distribution is used by McFadden not because of its intuitive appeal but because it offers computational conve-nience and does appear reasonable). The model derived, using these assumptions, predicts the probability that a household will choose a brand, given the characteristics of the chosen brand, the characteristics of competing brands, and the household's characteristics. There are several advantages to using McFadden's utility choice model for this work. First, the hypothesis that the effect of an increase in television advertising exposures results in a decrease in household price sensitivity, can be tested at the household level. Since no data aggregation is needed to test the hypothesis, this model provides a direct test. Operationally, a multiplicative variable of price times television advertising exposures along with a price variable is needed in the logit model (see next section for more detail). In such a model, the price variable coefficient can be used to measure household price sensitivity. In addition, the coefficient of the multiplicative variable can be used to test the hypothesis concerning the effect of telelvision advertising on household price sensitivity". Thus, if the sign of the multiplicative variable is positive, and statistically significant it will be taken as an indication that an increase in television advertising exposures is expected to result in a f The linkage between the theoretical model and the statistical model is provided in section 4.3.3. 77 lower price sensitivity. Similarly, a negative coefficient of the multiplicative variable is an indication (see elasticity expression in subsection 5.3.5) that increased advertising results in a higher price sensitivity. Second, McFadden's logit model is based on individual household brand choice be-haviour. Thus, the logit model that is used in this dissertation for hypothesis testing purposes, follows a similar structure of assumptions (for households) to the one followed in chapter III for the theoretical model. This provides a opportunity to maintain consistency between the theoretical model and the hypothesis testing model. Finally, Oum (1979) reports that the market share based models, especially those involving the ratio of market shares, impose stronger restrictions on the household utility function than McFadden's logit model. Third, although McFadden's model is nonlinear in several unknown parameters (see below), the model is computationally easy to estimate. There are good reasons for this. First, the likelihood function that must be maximized to estimate parameters is strictly concave (Chow 1983 and Maddala 1986). Thus, any properly written iterative search procedure, especially those that use second derivatives of the likelihood function to estimate parameters, will likely converge in a finite number of iterations (usually less than 10). This has facilited development of computer programs to estimate parameters of the logit model using maximum likelihood procedures for both the mainframe and micro computers (Amemiya 1981). Lastly, Maddala (1986) and Amemiya (1985) report that it is doubtful whether the probit model is worth the associated computational troubles when the number of choices is greater than four. In our case, the product categories such as dog food and detergent have more than 20 brands in the market. Thus, it appears that a model that restricts 78 computational feasibility when over four brands are present is not useful for the purposes of this dissertation. In this subsection, the advantages of McFadden's logit choice model to test hypotheses concerning the effect of television advertising on household price sensitivity have been discussed. A unique and desirable feature of the logit model is that it maintains consistency between the theoretical model and the hypothesis testing model. Finally, the logit model is computationally easy to estimate. In the next subsection, a detailed description and derivation of the logit model is given. 4-2.8. Derivation of the Statistical Model In this subsection, several similarities and differences between the basic equation used in the logit model is compared to the basic equation used in chapter III for households. Then, the distributional assumptions are used to derive the logit choice model. McFadden (1973) assumed the utility of brand i to be a function of the attributes of the available choices. He also assumed that a household makes choices that maximize its perceived utility. While the true utility u,- cannot be observed, McFadden approximated it by V{ (a function of the brand attributes) and a random component e,. For a given household, McFadden decomposed u,- into u,- + e,-. If a household chooses brand i, then we must have v,• + e,- > Vj + ej for all j ' ^ i or (4.1) e, + v, — Vj > €j for all j ^ i In comparison in the theoretical model of chapter III, we considered there a household that has the base reference price v which is increased by the advertising intensity of the target brand (a) and the competing brand's advertising efforts (a). Further, the target brand and the competing brand charge prices of p and p respectively. In addition, the spatial 79 distance between the household's most preferred product benefit and the target brand is z and the distance between the target brand and competing brand is x. Finally, t is a per unit cost associated with benefit loss. Thus, the theoretical choice model is based on the expression v + ak - (p + tz) > v + al - (p + t[x - z\), (3.1) where k and I are the marginal changes in the. base reference price as a result of advertising efforts of the target brand and the competitive brand respectively. We comment on similarities between equation (4.1) and equation (3.1) below. In chapter III, v was considered to be the consumer's reference price for a product category. Here, v,- is the utility of brand i. Both equations assume that a higher level of product benefits are preferred to a lower level. Furthermore, both equations implicitly assume that brand attributes, such as price, and also the number of advertising exposures linearly affect a brand's utility or reference price. In microeconomic terms, utility v, is considered an indirect utility function, since brand prices and advertising exposures and the household's income may be a subset of its determinants. In chapter III it is noted that v is the expected expense that a consumer is willing to incur to purchase the brand. There are, however, two major differences between equation (4.1) and equation (3.1). First, in this chapter utility (vi) is decomposed into its brand-related and household-related components whereas in chapter III, the reference price (v) is not decomposed. Specifically, Vi in expression (4.1) is equivalent to v + ak — (p + tz) in expression (3.1). Similarly, Vj in expression (4.1) is equivalent to v + al — (p + t[x— z\) in expression (3.1). Second, since there may be some determinants of choice (related to brand or household) unknown to us or otherwise unavailable for estimation purposes, equation (4.1) contains a random component (e,). In other words, if we are successful in obtaining a "perfect fit" between 80 the data and the estimated model, variation in the random component will be zero and equation (4.1) will be reduced to equation (3.1). In this sense, this model is consistent with the theoretical model reported in chapter III. Now, the observed dependent variable is denned as t/, = 1, if the household purchased brand i, otherwise t/,- = 0. Thus, probabilistically we may rewrite (4.1) as prob(?/i = 1) = prob(e, + - Vj > e;) for all . j i. (4.2) Note that Vj and are the components of utility that are determined by both brand characteristics and household characteristics. Hence, the prob(y,- = 1) depends upon the probability distribution function (PDF) of e, and. the cumulative distribution function (CDF) of each of £j. More specifically, each e;- must be greater than e,• + w, — Vj. This can be expressed mathematically as prob(</, = 1) = / f(€i)d€i / f(e1)de1 • / f(e2)de2 •••• f{en)den , J — o o ' ^ — o o J — o o ^ — o o or /o o (4-3) where F(-) and /(•) are the CDF and PDF mentioned above. At this point, it is necessary to make some distributional assumptions about the ran-dom variables e\, e2, • • • ,en. McFadden (1973) and others have found that a computationally convenient distribution is type I extreme-value (Johnson and Kotz 1970). Specifically, the random variables ei, e2, • • • , en a r e independently and identically distributed with a CDF of F(ei < e) = exp(-exp(-e)) and their PDF is /(e,-) = exp(-e,- - exp(-e,-)). Chow (1983) has remarked that compared to a normal distribution having a mean of 0.5 and variance of one, the type I extreme-value distribution with mean zero is skewed to the right and has a longer right tail than the normal distribution. 81 Continuing with the development, we may substitute values of the CDF and PDF of the type I extreme value distribution into (4.3), and note that JJ F((i + Vi - Vj)f(ei) - JJ exp -exp(-e,- - v{ + v;)exp(-e,- - exp{-e,}) Now if we write exp -e,-- exp(-e,-(l + ^ exp(vy) exp(t;,-) )) 9i = log exp(vy) ' exp(v.) exp(v,). then substituting this term into expression (4.3) as well, we obtain /oo exp (-e,- - exp(-e, + &))de,. -oo Now substituting e,- = e, - and de, = de,- we obtain prob(y, = 1) = / e x p ( - e , - e x p ( - e , ) ) d e , J-oo /oo exp(-e,- - exp(-e,))Se,. Note that the term under the integral sign sums up to one. Finally, substituting back for the value of gi and simplifying we obtain prob(j/, = 1) = exp(v,-) E"=i exp(v,-) (4.4) Now, we may decompose household utility (v,) into brand-related *nd household-related characteristics. The approach, used by McFadden, is to express the utility of a brand as a linear function of brand and household characteristics. If a household faces n brands at a given purchase occasion, then each brand's characteristics such as price, television advertising exposures, etc. are represented by a vector X, where there are K 82 different brand characteristics. The household's characteristics such as income, family size are represented by a vector Z. Each household can be summarized by M different characteristics. Household h's utility for brand i may be denoted by i^ f is decomposed as K M k=l m=l Note that the components of vectors and am are unknown parameters to be estimated from the data. Now, substituting values of t% and Vjh into expression (4.4), we may write e x P Efc= 1 Xkihfik + Em=l Zmh<*mi pvob(yih = 1) = — - r — — — — T (4.5) p £ J = 1 e x p This is a functional form associated with the conditional logit choice model. Note that the model is non-linear in the unknown parameters (3^, and ami. The parameters reflect the marginal utility an average household associates with the product characteristics k and these parameters are invariant across brands. Furthermore, we may also write (4.5) as prob(ytA = 1) = 1 (4.6) E"=iexp Ysk=\{Xkjh - Xkih)Pk + J2m=lZmh(amJ ~ ami) Note that (X^jh — Xfah) is a difference between characteristics of brand purchased and j01 competitive brand. In other words, all the variables in this model are treated on the relative basis and multiplying brand characteristics values by a constant (for example, an inflationary trend) does not affect the predicted probabilities. Parameters a,m, on the other hand, reflect the marginal utility that an average house-hold associates with brand i for the characteristic. These parameters vary across f For a given household, it is the brand characteristic which determines a household's utility for a brand. Since brand choices may be influenced by the household characteristics, this simple approach is used to account for household variation within the multinomial logit model structure (Maddala 1986). 83 brands. In addition, parameters of variable Zmh appear in the form of difference ( a m ; - a m t ) . For n brands, thus, one can only estimate ra — 1 differences uniquely. As a result, one of the brand's amj is set equal to zero. This assumption is similar to the one used in regression analysis with a categorical independent variable, where the regression coefficient of one of the categories is set equal to zero. To estimate parameters 3k, product characteristic information for the brand chosen (Xjah) as well as information for the competitive brands {Xkjh,i ^ j) is needed. This information is provided by the IRI database. To estimate a j m , household characteristic information is needed. In other words, to estimate n — 1 parameters (a j m) one variable is used. Thus, to estimate a ; m one may require fewer variables than 3k, but the number of a,m parameters estimated in the model is directly proportional to the number brands in the market. Thus, there is a trade-off between the number of input variables and the number of parameters estimated for this model. To conclude, in this section starting from assumptions about household brand choice, a conditional logit choice model is developed and described. In the next section, attention will focus on the operational definition of the dependent variables (y,-) and independent variables (Xkuij for the model. 4.8. VARIABLE DESCRIPTION and MODEL HYPOTHESES. Data for the product categories of aluminum foil and dry dog food are used to estimate the effect of television advertising on household price sensitivity. These two product categories are chosen because they exhibit marked differences in the number of brands in the market (three for aluminum foil and 39 for dry dog food) and the mean interpurchase time (once every 20 weeks for foil and once every four weeks for dry dog food). Further, 84 the number of brands advertised on television was one for aluminum foil and 18 for dry dog food. These quite different markets provide a test for the robustness of the estimated models and pose an interesting dilemma for the theoretical model. 4-8.1. Dependent Variables The purpose of this section is to describe procedures used in deriving dependent variables in the sample for the product categories dry dog food and aluminum foil. Since the unit of analysis in this dissertation is brand choice by individual household, the dependent variable, brand choice, is a categorical one. The dog food product market consists of three sub-markets: canned, dry, and semi-moist. Of these three sub-markets, the dry dog food accounts for about 60% of total consumer expenditure in this product category. In addition, of the 39 unique brands in the market, 18 brands were advertised on television with large differences in advertising intensity. Moreover, there were 186 unique brand-size combinations available in the market. Finally, only one brand had a market share greater than 10%. Since it is difficult to estimate and interpret a choice model with 186 brand-sizes, it was decided to combine minor brands into aggregate brand categories (Tellis 1988). The rationale for this decision and the procedure used to combine brands is outlined below. It was decided to concentrate on brand choice behaviour and not consider size (or brand-size) choice behaviour for four reasons. First, television brand advertising focuses on brand benefits and does not address package size. Thus, the effect of advertising appears to be directed toward influencing brand choice and not size. Second, since the theoretical model in chapter III is concerned with household brand choices, the statistical model must be consistent with this. Third, a household may decide on the package size based on its 85 consumption rate and inventory costs. Thus, package-size decision may not be a purchase-to-purchase decision for many households. Finally, for the IRI sample households, more than 70% of total purchases were for a package size of five pounds. Ignoring the package size designation still resulted in 39 available choice alternatives. To reduce the number of alternatives even further, brands were grouped into aggregate categories (see Tellis 1988). This was done as follows: if a brand had more than a 5% market share, then the brand was kept by itself. In addition, brands with market shares smaller than 5% were grouped according to manufacturer, with distinctions made between those brands that were advertised and those that were not. With these simplication rules, it was possible to reduce the number of alternatives to 11 and of these, 7 were advertised on television (see summary Table 4.1). A similar procedure was applied for aluminum foil, which belongs to the broader product category of household wrap. The aluminum foil market (in the test city) consists of three brands, Reynolds, private brands, and generic brands. Only the Reynolds brand is advertised on television. In the next subsection discussion is concerned about independent variables used in the model. 4-3.2. Independent Variables In this subsection, the procedures used to derive the independent variables for the model are described. In operationalizing these variables, procedures used by Guadagni and Little (1983) are maintained. This will facilitate comparison of results in this dissertation to the results reported in the literature. For variables of television advertising and brand loyalty, several alternative operational definitions are explored (see Table 4.2). In the next chapter, the effect of alternative definitions on the model parameters and its implications 86 on the effect of advertising on household's price sensitivity is evaluated. Description of these variables is given below. Price: This is the regular unit price of the product purchased. It is obtained by adding the deal amount and/or the coupon value to the actual purchase amount for the brand and then dividing that sum by the purchase volume. There are three reasons for using this definition of regular price. First the regular price is a managerially relevant variable and its implication should be long term in nature. Second, the consumer may well be focusing on an expectation of regular price before the purchase is made. Finally, others like Guadagni and Little have also used this definition of regular price. This will facilitate the comparison of our results with theirs. Competitive Brand Prices: These are also regular unit prices and were obtained by analyzing all purchase observations for a product class. Tables of regular unit prices were constructed for each of the brand-size combinations, for each of 10 stores, and for each of 88 weeks. However, during some weeks multiple purchases for some brand-size combinations did occur. In such instances, a simple average of those multiple purchases was used. Also during some weeks for some brand-size combinations, there were no purchases. In such instances, competing unit brand prices were estimated using either prior weeks of purchases or purchases for the brand in the other nine stores in the market. It was observed for the two product categories that within a given week, brand prices varied as much as 20% across stores. As a result, data from prior weeks were used for the calculation. Thus, in instances when there were no purchases in a store, the simple average of TO prior weeks of regular prices were used. This approach is similar to one used by Krishnamurthi and Raj (1988). Such imputations, however, were used for about 15% of purchase observation. Consumer Sales Promotional Variables : The IRI database contains four consumer sales 87 promotional variables: presence of store display (1,0), presence of feature in a newspaper or flyer (1,0), the value of a deal, and coupon value. A frequency count of these four variables for four product categories revealed that the occurrence of display and feature was infrequent (see Table 4.2). Since the display is an in-store promotional activity and feature activity is outside of the store, these two variables were kept separate. This, however, creates a problem for the aluminum foil product category, since only one brand is featured, and feature and display activities for that brand are perfectly correlated. As a result, it is not possible to compare the effect of these two variables for both product categories. Coupon and price-deal both lower a regular price but consumer action necessary to claim the price saving is different. Thus, to receive price saving associated with-coupons, the consumer is required to present a coupon. To receive a price saving associated with a price-deal, on the other hand, the consumer is only required to purchase the brand on price-deal during the deal period. While the IRI database contains information on whether a purchase was made using a coupon along with the face value of coupon, the database does not contain any informa-tion about the availability of competitive brand coupons. An attempt was made to infer competitive couponing activity, an approach similar to the one used to obtain the compet-itive brand prices for the product category dry dog food, but results were inaccurate. In addition, it is not possible to know which consumers actually received coupons and then decided not to use them. For these reasons, couponing activity in this research was ignored (see Neslin and Shoemaker 1983 for an analysis of of coupon promotions). To measure consumer responses to price-deals, two variables were created: the first variable indicated the presence(l) or absence(O) of a deal and the other variable measured 88 the amount of a deal's price saving. Television Advertising: The IRI database contains both the time of the day and the date that a given household was exposed to a television commercial. The database also contains information as to whether or not a household changed television channels during a commercial. Furthermore, each commercial exposure sequence is broken down into contiguous five second intervals. A household television advertising exposure is counted if the television set was tuned to a commercial for more than half of its total duration. For the dry dog food product category, 94.1% of television commercials were of 30 second durations, whereas all the commercials in the aluminum foil product category were of this durations. Thus, in general, if a household was exposed to a television commercial for more than 15 seconds, then the exposure was counted. The television advertising exposure variable for a brand was measured by counting the number of television advertisements that a household was exposed to in an interval before a purchase. Three alternative specifications of the advertising exposure interval were tested. In the first case, the time interval was defined as the number of days between two successive purchases. A problem with this measure is that the purchase interval varies across consumers. For example, a household making the brand choice every two months may be exposed to fifty television brand advertising exposures, whereas another household making the brand choice every two weeks may be exposed to five. If household brand choices are primarily sensitive to recent advertising exposures, then variability of purchase timing may increase the possibility of obtaining an insignificant parameter estimate for the television advertising variable. An alternative to test is whether the model fit would differ if a fixed advertising exposure interval was specified. Thus, in the next two cases the time intervals were fixed four weeks and eight weeks before a purchase. Using the fixed 89 ' duration has its own problems. For example, for a frequent purchaser a set of exposures may contribute to several purchases. Brand Loyalty: In the past, most studies have focused on some form of brand share as a measure of brand loyalty (see Guadagni and Little 1983 and Tellis 1988). For example, Guadagni and Little used an exponential smoothing of past purchase share for each consumer as a measure of brand loyalty. Thus, if a brand was purchased on a previous occasion, then the brand loyalty measure for the next purchase would be equal to one minus the smoothing constant. Guadagni and Little found that the smoothing constant of 0.875 best fitted the regular coffee data. Subsequent work (Ortmeyer 1985, Tellis 1988 and Gupta 1987), however, has demonstrated that varying the smoothing constant by ±0.1 does not affect model fit results. Given the simplicity and success of Guadagni and Little's measure, it is used as a comparison to the measures developed here. In deriving a measure for brand loyalty, insights from prior work on stochastic choice models are used. The measure of brand loyalty is defined to vary over competing brands as well as across households. For example, a popular brand is likely to have greater brand loyalty than a private label brand. Similarly, a consumer who purchases a product infrequently is likely to have lower brand loyalty than a frequent purchaser (Morrison 1966). Thus, the measure of brand loyalty proposed here weighs brand share inversely and exponentially with the number of days since a purchase occurred. Further, the functional form of the proposed procedure assumes that as the last purchase event becomes "older", the probability of repeating a brand purchase decays exponentially. It is also assumed that the decay constant (A) depends upon an individual household, i.e., it need not be constant across the sample of households. Finally, the weights assigned to past events may assume any convenient functional form, including linear, log-linear, etc. Thus, the probability p(x) 90 that the household will repeat the purchasing event, is p(x) = Aexp — A/(x) where A is the decay constant, and x is the number of days since the last purchase of the same brand. Thus, this measure of brand loyalty is a flexible exponential weighting of past choices. This measure is a general form which can be used to derive special cases. For example, it can be shown that when f(x) = x; then p(x) = Aexp(-Ax), the standard exponential decay curve. Further, it can be shown that for f(x) = log(x) and A = 1, p(x) = 1/x, which is the harmonic decay curve. Since p(x) is a random variable and A is a parameter to be estimated, then if m purchase observations are available, the maximum likelihood estimate of A is A = ^ mm , / — y . For the standard exponential decay curve, A is the probability of repeating the same brand purchase if another purchase is made on the same day (if x = 0, then p(x) = A). Similarly, if f(x) — log(x), then A is the probability of the repeating the same purchase on the next day. Since interpurchase time (x) is not normally distributed (Lawrence 1980 and Herniter 1971), a simple approach to transform this variable into one that is normally distributed with a constant variance is the logarithmic transformation (Box and Cox 1964). Hence, f(x) = log(x) is used in the present work. Here is a simple numerical illustration. Suppose the brand was bought 10, 40, and 100 days ago. Then, if f{x) = log(i), X is log(lo)+iog(2o)+iog(ioo) = ° " 3 0 2 9 -The first 32 weeks of data were used to make an initial estimate of A. Values of A were updated at each of the subsequent purchase occasions for a given household. For n brands and for each household, n values of A were estimated! Furthermore, these estimates are obtained independently. This may result in the measure of brand loyalty across all brands f If two households had purchased a brand on the same days, then estimated values of A for both household will be the same, even though two may made different number of purchases. 91 not adding up to one. To ensure that measures of brand loyalty ranged between zero and one, all cumulative probabilities were normalized to add up to one. Although it appears logical to update values of A after each purchase, recent evidence by Tellis (1988) suggests that updated measures may only provide slightly higher explana-tory power. On the other hand, if successive brand choices are dependent, then one would expect the updating procedure to provide a higher explanatory power. To test this hy-pothesis, two alternative measures of brand loyalty were used; one which was updated after each purchase occasion and the other based on the first 32 weeks of data. To summarize, three alternative measures of brand loyalty will be used. The first measure is based on Guadagni and Little's work and is an exponentially smoothing of past choices. The next two measures are based on flexible exponentially weighting past brand share as described above. In chapter V these measures will be analyzed to find out which of these is a "better" measure for the present data. Modelling Television Advertising and Price Interaction: To test the effect of television advertising on household's price sensitivity, a multiplicative variable of regular price and television advertising exposures will be used. Thus, if the sign of the multiplicative variable is positive (negative), statistically significant and different from zero, then it will be taken as an indication that an increase in television advertising exposures is expected to result in a lower (higher) price sensitivity. Several comments on the sign of coefficient for this variable and its link with the theoretical model is given in the next subsection. 