Investigating the Brand Choice Decision Using Hierarchical Models on Aggregated Panel Data

ABSTRACT - There exists a lot of research investigating the role of the brand name in the consumer purchase decision. Especially in the second half of the eighties a vast amount of experimental studies have been carried out on individual data of decision processes with mixed results. In this paper we use data from a representative consumer panel and a hierarchical modeling approach. The study is based on the hypothesis that the brand name is the key criterion in nondurable consumer goods. The method of hierarchical models is used to investigate the meaning of Abrand name@ versus Aproduct attributes@ as key criterion. The paper is organized as follows: After a review of the theoretical approaches to explain the choice decision, we describe the method of hierarchical model and show that this approach has some advantages to investigate the meaning of different choice criteria. As a result we prove that in a road diversified product category brand name is a more decisive decision criterion than other product criteria.

Citation:

Lutz Hildebrandt and Kerstin Kamlage (2001) ,"Investigating the Brand Choice Decision Using Hierarchical Models on Aggregated Panel Data", in E - European Advances in Consumer Research Volume 5, eds. Andrea Groeppel-Klien and Frank-Rudolf Esch, Provo, UT : Association for Consumer Research, Pages: 187-192.

European Advances in Consumer Research Volume 5, 2001      Pages 187-192

INVESTIGATING THE BRAND CHOICE DECISION USING HIERARCHICAL MODELS ON AGGREGATED PANEL DATA

Lutz Hildebrandt, Humboldt-University of Berlin, Germany

Kerstin Kamlage, Humboldt-University of Berlin, Germany

ABSTRACT -

There exists a lot of research investigating the role of the brand name in the consumer purchase decision. Especially in the second half of the eighties a vast amount of experimental studies have been carried out on individual data of decision processes with mixed results. In this paper we use data from a representative consumer panel and a hierarchical modeling approach. The study is based on the hypothesis that the brand name is the key criterion in nondurable consumer goods. The method of hierarchical models is used to investigate the meaning of "brand name" versus "product attributes" as key criterion. The paper is organized as follows: After a review of the theoretical approaches to explain the choice decision, we describe the method of hierarchical model and show that this approach has some advantages to investigate the meaning of different choice criteria. As a result we prove that in a road diversified product category brand name is a more decisive decision criterion than other product criteria.

INTRODUCTION

To satisfy our daily needs we in general purchase the necessary product bundle, that means for each product class, we have chosen between two or more alternatives. In classical economic theory it is assumed that these decisions are the result of a rational decision process. But in reality this assumption is not tenable, since in most purchase decision processes the consumer does not have the capabilities to compare all advantages and disadvantages of each alternative in a multi-product market to identify the best choice. In real life, decision processes in general will be based on a limited set of comparable attributes or on the use of some key information and a simplified decision process. As such the knowledge about the use of specific key information is critical for the marketing manager.

Although we have considerable knowledge about consumer decision processes, empirical results are still limited. First, most of the earlier studies are rather small and solely based on experimental data. Second, later studies start to use real market data but are frequently concentrated only on single hypothesis or sometimes disregard the hierarchy of the decision process. In this study we use real market data, especially aggregated household panel data, but apply the method of hierarchical linear models to investigate the relative relevance of information like brand names and product attributes in a choice hierarchy. In the first section we describe some aspects of the brand choice decision. At first we review the results of experimental studies concerning choice criteria, with a focus on the brand name. Then we present some empirical studies with panel data that investigate the meaning of different choice criteria. Subsequently we will provide the alternative way of hierarchical models to describe and evaluate the meaning of these features.

THE MEANING OF BRAND NAME IN THE BRAND CHOICE DECISION RESEARCH

Experimental studies

In economic theory consumers are assumed to make rational purchase decisions by using and evaluating all available information. In consumer behavior research this approach has been rejected because it does not properly represent the decision process of the consumer. The key work of Bettman (1979) on his information-processing model already stressed that consumers only have a limited capacity for processing information.

Classical experimental studies have shown that consumers only process a few information dimensions (chunks) to make a buying decision for frequently purchased goods (e.g. Jacoby, Szybillo & Busato-Schach 1977, Alpert 1980). This behavior can be described by applying decision models based on cognitive algebra. Some studies investigated the meaning of brand name versus other product attribute information for the decision process. A focal point is the question if consumers make a choice by processing brand names or by processing attributes. The results of the experiments are mixed with a slight advantage for processing by attribute as the preferred evaluation strategy (Bettman & Kakkar 1977, Biehal & Chakravarti 1986), but the brand plays a key role as a product attribute.

