Modeling Household Brand Choice: the Impact of Customer Evaluations of Quality and Household Characteristics

ABSTRACT - The arrival of the scanner panel data has created a unique opportunity for researchers who are interested in examining the purchase behavior of households as a consumption unit. Besides taking into account the past purchase behavior models should also incorporate current marketing activity and other behavioral variables such as satisfaction and customer evaluation of the products performance.


Richard Tay and William J. Qualls (1994) ,"Modeling Household Brand Choice: the Impact of Customer Evaluations of Quality and Household Characteristics", in AP - Asia Pacific Advances in Consumer Research Volume 1, eds. Joseph A. Cote and Siew Meng Leong, Provo, UT : Association for Consumer Research, Pages: 246-250.

Asia Pacific Advances in Consumer Research Volume 1, 1994      Pages 246-250


Richard Tay, Nanyang Technological University

William J. Qualls, MIT-Sloan School of Management


The arrival of the scanner panel data has created a unique opportunity for researchers who are interested in examining the purchase behavior of households as a consumption unit. Besides taking into account the past purchase behavior models should also incorporate current marketing activity and other behavioral variables such as satisfaction and customer evaluation of the products performance.

In this paper we empirically test an extension of the multinomial logit model of brand choice. Specifically, household characteristics and consumer perceptions of quality are incorporated into a household model of brand choice. The results suggest that consumers' perception of quality enhances the predictive ability of the model, while adding household characteristics had minimal impact. It is our contention that future research should also address the impact of customer loyalties of future brand choice in the same manner in which researchers have examined the antecedents of customer loyalty and brand choice behavior.


Research on the brand loyalty behavior of households has increased dramatically in the last 10 years. One reason that may account for the growth of research in this area has been the development of modeling techniques that can take advantage of the large amount of purchasing data that is being collected by way of scanners in retail stores. The seminal work by Guadagni and Little (1983) has provided the impetus and foundation for continued research that addresses the issue of determining "what factors influence household brand choice behavior" and "why household loyalty is such a strong predictor of future consumption behavior."

The answers to such questions would provide invaluable insights to manufacturers and retailers who are interested in increasing the effectiveness of their pricing, advertising, and promotional programs. Yet despite the methodological advances achieved in previous research on most models of household brand choice remain incomplete due to the lack of customer oriented behavioral and attitudinal data.

Overwhelmingly, logistic regression models or some variation has been the modelling technique of choice by most researchers who have examined the issue of consumer loyalty and household brand choice behavior. Yet current logistic regression models do not incorporate in estimates of the effects of such variables as household characteristics and customer satisfaction; leading to unstable and inaccurate estimates of causal effects and choice probabilities. It appears that the majority of the research in this area has focused more on the methodological issues and less on the more substantive issues.

In this paper we propose that current methods for modelling household brand choice and household loyalty would be improved if customer oriented variables are considered and incorporated.

Logistic Models of Brand Choice

The conditional logit model developed by McFadden (1973) has been the cornerstone of marketing research methodology when analyzing household brand choice panel data,(Guadagni and Little 1983, Gupta 1988, Krishnamurthi and Raj 1988, and Lattin and Bucklin 1989). Logistic models of brand choice attempt to estimate the relationship between a household's choice probability and certain marketing mix variables and sometimes household demographic variables. The logistic regression model when used to estimate household brand choice probabilities, assumes a linear utility maximization model of consumer behavior.

The basic premise of the utility maximization model is that individuals make choices that are independent of one another by considering the full set of brand alternatives at the time of the choice. The consumer household c selects one item i from a finite set of discrete alternatives. Given a set of brand alternatives, the decision maker makes a brand choice that maximizes their utility function across the set of information sources used to make the decision. As such Uic= Vic +eic. Since the utilities are typically unknown, research using logistic regression treat such utilities as random variables. As a result brand choice can be estimated at the household or individual level. Violations of the utility maximization model have been documented however, in marketing and other decision contexts. The logistic models of brand choice although rich in predictive power, are often difficult to estimate when the impact of contextual variables is known.

The success of most logistic models in accurately predicting household brand choice is often predicated by the assumptions made by the researcher regarding the relevant set of brands that are considered, and the distribution of the choice probabilities. One problem in particular that logistic models have experienced has been the model's inability to simultaneously examine the impact of marketing mix variables, household characteristics, and the omission of customer oriented variables.

More recently multinomial logistic regression models have been utilized to analyze household level brand choice decisions and as a means of examining both marketing and consumer behavior phenomena (Guadagni and Little 1983, Lattin 1987, Lattin and Bucklin 1989, Ortmayer, Lattin and Montgomery 1991, Neslin et. al. 1985, Tellis 1988). These models share both a common objective; estimating brand choice today as a function of previous choices, and a common problem of estimating nonlinear parameters simultaneously with traditional linear parameter estimation logit procedures. To resolve the issue of non-linear estimations, previous research has relied upon setting arbitrary values or using simple assumptions to approximate the non-linearity.