4-S.S. Model Hypotheses The purpose of this subsection is provide a link between the theoretical model and the statistical model. In addition, in this subsection algebraic signs of coefficients for the 92 independent variables are discussed. The theoretical model reported in chapter III provides the conditions under which increased advertising results in a lower price sensitivity. Thus, in empirical work, if the price times television advertising variable has a positive coefficient, then this evidence will confirm the theoretical modelf. On the other hand, if the price times television advertising variable has a negative coefficient, then this evidence will partially disconfirm the theoretical model. The reasons for the latter conclusion is outlined below. The model proposed in chapter III generalizes Salop's (1979) work on brand differenti-ation by including the effect of advertising on brand choices. It is argued that the effect of advertising is to differentiate among the competing brands such that increased advertising results in a lower price sensitivity for an individual choice. At the market level, however, one may observe that increased advertising results in a higher or lower price sensitivity depending upon the reaction of competitors to advertising. In other words, the market level data may include instaneous adjustments of competitors to the actions of the "target" brand. The statistical models use household level brand choices to evaluate the effect of tele-vision advertising on household price sensitivity. If the price times television advertising variable has a negative coefficient, then one must attribute (according to the theoretical model in the chapter III) the sign of this coefficient to the competitive reaction to the advertising. The competitive reaction of advertising is based on the market level measure-ment. The theoretical model presented in the chapter III is concerned with the effect of t The relation between coefficient of the price times television advertising variable and price sensitivity change due to advertising is given in subsection 5.3.5 93 advertising on changes in brand prices. The statistical model, however, is structured to directly measure the effect of advertising on consumer price sensitivity (elasticity) with respect to brand purchase probability. This distinction is illustrated below. Let rji and r)2 be price elasticities when brand prices are p\ and p2 respectively. In addition, suppose price pi is less than price p2 and a higher advertising level is associated with price p\. In other words, a higher advertising level is associated with a lower brand price. One may then ask whether or not the above condition of a price difference is the sufficient condition to conclude that a higher advertising level results in higher price sensitivity. To answer this question one must examine the algebraic expression for price elasticity for the logit model. . As described in the chapter V (see section 5.3.5), since the logit model is probabilistic, estimated price elasticities at prices p\ and p2 will be n\ =  dPg°^ 1 • p rob 2 and rj2 =  9Pgp^ 2 " p rob 2 • In other words, one may not conclude that \r)i \ is more than l ^ l only based on prices p\ and pi- In conclusion, the logit model proposed here is used to test the effect of advertising on price sensitivity. It is now possible to summarize the expected algebraic signs of coefficients for the variables in the logit choice model. Table 4.3 is a summary of the alternative variable definitions discussed so far and presents the expected signs that coefficient of these variables will have when the brand choice model is estimated. First, higher brand price is perceived by a household less favourably with the result that the household is less likely to buy the brand. Thus, one would expect the price variable coefficient to be negative. Second, if a household is exposed to more television brand advertisements, greater is the likelihood that, the household will buy the brand. Hence, the television advertising variable parameter is expected to be positive. Third, in-store display and newspaper features affect brand choice positively. Fourth, one would expect that the higher the value of the deal, the 94 higher should be the utility of a brand. In addition, since the presence of a price discount improves (or does not decrease) a brand's utility, one would expect both the deal variables to be non-negative. Fifth, a higher level of brand loyalty is expected to increase the likelihood of repeating brand purchase. Thus, the brand loyalty variable will have a positive coefficient. Finally, the sign of the television advertising variable multiplied by the regular price cannot be determined a priori without making additional assumptions. The sign of this variable will indicate whether the television advertising increases or decreases household price sensitivity. This, of course, is the central focus of this research. 4.4. DEMOGRAPHIC COMPARISONS The target population supplying data for this study is households in the congressional district of Eau Claire, Wisconsin, who were buying frequently purchased, low priced, and packaged consumer goods through supermarkets. The purpose of this section is to ascertain the representativeness of the IRI sample to the target population in terms of several demographic variables such as family size, family income, etc. The target population estimates are obtained from Census of Population and Housing (Congressional District Data Book, 1984). In general terms the IRI data collection procedure is based on non-probability sampling. Census data provides the target population estimates for family size, family income, home ownership, type of home occupied, occupation, education, and race for the congres-sional district of Eau Claire and the state of Wisconsin. In Table 4.4 to 4.10, a comparison of all of these variables with the IRI sample estimates is provided. Table 4.4 reports the comparison between the IRI sample and the census estimates for household size. As can be seen in Table 4.4, the IRI sample households had 2.82 members per household while the 95 census reported 2.77 members per household in the congressional district. The differences between the two sample estimates are statistically not significant at p < 0.05. The pro-portion estimates reported in Table 4.4, are systematically different in several categories, however. More specifically, the IRI sample tended to have fewer one-member households and more three- and four-member households than the census estimates. The %2 statistic computed to test whether these differences are statistically significant revealed that the X2 test statistic is 140.61 with five degrees of freedom. Since the critical value of the x 2 is 11.07 at p < 0.05, it is concluded that the differences are significant. Thus, the IRI sample overrepresents three- and four-member households. In Table 4.5, a comparison between the estimated proportions for several categories of household income in the IRI sample and the census estimate suggests that the IRI sample underrepresents households with low income and overrepresents households with high income. The percentage of households earning income higher than $25,000 per year is 21.4% in the congressional district while the IRI sample percentage is 44.8% of households. These differences indicate that the IRI sample may be overstating household income by a substantial amount. Given that family income is overstated in the IRI sample, one would expect there would be similar differences between the IRI sample and the population estimates on factors that covary with family income. Thus, Tables 4.6 and 4.7 reveal that the IRI sample overrepresents families living in a single family, detached owner-occupied home. More specifically, the IRI sample estimates that about 83% of households live in a single family detached homes while the census estimate for the congressional district is about 75% and 69% for the state of Wisconsin. Given that family income covaries with occupation and education (Table 4.5 and 4.6), it is found that the IRI sample also tended to overrepresent 96 households whose members are employed in managerial and professional occupations. In addition the IRI households are headed by members who had more education than those in the population'. Finally, Table 4.7 reveals that the IRI sample slightly overrepresents households headed by racially white member. All of these differences are statistically significant at p < 0.05. In brief, it appears as though "upscale" households are overrepresented in the sample. No hypothesis can be made at present as to whether these households are unusual in terms of their response to advertising or price changes. The approach used to control for these unrepresentative characteristics of the sample is to include variables like household income and household size in the statistical models that will be used to test the hypotheses being considered. This is similar to the approach suggested by Hensher and Johnson (1981). 4.5. COMPARISONS of HOUSEHOLD and STORE LEVEL SALES ESTIMATES In the preceding section, comparisons between several demographic variables revealed that the IRI sample overrepresents higher income and larger households when compared to the census estimates. However, the IRI sample households purchase the same brands as the overall population. To find out whether or not the IRI sample households purchased the same brands would be to compare the estimate for total brand sales across all stores to the IRI sample sales estimates. Since the database provides total sales for each brand for each week of purchase history for most of the stores in Eau Claire, Wisconsin market, such a comparison is possible. The store level data, however, is recorded for each brand by aggregating across several sub-product categories and this causes s o r t i e problems in the comparisons. For example, in the dog food product category, Alpo is sold as both a dry and a canned food product. In the store level data, however, w e cannot distinguish 97 between these two product categories. In Tables 4.11 and 4.12, the brand sales comparisons for the broad product categories of household wrap and dog food are reported. To facilitate comparison, all brand sales are converted to market share estimates. Three different estimates of market share are obtained based on (1) total number of packages sold, (2) total volume sold, and (3) total dollar amount spent by households. Examination of Table 4.11 indicates that for the household wrap product category, most of the estimates were within one percent of each other. For the dog food product category, however, estimates are not that close for several brands (Table 4.12). For example, the estimate based on total dollar sales from the household level sample shows that the brand Dog Chow is the market leader, while based on the store level data, Tuffys is the market leadert. While there are several differences between these two estimates of brand sales, there is a strong positive association between the sample estimates and the store level estimates. A measure of association computed between the sample estimates and the store level estimates for the dog food product category, found correlations of 0.963, 0.932, and 0.896 for market shares based on units sold, volume sold, and dollar amount respectively. This would indicate that the observed differences between the estimates of share may be random and the IRI household sample is representative of the total sales through the supermarkets. Table 4.12 also reveals an inconsistent pattern of market share estimates based on f It is believed that the store level may contain some inaccuracies. According to the industry estimates Dog Chow, Alpo, and Puppy Chow were among leading brands in 1983 to 1985 (Maxwell 1984, 1985 and 1986). Tuffys, which is not listed in the leading brands, is a regional brand. As confirmed through personal communication with a marketing manager of Tuffys, this brand is one of the two leading brands in the states of Wisconsin and Minnesota, its main marketing area. Thus, the individual household level data may "better" reflect the target population purchase behaviour. 98 the total number of packages sold, volume sold, or dollar sales for the dog food product category. This inconsistency occurs primarily because two distinct product categories (dry and canned) are aggregated together to provide brand sales estimates. We will illustrate this further. The canned dog food comes in packages weighing about six, fourteen, and twenty three ounces whereas the dry dog food is sold in packages weighing from four to forty pounds. The modal canned dog food purchase is two cans at an average price of about $0.90 per pound. On the other hand, the modal volume of dry dog food is five pounds at an average price of $0.50 per pound. If the dog food consumption rate is constant for a household, then the canned dog food may have more units purchased than the dry dog food. On the other hand, if the total number of purchases of both the categories of dog food is the same, then the dry dog food would have higher dollar sales and volume sales than the canned dog food. As a result of these comparisons, we conclude that one must be cautious in using the store level sales data. For this reason, the analysis conducted in this dissertation is confined to the household level data. Comparisons between the IRI household sample and the store level estimates of brand sales are found to be close. For some regional brands, however, there are some differences. IRI in its attempts to collect a representative sample, appear to have been modestly successful at this task. 4.6. DATA MANAGEMENT ISSUES. The IRI sample data for the product category dry. dog food contains 14,728 pur-chase observations, about 150,000 television exposure observations, 286 different universal product codes, and extensive demographic information for 2,184 households. As might be expected, none of the standard statistical program packages can handle such a mas-99 sive database. The purpose of this section is to describe our method of simplifying data management tasks. Figure 4.1 outlines the process flow diagram by which parts of the database were ul-timately converted so as to be compatible with the program package Statistical Analysis System (SAS 1984). The model estimation reported in the next chapter was performed us-ing the PROC MLOGIT procedure of SAS. Several conceptual issues involved in organizing the database are pointed out below. The purchase observation data file contains a variable indicating the number of pack-ages bought by a household. These purchases, generally, consisted of one package. For purchase occasions where a household purchased more than one package of the same brand, multiple observations were combined and treated as one observation. This reduced the to-tal number of purchase observations by about 1% for the dry dog food product category and less than 1% for aluminum foil. Combining records as described above poses a problem, however, if a household buys two or more packages of different brands on the same purchase occasion. Namely, what brand do we indicate in the single purchase observation? (For the product category dry dog food and aluminum foil, such purchases only occured less than 1% of times). To avoid this problem, multiple brands purchased at the same occasion were not combined. Instead they are treated as two separate but simultaneous (for the purpose of the brand loyalty measure) purchases. . Of the 14,728 purchase observations of the product category dry dog food, 10,693 purchases were made by the 621 households with five or more purchases during the 88 week sample period. Given that we need to estimate brand loyalty, we decided that 100 households making fewer than five purchases during the 88 weeks should be excluded. Guadagni and Little (1983) and Tellis (1988), in their study of brand choice, also excluded households with less than five purchases. Of these 621 households, only 250 households are included in the television exposure measurement sample. This latter group made 4,167 purchases which were retained for further .analysis. We also found that these purchase observations were not different for the total purchase database for this product category (see Table 4.11). Finally, as mentioned earlier, the first 32 weeks (of the 88 weeks in total) of purchase observations (1,615) were used to initiate the measure of brand loyalty. The remaining observations (2,552) were divided into two equal subsamples: the first and second half of sample. The first half of sample was used to estimate several alternative models to predict the probability that a household would purchase one of 11 brands of dry dog food. The second half of sample was used to test the results of the "best" model from the first half of sample. For aluminum foil, the same procedure was used but restricted to those households who made two or more purchases in the 88 weeks. This change in the procedure was necessary since aluminum foil is infrequently purchased. This resulted in 975 purchase observations made by 301 households (with television exposure). 101 CHAPTER V Model Estimates and Interpretations. 5.1. Introduction 103 5.2. Approaches to Compare Models 104 5.3. Brand Choice Models for Dry Dog Food 106 5.3.1. Alternative Model Specifications: Summary . . . . . . . . . . . . . 107 5.3.2. Alternative Model Specifications: Estimation Result 108 5.3.3. Model Cross-validations 123 5.3.4. Price Sensitivity Interpretation 125 5.3.5. Model Interpretations . 130 5.4. Brand Choice Models for Aluminum Foil 134 5.4.1. Alternative Model Specifications 134 5.4.2. Model Interpretations 138 5.5. Model Comparison across Product Categories . . : 139 102 5.1. INTRODUCTION The purpose of this chapter is to present and interpret the results of the analysis conducted to measure the effect of television advertising on household price sensitivity. Household brand choices are examined for two product categories, dry dog food and aluminum foil. The chapter is organized in four sections. In section 5.2, various approaches used for comparing brand choice models are described. In section 5.3, several models of brand choice for the product category of dry dog food are reported. These models suggest that television advertising, display, presence of a deal, deal amount, and brand loyalty have a positive influence on brand choices. Furthermore, the models predict that higher brand price results in lower likelihood of choosing a brand, and the effect of television advertising is to increase household price sensitivity. In addition, models in which the television advertising variable is disaggregated by brands, better explain brand choices than models utilizing an aggregate variable. These conclusions appear to hold for a variety of model specifications. In section 5.4, brand choice models for aluminum foil are reported. Although this product category has fewer brands than dry dog food, the analysis show that on the average the effect of television advertising on household price sensitivity is similar to those observed for dry dog food. In section 5.5, models are compared for the product categories of dry dog food and aluminum foil, based on the t-statistics of parameter estimates and elasticities are com-pared. This comparison suggests that the effect of television advertising exposures and sales promotional variables on the average household's brand choices varies across product categories. 103 5.2. APPROACHES to COMPARE MODELS The purpose of this section is to summarize and comment on four common measures that will be used to compare the statistical "fit" of competing models. The measures are the percentage of brand choices correctly predicted by the model, goodness-of-fit, a x 2 based measure, and the t-statistic associated with parameter estimates. A way to make model comparison is to compare the percentage of choices correctly predicted by the models. This measure is based on the cross-tabulations of actual household choices and model predictions of those choices. The ratio of the total number of purchases along the main diagonal in the cross-tabulation to the total number of purchases observed times 100 gives the percentage of purchases correctly predicted for each brand. A weakness of this measure is that it lacks mathematical rigour necessary for statistical comparison. An alternative and more formal approach is goodness-of-fit, a summary measure statistic indicating the accuracy with which a model approximates the observed data. A measure commonly used with multinomial logit choice models is p2; a statistic which reflects the fraction of uncertainty empirically explained by the model relative to some other competing null model (McFadden 1973 and Hauser 1978). It is common in the brand choice models to use the multinomial logit market share model as the null model (Gensch and Recker 1979 and Guadagni and Little 1983). Formally, if L(f3) is the maximum of the logarithm of the likelihood (log-likelihood) function maximized with respect to all parameters, 01,02, • • • ,/3m, and L(0) is the maximum value of the log-likelihood function, setting all parameters with respect to all independent variables equal to zerof, then p2 is t Some computer programs such as MLOGIT provide estimates of the log-likelihood function when all coefficients are zero. In other words, this approach assumes that all brands have equal prior probabilities. This, generally, tends to have effect of overstating p2. A more acceptable procedure (Hensher and Johnson 1981) would be to use market shares as estimates of prior probabilities. In the present work, market share estimates are 104 one minus the ratio of L(j3) and L(0). Thus, p2 = 1 - If we can obtain the perfect prediction for household's brand choices using the estimated model parameters then the logarithm of the likelihood function will be zero, and p2 will be one. On the other hand, if brand choices do not depend upon any of the independent variables such as price, deal amount and so on, then L(/3) will be equal to L(0), and this would result in p2 equal to zero. Further, p2 is similar to the R2 statistic in regression analysis and ranges from zero to one. Since factors influencing brand choices may vary across a sample of households, values of p2 between 0.2 and 0.4 can be interpreted as representing a good fit of the model to the data (Hensher and Johnson 1981). The third approach used to compare two alternative models is the x 2 test statistic. If L(/3m) is the log-likelihood function value with m parameters in the model, and is the log-likelihood function value for the model with k parameters (k > m and m is a subset of k), then -2[L(/3m) - £(/?*)] is distributed according to x 2 with (A- m) degrees of freedom (Maddala 1983). A significant value for the x 2 test statistic indicates that the model with more parameters is statistically preferred to the model with fewer parameters. Although the x 2 statistic provides an approach to compare alternative models with differing number of parameters, the statistic is less useful when two competing models have the same number of parameters. This possibility arises when a variable such as brand loyalty is measured by alternative definitions. In such instances the asymptotic t-statistic associated with the variable becomes a viable alternative measure to compare models. The t-statistic also indicates the relative importance of the variables in explaining a household's brand choice behaviour (Guadagni and Little 1983, and Johnston 1984). If the sample size is fixed, then the absolute value of the t-statistic provides an indication of the explanatory used for prior probabilities. 105 importance of the independent variable. As might be expected, the numerator of the t-statistic is the coefficient itself and the denominator is the standard error of the estimated coefficient. Thus, an increase in the absolute value of the coefficient, and/or a decrease in the coefficient's standard error, will result in an increase in magnitude of the t-statistic. 5.8. BRAND CHOICE MODELS for DRY DOG FOOD The purpose of this section is to summarize and comment on the analysis conducted to assess the effect of television advertising on a household's brand price sensitivity for the product category dry dog food. This section is organized into six subsections. In subsections 5.3.1, and 5.3.2, several alternative model specifications are summarized and compared. It is found that the model in which the television advertising variable is disaggregated by brand is preferred to the specification in which the television advertising variable is aggregated. It is also found that for a variety of specifications, the effect of television advertising is to increase a household's price sensitivity. In subsection 5.3.3, the selected models from previous subsections are cross-validated. The cross-validation analysis indicates that the models do fit the dry dog food data well and reinforces findings about the television advertising variables. In subsection 5.3.4, the theoretical model is used to interpret whether the observed effect of television advertising may be due to competitive reaction effects associated with price and advertising. It is found that the effect of an increase in advertising is an increase in price sensitivity, but the effect is likely to be small. The model parameters are then interpreted in terms of elasticities with respect to the marketing decision variables in subsection 5.3.5. 106 5.3.1. Alternative Model Specifications: Summary Model specification comparisons are based on the work of Guadagni and Little (1983). In their work, the final model specification was chosen by observing changes to parameter estimates resulting from adding or removing variables from the model. Before describing the specification comparisons, an overview of this subsection is presented below. The data analysis in this subsection proceeds in four phases. In the first phase, eight model specifications are investigated. In the second phase, alternative measures of the television advertising are investigated. In the third phase, alternative measures of brand loyalty are investigated, and finally the household income variable is included as an explanatory variable of household's brand choices. These four phases are described in more detail overviewed below. Phase I: The base specification contains no independent variables. Nine independent variables (brand loyalty, feature, display, presence of a deal, deal amount, television ad-vertising, price, price times television advertising, and price times feature) are sequentially added to create the next five specifications. When all variables are included, the model is called the "reference" model. In specification seven, the brand loyalty variable is excluded. In the last specification, all the variables except the television advertising variables are included. It is found that the reference model M6, with nine independent variables, is "better" than the other seven specifications. It is also found that the effect of television advertising on brand choices is to increase a household's price sensitivity. Phase II: Variations of the reference model are compared by varying the measurement definition of the television advertising variable. In this phase, the reference model is compared to the model in which the television advertising is "disaggregated" by brands. 107 It is found that the model in which television advertising is disaggregated by brands is preferred to the model in which television advertising is aggregated. Phase III: The analyses and comparisons of the eight specifications are repeated by replacing the proposed brand loyalty variable definition by the one used by Guadagni and Little (1983). The proposed brand loyalty variable is found to predict the brand choices better than the measure proposed by Guadagni and Little. On the other hand, the model specification that includes the measure of brand loyalty proposed by Guadagni and Little has higher t-statistics for the parameter estimates of the television advertising variables than does the reference model. With either measure, it is concluded that increased television advertising is associated with increased household's price sensitivity. Phase IV: The reference model is compared to a model in which the household income variable is included as an explanatory variable for brand choices. A comparison of these two models indicates that the effect of household income is very small, hence, the income variable is not considered in the subsequent analysis. Unique model specifications are labelled M l through M8. The t-statistic or the x 2 is reported in the text when there is a need for comparison between these specifications. When the differences are obvious, the test statistics is not reported. Also, since the estimated values of the brand specific constants tend to be unstable, changes to these coefficients are not discussed. 5.3.2. Alternative Model Specification: Estimation Results Phase I: Table 5.1 shows the effects of alternative model specification on maximum likelihood parameter estimates for the dry dog food. To measure the relative importance of independent variables and to investigate the stability of the coefficients to changes in 108 model specification, the reference model is modified according to the variables included with each specification. Model development is analyzed through changes in p2 and/or by changes in the log-likelihood statistic. It is found that the reference model including nine independent variables is the "best" model, as discussed below. As shown in the lower half of Table 5.1, the base model specification (Ml) contains only brand specific constants. These constants reflect estimates of the market share of brands relative to one particular brand (Brand K). In addition, p2 of this model is assumed to be zero. In this sense, the power of other explanatory variables lies in explaining the purchase probability fluctuations over and above the average market share of a brand. The maximum likelihood estimates for brand specific constants for this specification may be obtained as follows. If pj is an estimated market.share for brand j and pk is the estimated market share for some other competitive brand, then the maximum likelihood estimate of the brand specific constant for brand j is log(pj/pk). As might be expected, if pj is equal to pk, then the ratio\pj/pk) is one and the brand specific coefficient is zero. Furthermore, if brands have equal market shares, then the brand specific coefficients will be equal for these brands. In the second specification (M2) in Table 5.1, the addition of the brand loyalty variable to the model produces a major change in p2. This indicates that the brand loyalty variable accounts for major variations in brand choices. A large value of the t-statistic for the variable brand loyalty implies that the brand loyalty coefficient is highly significant and an important predictor of brand choices. In the next specification (M3), sales promotional variables such as deal amount, presence of deal, feature and display are added to the specification M2 with another major increase in p2. Thus, the sales promotional variables have a significant influence 109 in accounting for variation in the household's brand choices. In addition, the variable deal amount has the largest t-statistic among the sales promotional variables. This implies that deal amount is the most important variable among the sales promotional variables. Note also that although the feature variable has a positive influence on brand choice, the variable is not statistically significant at p < 0.05. In the next specification (M4), the television advertising variable is added to the specification M3 with no change in p2. The t-statistic of -0.95 indicates that the effect of the variable is not significant at p < 0.05. Furthermore, the negative sign of the coefficient is in the wrong direction given our expectations. In the next specification (M5), the price variable is added to the previous specification bringing with it about a one percent increase in p2. Note however, that inclusion of the price variable has changed the magnitude of the parameter estimates for display, feature, and presence of deal variables. Finally, comparing the log-likelihood values for the specification M4 with the specification M5, the y2 test statistic associated with their difference is 46.8 with one degree of freedom. Since the critical value of x 2 at p < 0.05 is 3.84, the specification M5 is statistically preferred to the specification M4. In the reference model specification (M6), the variables price times television adver-tising and price times feature are added with a small increase in p2. In comparing the log-likelihood values for the specification M6 with M5, the x 2 test statistic associated with their difference is 10.6 with two degrees of freedom. The critical value of the x 2 statistic is 5.99 at p < 0.05, suggesting that the specification with the multiplicative variables (M6) is statistically better than without (M5). In addition, the following four observations are made comparing M5.with M6. First, the parameter estimate for the television advertising variable now has a positive sign and so affects brand choice. The t-statistic indicates that 110 the effect of television advertising variable is statistically significant at p < 0.05. Second, a negative and statistically significant coefficient for the variable price times television adver-tising indicates that an increase in television advertising results in increased household's price sensitivity. Third, there is a small drop in the absolute parameter value for the price variable in the specification M6 from that in M5. Finally, inclusion of the variable price times feature changed the sign of the feature variable from negative to positive in M6. The main effect of the feature, however, is not statistically significant at p < 0.05. The negative coefficient for the variable price times feature suggests that feature advertising . increases brand price sensitivity. It is worth noting that if the "true" model contained the price times advertising inter-action, then models containing the main effect of advertising only, may either overstate or understate (i.e. the effect is indeterminate) parameter estimates (Neter and Wasserman 1974). For example, the specifications M4 and M5 contain main effects for the variables television advertising and price, and it is found that coefficients of both variables are un-derstated. This reinforces the importance of including the variables price times advertising and price times feature in the model. In addition, since the models are estimated using household level data, the specification M6 provides a direct test for a hypothesis concerning the effect of advertising on household's price sensitivity. The brand specific constants in Table 5.1 capture uniqueness associated with brands that is not explained by the above mentioned independent variables. If there is a variable. or set of variables that explain all the choice differences among brands, then the brand specific coefficients will become zero. Obviously such variable(s) is difficult to find. In examining the t-statistic for the brand specific coefficients across M l to M6 specifications, no systematic pattern is observed. As a result, the brand specific constants are not 111 interpreted for their implications in the subsequent discussion. A question one may ask is whether a household level model (particularly, one that includes the variable of brand loyalty) offers an advantage over a market level aggregate model. The model specification M7 investigates this issue. Clearly, variables of television advertising f, price, display, feature, and deal amount are all market level variables. The brand loyalty variable, however, captures household specific brand choice variation. In the specification M7, the brand loyalty variable is dropped with all the market level variables remaining in the model. Note that there is a major drop in p 1. In addition, the exclusion of the brand loyalty variable affects the coefficients of other variables. More particularly, the display, feature, and price times feature coefficients are changed by as much as 40% from their values in the specification M6. Thus, it is concluded that estimation of the effects of feature and display variables are conditional on the presence of the brand loyalty variable. It is also found that seven out of ten brand specific constants increase in absolute value compared to the specification M6. In the specification M7, the brand specific constants may be providing brand specific information previously contained in the brand loyalty measure. In addition, the quality of model fit deteriorates and marginally significant parameters become insignificant at p < 0.05.. All the above findings reinforce conclusion that brand loyalty is an important variable for explaining brand choices across households, and therefore should be included in the brand choice model as an explanatory variable. It is next assessed whether model specification M8, which excludes the television adver-tising variables, is statistically preferred to M6, which includes these variables. Exclusion of the television advertising variables decreased the log-likelihood value by 3.7 units. The f In this particular database advertising is a household specific variable. However, to maintain similarity in comparison with the Guadagni and Little study, we keep advertising variables in the model. 