The results of experimental studies show that brand names as well as price information are used as indicators for quality in the purchase decision process (Obermiller & Weathley 1985, Dodds & Monroe 1985, Obermiller 1987, Lefkoff-Hagius & Mason 1993). Some researchers also assume that the use of brand names is a choice tactic for inexperienced consumers to choose among different brands and that positive experience with a brand increases the ability to remember the brand (Biehal & Chakravarti 1986, Hoyer & Brown 1990).

Household panel data studies

In contrast to the use of experimental studies many researchers in later studies have analyzed consumer panel data (or market data) to investigate the brand choice decision. Most of the studies examine the hierarchy of the evaluation procedure. The results of these studies depend in general on the examined product category. But even for the same product category the choice criteria may differ. The reported studies on switching and brand choice behavior for soft drinks or coffee characterize the status of the research field.

Rao and Sabavala (1981) applied hierarchical cluster analysis on switching data for 17 soft drink brands. The results show that national vs. regional distribution seems to be the most important criterion, followed by brand names for the regional products, while four of the national brand names formed own segments. For the national brands the diet vs. non-diet attribute appeared as the next most important attribute, followed by the flavor (lemon vs. cola) for the diet segment. Kumar and Sashi (1989) analyzed switching data of eight soft drinks and compared different market structures by using different sequences of choice criteria. They came to the result that diet (vs. non-diet) is the most important criterion in the case of brand switching. A consumer already drinking a diet drink will switch to another diet drink no matter of cola or non-cola. An alternative approach by Moore and Lehmann (1989) used dissimilarity ratings and constant-sum buying intentions to rate twelve soft drinks. They found that either flavor, diet (vs. high calorie) or caffeine (vs. none) is the first decision criterion and no more than two criteria are necessary for the decision.

Currim, Meyer and Le (1988) investigated the consumer choice of coffee with so called concept learning system (CLS) algorithm. CLS is a classification procedure that relates a set of predictor variables (like product attributes) to a discrete outcome variable (like choice) and generates decision trees. The sample contained data of five regular ground coffee brands with the variables price levels, features, displays as well as brand names and store chains. They found three major decision trees. 35% of the consumers consider price in the first step, 33% starts with the brand name and 27% with features. 60% of the consumers who consider brand name as the first criteria use price as the second attribute. In the same product category Vilcassim (1989) tested different choice hierarchies by comparing corresponding regression models of choice criteria and came to the result, that the first decision is made between ground and instant coffee, then between caffeinated and decaffeinated, followed by the choice between regular and freeze dried.

Although the previous studies have contributed significantly to the research of decision processes, they suffer from several problems. Experimental studies miss external validity because of the artificial buying or choice situation. The main problem with studies using household panel data is the use of only product category specific choice criteria like caffeinated vs. decaffeinated for coffee or diet vs. non-diet for soft drinks. Therefore the results are limited to the selected product category. More general product attributes like brand name or price are rarely part of the analysis. Second, the applied methods (like hierarchical cluster analysis) are not able to measure the size of the effect of the product attributes.

AN ALTERNATIVE APPROACH APPLYING HIERARCHICAL MODELS

The study conducted here differs from the existing research in three aspects. First, instead of individual choice data we analyze aggregated household panel data which from the practical perspective are available in every retail chain, second, we choose general choice criteria like brand name and price so we can in a further step compare different product categories, third we apply the method of hierarchical models to investigate the meaning of brand names vs. product attributes for the choice process. An advantage of the hierarchical method is that we can measure the size of the influence of the choice criteria.

Based on the results of the experimental and individual household panel data studies we investigate the following two hypotheses:

Hypothesis 1: The brand name is more important than the product attribute in nondurable consumer goods.

This hypothesis is coincide with many experimental studies and also with the panel study by Currim, Meyer and Le (1988), where, apart from price, the brand name is the most important variable.

Hypothesis 2: The size of the influence of the brand name remains high while controlling for further criteria.

This means, that the brand name is a relative robust and independent attribute in the brand choice process. After a brief introduction in the method of hierarchical models, we describe our data set and specify the model and its extensions. Subsequently we will present and discuss the results.