In addition, the majority of the panel data that has been used to model household brand choice, have been based on a limited number of household observations. As a result previous research has focused almost exclusively on individual brand choice behavior. Hence there is not only a need for developing a deeper understanding of the antecedents of brand choice, but also a methodology that captures brand choice behavior at the household level.

Household Brand Loyalty

The concept of brand loyalty has been one of the key principles in marketing used to explain and understand household brand choice behavior. Brand loyalty is typically conceptualized as a form of repeat purchase behavior. The purchase behavior is believed to be motivated by the consumer's decision to continue buying the same brand (Aaker 1991). For example, assume that in a small town there are only two Italian restaurants (A & B) that households can select from; with each restaurant carrying the same menu, prices, and atmosphere. Utilizing the Guadagni and Little (1983) methodology for determining consumer loyalty, you would find that household with a strong loyalty to restaurant A would have a higher probability of having patronized restaurant A on the most recent purchase occasion. As such past purchase behavior serves as a strong indicator of future purchase behavior. Loyalty is then based on what you do as a consumer, instead of what you think or feel. What is not clear from this conceptualization of loyalty is the reasons underlying the behavior.



One school of thought contends that it is the positive feedback from the most recent purchase occasion that increases the probability of the household patronizing the same restaurant in the future. Fueled by the work by Guadagni and Little (1983), researchers have focused on the improvement of the methodology that estimates consumer purchase probabilities. Two views have emerged from the research that suggests 1) that consumers have similar purchase probabilities and 2) that consumers do in fact have different purchase probabilities and by understanding how to estimate the purchase probabilities we can better understand household brand loyalty.

A Theoretical Model of Household Brand Choice

Household brand choice can be broadly defined as an individual representative or multiple individuals from a household that participate in the selection of one brand from a set of available brands. The decision to purchase a specific brand is influenced by numerous factors that include, customer need requirements and purchase experiences, manufacturers marketing programs, retailers' programs, and household characteristics. Given the interactions between these variables' households make purchase decisions regarding future product/brand choices. Figure 1 illustrates how household purchase and loyalty behavior are hypothesized as moderating the effects of marketing mix variables and household characteristics on household brand choice behavior.

Briefly the model is based on the premise that household purchasing behavior and household loyalty is hypothesized as key determinants for predicting future household brand choice behavior. In turn household purchase behavior and loyalty are determined by the promotional programs and marketing mix elements of the manufacturer and retailer that interacts with household characteristics and customer evaluations.

Equation (1) presents a general expression for household brand choice :

HBCijkt=f1(HLijkt) +f2 (PBijkt) + f3 (PSijkt,SEijkt,QEijt,MMijkt,HCit)    (1)


HBCijkt = household choice of brand j of size k for customer i at time t,

HLijkt = household loyalty for brand j for customer i at time t,

PBijkt = household purchase behavior of brand j for customer i at time t,

PSijkt = household purchase Practice of brand j for customer i at time t

SEijkt = store marketing environment for brand j for customer i at time t

QEijt = Consumer perception of quality for brand j for customer i at time t

MMijkt = Marketing Mix variables for brand j for customer i at time t

HCit = household characteristics for customer i at time t

f1= concave function for the impact of loyalty on brand choice

f2= constant linear function

f3= concave function for the impact on purchase behavior and loyalty

The first term in equation 1, represent the direct impact of household loyalty on brand choice. If all other things are equal, then household loyalty should be a strong predictor of brand choice. As suggested by previous research (Guadagni and Little 1983, Lattin 1987, 1992), the direct impact of loyalty on household brand choice is expected to be a concave function as different marketing related and household factors change over time. Similarly, when household purchases occur, the purchase behavior is expected to be constant and linear over time.

The loyalty coefficient in the present model plays a dominate role because in previous research the loyalty construct has been the means by which brand choice made in the past is captured as an effect upon future brand choices. For example Guadagni and Little (1983) incorporate a multiple measure household loyalty based on "brand loyalty" and "size loyalty" in the analysis of coffee data. Similar to the loyalty coefficient employed by Guadagni and Little( 1983) and more recently Lattin ( 1987, 1992), the loyalty coefficient (HLijkt) is operationalized as

HLijkt = HLijkt-1 + (1-l)yijkt    (2)


HLijkt = the loyalty of household i to brand j and size k at time t

yijkt = indicator of brand choice

l = a non-stationary parameter that indicates the relative weight of loyalty on the purchase at t-1 in determining loyalty at time t.

As such household loyalty is obtained by exponentially smoothing observed purchase behavior over the last household purchase occasion.