112 corresponding x 2 test statistic associated with this value is 7.4 with two degrees of freedom. The critical value of the x2 statistic at p < 0.05 is 5.99, suggesting that M6, in which the television advertising variables are included, is statistically preferred to specification M8, that excludes them. It is noteworthy that the inclusion and the exclusion of variables in the eight speci-fications described above, produced relatively stable parameter estimates for most of the variables. The stability of parameter estimates is an indication that collinearity is not a problem for the reference model M6. In addition, the parameter estimate for the televi-sion advertising variable depends upon the presence of price times television advertising variable. This dependence is consistent with the theoretical and empirical literature re-viewed in chapter II. Finally, the effect of feature and price times feature on household's brand choices are weak. For this reason, subsequent discussion will not comment on the parameter estimates for feature variables. Alternative Definition of Advertising Variable In Phase II of the analysis, three different procedures are used to measure the television advertising variable. The question to be answered is as follows: Which measure is most consistent with the observed data, and how do different measurement procedures affect conclusions about the effect of television advertising on a household's price sensitivity? The television advertising exposure variable for a brand is measured by counting the number of television advertisements that a household is exposed to in an interval before a purchase. The interval before the purchase is varied in the specifications M6, M6-4W (four weeks before purchase) and M6-8W (see Table 5.2). In the first case, the time interval is measured by the number of days between two successive purchases (specification M6, see Table 5.2). In the next two cases the time interval is a fixed four week interval 113 (specification M6-4W) and eight week interval (specification M6-8W) before the purchase. Comparing these three specifications, it is concluded that the effect of varying the time interval appeared to have minimal impact on (1) estimated model parameters and (2) the fit between the model and the observed choices. In addition, each of these three measures of television advertising have a strong positive correlation with the other two measures. Moreover, the effect of varying the time interval is to scale the values of the television advertising variables. Finally, since the scaling of variables does not affect the predicted brand choices in the multinomial logit choice model (see chapter IV), the log-likelihood statistics for M6 and the competing specifications M6-4W and M6-8W are almost equal. Four weeks is an average iriterpurchase interval for this sample but the t-statistic for both the advertising variables (television advertising and price times television advertising) is not significant at p < 0.05. Furthermore, there are minor differences in the parameter estimates between the specification M6 and the specification M6-8W. In general, however, the number of television advertisements a household is exposed to in eight weeks prior to purchase and the number of television advertisements the same household is exposed to between two successive purchases has statistically similar impacts on the parameter estimates. Since the measure based on the time interval between two successive purchases is a conceptually better measure, this measure of television advertising is used for all subsequent analysis. Disaggregate Form of Advertising Variable An important question concerning the definition of the advertising variable is whether the marginal effects of television advertising varies across brands. One approach to help answer this question is to estimate one parameter for every advertised brand. This form of advertising variable is referred to as "disaggregate". A related question is whether this 114 form of advertising variable will affect conclusions about the effect of television advertising on household price sensitivity These questions will be addressed in the following discussion. Specification M6-D7 is formulated to answer these questions. In this specification seven pairs of parameters are added to the.model; one set for the television advertising variable and the other set for the price times television advertising variable. The %2 test statistic used to compare the specification M6 and the specification M6-D7 is 29.2 with 12 degrees of freedom, and the critical value of the x 2 at p < 0.05 is 21.03. This suggests that specification M6-D7 is statistically preferred to the specification M6, and also suggests that disaggregate modelling of television advertising variables is preferred to the aggregate approach. It also suggests that the effectiveness of television advertising in influencing brand choices varies among the seven advertised brands. A closer examination of the parameters of the television advertising, and price times television advertising variables, reveals that for the specification M6-D7, three pairs of pa-rameters are statistically significant at p < 0.05. Furthermore, six out of seven parameters for the variables of television advertising are positive and the same number of parameters for the variable price times television advertising are negative. This latter observation reinforces an earlier conclusion; that for the product category dry dog food, increased television advertising increases household price sensitivity. It is evident that the television advertising variable disaggregated by brand provides a "better" explanatory model. To explore and compare the various specifications given in Table 5.1 to those where the television advertising variables are replaced by the disaggre-gated variables, several model specifications are re-estimated and reported in Table 5.3. In addition, since parameters for Brands A, B, and E are significantly different from zero at p < 0.05, in the subsequent analysis television advertising variables are restricted to these 115 three brands (see Table 5.3). All other variables in Table 5.1 and Table 5.3 are identical. A summary of the comparison is reported below. First, specifications where neither price nor price times television advertising vari-ables are included are compared. Comparing log-likelihood values for the specification M4 (Table 5.1) with the specification M4-D (Table 5.3), the x 2 test statistic associated with their difference is 4.0 with two degrees of freedom. The critical value of x 2 at p < 0.05 is 5.99. Thus, the simpler specification M4 is statistically preferred to the specification M4-D. When the price variable is added, a similar comparison between the specification M5 and the specification M5-D, implied that the specification M5 is preferred to M5-D. However, a comparison between the aggregate specification M6 (with price times adver-tising variable included) and the corresponding disaggregate specification M6-D, reveals that the latter, M6-D, is statistically preferred to M6. The x2 test statistic for this com-parison is 20.4 with four degrees of freedom. The critical value of x2 at p < 0.05 is 9.49. Comparing the log-likelihood statistics for the specification M7-D with the specification M7, results in the x 2 test statistic 30.2 with four degrees of freedom. The critical value of X2 at p < 0.05 is 9.49. These comparisons suggest that the specifications involving disag-gregated forms of television advertising and presence of price times television advertising are statistically preferred to the specifications involving the aggregate form of television advertising variables. Alternative Definition of Brand Loyalty Variable: Phase III In Phase III analysis, three measures used to operationalize the concept of brand loyalty are compared. In addition, since the brand loyalty variable is the strongest predictor of brand choice, a natural question one may ask is whether the alternative definitions of the brand loyalty variable affect our conclusion about the effect of television advertising 116 on a household's price sensitivity. This question is addressed below. The first measure of brand loyalty is derived by (used in preceding analysis) weighting past brand choices exponentially and inversely by the number of days since a brand choice occurred. The second measure of brand loyalty is derived by exponentially weighting past brand choices. This is the approach suggested and used by Guadagni and Little (1983). Since the first 32 weeks of data is used to initialize the loyalty measures, the measures described above are up-dated until after the first 32 weeks. For the third measure of brand loyalty the measure is not up-dated after the first 32 weeks. Table 5.4 summarizes model estimation results from varying of the brand loyalty definition in brand choice models. In the specification M6-BL32, the brand loyalty measure remains the same for a household after the first 32 weeks. In the specifications M6 and M6-GL, brand loyalty measures are up-dated as household purchases occur after the first 32 weeks. In the specification M6-GL, past purchases are exponentially smoothed using a smoothing constant of 0.875J. Thus, the brand loyalty measure in the specification M6-GL is the measure proposed by Guadagni and Little (1983). A comparison of p 2 across models (see Table 5.4) reveals that the exponentially weighted and updated measure outperform the other two measures. It is interesting to note that parameter estimates and their corresponding t-statistics are similar across the model specifications even though the log-likelihood estimate increases from -1783.2 to — 1480.9. These three specifications are compared below. A closer examination of the parameter estimates in Table 5.4 indicates that the specification M6 and the specification M6-GL estimates are very close with one exception. That exception is for the parameter estimates for variables of brand loyalty and advertising, t The optimal value of the smoothing constant is determined below. 117 In the specification M6, the brand loyalty variable parameter is estimated with a smaller standard error than in the specification M6-GL. In the specification M6-GL, however, parameter of the television advertising variable is 31 percent higher than it is in M6. The same pattern is true for the variable price times television advertising parameter whose estimate is higher by 27% compared to its value in the specification M6. We will comment on a possible reason for these differences below. In the final comparison involving the specifications M6 and M6-BL32, we find that the brand loyalty variable parameter is lower by about 40% in M6-BL32. The parameters of most other variables in these two specification are very close to each other. To explore and compare various specifications given in Table 5.1 to those where the Guadagni and Little measure of brand loyalty is used, several specifications are re-estimated and reported in Table 5.6. Subsequently, a comparison of models when the television advertising variables are replaced by the disaggregated variables is reported in Table 5.7. Before we present and comment on various comparisons, a summary of the procedure used to obtain the optimal smoothing constant for the brand loyalty measure proposed by Guadagni and Little is presented below. An interesting question is how to pick the smoothing constant for measuring the loyalty variable. Since Guadagni and Little (1983) found the smoothing constant of 0.875 to be the optimal value for the regular coffee data, this value is intially used. Following this, smoothing constants of 0.1, 0.2, 0.5, 0.7, 0.8, and 0.9 are used to re-estimate the specification M6-GL. The estimated parameters and the values of the log-likelihood statistics for all these re-estimated model are reported in Table 5.5. It is interesting to note that the log-likelihood function is concave with respect to the smoothing constant. Moreover, the function peaks around 0.8 and is relatively "flat" around the peak. To 118. facilitate the comparison of our results with Guadagni and Little, then smoothing constant of 0.875 is used in the subsequent analysis. A comparison of the specifications reported in Table 5.6 (Guadagni and Little measure of the brand loyalty) to those in Table 5.1, reveals that the t-statistic associated with the brand loyalty parameter estimate is always higher in Table 5.1 than in Table 5.6. This means that the brand loyalty measure used in Table 5.1 has a lower standard error than the measure used in Table 5.6. To test whether the brand loyalty parameters in Table 5.1 are equal to the parameters in Table 5.6, the t-statistics of the differences between parameter estimates in both tables are usedf. The tests reveal that the brand loyalty parameters in Table 5.1 are lower, and significantly different at p < 0.05. In addition, the log-likelihood values in Table 5.6 are generally lower than those in Table 5.1. Since all the other variables are the same for the models in both tables, it is concluded that the brand loyalty measure proposed herein predicts brand choices better than the brand loyalty measure proposed by Guadagni and Little. Additionally, as reported above, parameter estimates for the television advertising variable in the specification M6-GL are higher by 31%, and the price times television advertising parameter estimate is higher by 27%, compared to the estimates in the specification M6. We comment on these findings below. The measure of brand loyalty proposed in this dissertation weighs past brand choices inversely by the number days since a choice occurred. This measure thus captures the possible effect of interpurchase duration on brand choice. To measure television advertising exposures, the number of exposures a.household is exposed to between two successive t Since the logit model is non-linear in estimated parameters, to compare parameters of two competing models, one must use the variance-covariance matrix of two estimated model parameters Chow(l983). Since the two alternative measures of brand loyalty variable have small impact on the other parameter estimates with the exception of the television advertising parameters, the procedure is a simple and reasonable approximation (Neter and Wasserman 1974). 119 purchases are counted. The interpurchase duration is an implied variable in this measure also. Guadagni and Little's measure of brand loyalty, on the other hand, weighs all past purchases equally. Thus, one may argue that the measure of brand loyalty proposed here may be correlated with the measure of television advertising. An examination of the correlations between these measures found no evidence for this argumentf. In addition, the observed differences between the parameter estimates of advertising variables are statistically not significant at p < 0.05. Thus, it is concluded that the observed differences are random and may be ignored. In addition, it is concluded that the brand loyalty measure proposed in this dissertation predicts brand choices consistently better than the other measures of brand loyalty. Comparing the specifications in Table 5.6 (the aggregate form of the television adver-tising variable) to the specifications in Table 5.7 (the disaggregate form of the television advertising variable), the conclusions reported above concerning the disaggregate form of television advertising are further reinforced. Although most of the parameter estimates in Table 5.3 and the equivalent specification in Table 5.7 are similar, there are exceptions for the variables involving Brand E's advertising variable. More specifically, comparing the specification M6-D with the specification M6-GL-D, it is found that the coefficient for Brand E's advertising variable in the specification M6-GL-D is 27% higher than the same coefficient in M6-D. Similarly, the coefficient of price times Brand E's advertising variable in the specification M6-GL-D is 22% higher than the same coefficient in the specification M6-D. When these variables are examined for collinearity, the observed correlations are t The estimated correlation between the brand loyalty measure proposed in this disserta-tion and the television advertising variables are -0.0315, -0.0314, and -0.0819 for Brand A, Brand B, and Brand E respectively. Similarly, the estimated correlations between the Guadagni and Little's measure of brand loyalty and the television advertising variables are -0.0241, -0.0302, and -0.0672 for Brand A, Brand B, and Brand E respectively. All of these correlations are not significantly different from zero. 120 small and statistically insignificant at p < 0.05. In addition, the observed differences be-tween the parameter estimates for the advertising variables are statistically not significant at p < 0.05. Thus, it is concluded that the observed differences are random and may be ignored. Several comparisons summarized above suggest that the brand loyalty measure pro-posed in this dissertation is a better measure than the brand loyalty measure proposed by Guadagni and Little, at least for this data set. However, even if the brand loyalty measure proposed by Guadagni and Little is used in the brand choice model, the observed effect of television advertising is an increase in household price sensitivity. Impact of Household Income on Brand Choices: Phase IV In Phase IV the effect of a demographically up-scale sample on brand choices is investigated by including demographic variables in the model. Not surprisingly, none of the coefficients associated with household size were significantly different from zero. Thus this variable is excluded from further consideration. Household income had only a small influence on brand choices, and parameters of independent variables such as advertising, price, and so on are not affected when household income is included in the brand choice model (see Table 5.8). On the other hand, several brand specific constants are significantly affected by including household income in the brand choice model. In the specification M6, seven brand specific constants are significantly different from zero at p < 0.05. In specification M6-Inc, where the income variable is included as an explanatory variable of brand choice, only three brand specific constants are significantly different from zero at p < 0.05. It is also noted that for the specification M6-Inc, television advertising increases household price sensitivity, as evidenced by a negative coefficient for the variable price times television advertising. This conclusion thus remains unchanged from that suggested 121 by specification M6. A closer examination of household income parameters across brands reveal that only one parameter out of ten is significantly different from zero (Brand E) at p < 0.05 and one more parameter is marginally significant (Brand A) at p < 0.06. It is concluded from this that Brand A (Dog Chow), the most popular brand in the market, and Brand E the most heavily advertised brand, appear to be chosen more often by households with a higher income than those with a lower income. Inclusion of the ten income associated parameters in the. model affects p2 by only 0.003. Hence, household income variable is excluded in subsequent analyses. In summary, a total of 26 alternative model specifications are examined using data from the product category dry dog food, in order to measure the effect of television advertising on household price sensitivity. It is concluded that two model specifications provided the "best" explanation of the household brand choices. These two best specifications are: the first M6-D, with the brand specific disaggregated form of the television advertising variable, and exponentially and inversely weighted past brand choices as the measure of brand loyalty. The second best specification is M6 with the same measure of brand loyalty as M6-D, but with the aggregate form of the television advertising variable. For both model specifications, the parameter estimate for the variable price times television advertising is negative and significantly different from zero. It is thus, concluded that the effect of television advertising, in the case of dry dog food at least, is to increase household's price sensitivity. 122 5.3.3. Model Cross-validation In the analysis described in the preceding subsection, all the models are estimated using the first half of the sample. The second half of the sample is used to test the stability of the estimated parameters for specifications M6-D and M6. Parameter stability is tested using brand choice model parameters from the first sample to predict brand choices for the second sample and vice versa. It is found that parameter estimates for most of the independent variables are stable, in particular for the model in which the television advertising variable is the disaggregated form. In addition, the fit between, the model and the data is very high. Details of this analysis are provided below. The specification M6-D in which the television advertising variable is disaggregated for three brands is cross-validated. The parameter estimates and goodness-of-fit measures are given in Table 5.9. Since the television advertising parameters for three brands only (A, B, and E) have statistically significant parameter estimates at p < 0.05, the cross-validation models are restricted to these, and individual parameter estimates for the television advertising variable are only calculated for these three brands. Of the 13 parameters estimated for independent variables, 12 have the same sign across samples. This is reassuring. The magnitude of the parameter estimates, however, have remained in the same range ( ± 2 5 % ) for five variables (price, Brand E's advertising variable, presence of a deal, deal amount, and brand loyalty). All other parameter estimates have changed more than 25% across the samples. Additionally, the parameter estimates for the Brand B's advertising variable, and price times Brand E advertising variable, are not significant at p < 0.05 in the second half of the sample. Estimated parameters for the feature variables are not significant at p < 0.05 for either sample. While we can find no systematic differences between the two halves of the sample, the magnitudes (in absolute 123 value) of the significant coefficients are generally higher in the second half of the sample. The overall fit between the model and the data is very good for both samples. The observed cross-validated p2's of 0.485 and 0.517 are better than those reported by Guadagni and Little (p2 = 0.48). Because there are 11 brands in the market and as reported in chapter IV, no one brand has more than 25% of the market, an achievement of 53% and 54% of correctly predicted choices is very impressive indeed. Parameter estimates for the specification M6 are reported for both halves of the samples in Table 5.10. Eight out of nine independent variables have the same signs for their parameter estimates in both samples. The magnitude of parameter estimates, however, has remained in the same range (± 2 5 % ) for five parameters (price, brand loyalty, two deal variables, and television advertising). More disturbing is the fact that the t-statistics for three parameter do not match in their levels of significance across samples. This point is elaborated on below. Television advertising and price times television advertising have parameter estimates in the same direction for both samples. However, the associated t-statistics for the second half of the sample indicate that parameters of these variables are statistically not significant at p < 0.05. The variable feature has a positive sign for the second half sample, but the parameter is not statistically significant at p < 0.05. Additionally, the variable price times feature is significant in the first half of the sample, but not in the second, and the sign for the parameter estimate changed from negative in the first half to positive in the second. While all the television advertising variables in Table 5.9 are significantly different from zero at p < 0.05, this is not the case for the same set of variables in Table 5.10. This finding reinforces the conclusion that television advertising disaggregated by brands is a 124 better specification than the aggregated specification. In summary, two sets of sample data are used to test the stability of estimated model parameters. It is found that the fit between the model and data is very high for both samples. In addition, most of the estimated parameters for independent variables have the same sign across the two samples. Finally, the model in which the television advertising variable appears in disaggregated form is statistically better than the aggregate form model. 5.8.4- Price Sensitivity Interpretation In the preceding analysis, it is consistently observed that the predictor variable price times television advertising for household brand choices has a negative coefficient. This is interpreted (see elasticity expression in subsection 5.3.5) to mean that the effect of television advertising is to increase a household's price sensitivity for the product under consideration. An important theoretical question is whether the effect of television adver-tising may be due to a competitive reaction associated with price and advertising. The purpose of this subsection is to use the theoretical model developed in chapter III in an attempt to answer this question. In chapter III, three conditions were discussed that must be met in order to observe the effect of television advertising on household price sensitivity. First, the brand prices of competing brands must covary in the same direction. Second, the television advertising competitive reactions must be positive. Finally, the average increase f in the reference price due to advertising for the brand chosen must be less than the average increase in the reference price due to advertising for all other brands. Each of these conditions is assessed and discussed below. f Weighted by the measure of loyalty. 125 To assess the above conditions, purchase observations from the first half of the sample are used. The regular brand price at the purchase occasion is used as a measure of the reference price. Then, the actual prices may be used to assess competitve reaction with respect to the price variable. The advertising variable is measured by counting the number of television advertisements for a particular brand a household has been exposed to since the last purchase. Since household purchase behaviour is basic unit of measurement, purchase observations are used to assess these conditions. There are potential problems, however, in using actual prices to assess the covariations in brand prices. First, the price may have an inflationary time trend. Second, the actual price is likely a function of the marginal cost of production and the retailer markup. One approach to account for these variations is to estimate the time trend and the marginal cost and then correlate the residuals of the actual price. These residual correlations will provide an indication of whether or not competing brand prices covary in a similar pattern. Although obtaining an estimate of the marginal cost is difficult, one may use a proxy for it. The procedure used to obtain both the time trend and the proxy for the marginal cost is described below. Let the actual price of brand i on purchase week t be denoted by P,t, such that P, = Pc, - a,t • e„ where a, is the time trend estimate for brand i, PQ,- is the regular price for brand i when t is equal to zero, and e,- is a random error component not accounted by the time trend. We may think of Poi as the proxy for marginal cost. Since the measurement units of time are weeks, the linear form may be a reasonable approximation. In addition, if the price changed by 5% per annum for a brand, then the corresponding weekly price change will t For clarity of discussion, the purchase observation subscripts are not written. 