Hierarchical linear models

Hierarchical models are able to analyze data structures in which units are nested within larger clusters (like people which can be grouped into families) (Bryk & Raudenbush 1992). In our kind of data some units should be more similar than others and the assumption of independent observations, which is a basis for classical statistical analysis, can not be uphold. The idea of hierarchical models is to consider that units in the same group are more related than units in different groups, because they share common, possibly unobservable characteristics, which are causing correlation between disturbances (Bryk & Raudenbush 1992). In hierarchical models this disturbances are covered by random effect variables for the groups. Consequently, hierarchical models must consist of at least two levels. The level-1 is the lowest level and represents the individual unit, like people or products. The level-2 is the grouping level, like families. For each level an equation is set up, where one or more of the coefficients of the level-1 are functions of level-2 effects.

The most simple case of a hierarchical model is a unconditional two-level model:

Level-1: Y_ij = ¯j + r_ij

Level-2: ¯_0j = y₀₀ + u_0j

with

Y_ij : response variable, i = 1,, N_j j = 1,M

¯_oj : intercept, determined by a higher level

r_ij : random error, r_ij ~ N(0, s²)

y_oo : overall mean

u_oj : random effect parameter, u_oj ~ N(0, t)

The substitution of the level-2 model into the level-1 model leads to the combined hierarchical model:

Y_ij = y₀₀ + u_0j + r_ij

where u_oj is a random variable with normal distribution, mean of zero and variance of t. This specification allows to take into account the heterogeneity which is used as an indicator for unobservable characteristics of the level-2 groups. The random effect parameter u_oj explains part of the variability in level-1 (Littell et al. 1996).

The basic model can be extended to a conditional model by the introduction of variables in level-2:

Level-1: Y_ij = ¯_oj + r_ij

Level-2: ¯_oj = y₀₀ + y₀₁W_j + u_oj

Substitution of level-2 model yields to the combined conditional model:

Y_ij = y₀₀ + g₀₁W_j + u_oj+ r_ij

The same is possible in level-1. The introduction of covariates in level-1 leads to a random-coefficient model (Longford 1993):

Levl-1: Y_ij = ¯_0j + ¯₁X_ij + r_ij

Level-2: ¯_oj = y₀₀ + y₀₁W_j + u_oj

¯_1j = y₀₀ + y₁₁W_j + u_1j

Substitution of level-2 model yields to the combined model:

EQUATION

with

r_ij ~ N(0,s²)

EQUATION

This model contains fixed and random effects and the estimation of the combined model is here done by PROC MIXED of SAS with the maximum likelihood estimator. Criteria used for the goodness of fit of the models are AIC (Akaike’s Information Criterion), SBC (Schwarz’s Bayesian Criterion) and the intraclass correlation coefficient. They are defined as follows:

AIC = ln(L) - q

with ln(L): log likelihood function

q: number of covariance parameters

SBC = ln(L) B qln(n)/2

with n: number of observations

Intraclass correlation: EQUATION

A higher value for AIC and SBC indicates the better model. Basic prerequisite of using these criteria for model comparison is that the models contain the same fixed effects (Latour 1994). The intraclass correlation coefficient (with values between 0 and 1) measures the proportion of variance in Y between the groups (Bryk & Raudenbush 1992). The higher the intraclass correlation the more variance is explained by the level-2 groups.

In our study we want to compare the influence of brand name and of product attribute on product outcome. For this reason we analyze products as level-1 units. The level-2 grouping variables are in one specification the brand names (in this study the name of the firm) and in the other specification a product attribute (in this study the flavor). We analyze which specification of the second level is better to explain the differences between the market share of the products. So far thee are hardly any studies comparing the importance of brand name and product attribute. We compare the following structures:

Alternative 1:

Alternative 2:

In the extended model two other explaining variables will be introduced.

Data base and model specification

The set of aggregated household panel data is provided by the GfK, Nnrnberg and contains information regarding yogurt with three product flavors: natural, fruit and drink. In the selected data set are 62 products from six firms, that means six different brand names. It is a time series with 25 months (September 1995BSeptember 1997) and 1360 observations. As the dependent variable we use market share.