126 be approximately 5/52 or 0.096% per week. This implies that values of coefficients a,- will be very small. The estimated parameter values of P 0i, a rid a goodness-of-fit measure for 11 brands of dry dog food appear in Table 5.11. The results of regression analysis indicate that the time trend is significantly different from zero at p < 0.05 for eight of the 11 brands. Since it is expected that the time trend to have small coefficient, it also accounts for a very small amount of variation in the brand prices. As a result, it is suspected that remaining variations in the competitive brand prices are due to a component not accounted for in the regression analysis. The estimates in Table 5.11 are used to obtain residuals for 1276 purchase observations and 11 brands. The lower half of the estimated correlation coefficient matrix, using the actual brand price residuals in conjunction with the actual price means, appears in Table 5.12. All correlations are positive, statistically significant, and generally of modest size. The highest observed correlation between brand A and brand B is 0.510 while the lowest observed correlation is 0.161. This indicates that the condition of positive price covariation has been met for the product category dry dog food. To assess the direction of competitive reaction to television advertising, advertising variables are correlated for the seven advertised brands. Since the advertising variables are measured by counting the number of television advertisement exposures to a household")", it is unlikely that this variable will contain a time trend. As expected, no time trend is found. Table 5.13 reports observed correlations between exposure levels for seven advertised brands. Advertising exposure intercorrelations are considerably smaller in magnitude than the price intercorrelations, and all are (again) positive. Furthermore, of the 21 correlations t The use of aggregate frequency counts (summed over a week and across households) of the number of exposures, appeared to give similar results for the advertising reactions. These correlations are also reported in Table 5.13. 127 reported in Table 5.13, 15 are significantly different from zero at p < 0.05. It is thus concluded that the direction of competitive reaction to advertising is positive for this product category. The third condition involves comparing a weighted average (by proportion consumers loyal to a brand) of the chosen brand's reference price change due to advertising, with the weighted average of the reference price change due to advertising for brands not chosen. The procedure used to assess this condition is described below. Let ki and kc denote the marginal change in the reference prices due to the advertising efforts of the brand chosen and the competitive brand respectively. Further, let 9 and 9 denote the proportion of households loyal to the chosen brand and the competitive brand respectively. Since there are modest positive correlations between competitive brand prices, and between competitive brand advertising variables, this indicates that there are positive competitive reactions to price and advertising. From chapter III (equation 3.11), substituting the positive reactions, we may write: dp _ kfi - k~9 da" 2 ' [ b - 1 } where dp and da are changes in the reference price and advertising respectively and the ratio ^ is a proxy for a household's price sensitivity. A positive value of the right hand side of (5.1) indicates that increased advertising is expected to result in decreased household price sensitivity. In addition, the effect of price sensitivity is incremental over and above the competitive effects of advertising. To apply the same model for the individual household choices, a measure of brand loyalty for the households may be used instead of using the aggregate proportions across households. The procedure used for this comparison replaces the aggregate measure of 128 brand loyalty by a household level measure of brand loyalty. Thus, if there are n brands in a market, and A, denotes an index of brand loyalty for brand i and E£=i A, = 1, then expression (5.1) can be re-stated as dp feA, - E,^/ *i-Ai Ta = 2 . ( 5-2 ) The measure developed in chapter IV that weighs past brand choices exponentially and inversely, can be a useful measure of brand loyalty for this purpose, and is used here. Finally, to complete the analysis, an estimate of reference price change (&,) for the seven advertised brands of the dry dog food is needed. The simple procedure used to obtain these is summarized below. The regular price may be used as a proxy measure for the reference price. It may be regressed with the measure of advertising intensity, (say number of exposures) and the regression coefficient of the advertising variable becomes a measure of the marginal change (&,) in the regular price due to advertising intensity. We know that a major proportion of variation in the regular prices is accounted by competitive prices. Thus, a variable is needed that contains the competitive price variations but does not contain any premium due to advertising. A simple average of non-advertised brand prices satisfies this condition. Let RPi denote the regular price for a brand i, AP the average regular price of non-advertised brands, and NEXPi the number of television advertisements a household is exposed to since its last purchase. The regression model is R Pi = 70,- + luA P + kiNEXPi + m where 7 c , 7i,- and A;, are unknown parameters to be estimated and /z, is the random error component. The parameters are estimated using the first half of sample (n = 1276) purchase observations. The estimation results appear in Table 5.14. It can be seen from the results 129 that there is considerable variation in the goodness-of-fit measures [ R 2 ) , as well as in the parameter estimates for the effect of advertising on regular prices. Using the estimate for ki, the weighted average price change is computed for each purchase observation, and the mean difference of the weighted averages are then estimated (for 1276 purchase observations). The observed mean difference (the numerator of equation 5.2) is -0.103 with a standard deviation of 0.145. The observed difference is small but the t-statistic estimated to test whether this mean difference is equal to zero, revealed the test statistic of -0.103/(0.145A/1276) = -25.4 (where 1276 is sample size). Thus, these estimates of the mean difference suggest that the effect of advertising is to increase household price sensitivity, but the effect is likely to be small. This is consistent with the results reported in subsections 5.3.1 and 5.3.2. To conclude, there is a positive correlation across advertising intensities and a small negative difference between the weighted average prices. These findings suggest that the observed effect of television advertising may be due to competitive reaction effects. 5.8.5. Model Interpretations In this subsection estimated model parameters are interpreted for qualitative and quantitative estimates of the effect of television advertising on brand price sensitivity. Before deriving the measure of price sensitivity, a qualitative discussion of the model parameters is provided. The effect of television advertising on price sensitivity is then quantified using a price elasticity measure. The t-statistic for brand loyalty indicates that it is the most important determinant of a household's brand choice. This variable alone explains about 31% of the variation in the brand choices for the first half of the sample. This is consistent with the work reported 130 by Guadagni and Little (1983) and others. The variables, deal amount and presence of a deal, also have a strong influence on a household's brand choice. As expected, a higher regular price decreases the probability of brand choice. The strong explanatory power of these three price-related variables is evident from the fact that inclusion of these variables account for an additional 18% of the variation in a household's brand choice behaviour for dry dog food product. One of the key questions that this dissertation set out to address is whether the effect of television advertising is an increase or decrease in brand price sensitivity Estimation results reported for brand choice models in Tables 5.1 to 5.10 suggest that the effect of television advertising is an increase in a brand's price sensitivity (represented by a negative statistically significant coefficient for the price times television advertising variable). A more formal and interesting approach of addressing this question is to compare changes in price elasticity measures as a result of increased television advertising. Independent variable (such as price, display etc.) elasticity can be evaluated at the sample means, or at each purchase observation. Hensher and Johnson (1981) argue that the multinomial logit choice model is non-linear in estimated parameters, and the estimated logit function need not pass through the point defined by the sample means. Hence, a better approach is to evaluate elasticity estimates for the individual purchase observations and weight these elasticities by the estimated choice probabilities. This approach is described below. Let p, be the predicted probability that a household will choose brand i whose variable k is represented with a value, If the marginal utility of the variable is 3k, then the 131 elasticity of the attribute k with respect to the probability is r)ik = §~~ = 8kXtk(l-p,). (5.3) This means that if there is a one percent change in the variable A7,*, then as a result of this change, the expected change in the probability of choosing the brand i will be mk percent. Fur.thur, if there are ./V purchase observations and p^ is an estimated choice probability of brand i at the purchase occasion h, then the weighted average elasticity (rf^) is calculated by - _ l^h=l PihVihk The weighted average elasticities for 11 brands of dry dog food with nine independent variables of choice appear in Table 5.15. Parameter estimates from the total sample (Table 5.9 and 5.10) are used for deriving these elasticities. Several comments about elasticities are made below. Although the parameter estimates for the multinomial logit choice model are the same for all brands (the exception, of course, being the television advertising variables in Table 5.9), the elasticity estimates differ across brands. This is because the variable brand loyalty differs across households. In addition, for each purchase occasion the values of variables such as price, deal amount, etc. will also differ. For instance, the regular price elasticity ranges from a high of -0.50 to a low of -1.35 for the eleven brands. It is known from the analysis (sections 5.3.1 and 5.3.2) that the effect of television advertising is an increase in a brand's price sensitivity. For Brand K. the increase in the price sensitivity is about 20%. On the other hand, if estimates from the model with the disaggregated form of television advertising is used, then price sensitivity of Brand A increases by 70%. For other brands, the increase differs from a low of 1% to a high of 50%. 132 In general the sign of the elasticity measures depends on the sign of the parameter estimates. This is because the right hand side of equation (5.3) is f3kXik(l — p,), where values of the independent variables (X^) are always non-negative in the data, and predicted probabilities (p,) are always between zero and one. In examining the absolute magnitude of elasticities, the price variable appears to cause the largest change in the probability of choosing any particular brand. This is true even though the t-statistic for this variable is only the third largest (in absolute magnitude) among the variables affecting a household's brand choices. A possible cause for this result follows. An examination of equation (5.3) reveals that the unweighted elasticity estimate will be equal to zero when p,-, the estimated choice probability for brand i is one and/or X^, the value of variable k for brand i is equal to zero. Since the independent variables such as display and presence of a deal are either one for a small fraction of purchases, or zero for others, the present procedure does not indicate the "true" impact of these variables on the brand choice probabilities. This, however, is not a problem for the price variable. As a result, one should be cautious in interpreting the order of variable importance through the use of elasticity estimates. An approach to correct this problem is to evaluate elasticities at the sample means. These are summarized in the Table 5.16. A comparison of estimated elasticities between those in Table 5.15 and 5.16 reveals that the elasticities are generally scaled higher in Table 5.16 than in Table 5.15. These differences in estimated elasticities are consistent with the work of Hensher and Johnson (1981). In addition, the problem of low elasticity estimates for sales promotional variables is also evident here. Thus, one should be cautious in using these elasticities also. 133 In conclusion, it is found that the effect of advertising is to increase household price sensitivity. The increase differs from a low of 1% to a high of 70% for the advertised brands. To conclude and summarize, in section 5.3, 51 brand choice models for the product category dry dog food are analyzed. It is found that all variables in the model displayed, the signs expected. More specifically, the estimated models suggest that the television advertising, display, feature, presence of a deal, deal amount, and brand loyalty variables positively influence brand choices. As expected, a higher brand price result in a lower likelihood of choosing a brand, and television advertising results in increased consumer price sensitivity. These conclusions hold for a variety of model specifications. 5.4. BRAND CHOICE MODELS for ALUMINUM FOIL A similar analysis to that of dry food is conducted for the product category aluminum foil. This market consists of three brands, Reynolds (R), private brands (P), and generic brands (G). Only the first (Reynolds) is advertised on television. The variable feature is perfectly correlated with the variable display, thus, the variable display also measures the effect of the feature variable. With these exceptions, the reference model development for this product category is similar to the product category for dry dog food. 5.4.I. Alternative Model Specifications Several alternative specifications for the sample of aluminum foil brand choices are summarized in Table 5.17. The development of the reference model and its parameter stability can be judged as variables included or excluded from the model as described below. For this product category, the specifications are labelled from Ml-F to M8-F. 134 A comparison ;df she various models reveals that the brand loyalty and price variables are better predictors of aluminum foil brand choices than all other predictors used. In addition, the-effect of television advertising is found to increase household price sensitivity. The analysis summary leading to these conclusions is summarized below. In the specification Ml-F, only brand specific constants are included and p2 of the model is assumed to be zero. In the second specification (M2-F), the addition of the brand loyalty variable tto the model increased p2 from 0 to 0.218. The best specification in this Table has a p2 of 0.371. Thus, the brand loyalty variable accounts for more than half of the variations in ibrand choices for aluminum foil. In the next specification (M3-F), sales promotional variables such as deal amount, presence of a deal, and display are added resulting in a small increase in p2. Although the improvement in p2 is 0.007, the x2 t est statistic of 13.2 with four degrees of freedom is significant at p < 0.05 when comparing the specification M2-F with the specification M3-F(the critical value of x2 is 9.49). Note, however, that none of the sales promotional variables are sigaaificantly different from zero at p < 0.05. As a result, it is concluded that the sales 'promotional variables play a minor role in household's aluminum foil brand choices. In the specification M4-F, the television advertising variable is added to the speci-fication M3-F. Although the television advertising variable has a positive parameter, it is not statistically significant at p < 0.05. In addition, comparing specification M3-F to specification M4-F, the x 2 test statistic is 3.0 with one degree of freedom. The critical value of the x 2 is 3.84 at p < 0.05. Thus, the specification M3-F is statistically preferred to the specification F3. This observation is consistent with the product category dry dog food data. 135 In the specification M5-F, the variable price is added to the specification M4-F resulting in a major improvement in p2. This suggests that the price has a somewhat more important role in household brand choices of aluminum foil than most other variables. A large absolute value of the t-statistic also supports this observation. In addition, with the exception of the brand loyalty parameter estimate, all other parameter estimates are changed from the specification M4-F. Two major changes are discussed below. First, the television advertising parameter is increased by 34% and the parameter is significantly different from zero at p < 0.05. Second, changes to parameters of sales promotional variables (display, presence of a deal and deal amount) are small, all these parameters are still not significantly different from zero at p < .0.05. In addition, these changes reinforce the importance of the price variable in the brand choice model. In the specification M6-F, the variable price times television advertising is added to the specification M5-F. Note that from the previous specification the parameter value of the television advertising variable is increased by more than ten times and both the television advertising variables are significantly different from zero at p < 0.05. A negative coefficient for the variable price times television advertising implies that increased television advertising results in increased price sensitivity. All other parameters, however, have remained stable from the specification M5-F. To test the inclusion of the price times advertising variable, the x 2 test statistic is calculated to be 5.2. Since the critical value of X2 is 3.84 with one degree of freedom at p < 0.05, the specification with the multiplicative variable is statistically preferred to the one without it. Finally, since M6-F is statistically the best specification, it is used to derive elasticities with respect to all the independent variables. Since the brand loyalty variable is a strong predictor of brand choices, it is expected 136 that exclusion of it (the specification M7-F) from the model should affect all other pa-rameter estimates. This is evident in comparing parameter estimates in the specification M7-F with those from the specification M6-F. Although changes are small for most of the variables, parameters of sales promotional variables have changed more than 40% from the specification M6-F. This suggests that the effect of sales promotional variables may be conditional on the brand loyalty variable for this product category. In the specification M8-F, both the television advertising variables are excluded from the specification M8-F. A comparison of the specification M8-F with the specification M6-F results in the y2 test statistic of 11.6 with two degrees of freedom. The critical value of the \2 is 5.99 with two degrees of freedom at p < 0.05, which implies that the specification with advertising variables included is statistically preferred to a specification which excludes them. Note also that parameters of sales promotional variables in the specification M8-F are changed from those in the specification M6-F. This suggests that sales promotional variables may be dependent on the television advertising variables for this product category. To conclude and summarize, various model specifications for the product category aluminum foil suggest that the variables of brand loyalty, price, price times television advertising, and television advertising consistently appear as important predictors of brand choices. The impact of sales promotional variables (presence of a deal, deal amount and display) is small and statistically not significant at p < 0.05. Finally, a negative sign for the variable price times television advertising indicates that the effect of increased television advertising is an increase in household price sensitivity. 137 5.4-2. Model Interpretations The estimated multinomial logit model fits the aluminum foil data reasonably well and indicates that the effect of television advertising is an increase in brand price sensitivity. The effects of brand loyalty and price are stronger on brand choices than all the other predictors used in this study. While market variables such as price, presence of deal, deal amount, display, and household characteristic variables such as brand loyalty have the same signs as those reported for dry dog food, their magnitudes are different. The brand loyalty variable has the largest t-statistic but the absolute value of the t-statistic for the price variable is very close to the brand loyalty variable. This suggests that price is an important determinant of household brand choices. In addition three sales promotional variables — display, presence of a deal, and deal amount have expected signs but statistically are not significant at p < 0.05 across all six specifications. Thus, it is likely that, in this market, sales promotion has a small effect on brand choices. Th effect of television advertising on brand choice is small, positive, and statistically significant at p < 0.05. Furthermore, the effect of television advertising is an increase in brand price sensitivity. This is concluded from a negative parameter estimate of the price times television advertising variable. If only one brand is advertised on television, then the model in the chapter III predicts that increased advertising results in lower price sensitivity. This prediction is not supported by aluminum foil dataf. Thus, the theoretical model is partially disconfirmed from this f When the regular price is regressed with the measure of advertising intensity (say number of exposures), then the regression coefficient of the advertising variable is a measure of marginal change (A,-) in the regular price due to advertising intensity. For Reynolds' brand, estimated value of fc, is 0.103 with corresponding t-statistic of 1.92. 138 evidence. In order to quantify the effect of television advertising on brand price sensitivity, the weighted average elasticities are derived for all independent variables in the specification M6-F. The computational procedure used to obtain the weighted average elasticities is fully described in subsection 5.3.4. The weighted average elasticities for the product category aluminum foil are in Table 5.18. As might be expected, the effect of price dominates all other variables. A one percent increase in the regular price of Reynolds brand aluminum foil is expected to result in about two and half percent decrease in the probability of choosing the brand. The effect of television advertising is an increase in the brand's price sensitivity, accounting for about 12% in the increase. The effect of television advertising on the brand choice is small and positive. Finally, the sales promotional variables have very small impacts on the brand choices. 5.5. MODEL COMPARISON across PRODUCT CATEGORIES In this chapter several alternative model specifications are reported for the product categories of dry dog food and aluminum foil. The purpose of this section is to comment on applicability of the theoretical model for determining household price sensitivity. Several comments are also provided concerning similarities and differences across the product categories. The main focus of this dissertation is whether household price sensitivity increases or decreases as the number of exposures to television advertising increases. The empirical results for both product categories studied indicate that increased advertising is associated with higher price sensitivity. For the dry dog food product category, this empirical finding can be explained by the advertising reaction of competitors, and this partially confirms 139 the theoretical model. However, since the aluminum foil product category has only one television advertised brand, the empirical result cannot be fully explained by the theoretical model of the chapter III and shows an important limitation of the model's applicability f. There are three broad similarities in the estimated models across the product cate-gories. Firstly, the brand loyalty is the most important predictor of brand choice and accounts for about 31% and 21% of brand choice variations in the product category dry dog food and aluminum foil respectively. This is consistent with the literature. Secondly, as mentioned above, it is found that the effect of television advertising is to increase brand price sensitivity in both categories. The existing literature has not been consistent on this. The magnitude of the increase varies across brands as well as between product categories. Finally, the effect of television advertising on brand choices is positive and small. Again, the magnitude of the effect varies across brands. This finding is also consistent with the literatue. There are two major differences across the product categories. Although the regular price has the third largest t-statistic in the reference brand choice model for the dry dog food product category, the price variable only accounts for about one percent variation in brand choice. The weak effect of the price variable is also reflected in the low price sensitivity for this product category. The sales promotional variables, on the other hand, account for about 15% variation in brand choice for the same product category. The effect of regular price on brand choices of aluminum foil is strong and accounts for 12% f This conclusion is consistent with Steiner's (1973) model, which specifies possibly distinct effects at the consumer and retailer level. Steiner argues that increased brand price comparisons by consumers take place because of improved brand identification brought about by brand advertising. In this case, even if consumers are "pre sold" (brand loyal) by advertising, competition between retailers will exert a downward pressure on prices which, in our models, is reflected by the negative coefficient estimated for the price times television advertising variable. 140 of the variation in brand choices. The sales promotional variables, on the other hand, account for less than one percent of the variation in brand choices for the product category aluminum foil. Since the price deals occur with about equal frequencies across these two product categories, it.may be that in the dry dog food product market households are "deal conscious", but in the aluminum foil product market, they are "price conscious". 141 CHAPTER VI Contributions, Implications, Limitations, and Future Work 6.1. Contributions 143 6.1.1. Theoretical Contribution 144 6.1.2. Empirical Contributions 145 6.2. Managerial Implications 149 6.2.1. Implications from Theoretical Model 149 6.2.2. Implications from Empirical Work . . . ._ . . 150 6.3. Limitations 153 6.4. Future Work 156 6.5. In Conclusion 158 142 In this study, the effect of television advertising on household price sensitivity is critically analyzed and modelled. Specifically, three questions concerning price sensitivity are explored in this dissertation. These are as follow: (1) How does an increase in television advertising efforts of a brand affect consumer price sensitivity? (2) Since the brand loyalty variable is the best predictor of consumer brand choices, to what extent do alternative definitions of the loyalty variable affect predictive ability of the multinomial logit models and, most particularly, affect consumer price sensitivity? (3) Is the effect of sales promotional variables stronger than the effects of television advertising variables on household brand choices? The purpose of this chapter is to summarize answers to these questions in the form of major contributions and managerial implications. This chapter is organized in four sections. In section 6.1, empirical and theoretical contributions are summarized. In section 6.2, the major findings and the managerial implications with respect to price sensitivity are commented on and summarized. In section 6.3, the limitations of the study are acknowledged and summarized. A set of suggestions for future research is given before the chapter is concluded. 6.1. CONTRIBUTIONS This research contributes both to the growing theoretical body of knowledge concerned with the effect of television advertising on a household's price sensitivity as well as to the empirical literature addressing this subject. Each of these contributions are discussed in detail below. 143 6.1.1. Theoretical Contributions In chapter III a theoretical model is developed by incorporating the effect of advertising on consumer brand choices. The theoretical structure, based on earlier work by Salop (1979) models both consumer brand choice and manufacturer behaviour. In addition, it incorporates the variables of price and television advertising and predicts the direction of consumer (household) price sensitivity as a function of these variables. Specifically, the behaviours of consumers loyal to the target brand, loyal to the competitive brand, and brand switchers are used to derive a demand function for the target brand. The derived demand function is such that the target brand's own price and the advertising of the competing brand negatively influence the target brand's demand. In addition, target brand demand is positively influenced by the brand's own advertising and the (increased) price of a competing brand. This demand function is then used to obtain the profit maximizing price of the target brand. The effect of television advertising on the profit maximizing price in Salop's generalized model is dependent on the proportion of consumers loyal to the target brand and a reaction function involving advertising intensity. More specifically, if two brands make advertising decisions independently of each other, the more advertising there is by either or both brands, the lower is price sensitivity. If, on the other hand, the competitive brand reacts to the advertising decisions of the target brand, then higher advertising may overcome advertising's direct effect and result in higher price sensitivity. Thus, in chapter III, it is demonstrated that the effect of advertising on a brand's price sensitivity depends upon the advertising reaction of competitive brands. This is an answer to the first question. 