At the lowest level, the level-1, we have time series for each of the 62 single products. [We could also specify a three-level model with time series for each product as level-1, products as level-2 and product attributes resp. Brand names as level-3.] For the higher, second level, we compare two different alternatives. In the first specification brand names build the second level, in the other specification the product attribute flavors build the second level.

According to these models we specify our model as follows:

Level-1: Products: Y_ij = ¯_0j + r_ij

Level-2: Brand names respectively product attribute:

¯_0j = y₀₀ + u_0j

Substitution of level-2 model leads to the unconditional hierarchical model:

Y_ij = y₀₀ + u_0j + r_ij

i = 1,, Nj products

j = 1,,M brand names or product attributes

r_ij ~ N(0, s²)

u_0j ~ N(0, t)

RESULTS

The estimation of the parameters of the unconditional hierarchical model are presented in table 1 and 2. The results for the fixed effects show that only the overall mean for the brand name is significant and the maximum likelihood estimate is 0.9308. In the specification with brand name as grouping (level-2) variable most of the variance in the outcome is at the product level, =1.7487, but a great part of the estimated variability is between the different brand names, =0.4091. In the other specification the variance of the level-1 is similar, =1.8957, but the variability between the product attributes is much less, =0.1584, than in the other specification and not significant, so we may assume that all flavors have the same mean. The same is evident in table 3: The proportion of variance between brand names is much higher than between product attributes: 19% vs. 8% which means that the brand names are more different than the product attributes. AIC as well as SBC is higher for brand name than for product attribute. It seems that the brand name is more important for the product choice than the product attribute flavor. Therefore in this analysis our hypothesis 1 cannot be rejected.

The next step is the integration of further variables to test the meaning of other choice variables. We introduce the level-2 variable "number of products" W₁₀. For the first alternative this variable measures how many products a firm offers in the product category. The more products the firm offers (means broader product line) the more power the firm has. Similar for product attribute: The more products are in the single category, the more important is this category.

TABLE 1

ML ESTIMATION OF RANDOM EFFECTS FOR THE UNCONDITIONAL MODEL

TABLE 2

ML ESTIMATION OF FIXED EFFECTS FOR THE UNCONDITIONAL MODEL

TABLE 3

MODEL COMPARISON OF THE UNCONDITIONAL MODEL

The conditional hierarchical model is:

Y_ij = y₀₀ +y ₀₁W₁₀ + u_0j + r_ij

The tables 4 to 6 provide the results of the maximum likelihood estimation. The variable "numbers of products" has a significant effect on market share in both model variations. But the direction of the influence is not clear. If brand name is the level-2 variable then the market share decreases with an increasing number of products of one firm, but if product attribute is the level-2 variable then the relationship is positive. A reason might be that the products in the shelves are sorted by flavor not by brands. And therefore the probability of buying a flavor with more products is greater than for a flavor with less products. However, the width of the product line of the firms is not obvious for the consumer. The covariance parameters for both cases are not significant, but the goodness criteria still favor the brand name model. Therefore the next step only concentrates on the better model, the brand name model.

Now we will examine the meaning of price (similar to Currim, Meyer & Le 1988). The integration of the variable "price" X_ij as a level-1 variable leads to a random-coefficient model:

Y_ij = y ₀₀ + y ₀₁W_j + y ₁₀X_ij + y ₁₁W_jX_ij + u_1j X_ij + u_0j + r_ij

The covariance parameter estimations in table 7 show significant effects (at the 10% level) for the variance of the intercept (t₀₀) and the slope (t₁₁). That means that intercepts and slopes vary significantly between brand names. The fixed effects in table 8 show a significant positive overall mean which is estimated to be 3.3968. The number of products has still the negative effect on market share. The price is not significant so it does not have a meaning for product choice. A reason could be that 11 of the 62 products are very successful new products with a high price. A separation between the new and the old products might be meaningful. The cross-level effect "product*price" is significant (at the 10% level) with a negative sign which means that firms with more products have lower price slopes on average.

These results support the importance of the brand name in the decision process. Even after controlling for other choice criteria like "price" and "number of products" the influence of brand name is still the highest. So, we cannot not reject our hypothesis 2. In sum, our analysis shows the relevance of brand name in comparison to other choice criteria and thus supports the results of many experimental studies.