144 6.1.2. Empirical Contributions An IRI scanner panel database for the dry dog food and the aluminum foil product categories is used to assess household price sensitivity. Only the households (with related television exposure data) who made more than five purchases of dry dog food and two purchases of aluminum foil during the 88 weeks of the sample are used for estimating several alternative models. The sample used in estimating these models is generally larger than those previously available for sales-advertising models. The model parameter estimation results generally confirm the hypothesized signs for the variables in the models. Several important empirical contributions are commented on and summarized below. A Summary of Empirical Findings: The brand loyalty variable is found to be the best predictor of the brand choices. In addition, the effect of television advertising is found to increase brand price sensitivity. The empirical results also suggest that a higher brand price results in a lower likelihood of a consumer choosing that particular brand. Finally, the effects of television advertising, in-store display, presence of a deal, and deal amount variables have a positive influence on brand choices. The magnitude and statistical significance of these effects, however, are found to be dependent on the product category and broad generalization of the findings are not justified. Empirical Contribution I: This is the first study that measures the effect of television advertising on a household's individual brand choices at the micro level. McDonald's (1971) pioneering study analyzed the effect of television advertising exposures on individual brand choice in a competitive field setting. He, however, focused only on brand switching and excluded repeat purchases and did not control for other marketing variables such as price and dealing. McDonald also relied on self-reports of television viewing activity to derive his measure of advertising exposure. Tellis (1988) included various purchase level marketing 145 variables but television advertising exposures were only available as weekly aggregates for a household. In this study, the effect of television advertising on brand choices is found to be small, positive, and statistically significant when measured at the level of a household. The finding that television advertising variables account for only small variations in brand choices is consistent with the existing literature (Assmus, Farley and Lehmann 1984). Empirical Contribution II: In this study, increased television advertising is found to in-crease consumer's price sensitivity. This is found to be the case for both product categories studied. Although the television advertising variables account for a small proportion of the variation in household's brand choices, the parameters of these variables are statistically significant, and different from zero. As was noted, previous empirical work produced mixed results with, for example, Eskin and Baron (1977) reporting increased price sensitivity and Krishnamurthi and Raj (1985) reporting decreased price sensitivity. The present results, being obtained across two product categories and measured at a micro level while controlling for other variables as discussed below, is worthy of careful attention. It is interesting to compare the values of price sensitivity estimated here with those of studies reviewed in chapter II although there are difficulties in comparing the estimates of price sensitivity with those from earlier studies. In the present work, price sensitivity is measured for a household's brand choice. Earlier studies focused on brand sales (Prasad and Ring 1976, Eskin and Baron 1977, and Krishnamurthi and Raj 1985). In addition, the present work focuses on multiple brands competing in the same product market and accounts for competitive, effects. Earlier studies have examined single brands in multiple product markets, with less attention paid to the competitive effect (Woodside and Waddle 1975, and Eskin and Baron 1977). The price sensitivity (as result of advertising) estimates 146 in this research are generally lower than the previous research estimates. For example, Eskin and Baron (1977) reported that for three newly introduced brands, increases in price sensitivity of 1.35, 2.01, and 2.4 were observed when the advertising expenditure was increased by 170, 200, and 230% respectively. These estimates are compared to estimates in this study as follows: Consider a brand that has the regular price of 50^ per pound, and a base advertising level of two exposures per household. If the base level of advertising is increased by 200%, and estimates from Table 5.16 are used, then the expected increase in price sensitivity will be 1.25, 1.08, and 0.18. for Brand A, Brand B, and Brand E respectively. Neverthless, the effect of advertising is found to increase price sensitivity and is supported by empirical results in two product categories. Empirical Contribution III: The brand loyalty measure developed in this study weighs past brand shares inversely and exponentially with the number of days since a purchase occurred. This measure of brand loyalty is found to offer a better explanatory model than the measures based on brand shares (weighted only by purchase sequence - Guadagni and Little 1983). In addition, this measure allows one to vary weighting constants across households and across brands for increased sensitivity. Finally, the measure of brand loyalty variable introduced here appears to have a smaller influence on the magnitude of advertising's effect on household price sensitivity. Empirical Contribution IV: This is also the first study in which the effects of sales promo-tional variables and television advertising variables on household brand choices are both measured at the purchase occasion level. Earlier studies (Guadagni and Little, 1983 and Tellis 1988) reported that the sales promotional variables account for a substantial pro-portion of the variation in brand choices. This finding is supported by the brand choice models for the dry dog food product category. The effect of sales promotional variables, on .147 the other hand, accounts for only a small amount of variation in the brand choices for the product category of aluminum foil. The magnitude of the effect of television advertising variable is stronger for this product category. Other Empirical Contributions: This work found that estimating the effect of television advertising in the disaggregate ' form (i.e., by brand) provides more stable parameter estimates than using an aggregate form (response to advertising being the same across brands). A better explanatory model is also obtained through this method. Conceptually, this implies that the marginal utility of a brand advertisement varies across brands for an average household. As a result of this finding, one may conclude (to illustrate) that the advertising exposures from the Dog Chow brand are likely to produce a different household response than the same number of advertising exposures by Ken-L-Ration. This probably reflects the influence of differences in advertising copy quality and timing of the exposures. The study replicates several findings from previous scanner panel studies (Guadagni and Little 1983 and Tellis 1988). It is found that higher variable values of television advertising, display, feature, presence of a deal, deal amount, and brand loyalty result in a higher likelihood of a consumer choosing the brand. Moreover, a higher regular brand price results in a lower likelihood of a consumer choosing the brand. The Guadagni and Little study grouped several sales promotional variables into three variables and included no measure of television advertising activity. Tellis, on the other hand, included weekly household level television advertising exposures, but did not include the deal amount variable. Even with these differences, both studies reported findings similar to this study. That is, the brand loyalty is the most important variable in predicting household brand choices. In addition, the t-statistic for the price variable across all of these studies, and the present study, are comparable (Guadagni and Little —6.6, Tellis — 4 and the present 148 study —5.6). Tellis also reported that the t-statistic for the advertising variable was 2.2. In the present study, estimates for the comparable t-statistic is 1.72 (p < 0.09, Table 5.10). These findings reinforce our assertion that.the estimated models in this study are robust, meaningful, and comparable with the existing literature on variables measured by others. 6.2. MANAGERIAL IMPLICATIONS The purpose of this section is to present several managerial implications that follow from the present study results. First, the theoretical model is used to comment on brand differentiation. Then, several implications from the empirical work including the allocation of resources for sales promotional activities and television advertising programs are summarized. 6.2.1. Implications from Theoretical Model The theoretical structure of the market model is based on the assumption that adver-tising serves to make a brand more valued by consumers. While the impact of advertising is conditioned by the loyalty of a consumer to a brand (or its competitors), increased ad-vertising by one brand only will result in increased prices by an optimally behaving firm. However, when competitive reaction to a brand is incorporated in the model, then the re-sult of competitive reactions to a firm's advertising may, under certain derived parametric conditions, result in decreased prices in the market. These conditions, in informal terms, depend on the relative effectiveness and spending levels of competitors, as well as brand loyalty. The theoretical model reinforces a need to use advertising to help differentiate and to set as much as possible brand advertising policies that avoids competitive reactions. Such policies are, however, difficult to implement for two major reasons. First, to achieve meaningful brand differentiation via advertising activity, a long term commitment from 149 brand management is often required at the cost of short term profits (Lodish 1986). Sec-ond, if brand differentiation involves a mere advertising claim that the brand is different, then a competitor can also make similar advertising claims that "me-too" is different. As a result of these competitive actions, the model in chapter III suggests that both brands may experience increases in their price sensitivity measures. It is noted that competitive intensity is measured in terms of consumer perceptions, not advertising spending levels. 6.2.2. Implications from Empirical Work Effect of advertising on price sensitivity: Managers know that effective television advertis-ing attracts consumers but can be very expensive. What managers may not have realized is that increased efforts of television advertising can, in the short run, increase price com-petition among brands.. Thus, brand managers should weigh the benefits of attracting consumers and the likelihood of increasing price competition when making decisions about television advertising expenditures. Brand Loyalty: The brand loyalty variable is the best predictor of future brand choices for a household. This finding reinforces a need to develop a brand franchise for the long term profitability of the brand. In addition, it is found that the brand loyalty measure developed in this dissertation offers better explanatory models than the measures proposed in previous research. The brand loyalty measure developed in this research weighs past brand shares inversely and exponentially with the number of days since a purchase occurred. Aside from providing a better fit, the new measure resulted in a lower value for the coefficient of advertising. As advertising can have an important influence on brand loyalty, management should be aware of the interrelatedness of these two factors. Relative Impacts of Sales Promotion and Television Advertising: Managers also know 150 that sales promotion activities, such as deals and coupons, often increase brand sales at the (possible) cost of lost revenues. What managers may not have realized is that both sales promotional activities and television advertising can, in the short run, have the effect of increasing price competition among brands. Price-deals increase brand price elasticity by affecting price directly (and perhaps by attracting marginal customes - Beardon, Teel and Williams 1982), whereas the net effect of advertising may increase consumer price sensitivity among consumers. Disaggregated form of Advertising: It is intuitive to believe that the effect of television advertising varies, across advertised brands. What was not available before this research was completed is a method to understand whether advertising by a given brand in a competitive market is affecting brand choices at the household level. Modelling television advertising in the disaggregated form provides the direct method to disentangle the effect of advertising on household brand choices within a competitive environment. For example, three out of seven advertised brands have significant estimated parameters for television advertising variables for predicting household's brand choices. In addition, these brand coefficients are different across the brands. This suggests that such factors as copy quality influence the impact of advertising exposures. Resource allocations Implications: The effect of promotional deal variables is stronger than the effect of television advertising variables on brand choices for the product category dry dog food. Thus, there is a temptation to conclude that if there is a trade-off in allocating scarce resources between deals and television advertising, a greater amount of the resources should be allocated to deals. Without considering the costs (or lost revenues) associated with these activities, this conclusion is incorrect. Consider a simple choice of either to offer a 25c7 deal or to spend $25 per 1000 households on television advertising. If the reach of 151 advertising is 50%, then the average cost of television advertising per household reached is 5^. On the other hand, the 25c7 deal measures directly the lost revenue. Thus, at the margin, for a brand manager to be indifferent between these two decisions, the effect of a deal on brand choice should be five times larger than the effect of television advertising. In short, the present model should be cautiously used to allocate resources across the brand promotional activities. The effect of television advertising measured by this model, (and by Tellis 1988), on brand choice is small. For example, if advertising exposure is increased by one percent for a particular household, then the probability of that household choosing the advertised brand is increased by only 0.062% (the simple average of advertising elasticities for seven advertised brands). This, however, represents only the short term effect of television advertising. The long term effect, on the other hand, must consider the effect of a change in brand loyaltyf. The logit model, however, is non-linear in parameter estimates, and as Guadagni and Little pointed out, simulation is necessary to understand changes in brand loyalty that are associated with advertising. Thus, a small direct effect for advertising does not necessarily mean that advertising is not an important determinant of brand sales for these product categories. f If MSt is brand's market share in period t and at is advertising expenditure in period t, then for log-log model, log(MSj) = Alog(MSf_i) + /31og(at), the long term advertising elasticity is 8/(1 — A) and 8 is the short term advertising elasticity (Lambin 1976). 152 6.3. LIMITATIONS The purpose of this section is to comment on the limitations of this study. As is common in the literature, the issues of brand choice (Guadagni and Little 1983, and Tellis 1988), purchase timing (Morrison 1966), and purchase quantity (Neslin Henderson and . Quelch 1985) are treated seperately. While this thesis introduced new measures to reflect the timing of advertising impressions and past purchases, the focus throughout has been on brand choice. First, the limitations of the theoretical model are mentioned. Then, it is pointed out that the thesis models of price sensitivity and brand choice are primarily concerned with short term impacts. It is also noted that the present study does not consider any lead and/or lagged effects of sales promotional variables on brand choice. Finally, we mention that the estimated results for the models depend upon the product category investigated. One limitation of the theoretical model is that it does not provide a simple and direct procedure to perform statistical tests of its predictions, either at the individual or market level. Thus, with the present theoretical model, one is required to conduct several independent tests to demonstrate the consistency between the empirical data and the theoretical model. The procedures used here to demonstrate such consistency are, at times ad-hoc and less elegant than would be optimally desired. A second difficulty concerns the product category of aluminum foil. In this product category only one brand is advertised. The theoretical model for such a case predicts that increased advertising should result in a lower household price sensitivity. However, increased advertising resulted in higher price sensitivity for this product category. Thus, the theoretical model is partially disconfirmed. In other words, although the theoretical model generalizes Salop's (1979) model of brand differentiation, it is still not general enough 153 to accomodate such a case. In particular, the theoretical model does not indicate the necessary conditions to observe an increase in household price sensitivity when only one brand is advertised and its advertising efforts are increased. The empirical study also addresses the issue of the short term impact of promotions and advertising. In the short term, direct effects of display, deal amount, and television advertising are found to be beneficial to a brand. Thus, a brand is made more attractive to consumers as a result of these marketing actions. However, the long term effect of promotions may be related to a decline in consumer brand loyalty or reference price. For example, if a brand is on deal every four weeks, a consumer's perception of its image, and hence the reservation price associated with it, can be expected to decline over time. The models in this dissertation only provide estimates of the positive short term effect of promotion. The long term potential consequences on brand loyalty or reservation price, are not considered. A variable that is not included in the study, but is known to influence consumer choice, is coupon usage. Actual coupon used are included, but whether a household has possession of coupon for a competitor is not known. Since manufacturer and store coupons reduce purchase price for consumers, they affect consumer purchase decisions and consumer price sensitivity f. In addition, omission of a coupon usage variable can lead to statistical specification bias in the estimated parameters (Horowitz 1981). The problem, however, is larger than just the statistical bias. Manufacturer coupons are often distributed through magazines and newspapers with an expiry date of usually three months or less. Store coupons, on the other hand, are often distributed through feature-flyers and newspapers t The reference model was also estimated by replacing the price variable, defined as regular price, by a price variable defined as actual price paid by the household. The results obtained were virtually identical to those reported in chapter V. 154 and are valid for.an expiry date of usually a week or less. Thus, monitoring competitive couponing activity is a major data collection challenge that this study did not attempt to meet. In this study, the lead and the lag effects of deals and coupons are not considered in the brand choice model. It is very possible that some consumer decisions are affected by their anticipation of future deals. That is, some consumers may delay their purchases in anticipation of a good deal in the future. They may also stockpile product during the time a brand is on deal. To test the effects of such behaviour on brand choices, Guadagni and Little (1983) used the lagged promotional variable as an explanatory variable. They found that the effect of lagged promotional was small. Also, when lagged promotion was included the coefficient for current promotion was statistically significant. Thus, these variables may not contribute to statistical specification bias in the models estimated here. Although the results of the estimated models may vary by product category, the models and the procedures used in this dissertation are sufficiently general to accomodate a variety of consumer products. This is evident from the estimated models for two different product categories. One may, however, argue that the product categories of dry dog food and aluminum foil are relatively homogeneous and television advertising can achieve, at most, a low level of brand differentiation. As a consequence of this, increased levels of television advertising may have limited direct impact and so may result in higher household price sensitivity due to competitive factors. This may not be the case for a product category such as ready-to-eat cereals. In such a product category, the effects of advertising may be to increase brand differentiation such that increased television advertising results in a lower household price sensitivity. 155 6.4- FUTURE WORK The purpose of this section is to comment on useful future investigations regarding the effect of advertising on household price sensitivity that may be attempted. The first area of investigation concerns improving the theoretical model. The second area concerns demand price sensitivity with respect to package-size decisions. In addition, it may be possible to relate the effect of sales promotional variables to purchase frequency of the product category. It is reported in chapter II that some of the brand advertising efforts are directed towards informing consumers that each brand is unique. An increase in such advertising for a brand may result in a consumer better informed about the alternative brands. Furthermore, better informed consumers may have a higher price sensitivity than poorly informed consumers. The theoretical model proposed in chapter III does not include this aspect of advertising on consumer brand choices. Future research may use this focus to further understand how increased advertising efforts increase or decrease consumer price sensitivity. It is found in the empirical work that if the television advertising efforts are increased, then a household's price sensitivity increases. It is expensive to advertise on television, and a result of this expenditure by a brand and its competitors is increased price consciousness. Further, the more price conscious a household is, the more likely it is that more downward pressure is exerted on optimal prices. If such conditions prevail over a longer term, then a natural question one may ask is whether a brand should be advertised on television at all and to what consequence? In future work, a household's choice of package-size can be investigated. This can 156 be accomplished by proposing a theoretical model similar to the one reported in chapter III. Using the theoretical model of package size choice (an ordinal variable), the effect of advertising on package-size price sensitivity may be derived. The empirical investigation may then use the dry dog food database to judge the validity of the theoretical propositions. A comparison of parameter estimates for the product categories of dry dog food and aluminum foil, reveals that the effect of the sales promotional variables is stronger for the dry dog food than the aluminum foil. A natural question then arises: is there a link between purchase frequency and the effectiveness of promotional tools? There are at least two alternative approaches which may be followed to answer this question. In the first approach, one estimates brand choice models for other product categories with varying purchase frequencies, and then links purchase frequency with the effectiveness of promotional tools. In the other approach, one estimates brand choice models for heavy product users and light product users, and then compares the effectiveness of promotional tools across these groups. Long term modelling (using scanner panel data) to develop strategic implications for advertising and sales promotion management is an area mostly unexplored at present. One variable that may be used for this exploration is brand loyalty. In this dissertation, a limited attempt is made to bridge the knowledge gap between stochastic choice models and random utility choice models approach. This is accomplished by using interpurchase timing as a component in the operational definition of brand loyalty variable. 157 6.5. In Conclusion The major goal of this dissertation has been to improve our understanding of the effect of advertising on household price sensitivity. To reach this goal, an economic model which incorporates the variables of price and advertising is developed. In addition, a closely related statistical model is identified to measure the effect of television advertising on household price sensitivity. Finally, price sensitivity parameters from the the statistical model are estimated for the two product categories. For estimation purposes, household level scanner panel data that included television advertising exposures was used. It is found, in the context of the empirical investigation, that if a household is exposed to a higher number of television advertisements for a brand than other competing brands, the household becomes more price sensitive. This effect is accentuated if the competing brands react with more advertising efforts of their own. 158 Table 2.1 A Summary of S t a d i a s from Economics Study Product or serv ice Primary data sources Dependent var tables Independent v a r i a b l e s R e s u l t s r e l e v a n t to a d v e r t i s i n g S c h r o t e r , Smith and Cox 198/ Haas-Wilson 1986 Kwoka 1984 Feldman and Begun 1978, a l s o 1980. Marvel 1979. and Maur tz1 1972. Cady 1976. Benham and Benham 1975, and Benham 1972. RoutIne lega 1 serv Ices Optometr1c servIces Optometr1c servIces Eye examlnatIon Regular and premium grades of gaso 11ne P r e s c r I p t I o n drugs Eye g lasses Survey sampled 17 markets for law of f ices 280 o p t o m e t r i s t s sampled from 12 c i t i e s . 147 o p t o m e t r i s t s sampled from 7 c t t l e s . A sample of 1195 optometr i s t s c o l l e c t e d in 1976 Monthly cross s e c t i o n and time s e r i e s from 10 m e t r o p o l i t a n areas for 1964 to 1971. Pr ice quoted for each three r o u t i n e l e g a l serv ices P r i c e , procedures used and prescr i p t I o n P r i c e and time 11 me spent on at tend ing to pat lent P r i c e of eye exam inatIon P r i c e of r e g u l a r and premium grades of g a s o l i n e 1848 drug s tores P r e s r l p t i o n information about drug p r i c e s , p r i c e s and amount s o l d for 10 common drugs. A n a t i o n a l sample of 1625 In 1970 and 291 In 1960 I n d i v i d u a l s who had eye check-up. P r i c e s for eye g lasses and eye examlnat i o n . ( F i r m ' s a d v e r t i s i n g •Market a d v e r t i s i n g • F i r m s i z e •Number of lawyers per c a p i t a •Media a d v e r t i s i n g observed, tNumber of d o c t o r s In the c i t y , •Per c a p i t a Income, •Input r e s t r i c t i o n . • S e v e r a l l e v e l s of a d v e r t i s i n g , •Number of d o c t o r s In the c i t y , •Per c a p i t a Income. • S e v e r a l measures of q u a l i t y of s e r v i c e , • A d v e r t I s l n g r e s t r 1 c t 1 o n ( 0 . 1 ) • P r i c e a d v e r t i s i n g r e s t r i c t Ion. • V a r i a b l e s about r e g i o n a l character 1 s t l e s •Drug s t o r e character 1st i c s • A d v e r t i s i n g r e s t r Ict1on(0.1) •Fami ly demograph i c s , • D o c t o r ' s profess iona 1 a f f l l a t l o n , • A d v e r t i s i n g r e s t r l e t 1on(0, 1) Market a d v e r t i s i n g had negat ive impact on p r i c e s . A d v e r t i s i n g serves to increase market c o m p e t i t i o n . Media a d v e r t i s i n g by o p t o m e t r i s t s r e s u l t e d In lower p r i c e s by 28-33%, w i t h c o n t r o l s for q u a l i t y . A d v e r t i s i n g c i t i e s r e p o r t e d lower p r i c e s . C i t i e s w i t h n o n a d v e r t i s i n g showed Improved q u a l i t y of s e r v i c e as measured by •ye examination t ime. R e s t r i c t i o n on a d v e r t i s i n g r e s u l t e d In higher p r i c e s by about I U . P r i c e a d v e r t i s i n g r e s t r l e t Ion had no Impact on p r i c e s of g a s o l i n e . R e s t r i c t i n g p r i c e a d v e r t i s i n g Increased p r i c e s on average by 2.9%. P r i c e s were lower by about 25% In s t a t e s where a d v e r t i s i n g was not r e s t r i c t e d . Table 2.2 A Summary of Studies from Marketing Study Product or Primary Dependent Independent Results relevant service data sources variables variables to advertising Krishnamurthi and Raj 1985. Eskin and Barron 1977. Wittink 1977. Prasad and Ring 1976. Woodside and Waddle 1975. Frequently purchased consumer product Four new consumer products: one cleaning item, and three food i terns ADTEL s p l i t cable sample data with 320 weekly diar ies for 52 control and 24 test weeks Store level data collected for 24 weeks. Frequently purchased branded products Packaged food i tem Coffee Metropolitan area sales data.' Purchase d i a r i e s with s p l i t cable 1 ike arrangment for 64 weeks. Store level data. Quant i ty purchased and implied consumpt ion rate. Number of packages sold norma 1ized by a measure of store s i z e . Monthly market share data. Market share for 64 success ive weeks. Unit sales •Relative p r i c e s , .Family s i z e , .Total grocery purchase, and •Total TV exposure. Experimental design involving two levels of advertising, and pr ices and other factors of point-of-sale advertising •Relative p r i c e s , • Relat ive advert i s i n g . • Relat ive pr ices •TV r a t i n g s , •Newpaper adv. •Magazine adv. • Exper imenta 1 design involving two levels of point-of-sales, two of couponing. Higher advertising resulted in lower absolute values of pr ice e last ic i ty. Across a l l four products, higher advertising resulted in higher absolute values of price e l a s t i c i t i e s . Higher advertising resulted in higher absolute values of pr ice elast ic i ty. Higher TV advertising resulted in lower absolute values of market share price e l a s t i c i t y . Stores with higher point-of-sale advertising tended to have higher absolute values of pr ice e l a s t i c i t y . T a b l e 2.3 P r i c e E l a s t i c i t y a t D i f f e r e n t A d v e r t i s i n g L e v e l s ( t - s t a t i s t i c s i n p a r e n t h e s e s , u n l e s s s p e c i f i e d ) S t u d y R a t i o o f h i g h t o low a d v . e x p e n s e s A d v e r t i s i n g l e v e l H i g h Low E f f e c t o f i n c r e a s e i n a d v . on p r i c e s e n s i t i v i t y K r i s h n a m u r t h i a n d R a j ( 1 9 8 5 ) 2: 1 F a r r i s a n d n a R e i b s t e i n ( 1 9 8 0 ) E s k i n a n d B a r o n ( 1 9 7 7 ) W i t t i n k ( 1 9 7 7 ) P r a s a d a n d R i n g ( 1 9 7 6 ) E s k i n ( 1 9 7 5 ) W o o d s i d e a n d W a d d l e ( 1 9 7 5 ) M a s s y a n d F r a n k ( 1 9 6 5 ) 2: 1 B r a n d A 1.7:1 B r a n d B 2.3: 1 B r a n d C 2: 1 B r a n d D n a 2:1 2: 1 -0.790 ( 1 1 . 6 ) -1.330 ( 1 5 . 5 ) 2.280* ( 6 . 4 1 ) -2.174 -0.160 ( 2 . 6 8 ) -2.167 - 0 . 8 1 5 ( 2 . 2 6 ) -3.810 - 1 .410 ( 1 . 7 8 ) -0.729 -0.975 ( n s ) - 3 . 3 4 2 * * ( 7 . 0 2 ) -0.00632+ -0.00722+ ( p = 0 . 0 0 1 ) . ( p = 0 . 0 3 ) -2.088 - 0 . 9 7 5 ( 6 . 6 5 ) -1.678 - 1 . 0 9 5 ( 4 . 9 5 ) -1.620++ -3.350++ ( 3 . 6 4 ) ( 2 . 4 4 ) d e c r e a s e d N/A i n c r e a s e d i n c r e a s e d i n c r e a s e d d e c r e a s e d N/A i n c r e a s e d i n c r e a s e d N/A * I n t h i s s t u d y e f f e c t o f a d v e r t i s i n g was m e a s u r e d o n r e l a t i v e p r i c e . * * I n t h i s s t u d y a d v e r t i s i n g was v a r i e d a c r o s s d i f f e r e n t s a l e s t e r r i t o r i e s . + T h e s e a r e r e g r e s s i o n c o e f f i c i e n t s when m a r k e t s h a r e was r e g r e s s e d o n t e l e v i s i o n a d v e r t i s i n g t i m e s p r i c e a n d s e v e r a l o t h e r p r e d i c t o r v a r i a b l e s •+ T h e e s t i m a t e u n d e r c o l u m n s h i g h a n d low a d . l e v e l s a r e p r i c e e l a s t i c i t y m e a s u r e s f o r b r a n d l o y a l a n d b r a n d s w i t c h e r c o n s u m e r s r e s p e c t i v e l y . 161 Table 3.1 Consumer Segments and their Responses to Selected Marketing Mix Variables. Behavioural Pattern Loyal to the target brand* (0) Brand comparison Limited. Loyal to competitor brand + (B) Limited. Brand switchers (0 = 1-0-3) Extensive. Price Brand Advertising Price deals and Coupons Consumers may be willing to pay higher price to the target brand than to the competitive brands. The target brand advertising may be used for post-purchase brand evaluation. Consumers may be willing to pay higher for the competitive brands than for the target brand. The target brand advertising may be ignored. Consumers highly sensitive to price. The target brand advertising may be used for brand comparison. Consumers may be Consumers may Highly prone to willing to use the target brand coupons and deals ignore the target brand deals and coupons if dollar amount saved is not attractive. deals and coupons. * The brand about which we investigate through a model is called the target brand. + All other brands are considered the competitive brands. 162 Table 4.1 Purchase Share of Major Brands i n Dry Dog Food Category Brand Category Largest brand Purchase share Purchase share i n category of t o t a l dry of the dog food* e s t i m a t i o n sampl A d v e r t i s e d brands Dog Chow Puppy Chow K i b b l e s - N - B i t s N e s t l e Purina - others Quaker - others Adv. others T u f f y s C y c l e l - 4 Purina - others A l l others Dog Chow 11.12 Puppy Chow 6.48 K i b b l e s - N - B i t s 8.40 New Breed 8.65 P r a i s e 18.14 Love-me-tender 10.52 Alpo 4.32 Non-advertised brands T u f f y s 10.96 C y c l e l - 4 7.45 Mainstay 6.99 Generic 6.97 13.47 4.18 8.74 9.41 20.10 7.42 3.40 12.92 7.14 6.36 6.95 * These are based on the t o t a l purchases of the dry dog food product category. Table 4.2 Incidence of Sales Promotion A c t i v i t i e s by Product Product Category Promotion type Aluminum f o i l P l a s t i c wrap Canned dog food Dry dog food % of purchase on coupon 5.22 % of purchase during temporary price cut 9.19 % of purchase with feature 0.70 % of purchase with display 1.13 % of purchase with one or more 14.40 sales promotion 10.25 12.22 1.49 5.50 24.10 0.66 7.04 0.81 1.29 8.85 28.12 8.13 1.68 2.53 35.47 Total purchases 8597 5221 17009 14728 Table 4.3 Expected Signs f o r Independent V a r i a b l e s i n Brand Choice Model Independent V a r i a b l e Observed V a r i a b l e Expected s i g n P r i c e T e l e v i s i o n a d v e r t i s i n g D i s p l a y ( l , 0 ) Feature(1,0) I f brand on deal(1,0) Deal amount Brand L o y a l t y Regular p r i c e Number of ad exposures w i t h i n a purchase i n t e r v a l Number of ad exposures f o u r weeks p r i o r t o purchase Number of ad exposures e i g h t weeks p r i o r t o purchase D i s p l a y Feature Deal Deal amount E x p o n e n t i a l l y weighted and updated p a s t purchases E x p o n e n t i a l l y weighted f i r s t 32 weeks purchases Smoothing of p a s t purchase + + + + + + + + P r i c e x T e l e v i s i o n a d v e r t i s i n g i n t e r a c t i o n P r i c e x T e l e v i s i o n a d v e r t i s i n g 165 Category Table 4.4 Demographic Comparison - Family s i z e Observed p e r c e n t p e r c e n t i n p e r c e n t i n count of t o t a l Cong, d i s t r i c t the s t a t e One Two Three Four F i v e Six or more 319 757 407 453 185 62 14.6 34.7 18.6 20.8 8.5 2.8 21.71 31.39 16.25 16.17 8.65 5.82 22.40 31.25 16.46 15.73 8.42 5.74 mean f a m i l y s i z e T o t a l v a l i d o b s e r v a t i o n 2.82 2183 2.77 2.80 • Table 4.5 Demographic Comparison - Family Income Category Observed percent p e r c e n t i n p e r c e n t i n count of t o t a l Cong. d i s t r i c t the s t a t e l e s s than $10,000 298 13.83 32 • 79 26.65 $10,000 - $14,999 309 14.34 17 .63 15.02 $15,000 - $19,999 265 12.30 15 .50 15.08 $20,000 - $24,999 317 14.71 12 .62 13.93 $25,000 - $34,999 496 23.02 13 .13 'r'"? ~~-r • 17.22 more than $34,999^ 470 21.81 8 .33 12.10 T o t a l v a l i d o b s e r v a t i o n 2155 166 T a b l e 4.6 D e m o g r a p h i c C o m p a r i s o n • • H o u s e t y p e C a t e g o r y O b s e r v e d c o u n t p e r c e n t o f t o t a l p e r c e n t i n C o n g , d i s t r i c t p e r c e n t i n t h e s t a t e S i n g l e d e t a c h e d 1810 8 2 . 9 5 7 5 . 8 3 6 9 . 15 D u p l e x 166 7.61 6.97 11.29 M u l t i p l e d w e l 1 i n g 162 7.42 11.18 16 . 39 M o b i l e home 44 2.02 6.02 . 3. 16 T o t a l v a l i d o b s e r v a t i o n 2 182 T a b l e 4.7 D e m o g r a p h i c C o m p a r i s o n - Home o w n e r s h i p C a t e g o r y R e n t e d Ownwer o w n e d O b s e r v e d c o u n t 352 1827 p e r c e n t o f t o t a l 16. 15 8 3 . 8 5 p e r c e n t i n p e r c e n t i n C o n g , d i s t r i c t t h e s t a t e 2 7 . 2 5 7 2 . 75 31 .77 6 8 . 2 3 T o t a l v a l i d o b s e r v a t i o n 2 1 7 9 T a b l e 4.8 D e m o g r a p h i c C o m p a r i s o n - E d u c a t i o n C a t e g o r y E l e m e n t a r y o n l y Some h i g h s c h o o l H i g h s c h o o l c o m p l e t e Some c o 1 l e g e O b s e r v e d c o u n t 2 1 5 272 1364 568 F o u r o r m o r e y e a r s o f c o l l e g e o r u n i v e r s i t y 849 p e r c e n t o f t o t a l 6.58 8. 32 41 .74 17.38 25.98 p e r c e n t i n p e r c e n t i n C o n g , d i s t r i c t t h e s t a t e 21 .52 10.57 39 .97 14.24 13.69 17 .96 12.43 4 0 . 42 14. 36 14.83 T o t a l v a l i d o b s e r v a t i o n 3268 167 Table 4.9 Demographic Comparison - Occupation Category-Managerial and p r o f e s s i o n a l occ, Observed count 1071 T e c h n i c a l , s a l e s and admin, support 619 S e r v i c e o ccupations 174 Operators, f a b r i c a t o r s and l a b o r e r s 396 P r e c i s i o n p r o d u c t i o n c r a f t s 118 p e r c e n t of t o t a l 45.04 26.03 7.32 16.65 4.96 pe r c e n t i n per c e n t i n Cong, d i s t r i c t the s t a t e 20.45 25.89 16.51 20.63 16.52 21.23 28.99 14.89 22.11 12.78 T o t a l v a l i d o b s e r v a t i o n 2378 Category White A l l other Table 4.10 Demographic Comparison - Race Observed p e r c e n t p e r c e n t i n p e r c e n t i n count of t o t a l Cong, d i s t r i c t the s t a t e 2167 5 99.77 0.23 99.10 0.90 94.42 5.58 T o t a l v a l i d o b s e r v a t i o n 2172 168 Table 4.11 Household Wrap Brand Sales Comparisons based on Market Share Estimates Based on the household Store L e v e l Purchase Brand IRI Purchase Sample Units * Volume* Amount** U n i t s Volume Amount Reynolds 34.07 22.10 39.08 34.64 22.31 38.77 Generic++ 13.82 11.40 9.66 14.54 11.40 9.88 Handi Wrap 8.71 18.59 9.57 8.31 18.49 9.20 Glad Wrap 9.85 17.07 8.78 9.44 16.64 8.57 Saran Wrap 6.43 6.22 8.40 6.89 6.88 9.00 P r i v a t e Lbl++ 6.87 4.73 6.45 7.84 5.61 7.35 20 Below 1.81 2.14 4.20 1.88 2.46 4.56 Papermaid 0.78 0.72 1.50 0.90 0.87 1.84 Freezer paper 1.12 0.67 1.24 1.10 0.69 1.21 Diamond 2.08 0.57 1.04 2.02 0.57 1.00 Bes Pak 0.06 0.05 0.06 0.04 0.03 0.04 Hefty 0.04 0.03 0.04 0.06 0.04 0.08 Wax papers 14.37 15.70 10.06 12.34 14.01 8.56 * U n i t s are measured i n terms of number of packages s o l d . + Volume i s measured i n terms of number of f e e t of wrapping paper s o l d . ** Amount is.measured i n terms of d o l l a r revenue to the s t o r e s . +• These brands may c o n t a i n other product v a r i e t i e s i n c l u d i n g aluminum f o i l , p l a s t i c wrap, wax paper and f r e e z e r wrap. T a b l e 4.12 D o g F o o d B r a n d S a l e s C o m p a r i s o n s b a s e d o n M a r k e t S h a r e E s t i m a t e s B a s e d o n t h e h o u s e h o l d S t o r e L e v e l P u r c h a s e B r a n d I R I P u r c h a s e S a m p l e U n i t s * V o l u m e * Amount* * U n i t s V o l u m e Amount D o g Chow 2 .89 10 .25 9 .17 3 .77 10 .68 10 .99 A l p o 21 .95 7 .52 8 .92 20 . 14 4 .93 6 .66 T u f f y s 2 . 18 13 .12 8 .75 4 .29 22 .65 17 .27 P u p p y Chow 1 .67 4 .85 5 .45 2 .58 5 .59 7 .21 G a i n e s B u r g e r s 2 .44 2 :20 4 • 1 9 1 .62 0 .94 2 . 13 G e n e r i c 3 47 6 .65 4 . 16 4 .47 9 .23 4 .93 V e t s 14 .13 5 . 14 3 . 34 19 .40 5 . 10 3 .79 K e n L R a t i o n 3 .92 3 .45 3 .26 4 .28 2 .61 2 .82 K i b b l e s B B 1 .49 1 .96 2 .89 1 .09 1 .06 1 .91 M i g h t y D o g 9 .85 1 .11 2 .84 7 .26 0 .54 1 .58 H i P r o t e i n 0 .89 2 .82 2 .72 1 . 17 2 .95 3 . 33 C h u c k Wagon 0 .82 2 .40 2 .54 0 . 74 1 .43 1 .83 P r a i s e . 1 . 13 1 .71 2 .47 0 .90 1 .02 1 .86 G r a v y T r a i n 0. .88 2 . 18 2 .27 0 .90 1 .84 2 . 30 C y c l e 4 5, .07 1 .72 2 .24 2 .66 0 . 72 1 . 10 Come N G e t i t 0. .97 1 .83 2 . 18 0 .84 1 .17 1 .70 M a i n s t a y 0. .87 3 . 10 2 .17 0 .99 2 .46 2 .05 G a i n e s T o p 1 .46 1 .23 2 . 15 0 .93 0 .53 1 . 10 K a l K a n 3. .87 1 .78 2 .13 4 .29 1 .24 1 . 79 C y c l e 2 1 . 74 1, .69 2. 0 2 1 , .84 1 . .35 1 .77 L o v e Me C h u n k s 1 . .04 1 .42 1, .98 0. .71 0 .67 1 . 18 New B r e e d 1 . . 30 1. .92 1 . .94 0. . 79 0 .84 1 , .48 K e n L R a t i o n B&B 0. 99 1 11 1. 85 0. 54 0 . .41 0 , 79 M e a t y M e a l 0. 73 2 . 14 1. .82 0. ,69 1, ,50 1 . .51 K e n L R a t i o n Chk 0. 81 1 . .57 1. , 75 0. . 77 1. 20 1 . .58 B u t c h e r s B l e n d 0. 79 1 . . 36 1 . . 75 0. 81 1. ,08 1 , .71 F i t N T r i m 0 . 45 1 . 18 1. 31 0. 45 0 . 78 1 . 04 F r i s k i e s 4 . 06 1 . 05 1 . 22 1. 91 0 . 32 0 . 44 C y c l e 3 1 . 43 0. .90 1 . 18 1. 52 Q. 45 0. 69 K i b b l e s & B i t s 0 . 40 1 . 04 1 . 16 0 . 35 0 . 72 0 , 95 C y c l e 1 1 . 12 0. 94 1 . 13 1. 57 1. 10 1. 48 M o i s t & C h u n k y 0 . 76 0 . 90 1 . 10 0 . 72 0 . 63 0 . 87 S e a D o g 0 . 66 0 . 94 1. 08 0 . 34 0 . 38 0 . 60 K e n L R a t i o n C u t 0 . 64 0 . 38 0 . 91 0 . 52 0 . 21 0 . 61 P r i v a t e L a b e l 0 . 37 1. 62 0. 81 0 . 73 2. 52 1. 38 M o r t o n 2 0 . 17 1. 87 0. 73 0 . 59 4 . 39 2. 02 K e n L R a t i o n B r g 0. 32 0 . 33 0 . 64 0 . 21 0 . 15 0 . 35 A l l o t h e r b r a n d s 2. 31 2. 78 2. 04 2. 66 4 . 70 3. 32 * U n i t s a r e m e a s u r e d i n t e r m s o f number o f p a c k a g e s s o l d . + V o l u m e i s m e a s u r e d i n t e r m s o f p o u n d s o f d o g f o o d s o l d . * * Amount i s m e a s u r e d i n t e r m s o f d o l l a r r e v e n u e t o t h e s t o r e s . 170 Table 5.1 A B a s i c Sequence of A l t e r n a t i v e Modet Spec I f Ica t Ions for Dry Dog Food Brand Choice Models Maximum L i v e l i h o o d Parameter E s t i m a t e s (Asymptotic t - s t a t l s t i c In parentheses) Independent var(ab le Model with brand constants Brand loya My inc luded Promot ional A d v e r t i s i n g P r i c e v a r i a b l e s v a r i a b l e v a r i a b l e added Reference Mode 1 added added Brand l o y a l t y exc luded Adver t i s I n g v a r l a b l e s exc luded Ml M2 M3 M4 M5 M8 M7 MB P r i c e -0 .0444 -0 0389 -0 .0431 -0.0438 ( -6.71) 1-5.64) (•; 7.24) ( 8 621 Telev Is1on -0 00803 -0 .00980 0 .0873 0 .08 10 adver t Is Ing ( 0.91) ( • 1 1 1 ) (2.25) (2 57) D i s p l a y ! 1 , 0 ) 1 114 1 . 109 1 .243 1 . 173 0 .733 1 120 (4 02) (4 .00) (4 431 (4 . 16) 1 3 01) (4.28) F e a t u r e ! 1 , 0 ) 0 .0195 0 .0379 0 . 149 1 688 0 971 1 691 (0 041 (0.09) 1 0 .'361 1 1 .76) < 1 .29) (1.771 If brand Is on 0 999 0 999 0 868 0 886 1 046 0.878 dea l t 1.0) (6 .01) (6 .03) (5 .29) (5 .30) (7 .86) 15.22) Deal i amount 0 226 0 225 0 .234 0 .238 0 . 198 0.236 (9 .61) (9 .62) (9 .85) (9 85) 110.1) (9.88) Brand l o y a l t y 4. 394 4 .582 4 .564 4 .587 4 .580 4.573 (38.4) (35.7) 135.7) (35.3) (35.3) (35.3) Pr Ice x T e l e v i s i o n -0 00183 -0 00158 a d v e r t I s l n g 1 2.53) 1 2.65) P r i c e K Feature -0.0484 -0.0282 -0.0505 1 1.98) 1 1 39) 1 2.06) Brand constant Brand A 0 873 0 256 0 319 0 335 0 469 0 407 0 960 0 444 (6. IB) (1.49) (1 76) 11 .84) (2 55) (2 19) 16 33) (2.42) Brand B 0. 331 -0.785 -0 665 -0 .648 -0 . 145 -0 208 0 339 -0.171 1 - 1.80) 1 3 58) 1-2 84) 1 2 .76) 1 0 58) (-0 83) 1 1 67) 1-0 69) Brand C 0.482 0.513 0 249 0 269 1 289 1 227 1 .296 1.268 (3.20) (3.01) ( 1 32) ( 1 41) (5 23) (4 96) (6 07) (5 17) Brand 0 0.533 0. 305 0 0488 0 .0498 0 BB7 0 792 1 103 0 876 (3.57) (1.82) (0. 26) to 27) (3 94) (3 48) (5 571 (3.89) Brand E 1 .275 0 582 0. 295 0 361 1 . 003 0 935 1 64 1 0.924 Brand (9.49) (3.65) (1 78) (2 001 (4 83) (4. 53) 19 38) 1 4 82) F 0.363 0 783 0 479 0 497 1 IB3 1 120 0 708 1. 157 (2.34) (4.50) (2. 541 (2 62) (5. 49) (5. 17) (3 67) (5.39) Brand G 0. 331 0.00522 •0 . 0979 -0 . 06 38 0 210 0. 126 -0 290 0 167 1 - 1.80) (0 26) 1-0. 17) ( -0 28) (0 90) (0. 54) (- 1 4?) (0.73) Brand H 0.B37 0.276 -0 . 0299 •0. 0328 -0 . 0194 0 00630 0. 508 -0 0135 (5.89) 11.65) 1 -0 17) 1-0. 18) 1 0. 1 1 ) ( -0 .04) (3 36) ( 0 08) Brand I 0. 180 -0 248 -0 . 184 -0 . 187 0 982 -0 908 -0 . 577 -0 .995 ( 1.12) ( -1.31) ( 0 93) (-0. 94) 1 4 28) 1 -3 . 89) 1-2. 94) ( 4.32) Brand J 0.259 0. 103 0. 0465 0. 0465 0. 497 0. 436 0. 619 0 490 I 1 .64) 10.55V to 23) (O. 23) 12 34) 12. 04) 13 47) (2 30) Brand K 0 0 0 0 0 0 0 0 rho -squared 0 0 309 0 482 0. 482 0 490 0. 492 0. 184 0 491 Log- 1 Heel Ihood -2914.3 -2015.1 - 1510.1 -1509.6 - I486 2 -1480.9 -2377.9 - 1484.6 Table 5.2 E f f e c t of A l t e r n a t i v e Measures and A l t e r n a t i v e Forms of A d v e r t i s i n g on Dry Dog Foocf Brand Choice Models Maximum L i k e l i h o o d Parameter Es t imates (Asymptot ic t - a t a t l s t l c In parentheses) Independent var1ab le A d v e r t i s i n g Exposure Dura t ion Interpurchase 4 weeks be fore Bweexs be fore D lsaaggregate form of a d v e r t I s I n g t ime purchase ' purchase v a r l a b le M6 M6-4W M6-8W M6-D7 Pr Ice -0 .0369 -0 .0410 -0.0389 -0 0391 (-! 5.64) ( -6.01) ( -5.68) ( 5 . 4 8 ) T e l e v I s l o n 0 .0873 0 . 138 0. 126 see below adver t I s ing (2.25) ( 1 .48) (2.40) D i s p l a y ! 1 . 0 ) 1 . 173 1 181 1 . 174 1. 153 (4.16) (4 .201 (4.17) (4.03) F e a t u r e ! 1 . 0 ) 1 688 1 682 1.672 1.611 ( 1 .78) ( 1 . 73) (1.74) II 661 If brand Is on 0 .885 0 .887 0 885 0 893 d e a l ! 1 . 0 ) (5.30) (5 .28) 15.25) (5.23) Deal amount 0 .238 0 .236 0 238 0 238 (9.85) (9 .90) (9.87) 19.80) Brand l o y a l t y 4 .580 4 .584 4.581 4.595 (35.3) (35.3) (35.3) (35.0) P r i c e x T e l e v i s i o n -0 .00183 -0 .00325 -0.00284 see below adver t I s Ing ( 2.53) ( -1.84) I 2 67) P r i c e x Feature •0 .0484 -0 0482 -0.0479 -0 .047S ( - 1.98) ( - I 1.951 (-1.95) ( -1 .92) Brand Advert IsIng P r i c e x adv. Brand constant constant parameters Brand .A 0 .407 0 431 0.406 0 400 0.714 -0.0168 (2 19) (2 331 (2 19) ( 1 921 (3.16) ( -3 .14) Brand e -0 .208 -0 183 -0 211 0 0698 0 967 -0 .0219 (-0 .631 ( 0 73) ( 0 841 (0 25) (2. IB) ( -2 .45) Brand c 1 227 1 271 1 234 1 . 189 0.0921 -0.00173 (4 981 (5 16) (5.00) (4 38) 10.25) ( -0 .31) Brand D 0 792 0 821 0. 790 0 865 0 .492 -0.00633 (3 48) (3 62) (3.48) (2 78) (0 .87) (-0.671 Brand E 0 935 0 975 0 926 0 964 0. 108 -0.00231 (4. 53) (4. 86) (4 54) (4 58) (2.53) ( -2 .88) Brand F 1, 120 1 153 1 121 0 .995 -0 .240 0.00442 (5. 17) (5 35) (5.19) (4 26) ( -1 .32) (1.52) Brand G 0 126 0. 170 0. 12B 0 103 0.218 -0.00461 (0. 54) (0. 73) (0.55) (0 391 ( 1 601 ( 1 . 5 1 ) Brand H -0.00630 -0 146 -0.00687 0 00686 (-0.04) ( 0. 08) ( -0.04) ( -0 . 04) Brand I -0 . 906 -0 . 948 -0.910 -0 914 ( -3. 89) ( -4. 081 (-3.90) ( 3 88) Brand J 0. 436 0 459 0. 4 38 0 425 (2. 04) (2. 15) (2.05) ( 1 96) Brand K 0 0 0 0 0 0 r h o - s q u a r e d 0.492 0. 492 0.492 0 497 Log-11ke 1thood • 1480.9 • 1481.8 1480.8 -1466.3 Table 5.3 Effect of Alternative Specification! with Dlssggregated Fore of Advertising on Dry Dog Food Brand Choice Models Mailaua likelihood Pareaeter Eatlaates (Asyaptotlc t-statisttc In parentheses) Independent variable Model elth brand constsnts HI Brand loyalty Included M2 Proeot 1onal variables added H3 Advertising variables added M4-0 Price variable edded M5-D Reference Model ME-0 Brand loyalty excluded M7-0 Advert lain variables excluded MS Price — — — — -0.0445 (-6.73) •0.0376 (-5.52) •0.0410 (-7.02) -0.043S (-6.62) Brand A ad. exposures — — — O.OO027B (0.07) 0.00202 (0.05) 0.724 (3.23) 0 (2 .484 .87) — Brand B ad. exposures — — — -0.146 (-2.06). -0.138 (-1.85) 0.656 12.12) 1 (2 .115 .97) — Brand £ ad. exposures — — — -0.0109 (-1.17) -0.0147 (-1.57) 0.107 (2.55) 0 (3 .104 .08) — Olaplayd.O) — — 1. 114 (4.02) 1.126 (4.21) 1.250 (4.46) 1.162 (4.10) 0 (2 .730 .99) 1. 120 (4.26) Feature!1,0) — — 0.0195 (0.04) 0.0379 (0.09) -0.140 1-0.34) 1.637 (1.68) 0 11 .938 .23) 1.661 (1.77) If brand 1s on deeld.O) — — 0.999 18.01) 1.011 (6.12) 0.682 (5.33) 0.689 (5.23) 1 (7 .029 .68) 0.676 (5.22) - Oeal aeount — — 0.226 (9.61) 0.225 (9.61) 0.234 (9.85) 0.238 (9.811 0.200 (10.1) 0.236 (8.66) Brand loyalty — 4.394 (38.4) 4.562 (35 7) 4.566 (35.7) 4.576 135.3) 4.593 (35.1) — 4.573 (35.3) Price x Brand A ad. exposures — — — — — -0.0169 1-3.22) -0 {• .0119 2.95) — Price x Brand B ad. exposures — — — — • — -0.0217 (-2.41) -0 (-• .0252 1.05) — Price x Brand E ad. exposures — — — — — -0.00231 (-2.93) -0 (-.00204 3. 17) — Price x Feature — — — — • — •0 0481 (-1.941 -0 (-.0279 1.36) -0.0505 (-2.06) rho-squared 0 0.309 0.462 0.483 0.491 0 495 0 . 189 0.491 Log-1Ifcelihood •2914 .3 -2015.1 -1510.1 -1507.6 -1483.6 -1470.7 -2362.8 -1484 6 Grand constant Brand A 0. 673 0.256 0 .318 0 .261 0 .437 0 404 1 .013 0.444 16 16) (1.49) (1 .76) (1 .37) (2 . 10) I 1.94) (6 . IE) (2.42) Brand B -0. 331 -0.785 -0 .665 -0 .396 0 112 0.0566 0 .606 -0.171 (-1. BO) 1-3.66) (-2 .64) (-1 .51) (0 .40) (0.20) (5 .53) <-0.69) Brand C 0. 462 0.S13 0 .249 0 .223 1 264 1. 100 1. 163 1.266 (3. 20) (3.01) (1 .32). (1 . 17) (5 16) (4.44) (5 .28) (5.17) Brand 0 0. 533 0.305 0 .0486 0 .0140 0. 877 0 727 1 043 0.876 (3. 57) 11.82) (0 .26) (0 .06) (3 89) (3. 20) (5 .29) (3.89) Brand E 1. 275 0.562 0. .295 0 .358 1. 043 0.949 1 . 615 0.924 (9. 49) (3.85) (1. 78) (1 97) (5. 02) (4.56) (9. 221 (4 821 Brand F 0. 363 0.763 0 479 0 450 1. 157 1.035 0. 644 1 . 157 (2. 34) (4.50) (2. 54) (2. 36) (5. 40) (4.78) (3. 36) (5.39) Brand G -0. 331 0.00522 -0. 09 79 -0 123 0. 168 0. 112 -0. 270 0. 167 (-1. 80) 10.26) 1-0 17) (-0 55) (0. 74) 10.49) (-1. 35) (0 73) Brand H 0. 637 0.276 -0. 0299 -0. 0584 -0. 0229 -0.00707 0. 510 -0.0135 (5. 69) (1.65) (-0. 17) (-0. 32) (-0. 13) (-0.04) (3 37] <-0.08) Brand I 0. 160 -0.246 -0. 164 -0. 212 -0. 967 -0 666 -0. 545 -0.995 (1 . 12) (-1.31) 1-0. 93) (-1. 07) (-4. 30) (-3.62) (-2. 80) (-4.32) Brand J 0 259 0. 103 0. 0465 0. 0228 0. 498 0.410 0 . 564 0 . 490 (1. 64) (0.55) (0. 23) (0 11) (2. 34) (1.92) (3. 29) (2.30) Brand K 0 0 0 0 C 1 0 C l 0 173 T a b l e 5.4 E f f e c t o f A l t e r n a t i v e M e a s u r e s o f B r a n d L o y a l t y o n D r y Dog F o o d B r a n d C h o i c e M o d e l s Maximum L i k e l i h o o d P a r a m e t e r E s t i m a t e s ( A s y m p t o t i c t - s t a t i s t i c 1n p a r e n t h e s e s ) I n d e p e n d e n t A l t e r n a t i v e B r a n d L o y a l t y m e a s u r e s v a r i a b l e B a s e d o n w e i g h t i n g B a s e d o n f i r s t B a s e d o n o f p a s t c h o i c e s - 32 w e e k s o f p a s t cl M6 M6-BL32 M6-GL P r i c e -0.0389 - 0 . 0 3 6 5 -0.0353 (-5.64) (-5.73) (-5. 17) T e l e v i s i o n 0.0873 0 . 0 9 8 4 0.114 a d v e r t i s i n g ( 2 . 2 5 ) ( 2 . 6 9 ) ( 3 . 0 3 ) D i s p l a y ( 1 , 0 ) 1.173 1.116 1 . 104 ( 4 . 1 6 ) ( 4 . 1 7 ) ( 3 . 9 8 ) F e a t u r e ( 1 , 0 ) 1.688 1 .239 1.646 ( 1 . 7 6 ) ( 1 . 3 8 ) ( 1 . 7 2 ) I f b r a n d i s o n 0.885 0.880 0 . 8 8 8 d e a l ( 1 , 0 ) ( 5 . 3 0 ) ( 5 . 6 5 ) ( 5 . 3 5 ) D e a l a m o u n t 0.236 0.231 0 . 2 3 4 ( 9 . 8 5 ) ( 1 0 . 2 ) ( 9 . 8 8 ) B r a n d l o y a 1 t y 4.580 3.222 4 .974 ( 3 5 . 3 ) ( 3 3 . 2 ) ( 3 3 . 4 ) P r i c e x T e l e v i s i o n - 0 . 0 0 1 8 3 - 0 . 0 0 2 0 4 - 0 . 0 0 2 3 2 a d v e r t i s i n g (-2.53) (-2.99) (-3.29) P r i c e x F e a t u r e -0.0484 - 0 . 0 3 5 8 - 0 . 0 4 6 7 (-1.98) (-1.54) (-1.91) B r a n d c o n s t a n t B r a n d A 0.407 0 . 5 3 2 0 . 381 ( 2 . 1 9 ) - ( 3 . 1 6 ) ( 2 . 0 9 ) B r a n d B -0.208 - 0 . 1 1 9 - 0 . 3 1 4 (-0.83) (-0.52) (-1.25) B r a n d C 1 .227 1 .266 1.172 ( 4 . 9 6 ) ( 5 . 5 2 ) ( 4 . 8 4 ) B r a n d D 0 . 792 0 . 8 6 7 0.757 ( 3 . 4 8 ) ( 4 . 1 1 ) ( 3 . 4 0 ) B r a n d E 0 . 9 35 1 .095 0 . 8 9 5 ( 4 . 5 3 ) ( 5 . 7 6 ) ( 4 . 4 7 ) B r a n d F 1 . 120 1.037 1 .071 ( 5 . 1 7 ) ( 5 . 1 5 ) ( 5 . 0 2 ) B r a n d G 0 . 126 -0.0401 0 . 0 5 7 0 ( 0 . 5 4 ) (-0.18) (O.25) B r a n d H - 0 . 0 0 6 3 0 - 0 . 0 3 6 4 0 . 0 0 8 5 7 (-0.04) (-0.22) ( 0 . 0 5 ) B r a n d I -0.908 ^ 0 . 8 3 5 - 0 . 9 2 3 (-3.89) (-3.88) (-3.99) B r a n d J 0.436 0 . 3 7 5 0 . 322 ( 2 . 0 4 ) ( 1 .93) ( 1 .52) B r a n d K 0 0 0 r h o - s q u a r e d 0 .492 0 . 3 8 8 0 . 4 8 0 Log-1 i k e 1 i h o o d - 1480.9 -17 8 3 . 2 - 1515.8 174 Table 5 .5 E f f e c t of Smoothing Constant for QuadagnI and L i t t l e ' s Measure of Brand L o y a l t y on Dry Dog Food Brand Choice Models Maximum l i k e l i h o o d Parameter Est imates (Asymptotic t - s t a t i s t i c in parentheses) Independent v a r l a b le 0. 1 M6-GL1 m o o 0.2 M6-GL2 t h I n 0.5 M8-GI.5 g C o 0.7 M6-GL7 s t a 0.8 M6 - Gl. 8 0.875 MB-GL8. 0 .9 M6-GL9 P r i c e -0 .0369 -0 .0382 -0.0342 -0.0338 -0.0343 -0 .0353 -0.0358 ( 5 . 5 3 ) (-5 .41) ( -5.02) ( -4.93) ( -5.01) ( -5.17) ( -5.24) Telev i s l o n 0 . 115 0 ne 0 129 0. 130 0 124 0 . 114 0 109 adver 11s ing (3 . 19) (3 .28) (3.52) (3.49) (3.31) (3 .03) (2 89) D l s p l a y ( I . O ) 0 868 0 .873 0 932 1.009 1.059 1 . 104 1. 121 (3 . 19) (3 .20) (3.37.) (3.64) (3.82) (3 .98) (4.04) F e a t u r e t 1 . 0 ) 0 .969 1 .051 1. 359 1 540 1 .609 1 .648 . 1.651 ( 1 .09) ( 1 .17) (1.46) (1.62) (1.69) ( 1 . 72) (1.73) If brand Is on 0 .976 0 .982 0 927 0.908 0 897 0 .888 0.885 d e a l ! 1 . 0 ) (6 .22) (6 .08) (5.72) (5.52) (5.43) (5 . 35) (5 33) Dee 1 amount 0 .228 0 .230 0 233 0 233 0 233 0 .234 0 234 ( t o i l (10.1) ( 10.0) (9.91) (9.89) (9 .88) (9.88) Brand l o y a l t y 6^030 6 . 124 5.987 5.582 5.255 4 .974 4.866 (27.0) (27.3) (29.4) (31.4) (32.5) ( 33 4) (33.7) P r i c e x T e l e v i s i o n •0. .00233 -0. .00240 -0.00259 -0.00260 -0.00250 -0 .00232 -0.00224 adver t1s ing (-3.43) (-; 1.51) (-3.73) ( -3.72) ( -3.56) ( -3.29) ( -3 .17) Pr Ice x Fea tu re -0 0337 -0 0348 -0.0398 -0.04 36 -0.0455 -0 .0487 •0.0470 ( ' 1.44) (-1 .48) (-1.65) (- 1.79) I-1.B6I ( - 1.91) (- 1.92) Brand constant Brand A 0 594 0. 569 0.489 0.427 0.398 0 381 0.379 (3 45) (3 29) (2.75) (2.37) (2.18) (2 09) (2.08) Brand B -0 . 117 -0 . 128 -0.191 -0.274 -0.309 -0 314 -0.308 ( -0 48) ( -0. 51) ( -0.76) (- 1.08) (-1.22) <-! 25) ( -1 .23) Brand C 1 . 194 1. 195 1. 165 1. 169 1. 167 1 172 1. 178 (5. 04) (5. 03) .14.93) (4.84) (4.82) (4 84) (4.85) Brand D 0 669 0. 688 0.894 0.712 0. 732 0 757 0 788 (3. 19) (3 17) (3.15) (3.21) ( 3 . 30) (3 40) (3 451 Brand E 1. 048 1. 020 0.94 7 0.906 0.893 0. 895 0.899 (5. 46) (5. 30) (4.84) (4.58) (4.49) (4 47) (4 48) Brand F 1. 065 1. 074 1.078 1.06 3 1 .062 1. 071 1.078 (5. 09) (5. 11) (5.07) (4.99) (4 98) (5 02) (5 05) Brand G 0. 0956 0 107 0. 102 0. 722 0.0587 0 0570 0. 590 (0.42) (0. 47) (0.44) (0.31) (0.25) (0. 25) 10 25) Brand H 0. 169 0. 163 0 136 0.690 0 0467 0. 00857 -0.00405 ( 1. 00) (0. 97) (0.80) (0 51) (0.27) (0. 05) ( -0 .02) Brand I -0 . 751 -0 . 756 -0.801 -0.859 0.896 -0 . 923 -0.929 ( -3 . 39) 1-3. 41) (-3 53) ( -3.73) ( -3.88) ( 3 99) ( -4 .02) Brand J 0. 275 0. 281 0.295 0. 300 0.306 0. 322 0.332 ( 1. 30) ( t. 32) (1.38) (1.40) (1.44) ( 1 . 52) (1.57) Brand K 0 0 0 0 0 0 0 r h o - s q u a r e d 0. 444 0. 451 0.470 0.478 0. 480 0. 480 0.479 Log-1 Ike 1Ihood - 1620.60 - 1599.65 - 1545.65 - 1521.82 - 1515.74 - 1515.77 - 1517.28 Table 5 6 E f f e c t of A l t e r n a t i v e S p e c i f i c a t i o n for Guadagni and L i t t l e ' s Measure of Brand L o y a l t y on Dry Dog Food Brand Choice Models Maximum L i k e l i h o o d Parameter Es t imates (Asymptotic t - s t a t l s t i c In parentheses) Independent v a r l a b le Model with brand constants Ml Brand loya 1 ty Included M?GL Promotlona1 var lab les added Ml GL Advert IsIng var lab le added M4 - GL Pr Ice var lab le added M5 • GL Reference Model MB-GL Brand l o y a l t y exc luded M7 Adver t Is Ing var tab les exc luded MB GL Pr Ice 0.0426 ( 6 . 4 9 ) -0.0353 ( 5 . 1 7 ) -0.0431 (-7.241 -0.0402 ( -6 .46) T e l e v I s l o n -0.00714 •0.00875 0 114 0.0810 adver t IsIng (-0.83) ( -1.01) (3 03) (2.57) D i s p l a y ! 1 . 0 ) 1.056 1.051 1. 168 1 104 0.733 1. 132 (3 B3> (3.80) (4.24) (3 98) (3.01) (4 10) F e a t u r e ! 1 , 0 ) 0.0219 0.0370 -0.129 1 646 0.971 1 641 (0.05) (0.09) ( -0.32) ( 1 72) (1 29) ( 1 72) If brand Is on 0 993 0.994 0.885 0 888 1.048 0 877 d e a l ( l . O ) (6.10) (6.11) (5.39) (5 35) (7.8B) (5 32) Deal amount 0.224 0. 224 0.231 0 234 0. 198 0 233 (9.68) (9 68) (9.89) (9 88) (10.1) (9 90) Brand l o y a l t y 4.788 4 955 4 .957 4 .949 4.974 4.955 ( 38 11 ( 33 7 ) ( 33 7 ) ( 33 41 ( 33 4) 133 4) P r i c e x T e l e v i s i o n -0.00232 -0 00158 adver t Is Ing ( -3.29) ( -2 .65) P r i c e x Feature -0.0467 -0.0282 -0.0488 ( 1 . 9 1 ) ( - 1 .39) ( 1 98) Brand constant Br and A 0 873 0.278 0 320 0 333 0 461 0.381 0 .960 0 439 (6. 18) 11 661 ( 1 80) ( 1 .86) (2 541 (2 09) (6 33) (2 43) Brand 8 -0 . 331 -0 838 -0 . 730 -0 7 16 -0 229 -0.314 0 .339 -0 252 (- 1. B0) ( -3.79) (-3 10) (-3 03) (-0 91) ( -1.25) (1 .67) 1-1. 01) Brand C 0. 482 0.551 0 274 0 292 1 284 1.172 1 296 1 246 (3. 20) 13.33) ( 1 49) ( 1 .57) (5 .24) (4.84) (6 07) (5 18) Brand D 0. 533 0. 367 0 0850 0 0856 0 867 0.757 1 . 103 0 877 (3. 57) I 2.26 ) (0 47) (0 481 (4 03) (3.40) (5 57) (3 98) Brand E 1. 275 0 603 0 325 0 383 0 992 0.695 1 .641 0 922 (9. 49) (4.09) (2 05) (2. 12) (4 98) (4.47) (9 .38) (4 97) Brand F 0 363 0.783 0 4 80 0 495 1 154 1 .071 0 .706 1 132 (2. 34) (4.72) (2 60) (2 62) (5 44) (5 02) (3 .67) (5 36) Brand G -0 . 331 0.00523 -0 123 -0 09 36 0 167 0.0570 •0 290 0 129 1-1. 80) 10.26) t 0 56) ( 0 42) (0 , 73) (0 25) (-1 42) (0 57) Brand H 0. 837 0. 314 -0 0103 -0 0138 -0 00157 0.00857 0 508 0 00379 (5. 89) (1 93) ( 0 05) ( -0 08) ( -0 01) (0 05) (3 36) (0 02) Brand I 0 180 -0 294 -0 257 -0 260 - | . 024 -0 923 -0 577 - 1 037 ( 1 12) ( -1.57) ( 1 31) ( - 1. 32) ( -4 . 49) ( -3.99) ( -2 94) ( - 4 51) Brand J 0 259 0.00574 -0 0221 -0 . 0219 0. 405 0 322 0 619 0. 398 ( 1. 84) (0.31) ( 0 It) ( -0 . 11) ( 1. 92) (1.52) (3 47) ( 1 . B9) Brand K 0 0 0 0 0 0 0 0 rho- squared 0 0.294 0. 4 70 0. 470 0. 477 0.480 0. 184 0.477 Log-1 Ike) ihood 2914 3 -2057.3 - 1545.4 1545.1 - 1523.2 - 1515.8 -2377.9 - 1521 . 7 Table 5.7 Effect of Alternative Specification for Guadagni and Llttle'a Measure of Brand Loyalty and Disaggregated For* of Advertising for Dry Dog Food Brand Choice Models htaxteua Likelihood Parameter Estlaates (Asyaptotlc t-atatlatlc In parentheses) Independent variable Model with brand constants Hi Brand loyalty Included M2-GL Proaottonal •arlables added M3-GL Advertising variables added M4-GL-0 Price variable added W5-GL-0 Pries — — — ' — -0.0426 (-6.SO) Brand A ad. exposures' — — • — 0 (0 0177 45) 0.0171 (0.44) Brand B ad. exposures — — — -0 (-2 164 27) •0.157 (-2.14) Brand E ad. exposures — — — -0 (•1 00930 03) -0.0126 (-1.40) OisplayM.O) — — 1.056 (3.83) 1 (3 064 85) 1. 179 (4.27) Feature!1,0) — — 0.0219 (0.05) 0 (0 0422 11) -0.119 (-0.30) Xf brand 1s on deal!1.0) — — 0.993 (6.10) 1 te 001 16) 0.890 (5.42) Deal aaount — — 0.224 (9.68) 0 (• 224 88) 0.232 (9.89) Brand loyalty — 4.786 136. 1) 4.955 (33.7) 4.971 (33.7) 4.964 (33.3) Price x Brand A ad. exposures — — — — — Price x Brand B ad. exposures — — . — — Price x Brand E ad. exposures — — ' — — Price x Feature . — — Reference ttoOel Br end loyalty excluded Advertising variables excluded M6-GL-D M7-0 M7-GL -0.0344 -5.09) -0 (-7 0410 .02) -0.0402 (-6.46) 0.726 (3.27) 0 (2 494 87) — 1.002 (2.14) 1 (2 115 87) — ' 0. 136 (3.35) 0 (3 104 08) — 1.096 (3.92) 0 (2 730 99) 1. 132 (4.10) 1.606 (1.66) 0 (1 938 23) 1.641 (1.72) 0.692 (5.34) 1 (7 029 69) 0.877 (5.32) 0.236 (9.83) 0.200 (10.1) 0.233 (9.SO) 4.992 (33.2) — 4.955 (33.4) -0 0167 -3.18) -0.0119 (-2.95) — . •0.0229 -2.45) -0.0252 (-4.05) — -0.00282 -3.69) -0 (-: 00204 . 17) — -0.0465 - 1.86) -0 (-0279 . 36) -0.0466 (- 1.96) rho-squared 0 0.294 0 470 0 470 0.477 0 479 0 184 0 477 Log-1ikelihood -2914.3 -2057.3 -1545.4 -1542.1 -1520.2 -1505.1 -2362.6 -1621.7 Brand constant 439 Brand A 0 673 0.276 0 320 0 266 0 393 0 354 1 013 0 (6 16) (1.66) (1 80) (1 31) (1 92) (1 72) (6 16) (2 43) Brand B -0 331 -0.636 -0 730 -0 412 0 0596 -0 0131 0 606 -0 252 (-1 80) (-3.79) (-3 10) (-1 55) (0 21) (-0 05) (2 67) (-1 01) Brand C 0 462 0.551 0 274 0 271 1 241 1 041 1 163 1 246 13 20) 13.33) ( 1 49) 11 47) (5 17) (4 26) (5 53) 15 16) Brand D 0 533 0.367 0 0850 0 0755 0 677 0 697 1 04 3 0 877 (3 57) (2.26) (0 47) (0 42) (3 98) (3 13) (5 29) (3 9B) Brand E 1 275 0.603 0 325 0 399 1 023 0 906 1 615 0 922 (9 49) (4.09) (2 05) (2 29) (5 09) (4 501 (9 22) (4 97) Brand F 0 363 0.783 0 460 0 476 1 131 0 965 0 644 1 132 (2 34) (4.72) (2 60) (2 56) (5 35) 14 61) (3 36) (5 36) Brand G -0 331 0.00523 -0 123 -0 124 0 131 0 0613 -0 270 0 129 (• 1 80) (0.26) (-0 56) (-0 56) (0 56) (0 27) (-1 35) (0 67) Brand H 0 837 0.314 -0 0103 -0 0153 -0 00478 0 00577 0 610 0 00379 (5 89) (1.93) (-0 05) (-0 09) (-0 03) (0 03) (3 37) (0 02) Brand ! 