TABLE 4

ML ESTIMATION OF RANDOM EFFECTS FOR THE MODEL WITH "NUMBER OF PRODUCTS"

TABLE 5

ML ESTIMATION OF FIXED EFFECTS FOR THE MODEL WITH "NUMBER OF PRODUCTS"

TABLE 6

MODEL COMPARISON FOR THE MODEL WITH "NUMBER OF PRODUCTS"

SUMMARY

The study shows that hierarchical models are able to show the meaning of different choice criteria and how they can be used to measure the size of the influence. As expected the brand name turns out to explain more variation in the market share of the products than the product attribute flavor and therefore it seems to be a more decisive decision criterion. The number of products also has a significant effect, it represents a context variable in the buying situation. The price in this study is without influence, which could be explained by the low-price category or the number of innovative new products. This could be a way for further studies where old and new products are separated. Another extension could be the analysis of other variables beside brand name and flavor or a comparison with similar product categories.

TABLE 7

ML ESTIMATION OF RANDOM EFFECTS FOR THE RANDOM-COEFFICIENT MODEL

TABLE 8

ML ESTIMATION OF FIXED EFFECTS FOR THE RANDOM-COEFFICIENT MODEL

LITERATURE

Alpert, M.I. (1980) Unresolved Issues in Identification of Determinant Attributes. Advances in Consumer Research, Vol. 7, 83-88.

Bettman, J.R. (1979) An Information Processing Theory of Consumer Choice. Reading, MA: Addison-Wesley.

Bettman, J.R., Kakkar, P. (1977) Effects of Information Presentation Format on Consumer Information Acquisition Strategies. Journal of Consumer Research, Vol. 3, 233-240.

Biehal, G., Chakravarti, D. (1986) Consumers' Use of Memory and External Information in Choice: Macro and Micro Perspectives. Journal of Consumer Research, Vol. 12, 382-405.

Bryk, A.S., Raudenbush, S.W. (1992) Hierarchical Linear Models: Applications and Data Analysis Methods. Newbury Park: Sage Publications.

Currim, I.S., Meyer, R.J., Le, N.T. (1988) Disaggregate Tree-Structured Modeling of Consumer Choice Data. Journal of Marketing Research, Vol. 25, 253-265.

Dodds, W.B., Monroe, K.B. (1985) The Effect of Brand and Price Information on Subjective Product Evaluations. Advances in Consumer Research, Vol. 12, 85-90.

Hoyer, W.D., Brown, S.P. (1990) Effects of Brand Awareness on Choice for a Common, Repeat-Purchase Product. Journal of Consumer Research, Vol. 17, 141-148.

Jacoby, J., Szybillo, G.J., Busato-Schach, J. (1977) Information Acquisition Behavior in Brand Choice Situations. Journal of Consumer Research, Vol. 3, 209-216.

Kumar, A., Sashi, C.M. (1989) Confirmatory Analysis of Aggregate Hierarchical Market Structures: Inferences from Brand-Switching Behavior. Journal of Marketing Research, Vol. 26, 444-453.

Latour, K. (1994) Getting Started with PROC MIXED. Cary, NC: SAS Institute Inc.

Lefkoff-Hagius, R., Mason, C.H. (1993) Characteristic, Beneficial, and Image Attributes in Consumer Judgements of Similarity and Preference. Journal of Consumer Research, Vol. 20, 100-110.

Littell, R.C., Milliken, G.A., Stroup, W.W., Wolfinger, R.D. (1996) SAS System for Mixed Models. Cary, NC: SAS Institute Inc.

Longford, N.T. (1993) Random Coefficient Models. Oxford: Clarendon Press.

Moore, W.L., Lehmann, D.R. (1989) A Paired Comparison Nested Logit Model of Individual Preference Structure. Journal of Marketing Research, Vol. 26, 420-428.

Obermiller, C. (1987) When Do Consumers Infer Quality from Price? Advances in Consumer Research, Vol. 15, 304-310.

Obermiller, C., Weathley, J.J. (1985) Beliefs in Quality Differences and Brand Choice. Advances in Consumer Research, Vol. 12, 75-78.

Rao, V.R., Sabavala, D.J. (1981) Inference of Hierarchical Choice Processes from Panel Data. Journal of Consumer Research, Vol. 8, 85-96.

Vilcassim, N.J. (1989) Extending the Rotterdam Model to Test Hierarchical Market Structures. Marketing Science, Vol. 8, 181-190.

----------------------------------------