0 160 -0.264 -0 257 -0 263 -1 028 -0 911 -0 545 • 1 037 (1 12) (-1.57) (-1 311 (-1 34) ( - 4 51) (-3 95) (-2 60) < -4 64) Brand J 0 259 0.00574 -0 0221 -0 0211 0 407 0 300 0 564 0 396 ( 1 64) 10.311 (-0 11) (-0 11) ( 1 93) ( 1 42) (3 29) ( 1 69) Brand K 0 - 0 0 0 0 0 C C 177 Table 5.8 Ef f e c t of Household Income on Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates (Asymptotic t - s t a t i s t i c 1n parentheses) Independent Income excluded Income included variable M6 M6-Inc Pr ice -0.0389 -0. 0388 (-5.64) ( -5- 63) Tel e v i s i o n 0.0873 0. 0919 advert is ing (2.25) (2. 38) Display(1,0) 1.173 1 . 169 (4. 16) (4. 15) Feature(1,0) 1 .688 1 . 679 (1.76) (1. 74) If brand is on 0.885 0. 904 deal(1,0) (5.30) (5. 39) Deal i amount 0.236 0. 236 (9.85) (9. 98) Brand loyalty 4.580 4. 656 (35.3) (34 1.8) Pr ice x Tel e v i s i o n -0.00183 -0. 00186 advert i s i n g (-2.53) ( -2. 59) Pr ice x Feature -0.0484 -0. 0484 (- 1.98) ( - 1 . 99) Brand Brand Income constant constant parameter Brand A 0.407 -0. .318 0. .0269 (2.19) (-0. .78) (1. .92) Brand B -0.208 0. .408 -0. ,0216 (-0.83) (0. 78) (-1. .20) Brand C 1 .227 1. .071 0 00674 (4.96) (2. 40) (0. .45) Brand D 0. 792 0. 776 0. 001 16 (3.48) (1 . 72) (0. .07) Brand E 0.935 0. 102 0. 0303 (4.53) (0. 26) (2. 33) Brand F 1 . 120 1. 088 0. 00229 (5.17) (2. 56) (0. 16) Brand G 0. 126 -0. 492 0. 0237 (0.54) ( -0. 97) ( 1 . 40) Brand H -0.00630 -0. 505 0. 0200 (-0.04) ( - 1 . 31 ) ( 1 . 43) Brand I -0.908 -0. 898 -0. 0000676 (-3.89) ( - 1 . 97) ( -0. 04) Brand J 0.436 0. 241 0. 00842 (2.04) (0. 52) (0. 52) Brand K 0 0 0 rho-squared 0.492 0.495 Log-like!ihood -1480.9 -1470.1 178 Table 5.9 E f f e c t of A l t e r n a t i v e Samples on Dry Dog Food Disaggregated Form of Advertising Brand Choice Models Maximum Li k e l i h o o d Parameter Estimates (Asymptotic t - s t a t i e t i c in parentheses) Independent F i r s t h a l f of sample Second half of sample Total sampli Variables (n = 1276) (n = 1276) (n = 2552) Price -0.0376 -0.0464 -0.0415 (-5.52) (-6.49) (-6.42) Brand A ad 0.724 0.286 0.540 exposures (3.23) (2.20) (3.21) Brand B ad. 0.956 0.223 0.434 exposures (2.12) (1.62) (2.01) Brand E ad. 0.107 0.0885 0.0948 exposures (2.55) (2.00) (3.06) Display!1,0) .1. 162 0.934 1.025 (4.10) (3.33) (5.16) Featured.O) 1.637 0.445 0.906 (1.69) (0.52) (1.42) If brand is on 0.669 0.755 0.613 deal!1,0) (5.23) (4.10) (6.54) Deal amount 0.238 0.251 0.244 (9.61) (9.93) (14.0) Brand l o y a l t y 4.593 4.673 4.706 (35.1) (35.2) (49.8) Pric e x Brand A •0.0169 -0.00637 -0.0125 ad. exposures (-3.22) (-2.09) (-3.16) Price x Brand B -0.0217 -0.00599 -0.0108 ad. exposures (-2.41) (-2.13) .1-2.43) Price x Brand E -0.00231 -0.000336 -0.00140 ad. exposures (-2.S3) (-1.22) (-2.42) Price x Feature -0.0461 0.006 36 -0.0149 (-1.94) (0.32) ( -0.98) rho-squared 0.495 0.528 0.509 % purchases c o r r e c t l y 52.39 55. 32 53.69 predicted cross -va1idated 0.465 0.517 NA rho-squared cross-validated 52.98 54.09 NA X purchases c o r r e c t l y predicted Brand constant Brand A 0.404 0.0373 0.213 (1.94) (0.18) (1.48) Brand B 0.0568 -0.151 -0.0606 (0.20) (-0.55) (-0.31) Brand C 1 . 100 0.772 0.925 (4.44) (3.05) (5.25) Brand D 0.727 0.424 0.560 (3.20) (1.88) (3.51) Brand E 0.949 0.516 0. 729 (4.56) (2.54) (5.04) Brand F 1.035 0.778 0.891 (4.76) (3.63) (5.87) Brand G 0. 112 -0.763 -0.304 (0.49) (-2.96) (-1.7B) Brand H -0.00707 -0.564 -0.285 (-0.04) (-3.19) (-2.28) Brand I -0.88B - 1 . 163 - 1.022 (-3.82) (-5.21) (-6.36) Brand J 0.410 0.235 0.307 (1.92) (1.13) (2.08) Brand K 0 0 0 179 Table 5.10 Effec t of Alternative Samples on Dry Dog Food Aggregated Form of Advertising Brand Choice Models Maximum Likelihood Parameter Estimates (Asymptotic t - s t a t i s t i c in parentheses) Independent F i r s t half of sample Second half of sample Total sample Var iables (n = 1276) (n = 1276) (n = 2 5 5 2 ) Pr ice - o ; 0 3 8 9 -0.0476 - 0 . 0 4 2 4 (-5.64) (-6.51) ( -8. 4 5 ) T e l e v i s i o n 0 . 0 8 7 3 0 . 0 8 9 0 0 . 0 8 1 2 advert is ing ( 2 . 2 5 ) ( 1 . 4 2 ) ( 1 . 7 2 ) D i s p l a y ( 1 , 0 ) 1 .173 0.933 1 .032 ( 4 . 1 6 ) ( 3 . 3 3 ) ( 5 . 2 2 ) Feature(1,0) 1.688 0 . 4 0 2 0 . 9 0 9 ( 1 . 7 6 ) ( 0 . 4 7 ) ( 1 .43) If brand is on 0 . 8 8 5 0.752 0 . 8 0 9 dealt 1,0) ( 5 . 3 0 ) ( 4 . 0 8 ) ( 6 . 5 3 ) Deal amount 0.236 0.251 0 . 2 4 4 ( 9 . 8 5 ) ( 9 . 9 0 ) ( 14.0) Brand l o y a l t y 4 . 5 8 0 4.867 4.696 ( 3 5 . 3 ) ( 3 5 . 2 ) ( 4 9 . 9 ) Price x Tel e v i s i o n a d v e r t i s i n g Price x Feature Brand constant Brand A Brand B Brand C Brand D Brand E Brand F Brand G Brand H Brand I Brand J Brand K rho- squared % purchases c o r r e c t l y pred icted cross-validated rho-squared % puchases c o r r e c t l y predicted - 0 . 0 0 1 8 3 (-2.53) - 0 . 0 4 8 4 (-1.98) 0.407 ( 2 . 1 9 ) - 0 .208 (-0.83) 1 .227 ( 4 . 9 6 ) 0 . 792 ( 3 . 4 8 ) 0 . 9 3 5 ( 4 . 5 3 ) 1 . 120 ( 5 . 1 7 ) 0 . 1 2 6 ( 0 . 5 4 ) - 0 . 0 0 6 3 0 (-0.04) -0.908 (-3.89) 0.436 ( 2 . 0 4 ) 0 0 . 492 5 2 . 14 0 . 482 52.86 - 0 . 0 0 1 2 5 (-1.52) - 0 . 0 0 0 9 4 2 (-1.78) 0 . 0 0 7 1 4 ( 0 . 3 6 ) - 0 . 0 1 4 8 (-0.97) 0 . 102 ( 0 . 5 6 ) -0.297 - 1 . 2 1 ) 0 . 8 2 5 ( 3 . 2 5 ) 0 . 4 5 9 ( 2 . 0 3 ) 0.486 ( 2 . 4 0 ) 0 . 8 2 0 ( 3 . 8 0 ) -0.738 (-2.74) -0.566 (-3.21) - 1.180 (-5.26) 0.253 ( 1 . 2 1 ) 0 0 . 245 ( 1 . 9 0 ) - 0 . 2 7 0 (-1.55) 1.010 ( 5 . 7 4 ) 0 . 6 0 3 ( 3 . 7 7 ) 0 . 7 0 2 ( 4 . 8 7 ) 0 . 947 ( 6 . 2 3 ) 277 59) 286 (-2.28) - 1.035 (-6.41) 0.327 ( 2 . 2 0 ! 0 0.527 5 5 . 2 6 0 . 507 5 3 . 5 6 0.517 54.06 180 T a b l e 5.11 Estimates of Time Trend for the Actual Brand Prices 1n units of cents per pound ( t - s t a t i s t i c in parentheses) Sample size in a l l analysis i s 1276 purchases Brand ;Po a R-square A 42.53 0.00675 0.0004 (19.5) (0.73) B 47.69 0.0225 0.0047 (21.9) (2.46) C 46.27 0.0773 0.033 (19.5) (0.73) D 71.82 -0.0454 0.0085 (22.1) (-3.31) E 36.62 0.0758 0.0199 (10.3) (5.08) F 11.03 0.193 0. 123 (3.21) (13.4) G 35.79 0.0492 0.0173 (14.5) (4.73) H 50. 18 -0.0357 0.00274 (11.1) (-1.87) I 15.59 0.0308 0.0152 (9.42) (4.43) J 50.21 0.00693 0.0007 (28.8) (0.95) K 58.51 -0.0673 0.0181 (19.5) (-4.85) Brand A • - Dog Chow Brand B • - Puppy Chow Brand C • • Kibbles-N-•Bits Brand D • • Nestle Advertised Brands Brand E • • A l l other Purina Advertised brands Brand F • • A l l other Quaker" Advertised brands Brand G • • A l l other advertised brands Brand H • • Tuffys Brand I • • Cycle 1 - 4 Brand J -• A l l other non-advertised Purina brands Brand K • • A l l other non-advertised brands 181 Table 5.12 Intercorrelations (r) between the Residual of Actual Brand Prices and Mean Brand Prices Mean* s t d . dev Brand A 1 .00 43. 88 5. 48 Brand B .510 1 .00 53. 00 5. 31 Brand C .353 .446 1 .00 63. .96 7 . 47 Brand D . 258 .297 . 304 1 .00 60 . 28 8 .94 Brand E .424 .425 .331. . 341 1 .00 53 .49 9 .58 Brand F .268 . 336 . 320 .232 . 302 1 .00 56 .02 9 . 37 Brand G . 309 .372 . 320 .239 333 .267 1 .00 47 . 12 6 . 35 Brand H . 394 . 356 .212 .271 . 397 . 194 .260 1 .00 41 .06 11 . 36 Brand I . 358 . 246 . 161 . 185 .261 . 189 .215 .342 1 .00 22 .79 4 .05 Brand J . 292 . 326 .310 .250 .238 . 180 .314 . 190 .241 1 .00 51 .52 4 .45 Brand K . 330 . 308 .171 .171 .277 . 180 .270 . 355 .218 .250 1.00 42 . 35 8 . 36 Brand A - Dog Chow (Purina) Brand B - Puppy Chow (Purina) Brand C - Kibbles-N-Bits (Quaker) Brand D - Nestle Advertised Brands Brand E - A l l other Purina Advertised brands Brand F - A l l other Quaker Advertised brands Brand G - A l l other advertised brands Brand H - Tuffys Brand I - Cycle1-4 Brand J - A l l other non-advertised Purina brands Brand K - A l l other non-advertised brands * The actual prices are measured in cents per pound. Table 5.13 Intercorrelations (r) between Brand Advertising Exposures and Mean Number of Exposures per Purchase Mean Std. dev Brand A 1.00 1 .32+ 2.S3 Brand B .010*+ .116 1 .00 0 .92 2.03 Brand C . 104* .002 .087* .068 1 .00 4 .86 7.75 Brand D .097* .027 .010 .091 . 162* .211* 1.00 6 .69 13.3 Brand E . 118* .322' .056* . 141 . 185* .059 . 106* .053 1 .00 7 .42 11.2 Brand F . 104' .274* .139* . 145 .240* .074 .043 .258* .295* .261* 1.00 6. 43 9. 12 Brand G .131* .361* .044 • 107 .041 . 188 .090* .098 . 147* .536* .141* 1.00 2. .336* 64 8.05 * S i g n i f i c a n t l y d i f f e r e n t from zero at prob. of 0.05. + For 1276 purchases, the mean number advertising exposure for Brand A is 1.32 with the standard deviation of 2.93. •+ These c o r r e l a t i o n are based on 88 weeks of advertising exposure data across a l l the sample households. Brand A • - Dog Chow Brand B • Puppy Chow Brand C • Kibbles-N-Bits Brand D • • Nestle Advertised Brands Brand E • • A l l other Purina Advertised brands Brand F • • A l l other Quaker Advertised brands Brand G • - A l l other advertised brands 183 Table 5.14 Estimates of Marginal Change in the Regular Price as Result of the Television Advertising ( t - s t a t i s t i c in parentheses) Brand name Independent variable A + B C D E F G Average price of non-adv. brands 0. 527 (20.2) 0.467 (17.7) 0.493 (14. 1) 0.596 (15.8) 0.880 (20.4) 0.521 (10.5) 0.507 ( 16.0) Number of TV adver. exposures 0.568 (6.54) 0.984 (7.20) 0.428 (10.3) 0. 176 (10.1) 0. 145 (6.02) 0.465 (8.91) 0. 140 (5.07) constant 29.04 41 .58 55.43 48.53 21. 72 44.90 29.82 R-square 0.320 0.272 0.242 0.332 0.286 0. 149 0. 190 + For each brand regression analysis is conducted to estimate regression estimates for these two varia b l e s , hence iP r i c e of I = 29.04 + 0.527I Average price of I + 0.568I Number of I |Brand A | |non-adv. brands | |adv. expos.| This implies 0.568 is marginal increase in regular price as a result of increased level of advertising. Brand A • • Dog Chow Brand B • • Puppy Chow Brand C • • Kibbles-N-Bits Brand D • • Nestle Advert.ised Brands Brand E • • A l l other Purina Advertised brands Brand F • • A l l other Quaker Advertised brands Brand G -• A l l other advertised brands 184 Table 5.15 Weighted Aggregate E l a s t i c i t i e s using the Total Sample for 11 Brands of Dry Dog Food Brand A Brand B Brand C Brand D Brand E Brand F Brand G Brand H Brand I Brand J Brand K oo Pr ice -0. 6800+ - 1 . 1080 - 1 . 2769 - 1 . 1931 -0 .8426 - 1 .3491 - 1 . .1452 -0, .6653 -0, ,5048 -1 .0493 -1. , 1091 Television advert i s ing 0. .0363* 0 .0480 0 .0486 0 .0143 0. 1296 0, .0609 0. 0970 — — — — D1splay(1,0) 0. ,0139 0 .0393 0 .0028 0. 0034 0, .0013 0. .0405 0. ,0272 0. .0123 0 .0055 0 .0022 Feature(1,0) 0. ,0112 0 .0084 0 .0055 0 .0005 0 .0166 0 0042 0, .0144 0 .0002 0 .0085 If brand is on dea 1 0. 0433 0 .0288 0 .0748 0 . 1047 0. .0686 .0, .0978 0. , 1143 0 . 1004 0 .0640 0 .0650 0 .0823 Deal amount 0. .0489 . 0 .0226 0 . 1698 0 . 1945 0. , 1239 0 .2102 0 . 1894 0, , 1301 0 .0362 0 .1129 0 . 1243 Brand loyalty 0. ,5168 0 .8779 0 .4056 0 .5614 0. ,5921 0. .2353 0 .2778 0, .6030 0 . 7631 0 .4527 0 .6120 Price x Tele-v i s i o n adv. -0. .0362** -0 .0578 -0. .0727 -0 .0200 -0 , 1590 -0 .0789 -0, , 1044 — — — — Price x Feature -0. .0061 -0 .0087 -0. ,0050 -0 .0005 -0. .0094 -0, .0017 -0 .0039 -0 .0001 -0 .0042 + If the regular price of Brand A increases by one percent, then the pr o b a b i l i t y of choosing Brand A decreases by 0.68 percent. * If parameter estimates from Table 5.9 are used, then the e l a s t i c i t y estimates for t e l e v i s i o n advert ising wi11 be 0.489, 0.395, and 0.184 for Brand A, Brand B, and Brand E respectively. ** If parameter estimates from Table 5.9 are used, then the e l a s t i c i t y estimates for Price x Television advertising w i l l be -0.476, -0.491, and -0.229 for Brand A, Brand B, and Brand E respectively. Table 5.16 Aggregate E l a s t i c i t i e s Evaluated at the Sample Means for 11 Brands of Dry Dog Food Brand A Brand B Brand C Brand D Brand E Brand F Brand G Brand H Brand I Brand J Brand K Co cr\ Pr ice -1 . 6358+ -2. .1748 -2. 5063 -2, , 3600 -1 . 8048 -2. , 1668 - 1 . ,9371 - 1 . 5737 -0. 9023 -2. 0505 - 1 . 6862 Telev i s ion advert isin g 0 .0799* 0. , 1060 0. 1143 0 .0288 0. ,2705 0, 1000 0, , 1979 — — — — Display(1.0) 0 .0099 ' 0 .0262 0 .0011 0 .0048 0 .0007 0 .0132 0. ,0385 0. 0079 0 .0019 0 .0011 Feature(1,0) 0. .0084 0 ,0045 0. .0103 0, .0003 0, .0027 0. ,0073 0. ,0090 0, .0010 0 .0033 If brand is on deal 0 .0491 0. ,0129 0 ,0518 0 .0740 0 .0925 0 .0537 0 .0433 0. . 1203 0 .0504 0 .0410 0 .0337 Deal amount 0 .0543 0, .0092 0 , 1458 0 .1917 0 .2448 0 . 1480 0 .0773 o , 1501 0 .0277 0 .0663 0 .0481 Brand loyalty 0 .5523 0. .2942 0. .3033 0 .4505 0 .7420 0 , 1568 0 . 1072 0 .5729 0 . 3806 0 .2957 0 .3110 Price x Tele- -0 .0814* * -0. . 1299 -0. .1715 -0 .0411 -0 .3392 -0 . 1325 -0 .2176 vi s i o n adv. Price x Feature -0.0045 -0.0047 -0.0093 -0.0003 -0.0016 -0.0028 -0.0024 -0.0008 -0.0018 • If the regular price of Brand A increases by one percent," then the p r o b a b i l i t y of choosing Brand A decreases by 1.6358 percent. * If.parameter estimates from Table 5.9 are used, then the e l a s t i c i t y estimates for t e l e v i s i o n advertising w1ll.be 1.063, 1.133, and 0.631 for Brand A, Brand B, and Brand E respectively. ** If parameter estimates from Table 5.9 are used, then the e l a s t i c i t y estimates for Price x Television advertising w i l l be -1.081, -1.489, and -0.504 for Brand A, Brand B, and Brand E respectively. Table 5.17 A Basic Sequence of Specifications for Aluminum F o i l Brand Choice Models Maximum Likelihood Parameter Estimates (Asymptotic t - s t a t i s t i c in parentheses) Independent Model with Brand Promot iona1 Advert islng Pr ice Reference Brand Advert i s l n g var iable brand loya1ty var iables var iable var iable Model loya1ty variables constants included added added added excluded excluded M1 -F M2-F M3-F M4-F M5-F M6-F M7-F M8-F Price -4.512 -4.299 -4.275 - 4.501 _ ( -13.2) (-11.9) (-13.6) ( - 13.2) Telev is ion 0.0654 0.0878 1 .026 0.858 advert i s ing (1.72) (2.09) (2.37) (2.49) Display(1,0) 1 . 175 1.131 1 .062 - 1.033 0.531 1 . 109 (1.47) (1.41) (1.27) (1.24) (0.66) (1.33) If brand is on 0. 172 0.221 0.456 ' 0.458 -0.133 0.365 dealt 1,0) (0.52) (0.67) (0.99) (0.94) (-0.30) (0.80) Deal amount 1 .420 1. 302 1 .275 1.272 2.722 1 .556 (1.43) (1.32) (0.82) (0.76) (1.69) ( 1 .00) Brand loyalty 1 .943 1 .961 1 .957 1 .978 1 .985 1 .989 (18.3) (18.3) (18.3) ( 16. 1) ( 16.0) (16.2) Price x Television -0.342 -0.274 advertising (-2.73) (-2.61) Brand constant Brand Ft 0.789 0.189 , 0.113 0.014 1 .820 1 .704 2.264 1 .945 (10.5) (2.01) (1.15) (0.12) (9.22) (8.24) (12.6) ( 10.3) Brand P -0.573 -0.518 -0.515 -0.517 1.275 1.202 1 . 109 1 .273 (-5.54) (-4.53) (-4.48) (-4.50) (7.02) (6.46) (6.82) (7.02) Brand G 0 0 0 0 0 0 0 0 rho-squared 0 0.218 0.225 0.226 0.369 0. 371 0. 198 0. 365 Log-1 ike 1 ihood -926.5 -724.8 -718.2 -716.7 -585.1 -582.5 -743.0. -588.3 Table 5.18 Weighted Aggregate E l a s t i c i t i e s for Three Brands of Aluminum F o i l Brands Reynolds Private Generic Price -2 .6508+ -6 .4748 -4 .3053 Tel e v i s i o n advert is ing 0 .3444 Display .0010 0 .0059 If brand is on a deal 0. .0155 0 .0104 0 .0042 Deal amount 0. ,0114 0 .0089 0 .0020 Brand loyalty 0. 2699 0 .3806 0 .2595 Price x Tele-v i s i o n adv. -0. 3128 — ' + One percent increase in the price of Reynolds brand is expected to result in 2.6508 percent decrease in the p r o b a b i l i t y of choosing the brand. 188 Figure 3.1 Representation of Spatial Competition in One Dimension i—Location ef Consumer Proforoneo C Product Chtroctorlttlc Figure 4.1 Data Management Flow-chart f o r Dry Dog Food Count purchase observations by household i d . Exclude i d s . and purchase obs. with less than f i v e purchases Store remaining ids i n i d f i l e 1 and purchase obs. in purchase f i l e 1 Match i d s . i n two f i l e s and include common i d s . Include ids that were i n sample for more than 50 weeks Count TV obs. by household i d . Sort TV obs. by household i d and obs. date Store 22 TV obs. obs. f i l e s i n t o one data f i l e Obtain p r i c e f i l e Merge a l l f i les together to obtain purchase-to-purchase records • exclude id mismatches • exclude those ids not in demographic f i le Demographic f i l e f or income and other household va r i a b l e s Count purchase obs. by UPC 1 Group UPCs by brand- si z e combination Enter brand-size codes along with with UPCs Construct mean prices and promotional a c t i v i t i e s by week I Obtain promotional f i l e Use t h i s as input f i l e to SAS procedure MLOGIT 191 References Cited Aaker David A. and James M. Carman (1982) "Are You Overadvertising? A Review of Advertising-Sales Studies," Journal of Advertising Research, vol. 22(4), pp. 57-70. Amemiya Takeshi (1985) Advanced Econometrics, Mass: Cambridge, Harvard University press. Amemiya Takeshi (1979) "Qualitative Response Models: A Survey," Journal of Economic Literature, vol. 19, pp. 1483-1536. Arnold Stephen J. Tae H. Oum, Bohumir Pazderka, and Douglas W. Snetsinger (1987) "Advertising Quality in Sales Response Models," Journal of Marketing Research, vol. 24(1), pp. 106-13. Arnold Stephen J. Tae H. Oum, and Douglas J. Tigert (1983) "Determinant Attributes in Retail Patronage: Seasonal, Temporal, Regional, and International Comparisons," Journal of Marketing Research, vol. 20(May), pp. 149-57. Assael Henry (1987) Consumer Behaviour and Marketing Action. Third edition, Boston, Mass.: Kent Publishing. Assmus Gert, John U. Farley and Donald R. Lehmann (1984) "How Advertising Affects Sales: Meta-Analysis of Econometric Results," Journal of Marketing Research, vol. 21.(1), pp. 65-74. Bain Joe S. (1956) Barriers to New Competition, Cambridge, Mass.: Harvard University Press. Bayer Judy A. (1985) ADS - A Model of Brand Choice Response to Advertising and Price, Unpublished Doctoral Dissertation, New York: New York University. Beardon William O., Jesse E. Teel and Robert H. Williams (1982) "Consumer Response to Cents-Of Coupons," 1981 American Marketing Association Proceedings, AMA, Chicago, pp. 62-65. Benham Lee and Alexandra Benham (1975) "Regulating through the Professions: A Persepctive on Information Control," Journal of Law and Economics, Vol. 28(Oct), pp. 421-47. Benham Lee (1972) "The Effect of Advertising on the Price of Eyeglasses," Journal of Law and Economics, Vol. 25, pp. 337-52. Blattberg Robert D., Gary D. Eppen and Jeshua Lieberman (1981) "A Theoretical and Empirical Evaluation of Price Deals for Consumer Nondurables," Journal of Market-ing, vol. 45 (Winter), pp. 116-29. 192 Blattberg Robert C. and Abel P. Jeuland (1981) "A Micromodelling Approach to Investi-gate the Advertising-Sales Relationship," Management Science, vol. 27(September), pp. 988-1005. Borden Neil H. (1942) The Economic Effects of Advertising. Chicago, Illinois: Richard D. Irwin Inc. Boyer Marcel, Richard E. Kihlstorm and Jean-Jacques Laffont (1984) "Market Determi-nants of Misleading Advertising," in Bayesian Models in Economic Theory, Boyer Marcel and Richard E. Kihlstorm (eds.) pp. 241-70. Box George E. P. and D. R. Cox (1964) "An Analysis of Transformations," Journal of the Royal Statistical Society, vol. 26, series B, pp. 211-52. Bresnahan Timothy F. (1981) "Duopoly Models with Consistent Conjectures," American Economic Review, vol. 71(5), pp. 934-45. Brodie Roderick and Cornells A de Kluyver (1984) "Attraction versus Linear and Mul-tiplicative Market Share Models: An Empirical Evaluation," Journal of Marketing Research, vol. 21 (May), pp. 194-201. Butters Gerard R. (1976) "A Survey of Advertising and Market Structure," American Economic Review, vol. 66(2), pp. 392-97. Cady John F. (1976) "An Estimate of the Price Effects of Restriction on Drug Price Advertising," Economic Inquiry, vol. 24, pp. 493-510. Caves Richard E. and Peter J. Williamson (1985) "What is Product Differentiation, Re-ally?," Journal of Industrial Economics, vol. 34 (December), pp. 113-32. Capozza Dennis R. and Robert Van Order (1978) "A Generalized Model of Spatial Com-petition," American Economic Review, vol. 68(5), pp. 896-908. Carpenter Gregory S. and Donald R. Lehmann (1985) "A Model of Marketing Mix, Brand Switching, and Competition", Journal of Marketing Research, vol. 22 (August), pp. 318-29. Chamberlin Edward H. (1962) The Theory of Monopolistic Competition (8th edition). Cambridge, Mass.: Harvard University Press. Chow Gregory C. (1983) Econometrics, NY: New York, McGraw-Hill Book company. Clarke Darral G. (1987) Marketing Analysis and Decision Making, Calif: Redwood City, Scientific Press. Clarke Darrel G. (1976) "Econometric Measurement of the Duration of Advertising Effect on Sales," Journal of Marketing Research, vol. 13(4), pp. 345-57. 193 Comanor William S. and Thomas A. Wilson (1979) "The Effect of Advertising on Com-petition: A Survey," Journal of Economic Literature, vol. 17 (June), pp. 453-76. Comanor William S. and Thomas A. Wilson (1974) Advertising and Market Power, Cam-bridge, Mass.: Harvard University Press. Congressional District Data Book of the 98th Congress, Census of Population and Housing (1984), The United States Department of Commerce, Bureau of Census, Millwood, New York: Kraus Internal publication. Curhan Ronald C. (1974) "The Effect of Merchandising and Temporary Promotional Activities on Sales of Fresh Fruits and Vegetables in Supermarkets," Journal of Marketing Research, vol. 11 (August), pp. 286-94. Dixit Avinash and Victor Norman (1980) "Advertising and Welfare: Another Reply," The Bell Journal of Economics, vol. 11(2), pp. 753-54. Dixit Avinash and Victor Norman (1978) "Advertising and Welfare," The Bell Journal of Economics, vol. 9, pp. 1-17. Eaton Curtis B v and Richard Lipsey (1975) "The Principle of Minimum Differentiation Revisited," Review of Economic Studies, vol. 42(Jan), pp. 27-49. Ehriich Isaac and Fisher Lawrence (1982) "The Derived Demand for Advertising: A The-oretical and Empirical Investigation," American Economic Review, vol. 72 (June), pp. 366-88. Engel James F. and Roger D. Blackwell (1982) Consumer Behavior, New York, NY.: The Drydon Press. Eppen Gary D. and Yehoshua Liebermann (1984) "Why Do Retailers Deal:' An Inventory Explanation," Journal of Business, vol. 57(4), pp. 519-30. Eskin Gerald J. (1985) "Tracking Advertising and Promotion with Single-source Data", Journal of Advertising Research, vol. 25(1), pp. 31-39. Eskin Gerald J. (1975) "A Case for Test Market Experiments," Journal of Advertising Research, vol, 15(2), pp. 27-33. Eskin Gerald J. and Penny H. Baron (1977) "Effects of Price and Advertising in Test-Market Experiments," Journal of Marketing Research, vol. 14 ( N o v . ) . pp. 499-508. Farris Paul W. and David J. Reibstein (1979) "How Prices, Ad Expenditures, and Profits are Linked," Harvard Business Review, vol. 57(Nov.-Dec), pp. 173-8-1. Feldman Roger and James Begun (1980) "Does Advertising Reduce the Moan and Variance of Prices?" Economic Inquiry, vol. 18 (July), pp. 479-84. 194 Feldman Roger and James Begun (1978) "The Effects of Advertising Restrictions - Lessons from Optometry," Journal of Human Resources, vol.13 (supplement), pp. 247-62. • Gensch Dennis H. and Wilfred W. Recker (1979) "The Multinomial, Multiattribute Logit Choice Model," Journal of Marketing Research, vol. 16(February), pp. 124-32. Ghosh Avijit, Scott Neslin and Robert Shoemaker (1984) "A Comparison of Market Share Models and Estimation Procedures," Journal of Marketing Research, vol. 21 (May), pp. 202-10. Givon Moshe and Dan Horsky (1979) "Application of a Composite Stochastic Model of Brand Choice", Journal of Marketing Research, vol. 16, pp. 258-67. Greenhut Melvin L., George Norman and Chao-Shun Hung (1987) The Economics of Imperfect Competition: A Spatial Approach. New York: Cambridge University Press. Grossman Gene M. and Carl Shapiro (1984) "Informative Advertising with Differentiated Products," Review of Economic Studies, vol. 51, pp. 63-81. Guadagni Peter M. and John D. C. Little (1983) "A Logit Model of Brand Choice Cali-brated on Scanner Data," Marketing Science, vol. 2(3), pp. 203-38. Gupta Sunil (1987) An Integrated Model of Interpurchase Time, Brand Choice and Pur-chase Quantity. Unpublished Doctoral Dissertation, Columbia University. Gurumurthy K. and John D. C. Little (1986) "A Pricing Model based on Perception Theories and its Testing on Scanner Panel Data", Paper Presented at the 1986 ORSA/TIMS Marketing Sceience Conference, at Dallas, Texas. Haas-Wilson Deborah (1986) "The Effect of Commercial Practice Restrictions: The Case of Optometry," Journal of Law and Economics, vol. 29 (April), pp. 165-86. Hanssens Dominique M. (1987) "Marketing and the Long Run", Working paper no. 164, Center for Marketing Studies, John E. Anderson Graduate School of Management, University of California, Los Angeles. Hauser John R. and Steven M. Shugan (1983) "Defensive Marketing Strategies," Marketing Science, vol. 2(4), pp. 319-60. Hauser John (1978) "Testing the Accuracy, Usefulness, and Significance of Probabilistic Choice Models: An Information Theoretic Approach", Operations Research, vol. 26 (May), pp. 406-21. Hensher David A. and Lester W. Johnson (1981) Applied Discrete-Choice Modelling. New York: John Wiley and sons. 195 Herniter Jerome D. (1971) "A Probabilistic Market Model of Purchase Timing and Brand Selection," Management Science, vol. 18 (January), pp. 102-13. Horowitz Joel (1981) "Sampling, Specification and Data Errors in Probabilistic Discrete-Choice Models", in Applied Discrete-Choice Modelling, Hensher David A. and Lester W. Johnson (eds.), pp. 417-35. Houston Franklin S. and Doyle L. Weiss (1976) "An Analysis of Consumer Retention: An Individual/Aggregate Demand Function," in Marketing 1776 - 1976 and Beyond, Kenneth L. Bernhardt (ed.), American Marketing Association, pp. 111-14. Houston Franklin S. and Doyle L. Weiss (1974) "An Analysis of Competitive Market Behavior," Journal of Marketing Research, vol. 11 (May), pp. 151-55. Hoyer Wayne D.(1984) "An Examination of Consumer Decision Making for a Common Repeat Purchase Product," Journal of Consumer Research, vol. 11, pp. 822-29. Jeuland Abel P. and Steven M. Shugan (1983) "Managing Channel Profits" Marketing Science, vol. 2(3), pp. 239-72. Jeuland Abel P. and Chakravarthi Narasimhan (1985) "Dealing - Temporary Price Cuts - by Seller as a Buyer Discrimination Mechanism," Journal of Business, vol. 58(3), pp. 295-308. Johnson Norman L. and Samuel Kotz (1970) Distributions in Statistics: Continuous Univariate Distributions, volume 1., pp. 272-90, Boston: Houghton Mifflin. Jones John Philip (1986) What's in a Name? Advertising and the Concept of Brands, Lexington, Mass.: Lexington Books. Jones Morgan J. and Fred S. Zufryden (1982) "An Approach for Assessing Demographic and Price influences on Brand Purchase Behavior," Journal of Marketing, vol. 46(Winter), pp. 36-46. Jones Morgan J. and Fred S. Zufryden (1980) "Adding Explanatory Variables to a Con-sumer Purchase Behavior Model: An Exploratory Study," Journal of Marketing Re-search, vol. 17, pp. 323-34. Johnston J. (1984) Econometric Methods, Third edition, New York: McGraw-Hill. Kihlstrom Richard E. and Michael H. Riordan (1984) "Advertising as a Signal," Journal of Political Economy, vol. 92(3), pp. 427-50.. Kotler Philip and Ronald E. Turner (1979) Marketing Management: Analysis, Planning and Control. Canadian Third Edition. Ontario: Scarborough, Prentice-Hall of Canada. Kotowitz. Yehuda and Frank'Mathewson (1979) "Advertising, Consumer Information and 196 Product Quality," The Bell Journal of Economics, vol. 10(2), pp. 566-88. Krishnamurthi Lakshrnan. and S. P. Raj (1988) "A Model of Brand Choice and Purchase Quantity Price Sensitivity," Marketing Science, vol. 7(1), pp. 1-20. Krishnamurthi Lakshman and S. P. Raj (1985) "The Effect of Advertising on Consumer Price Sensitivity," Journal of Marketing Research, vol. 22 (May), pp. 119-29. Kuehn Alfred A. and Albert C. Rohloff (1967) "Consumer Responses to Promotions", in Promotional Decisions Using Mathematical Models, Robinson Patrick J. and Charles Hinkle (eds.) Boston: Allyn and Bacon, pp. 45-145. Kuehn Alfred A. (1962) "Consumer Brand Choice as a Learning Process," Journal of Advertising Research, vol. 2, pp. 10-17. Kwoka John E. Jr: (1984) "Advertising and the Price and Quality of Optometric Services," American Economic Review, vol. 74(1), pp. 211-16. Lambin Jean J. (1976) Advertising, Competition and Market Conduct in Oligopoly over Time. New York, New York: American Elsevier Publishing Co. Lancaster Kevin (1975) "Socially Optimal Product Differentiation," American Economic Review, vol. 65(4), pp. 567-85. Landwehr Jane T. (1986) Two Stage Model of Brand Choice. Unpublished Doctoral Dissertation, Chapel Hill: University of North Carolina. Lawrence Raymond J. (1980) "The Lognormal Distribution of Buying Frequency Rates," Journal of Marketing Research, vol. 17 (May), pp. 212-20. Lilien Gary L. (1974) "A Modified. Linear Learning Model of Buyer Behavior," Manage-ment Science, vol. 20 (March), pp. 1027-36. Little John D. C. (1979) "Aggregate Advertising Models: The State of the Art," Operations Research, vol. 27, pp. 629-67! Lodish Leonard M. (1986) The Advertising and Promotion Challenge, New York: Oxford University press. Maddala G.S. (1986) Limited-Dependent and Qualitative Variables in Econometrics, New York: Cambridge University press. Marshall Alfred (1923) Industry and Trade (second edition), London: Macmillan. Marvel Howard (1976) "The Economics of Information and Retail Gasoline Price Behavior: An Empirical Analysis," Journal of Political Economy, vol. 84 (Oct.), pp. 1033-66. 197 Mason Barry J. and Lehman M. Mayer (1981) Modern Retailing: Theory and Practice. Revised edition. Piano, Texas: Business Publications Inc. Massy William F., David B. Montgomery and Donald G. Morrison (1970) Stochastic Models of Buying Behavior, Cambridge: MIT press. Massy William F. and Ronald Frank (1965) "Short Term Price and Dealing Effects in Selected Market Segments," Journal of Marketing Research, vol. 2 (May), pp.171-85. Maurizi Alex (1972) "The Effects of Laws Against Price Advertising: The Case of Retail Gasoline," Western Economic Journal, vol. 10 (Sept.), pp. 321-29. Maxwell John C. (1986) "Pet Food Category off its Feed, but Snacks Gain", Advertising Age, August 18, 1986, p. 70. Maxwell John C. (1985) "Cats Trigger Pet-food Gains", Advertising Age, August 12, 1985. Maxwell John C. (1984) "Pet Food See Sales Gain", Advertising Age, June 11, 1984, p. 49. McAuliffe Robert E. (1987) Advertising, Competition, and Public Policy, Lexington, Mass:, Lexington Books. McDonald Colin (1971) "What is the Short-term Effect of Advertising?", reprinted in Effective Frequency, Naples Michael J. (1979), New York: Association of National Advertisers Inc., pp. 105-42. McFadden Daniel (1973) "Conditional Logit Analysis of Qualitative Choice Behaviour," in Frontiers of Econometrics, P. Zarembka (ed.), New York: Academic Press, pp. 105-42. Milgrom Paul and John Roberts (1986) "Price and Advertising Signals of Product Quality," Journal of Political Economy, vol. 94(4), pp. 796-821. Moore William L. and Russell S. Winer (1987) "A Penel-Data Based Method for Merging Joint Space and Market Response Function Estimation" Marketing Science, vol. 6 (Winter), pp. 25-42. Morrison Donald G. (1966) "Interpurchase Time and Brand Loyalty," Journal of Marketing Research, vol. 3 (August), pp. 289-91. Narasimhan Chakravarthi (1985) "A Price Discrimination Theory of Coupons," Marketing Science, vol. 3(2), pp. 128-47. Nelson Philip (1978) "Advertising as Information Once More," in Issues in Advertising: The Economics of Persuasion, David G. Tuerck (ed.), New York: New York University Press, pp. 133-61. 198 Nelson Philip (1975) "The Economic Consequences of Advertising," Journal of Business, vol. 48 (April), pp. 213-41. Nelson Philip (1974) "Advertising as Information," Journal of Political Economy, vol. 81 (July-Aug.), pp. 729-54. Nelson Philip (1970) "Information and Consumer Behavior," Journal of Political Economy, vol. 78(2), pp. 311-29. Neslin Scott A. and Shoemaker Robert W. (1983) "A Model for Evaluating the Profitability of Coupon Promotions", Marketing Science, vol. 2 (Fall), pp. 361-88. Neslin Scott A., Caroline Henderson and John Quelch (1985) "Consumer Promotions and the Acceleration of Product Purchases," Marketing Science, vol. 4(2), pp. 144-65. Neter John and William Wasserman (1974) Applied Linear Statistical Models, 111.: Home-wood, Richard D. Irwin Inc. Ortmeyer Gwen K. (1985) A Model of Consumer Response to Promotions. Unpublished Doctoral Dissertation, Stanford University. Oum Tae Hoon (1979) "A Warning on the Use of Linear Logit Models in Transport Mode Choice Studies," The Bell Journal of Economics, vol. 10(1), pp. 374-88. Prasad Kanti V. and Winston L. Ring (1976) "Measuring Sales Effect of some Marketing Mix Variables and Their Interactions," Journal of Marketing Research, vol. 13 (Nov.), . pp. 391-96. . Rothschild Michael L. and William C. Gadis (1981) "Behavioral Learning Theory: Its Relevance to Marketing and Promotions," Journal of Marketing, vol. 45 (Spring), pp. 70-78. Russell Gary J. and Ruth N. Bolton (1988) "Implications of Market Structure for Elasticity Structure." Journal of Marketing Research, vol. 25 (August). Salop Steven C. (1979) "Monopolistic Competition with Outside Goods," The Bell Journal of Economics, vol. 10(1), 141-56. Salop Steven C. and Joseph Stiglitz (1982) "The Theory of Sales: A Simple Model of Price Dispersion with Identical Agents," American Economic Review, vol. 72(Dec), 1121-30. Salop Steven C. and Joseph Stiglitz (1977) "Bargains and Ripoffs : A Model of Monop-olistically Competitive Price Dispersion," Review of Economic Studies, vol. 54, 493-510. Scherer Frederic M. (1980) Industrial Market Structure and Economic Performance, (Sec-ond Edition), Chicago: Rand McNally. 199 Schmalensee Richard (1978) "A Model of Advertising and Product Quality," Journal of Political Economy, vol. 86 (June), 485-503. Schmalensee Richard (1974) "Brand Loyalty and Barriers to Entry," Southern Economic Journal, vol. 40 (April), 579-88. Schroeter John R., Scott L. Smith and Steven R. Cox (1987) "Advertising and Competition in Routine Legal Service Markets: An Empirical Investigation" Journal of Industrial Economics, vol. 36(1), pp. 49-60. Shapiro Carl (1980) "Advertising and Welfare: Comment," The Bell Journal of Economics, vol. 11(2), pp. 749-52. Shilony Y. (1977) "Mixed Pricing in Oligopoly," Journal of Economic Theory, vol. 14 (April), pp. 373-88. Simon Julian and Johan Arndt (1980) "The Shape of the Advertising Response Function," Journal of Advertising Research, vol. 20 (Aug.), pp. 11-25. Steiner Robert L. (1973) "Does Advertising Lower Consumer Prices?," Journal of Mar-keting, vol. 37 (October), pp. 19-26. Stewart David W. and David H. Furse (1986) Effective Television Advertising: A Study of 1000 Commercials. Lexington Mass.: Lexington Books. Tellis Gerard J. (1988) "Advertising Exposure, Loyalty and Brand Purchase: A Two-stage Model of Choice," Journal of Marketing Research, vol. 25 (May), pp. 134-44. Tellis Gerard J. (1987) "The Price Sensitivity of Selective Demand: A Meta-Analysis of Sales Response Models," Paper Presented at 1988 ORSA/TIMS Marketing Science Conference at Seattle, Washington. Telser Lester G. (1962) "The Demand for Branded Goods as Estimated from Consumer Panel Data," Review of Economics and Statistics, vol. 44, pp. 300-24. Thaler Richard (1986) "Mental Accounting and Consumer Choice", Marketing Science, vol. 4(3), pp. 199-214. Varian Hal R. (1984) Microeconomic Analysis. Second edition, New York, W. W. Norton and Company. Varian Hal R. (1980) "A Model of Sales," American Economic Review, vol 70(4), pp. 651-59. Watkins Trevor (1986) The Economics of the Brand: A Marketing Analysis. London: McGraw-Hill Book company (UK) Ltd. Weinberg Charles B. and Doyle L. Weiss (1982) "On the Econometric Measurement of 200 Duration of Advertising Effect on Sales," Journal of Marketing Research, vol. 19, pp. 585-91. Wildt Albert R. and Russell S. Winer (1983) "Modelling and Estimation in Changing Market Environment" Journal of Business, vol. 56(3), pp. 365-88. Wilkenson J. B., Barry J. Mason and Christie H. Paksoy (1982) "Assessing the Impact of Short term Super Market Strategy Variables," Journal of Marketing Research, vol. 19(1), pp. 72-86. Winer Russell S. (1986) "A Reference Price Model of Brand Choice for Frequently Pur-chased Products," Journal of Consumer Research, vol. 13 (September), pp. 250-56. Winer Russell S. (1986) "A Price Vector Model of Demand for Consumer Durables: Preliminary Developments", Marketing Science, vol. 4(1), pp. 74-90. Wittink Dick R. (1977) "Exploring Territorial Differences in the Relationship between Marketing Variables," Journal of Marketing Research, vol. 14 (May), pp. 145-55. Woodside Arch G. and Gerald L. Waddle (1975) "Sales Effects of In-Store Advertising," Journal of Advertising Research, vol. 15 (June), pp. 29-33. 201 APPENDIX A The purpose of this appendix is to summarize and comment on the analysis repeated for low, medium and high brand loyalty groups for the product category of dry dog food. It is found that a household's brand price sensitivity varies across groups^ as does the effect of television advertising on price sensitivity. In addition, the effect of television advertising on household's price sensitivity is found to be strongest for the high loyalty group compared to the other two groups. This description is followed in two sections. A.l. MODEL ESTIMATIONS by BRAND LOYALTY GROUPS The literature review in chapter II indicated that household price sensitivity as well as the effect of television advertising on household price sensitivity may be dependent on brand loyalty. To investigate this proposition, models are estimated and discussed for the low, medium, and high brand loyalty groups. Before describing the analysis in detail, the procedure used for grouping purchases into the three categories of loyalty is described. Following this, alternative model specifications for the three loyalty groups are given. The measures of brand loyalty used in chapter V to predict brand choices are weighted averages of brand shares for 11 brands prior to a purchase occasion. For every purchase occasion, 11 variables were used to indicate the household's likelihood of buying each of 11 brands. The problem with such a measure of brand loyalty is that a procedure is still needed to identify whether the purchase is made by a low or a high loyalty purchaser. The procedure proposed here uses the maximum value of brand share to group purchases to categorize purchases into low, medium, and high loyalty groups. A maximum value of one indicates that at all prior purchasing occasions, only one brand is chosen. On the other hand, the maximum value of 1/11 indicates that at all prior purchasing occasions, purchases 202 are spread over all brands. These two extreme examples illustrate that the maximum value of brand share, as an index of brand loyalty, is a useful measure to segment product market. A cumulative distribution for the maximum value of brand shares for the product category dry dog food did not indicate a natural cut-off (see Figure 5.1) level to categorize purchases into low and high loyalty groups. Hence, purchases are divided into three categories of loyalty (low, medium and high) with about equal number of purchases in each category. Based on the observed cumulative distribution, cut-off levels of 0.513 and 0.72 provide equal number of purchases in three loyalty groups. Thus, if the maximum value of brand share is less than 0.513, the purchase is considered to be low loyalty. On the other hand, if the maximum value of brand share is greater than 0.72, the purchase is considered high loyalty. All the remaining purchases are considered to be made by medium loyalty households. Results of various specifications for the three groups are now dscribed. Several alternative model specifications for high, medium, and low loyalty group are summarized in Tables A.l, A.2 and A.3 respectively. For the high loyalty group, it is found that brand loyalty accounts for most of the variation in brand choices (Table A.l). Since this group has generally chosen only one or two brands, this result is expected. In addition, the presence of a deal variable is more important in choice of a brand than the deal amount variable. Finally, the effect of the price, television advertising, and price times television advertising variables on brand choices are conditional on the brand loyalty variable. Thus, in the presence of brand loyalty and the television advertising variables, the effect of the price variable is not significant at p < 0.05. On the other hand, in the absence of the brand loyalty variable, the effect of the price variable is significant while that of the advertising variable is not. Thus, the effect of these variables on brand choices is unstable for this high loyalty group. 203 For the medium loyalty group, the brand loyalty variable accounts for about 26% of the variation in brand choices (see Table A.2). In addition, the variable, deal amount, is more important than other sales promotional variables. Moreover, the effect of price times television advertising and display variables on brand choices is stable across model speci-fications. The effect of the price variable on brand choices depends upon the presence of the brand loyalty variable and/or the variable price times television advertising variables. Finally, the effect of television advertising variable on brand choices depends upon the presence of the price times television advertising variable. For the low loyalty group, the brand loyalty variable accounts for about 14% of the variation in brand choices (see Table A.3). In addition, the effect of deal amount, presence of a deal, display, and price variables on brand choices is stable across model specifications. As might be expected, the sales promotional variables account for more variation in brand choices than the brand loyalty variable. Finally, the effect of television advertising variables on brand choices is minimal, and statistically not significant at p < 0.05. A.2. COMPARISONS across BRAND LOYALTY GROUPS In Table A.5 and Table A.6 a summary of parameter estimates across loyalty groups is reported. A comparison of order of importance by variables (using t-statistic) in predicting brand choices across loyalty groups reveals an interesting pattern. Brand loyalty is the most important variable in predicting brand choices across all the loyalty groups. For the low loyalty group, the next three most important variables that influence brand choices are the deal amount, price, and display. For the high loyalty group, the order of variables is presence of a deal, deal amount, and price times television advertising. For the medium loyalty group, the order of variables is deal amount, display, and presence of a deal. This 204 pattern of responses is unchanged for the parameter estimates reported in Table A.5. These patterns suggest that the three loyalty groups focus on different variables in choosing brands. The effects of these variables on a hypothetical average brand are quantified below. To facilitate comparison of parameter estimates across loyalty groups, elasticities are derived for several variables. In addition, holding variable values of the average brand constant, responses of the three loyalty groups are studied. If m is the market share of the average brand and Xk is the value of the brand choice variable k and (3^ is the parameter estimates for variable k and loyalty group j, then Guadagni and Little (1983) show that elasticity t of variable k for the loyalty group j can be written as dm Xk dXk m The parameter estimates reported in Table A.5 and A.6 may be used to compute elasticities for all independent variables. To complete the calculations, however, values for the variables, Xk are needed. These are set as follows: The average brand is priced at 50^ per pound. When the deal is available, the price is cut by 5^ per pound. The average brand has two television advertising exposures per household for each purchase occasion. The definition used to categorize households into low, medium, and high loyalty groups suggests that the values of the brand loyalty variable must vary across groups. For example, for the low loyalty group, the value of the brand loyalty variable should be between zero and 0.513, for the medium loyalty group, the range must be 0.513 to 0.72,. and for the high loyalty group, the range must be 0.72 to one. For the average brand, mid-points of these ranges are taken as the level of the brand loyalty variable. Finally, the market f In deriving equation (5.4), Guadagni and Little assumed that there is homogeneity in the predicted choice probabilities within a group. This, of course, is a common assumption in any segmentation analysis. 205 share of the average brand is assumed to be 0.15f. To compare the elasticities, along with variable values, appear in Table A.7. For presenting elasticities for comparative purpose, only parameter estimates that are significantly different from zero at p < 0.10 in Table A.6 are used. Since the variable feature is significant at p < 0.10 for only the low loyalty group, elasticities for this variable are not reported. We w i l l discuss important distinctions across groups below. First , the low loyalty group responds to price changes more than the other two groups. For example, a one percent increase in the regular price of the average brand is expected to result in 2.5% decrease in the market share of the brand. O n the other hand, for the medium and high loyalty groups, a one percent increase in the regular price of the average brand is expected to result in about one percent decrease in the market share. This result is consistent with the empirical work reported by Massy and Frank (1965). Second, the effect of television advertising is small and positive on the choice prob-abilities of the average brand. A s one might expect, the effect is stronger for the high loyalty group than the medium loyalty. T h e effect of television advertising on the brand price sensitivity is the strongest for the high loyalty group. Thus, the price sensitivity of this group increases from 1.22 to 1.479 whereas the sensitivity of low loyalty group is 2.516. If we compare the values of price sensitivities across these two groups, these results show consistency wi th those by K r i s h n a m u r t h i and Raj (1985). They report that the high loyalty group is less price sensitive than the low loyalty group. However, the finding that the effect of television advertising is a larger increase in price sensitivity among the high loyalty group than the other groups, is surprising and deserves comment. f The effect of market share is to scale the elasticities and as such no substantive conclusions would be altered by choosing some other value of the market share. 206 When parameters are estimated separately, it is implicitly assumed that parameters are estimated with the same precision. This is not the case here. It is noted earlier that for the high loyalty group, the parameters of price, television advertising, and price times television advertising are unstable. Thus, it can not be determined whether the observed effects of television advertising are product specific or artifacts of estimation. In general, however, the effect of television advertising is an increase in brand price sensitivity. Continuing the comparisons, it is noted that in presence, of display, the market share of the average brand is expected to increase by about one percent for the low loyalty group. The market share change for the medium loyalty group due to the display is 1.17%. When the deal of 5c7 per pound is offered, the market share changes of the average brand has two components. The first component, the deal amount, has a strong influence on the low loyalty group. On the other hand, the second component, the presence of the deal, has a strong influence on the high loyalty group. Finally, the joint effect of these two components is to increase the market share of the average brand by 1.7%, 1.5% and 2.1% for low, medium, and high loyalty groups respectively. As might be expected, the effect of the brand loyalty variable is strongest for the high loyalty group, even though the parameter estimate for the low loyalty group is the highest. We may ask a question, whether it is possible to find an extreme example, where the impact of the brand loyalty will be greater for the low loyalty group than the high loyalty group, given the parameter estimates for the brand loyalty variable in Table 5.15? This requires finding values of the brand loyalty variable for the high loyalty(Xft) and the low loyalty (Xi) groups that satisfy /3;X/(1 — m) > 3^X^(1 — ra) where 3\ and 3^ are parameter estimates for low and high loyalty groups respectively. If the market share is held constant, then substituting values of 3i = 6.108 and 3h = 4.489 from Table A.5, gives Xt/Xh > 4.489/6.108. • 207 However, we know the ranges of Xh and Xi for the average brand. Thus, the feasible ranges are 0.72 < Xh < 1 and 0 < X, < 0.521. Substituting the lower bound of Xh (0.72) and the lower bound of Xi (0.521), it is possible to conclude that the impact of the brand loyalty variable will always be greater for the high loyalty group than the low loyalty group. This is an additional assurance that the estimated models are reasonable. In this section, brand choice models for the three brand loyalty groups are reported and are interpreted. The analysis reinforced earlier findings that the brand loyalty variable is the most important variable predicting brand choices. The effect of other variables, however, depends upon the loyalty groups. Finally, it is found that the effect of television advertising is an increase in the brand's price sensitivity. 208 Table A. 1 E f f e c t of A l t e r n a t i v e SpectTIcatlon for High Lo y a l t y Group for Dry Dog Food Brand Choice Models Maximum L i k e l i h o o d Parameter Estimates (Asymptotic t s t a t i s t i c in parentheses) Independent Model with Brand Promot lona 1 Adver t I s Ing Pr Ice Reference Brand Adver t i s it v a r i a b l e brand l o y a l t y var lab les v arlab le var i a b l e Model loya Ity var tables constants Included added added added excluded exc luded Ml -H M2-H M3II W-H M5 II MB-H M7 -11 MB tl P r i c e -0.0375 -0.0264 -0.0594 -0.0359 (-2.46) (-1 721 ( 5.34) ( 2 . 4 0 ) T e l e v i s i o n -0.0497 -0.0520 0.170 0.0655 advert i s I n g (-2.38) I 2.47) (1.75) (1.16) D i s p l s y l 1 . 0 ) 0 752 0 881 0.943 0 891 0.111 1 .028 (1.071 (0.96) (1.32) (1.22) (0.22) (1 46) Feature!1.0) 0 433 0 896 0 571 -0, 0358 -0 637 0. 141 (0 .53) (0 87) (0 . 70) (-0.02) ( -0 53) (0.07) If brand Is on 1 . 759 1 745 1 642 1 696 1 429 1 666 d e a l ! 1 . 0 ) (4 .92) (4 .92) (4 .54) (4.64) (6 26) (4. 58) Deal amount 0 . 148 0 153 0 158 0. 157 0 126 0 151 (3 . 10) (3 33) (3 .40) (3.40) (3 51) (3. 14) Brand loya1ty 4.223 4 420 4 .482 4 453 4. 489 4. 388 (25.7) (23 1) (22 8) (22.5) (22.3) (22 8) Pr Ice x T e l e v i s i o n -0 00420 -0 .00133 advert IsIng ( 2.31) ( 1.21) Pr Ice x Feature 0. 0202 0 0312 0. 0 0 3 4 ; (0 47) 11.01) 10.08) Brand constant Brand A 1 956 0.573 0 648 0 714 0 822 0. 680 2. 158 0. 747 (6 85) (1.41) ( 1 .50) (1 66) ( 1 90) (1. 55) (7. 00) ( 1. 73) Brand B 0 .452 -0.878 -0 .840 0 767 -0 352 -0. 511 1 344 -0. 442 (1 .32) (-1.83) (-1 61) (-1 47) ( 0 84) (-0. 92) (3 49) ( 0. 80) Brand C 0 .693 0.54 7 0 514 0 857 1 472 1. 334 2 093 1. 287 (2 .11) (1.23) ( 1 05) ( 1 34) (2 49) (2. 26) (4 881 (2 19) Brand 0 0 .0 -0 488 0 639 -0 511 0 .204 0 0222 1 139 0. 145 (0 0) (-0.94) (-0 95) (-0 90) (0 32) (0 04) (2 521 (0. 23) Brand E 1 651 0.827 0 423 0 . 848 1 329 1. 142 2 323 0. 863 (5 66) ( 1.54) (0 .98) (1 85) (2 66) (2 28) (6 51) ( 1 84) Brand F 0 496 0.844 0 740 0 784 1 310 1 180 1 234 1 244 (1 46) (1.95) (1 .55) (1 84) 12 52) (2 26) (3 13) (2. 39) Brand G 0 .0 -0.242 -0 .352 -0 151 -0 0211 -0. 240 0 259 -0. 232 (0 0) (-0.47) ( -0 63) (-0 27) ( 0 04) (-0. 41) (0. 63) (-0 41) Brand H 1 .273 -0 114 -0 .469 -0 531 -0 488 0. 529 0. 881 -0. 426 (4 21) (-0.27) (-1 0 0 ) (-1 I i ) (- 1 041 (- 1. ( t l (2 75) ( 0 90) Brand I 1 074 0.2B5 x 0 .403 0 357 -0 334 -0 136 0 132 -0 253 (3 47) (0.65) (0 .86) (0 76) ( 0 82) (-0. 25) (0 32) ( -0 47) Brand J 1 189 0.352 0 318 0. 303 0. 666 0. 511 1. 812 0. 666 (3 90) (0.80) (0. 67) ( 0 . 64) (1. 34) (1 02) (5. 25) ( 1 34) Brand K 0 0 0 0 0 0 0 0 rho-squared 0 0.815 0 688 0. 691 0 . 695 0. 698 0. 144 0. 692 Log-1Ikelihood -928.7 -357.3 -289.6 - 286.8 -283.5 -280.7 - 794 7 -286 5 Tab Is A .2 Effect of Alternative Specification for Medium Loyalty Group for Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates (Asymptotic ( s t a t i s t i c In parentheses) Independent var lab le Model with brand constants Ml-M Brand loyaIty tncluded M2- M PromotlonaI var tables added Ml M Adver tIsIng varlable added M4 M Pr Ice varlab le added MS M Reference . Brand Model loyalty exc luded MB-M M7-M Advert IsIng var tables excluded M8-M Price •0.0305 •0.0241 -0.0187 •0 0304 (2.91) (2.22) (2.04) (-2 89) Te lev Is Ion 0.000921 •0 00285 0. 113 0. 127 adver tIsIng ( 0.07) ( 0.21) (2.04) (2 56) Display!1,0) 1.258 1 255 1.425 1.373 1 015 1.385 (2.841 (2.84 1 ( 3.191 (3.08) (2 55) (3 09) Featured.0) _ -0.830 -0.0825 0 948 1.979 I 343 1.880 ( I I I ) (1.10) (-1 25) (0 94) (0 75) (0.901 If brand Is on 0 698 0 695 0 .655 0 623 1.011 0. 6 36 deal!1.0) (2 .65) (2 .65) (2 .50) (2 34) (4 50) (2 41) Deal amount 0 229 0 229 0 235 0 240 0. 172 0. 237 (6. . 121 (6. 13) 16 25) (6 24) (5.34) (6. 25) Brand loyalty 4.213 4 .445 4 445 4 463 4. 481 4 . 468 (213) (20.2) (20 2) (20.I) 120 11 (20 .2) Price x Television -0 00227 -0.00257 adver tIsIng ( 2.10) (-2.69) Price x Feature -0 0749 -0 0528 -0 0739 (-1 1.39) (- 1.10) ( 1 .39) Brand constant Brand A 0 102 -0.00821 -0 . 178 -0 174 -0 0981 -0 157 -0.0358 -0 . 109 (0 39) (-0.21) ( 0. 57) ( 0 57) (0.32) (-0. 50) (-0.13) (-0 .35) Brand B -0 288 -0.495 -0 562 -0. 560 0 211 -0. 263 •0.0178 0 .219 (-1 00) ( 1.53) (•1 68) ( - 1 . 67) ( 0 59) t 0 74) (-0.0B) (-0 .62) Brand C 0. 429 0.501 0 103 0 105 0 783 0 713 0 650 0 .778 (1. 77) ( 1 83) (0 35) (0. 36) 12 07) ( 1 88) (1.98) (2 07) Brand 0 0. 429 0.00895 -0 307 •0. 307 0. 271 0 165 0.497 O 268 (1. 77) (0.33) I-1 04) (-1 04) (0 76) (0 46) (1.60) (0 . 75) Brand E 1. 350 0.664 0 325 0. 332 0 767 0 730 1. 358 0 . 762 (8. 35) (2.74) (1 25) (1. 21) (2 46) (2. 33) (5 03) (2 .62) Brand F 0. 223 0.762 0 391 0. 394 0 883 0 830 0 198 0 .876 (0. 88) (2.77) (1 32) (1. 32) (2 60) (2. 43) (0.64) (2 .59) Brand G - 1 135 -0.574 -0. 780 -0. 778 -0 599 -0 674 - 1.358 -0 .615 ( -2 96) ( 1 43) (-1 84) ( 1 81) ( 1. 38) 1-1. 55) I 3 27) ( • 1 43) Brand H 0 916 0 308 -0. 184 -0. 184 -0. 201 -0. 168 0 519 •0 . 201 (4. 10) (119) ( 0 67) ( -0. 67) < 0. 73) ( -0. 61) (2.21) ( -1 0.731 Brand I -0. 388 -0 770 0 899 •0. 899 - 1 . 457 -1. 354 -0 858 - 1 486 ( - 1. 30) (2.34) ( 2 . 63) ( 2 . 63) ( 3 75) ( -3 44) (-2.501 ( -3 .80) Brand J -0. 388 -0.00863 •0. 253 -0. 252 0. 0'469 -0. 0167 •0 384 0 0383 ( - 1 . 30) ( 0.271 ( -0. 76) ( 0. 76) (0. 13) (-0. 051 (- 1.20) (0 .11) Brand K 0 0 0 0 0 0 0 0 rho-squared 0 0.261 0. 387 0. 3B7 0. 391 0. 395 0. 151 0. 392 Log-1Ikel(hood -932.9 •689.7 -572. 1 -572.1 -567.9 -564,8 -792.3 -566.8 Table A 3 Effect of Alternative Specification for Low Loyalty Group for Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates (Asymptotic t statistic in parentheses) Independent var lab le Model with brand constants MIL Brand loyalty I nc luded M2-L Promotlona1 var lab les added M3-L AdvertIsing var(able added M4 - L Price varlabia added M5- L Reference Model M6-L Brand loyalty excluded M M Adver tIsIng varlables exc luded M8-L Price Telev 1s Ion edver 11sIng -0.0611 -0 .0592 -0 .0588 -0.0608 (-5.81) (-5 . 14) (-5 52) (-5.54) 0.00569 0.00466 0 0288 0 278 (0.41) (0.341 (0 44) (0 45) Display!1,0) Feature!1.0) If brand Is on deal!1.0) Deel amount Brand loyalty Price K Television adver tIsIng Pr Ice x Feature 5.838 (16.5) I 273 1.270 1.251 1.207 0 826 1.211 (2.65) (2.97) (2.97) (2.77) (2 64) (2.65) 0.534 0.544 0.313 2.280 2 152 2.256 (0.90) (0.92) 10.53) (1.77) ( 1 79) I 1 76) 0.979 0.975 0.744 0 739 0 .770 0 738 (3.57) (3.55) 12.63) (2.59) (2 91) (2.59) 0 252 0.253 0.271 0.275 0 266 0 274 (6.40) 16.42) (8.61) (6.59) (6 89) (6.59) 6.078 6.075 8. 105 6. 108 B. 107 (15.2) (15.2) (15.0) 115.01 ( 15 0) -0.000464 -0. 000370 (-0.39) ( -0. 33) -0.0592 (-1.81) -0.0574 (-1.72) -0.0590 (-1.61) Brand constant Brand A 0.322 0 399 0.640 (1.32) ( 1 54) (2.28) Brand a -1.286 -0 953 •0 598 (-3.22) ( 2 321 I 1 39) Brand c 0.417 0.559 0.313 (1.74) (2.21) 11.05) Brand 0 0. 792 0.586 0.474 (3.54) (2 48) (174) Brand E 0.923 0.426 0 172 (4.21) (1.85) (0.66) Brand F 0.417 0.865 0.648 (1.74) ( 3 44) 12 20) Brand G -0.0351 0.667 0.661 (-0.13) (2 41) (2.10) Brand H 0.417 0.488 0.396 (1.74) ( 1 94) 1 I 43) Brand I -0.148 -0.00307 0.229 (-0.54) (-0.011 (0.76) Brand J -0.715 0 223 0.325 (-0.27) (0.79) (1.04) Brand K 0 0 0 0 626 0 .838 0 .821 0 820 0 .846 (2 20) (2 86) (2 77) (2 27) (2 90) -0 811 0 .565 0 .0452 -0 287 0 0712 - 1 41) (0 12) (0 . 10) (-0 66) (0 15) 0 , 300 1 .771 1 . 767 1 513 1 798 ( 1 00) (4 .351 (4 30) (3 98) (4 .43) 0 .473 1 621 1 585 1 608 1 .810 ( 1 74) (4 63) (4 45) (4 91) (4 59) 0 121 1 030 1 006 1 440 1 060 (0 42) (3 .04) (2 .93) (4 53) (3 .35) 0 634 1 .584 1 .559 0 958 1 586 (2 14) (4 58) (4 48) (2 99) (4 60) 0 637 1 .085 1 063 0 220 1 103 ( 1 99) (3 23) (3 11) (0 681 (3 33) 0. 398 0 413 0 .417 0 302 0 413 ( 1 44) (1 47) (1 49) (1 15) 1 1 47) 0 231 -0 809 -0 821 - 1 060 -0 .848 (0 76) ( 2 27) ( 2 26) (-3 09) (-2 .36) 0 327 0 960 0 940 0 553 0 954 ( 1. 05) (2 84) (2 77) (I 74) (2 .82) 0 0 0 0 0 rho-squared 0 0.138 0.369 0 369 0.387 0.388 0 270 0.388 Log-1IkelIhood -975.7 -841.1 -615.5 -615.4 -598.5 -597.0 -712.6 -597.1 T a b l e A.4 C o m p a r i s o n o f S p e c i f i c a t i o n s f o r T h r e e L o y a l t y G r o u p s a n d D i s a g g r e g a t e d Form o f A d v e r t i s i n g f o r D r y Dog F o o d B r a n d C h o i c e M o d e l s Maximum L i k e l i h o o d P a r a m e t e r E s t i m a t e s ( A s y m p t o t i c t - s t a t i s t i c In p a r e n t h e s e s ) I n d e p e n d e n t Low L o y a l t y Medium L o y a l t y H i g h L o y a l t y V a r i a b l e s ( n = 426) (n = 426) (n = 424) P r i c e -0.0565 -0.0238 -0.0256 (-4.87) (-2.21) (-1.68) B r a n d A a d . 0.554 1.277 0.638 e x p o s u r e s ( 1 . 6 7 ) ( 2 . 7 2 ) ( 1 . 3 5 ) B r a n d B a d . 1.137 0.293 1.626 e x p o s u r e s ( 1 . 5 7 ) ( 0 . 2 9 ) (1 .52) B r a n d E a d . 0.0444 0.135 0.212 e x p o s u r e s ( 0 . 6 1 ) ( 2 . 2 7 ) ( 1 . 9 7 ) D i s p l a y ! 1 . 0 ) 1.131 1.359 1 .047 ( 2 . 4 6 ) ( 3 . 0 1 ) ( 1 . 4 1 ) F e a t u r e ( 1 . 0 ) 2.294 1.763 -0.184 ( 1 . 7 7 ) ( 0 . 8 3 ) (-0.09) I f b r a n d i s on 0.757 0.62B 1 .653 d e a l t 1.0) ( 2 . 6 5 ) ( 2 . 3 4 ) ( 4 . 4 4 ) D e a l amount 0.274 0.242 0 . 163 (6.56 ) ( 6 . 18) ( 3 . 4 3 ) B r a n d l o y a l t y 6.119 4.449 4.513 ( 14.9) ( 1 9 . 9 ) ( 2 2 . 2 ) P r i c e x B r a n d A -0.0130 -0.0309 -0.0132 a d . e x p o s u r e s (-1.67) (-2.76) (-1.21) P r i c e x B r a n d B -0.0219 -0.00964 -0.0348 a d . e x p o s u r e s (-1.49) (-0.51) (-1.70) P r i c e x B r a n d E -0.000814 -0.00289 -0.00500 a d . e x p o s u r e s (-0.62) (-2.48) (-2.43) P r i c e x F e a t u r e -0.0594 -0.0736 0.0247 (-1.62) (-1.34) ( 0 . 5 6 ) r h o - s q u a r e d 0.390 0.401 0.701 L o g 1 i k e l i h o o d ( O ) -975.7 -932.9 -928.7 L o g 1 i k e l i h o o d ( b ) -594 .8 -558.7 -277.5 B r a n d c o n s t a n t B r a n d A 0.B36 -0.0753 0.475 (2.54) (-0.21) ( 1 . 0 0 ) B r a n d B -0.0223 0.0638 -0.260 - (-0.04) ( 0 . 1 6 ) (-0.39) B r a n d C 1.681 0.622 0:998 ( 4 . 0 7 ) ( 1 . 6 4 ) ( 1 . 6 8 ) B r a n d D 1.517 0.115 -0.152 (4.27) ( 0 . 3 2 ) (-0.24) B r a n d E 0.979 0.816 1 . 123 (2.81) ( 2 . 6 0 ) (2.20) B r a n d F 1.509 0.739 1 .020 (4.33) ( 2 . 1 6 ) ( 1 .92) B r a n d G 1.056 -0.64B -0.359 (3.19) (-1.51) (-0.62) B r a n d H 0.411 -0.162 -0.501 (1.46) (-0.59) (-1.05) B r a n d I -0.787 -1.331 -0.107 (-2.17) (-3.38) (-0.20) B r a n d J 0.897 -0.0209 0.471 ( 2 . 6 5 ) (-0.06) ( 0 . 9 3 ) B r a n d K 0 0 0 212 Table A.5 Comparison of Specifications for Three Loyalty Groups and Aggregated Form of Advertising for Dry Dog Food Brand Choice Models Maximum Likelihood Parameter Estimates (Asymptotic t - s t a t i s t i c in parentheses) Independent Low Loyalty Medium Loyalty High Loyal Var iables (n = 426) (n = 426) (n = 424 ) Price -0.0592 -0.0241 -0.0264 (-5.14) (-2.22) (-1.72) Telev i s ion 0.0288 0.113 0. 170 adver t i s i n g (0.44) (2.04) ( 1 .75.) Display(1,0) 1 .207 1 . 373 0.891 (2.77) (3.06) (1.22! Feature!1.0) 2. 280 1.979 -0.0358 (1.77) (0.94) (-0.02) If brand is on 0.739 0.623 1 .696 deal!1.0) (2.59) (2.34) (4.64! Deal amount 0.275 0.240 0. 157 (6.59) (6.24) (3.40) Brand loyalty 6.108 4.461 4 .489 (15.0) (20 . 1 ) (22.3! Price x Tele v i s i o n -0.000464 -0.00227 -0.00420 adver t is ing (-0.39) ( -2.10) (-2.31) Price x Feature -0.0592 -0.0749 0.0202 (-1.61) ( - 1 . 39) (0.47) Brand constant Brand A 0 .821 -0 . 157 0 . 680 (2 .77) ( -0 . 50) ( 1 . 55) Brand B 0 .0452 -0 . 263 -0 .511 (0 . 10) ( -0 .74) (-0 .92) Brand C 1 . 767 0 .71? 1 . 334 (4 .30) ( 1 .88) (2 .26) Brand 0 1 .585 0 . 165 0 .0222 (4 .45) (0 .46) (0 .04 ) Brand E 1 .006 0 .730 1 . 142 ( 2 .93) ( 2 ; .33) (2 .26! Brand F 1 .559 0 .830 1 . 180 (4 .46) (2. 43) (2. .26) Brand G 1 . 063 -0. 674 -0. 240 (3 .11) ( - 1 . 55) ( -0. .41 ) Brand H 0 .417 -0 . 168 -0. 529 ( 1 . 49) ( -0. 61 ) (-1. 1 1 ! Brand I -0. .821 - 1 . 354 -0. 136 ( - 2 25) ( -3. 44) ( -0. 25) Brand J 0. 940 -0. 0167 ' 0. 511 ( 2 . 77) ( -0. 05) (1. 02) Brand K 0 0 0 rho-squared 0.388 - 0.395 0.698 Log 1ikelihood(O) -975.7 -932.9 -928.7 Log 1ikelihood(b) -597.0 -564.6 -280.7 213 Table A.6 E l a s t i c i t y Estimates for Selected Brand Choice Variables by the Brand Loyalty Groups Independent Variable Brand Loyalty Group variable value Low Medium High Price 50 cents per pound -2.516+ - 1 .024 - 1 . 122 Telev is ion advert is ing 2 exposures — 0 . 192 0 .289 Display 0 or 1 1.026* 1 . . 167 If brand is on deal 0 or 1 0.628 0 .530 1 .442 Deal amount 5 cents per pound 1 . 169 1 .020 0 .667 Brand loyalty varies 1.332 2 . 338 3 . 282 Price x Tele-v is i o n adv. 100 cents-exposures per pound — -0 . 193 -0 . 357 * A one percent increase in the price of the average is expected to result in a 2.516 percent decrease in the market share for the low loyalty group. * When the display is present (1), the market share of the average brand is expected to increase by 1.026 percent for low loyalty group. 214 F i g u r e A.1 Cumulative Distribution of Maximum Loyalty High Loyalty Group 0 20 40 80 00 100 120 Index of Loyalty 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0098302/manifest

Comment

Related Items