Primary data for this analysis are taken from the IRI academic scanner panel database that covers a panel of some 2,000 households for 108 weeks beginning in April 7,1980. The data set includes information on households from Pittsfield, Massachusetts and their coffee purchases. In addition the data set provides information on the store environment( such as display, feature and coupons), pricing data, and household purchase behavior for those items of coffee that were purchased by at least one of the customers in the database during the particular week. Secondary data on quality of the products were taken from consumer reports between 1983 and 1985.

Although over 300 items were available, only those households that had purchased at least one item during the above data collection period was included in the sample. Since the majority of the remaining data were from the two largest stores in Pittsfield, households visiting these stores were used as the sampling base. In addition, several generic, private labels and specialty coffees were also eliminated due to the lack of information on their quality ratings. The final choice set contains 42 items of various sizes from 10 brands; of which 21 items were ground coffee from 8 brands and the other 21 items of instant coffee were also from 8 brands.

From the remaining sample of 267 households who purchased at least one of the 42 items, we randomly sampled half for estimation. Household purchases in the first 30 weeks were used to initialize the brand and size loyalty measures and the following 52 weeks of data were used for estimating the model. The remaining 26 weeks were used to forecast household decision to compare the actual with the prediction. The final sample consists of 118 households and 1321 observations.

Although it is feasible to estimate the model with the complete set of alternatives available to the household, consistent estimates can be obtained with a reduced set of alternatives (Ben-Akiva and Lerman, 1985). A random sample of nine other alternatives was added to the chosen item to form a set of 10 alternatives was added to the chosen item to form a set of 10 alternatives for estimation.

Several factors were hypothesized as influential to the household purchase decision. The regular price ( $/oz ) of the product is expected to be negatively correlated to the probability of purchase. On the other hand, the quality of the product is expected to increase the probability of a purchase. Both ground and instant coffee were rated by the consumer report and classified into four main categories of overall sensory index. Since only a small proportion of the items in our data set was rated in the least favorable class, two indicator variables, for the top two categories, were used to capture the effects of good ratings by consumer report on household choice decision, with the remaining items used as normalization. Another indicator variable, instant, is used to test for any systematic differences in choice between ground and instant coffees.

Consumer brand and size loyalties are important components of the household purchasing model and have been widely found to positively influence purchase decisions. Following Guadagni and Little (1983), we defined both the brand and size loyalty variables as the exponentially weighted average of past purchases of the brand or size. Carry-over constants of 0.875 and 0.812 from their study were also used in our model.

Store promotions include the presence of feature and/or display for the items and the amount of store coupons available. Both feature and display were captured by dichotomous (0-1) variables while store coupon is measure by monetary units (cents). Items on promotion were expected to have higher probabilities of being purchased.

Lastly, eight brand specific constants were included to measure any systematic preference for a particular brand that was not captured by other explanatory variables. Chock Full O' Nuts and Maxim were used for normalization.


Logistic Regression Analysis

The parameter estimates, standard errors and t-ratios are reported in Table 1. With the exception of store coupon that is insignificant, all estimates are consistent with prior expectations. Price is negative and significant whereas quality loyalty, store feature and displays are positive and significant.

To test the impact of brand and size loyalties of consumers who bought items during promotional periods versus those who purchased the items without promotion, we estimated eight additional models in which the loyalty variables were differentiated accordingly and conducted likelihood ratio tests. Table 2 reports the chi-squared statistics for these tests. All tests showed significant difference in customer loyalties depending on whether the purchase was made during promotion.





All the estimated coefficients of the brand loyalty variables were smaller in magnitude for purchases that were made during promotions than those made without promotions, implying that an average consumer derived less utility from being brand loyal when purchasing coffee during promotions. More substantively, no significant brand loyalty was detected for customer purchasing items on feature or with store coupons.

Similarly, all the estimated coefficients for the size loyalty variable were smaller in magnitude for purchases that were made during store promotions than those made with store promotions. The estimated coefficient for the size loyalty variable was however larger in magnitude for purchases made with manufacturer's coupon, indicating that the use of manufacturer's coupon increased an average consumer's utility for being loyal to a specific size.

Next we tested for differences in customer loyalty to items on promotion versus items not on promotion. Again we estimated six more models in which the loyalty variables were differentiated by the presence/absence of promotion in the store. As evident in Table 2, customer loyalties to both brand and size were positive and significant, regardless of whether the items were on feature. However , the estimated coefficients for the loyalty variables were smaller for items not on feature implying that store features had the effect of strengthening customer loyalties. Similar results were also found for store display. These results suggested that store feature and display had the effect of reinforcing customer loyalty. On the other hand, customers were found to be more brand and size loyal to items that did not have store coupons. In addition, no significant brand loyalty to items with store coupon was found. Therefore, we surmised that store coupons had the effect of reducing customer loyalty.

Household brand and size loyalties were also hypothesized to be indirectly related to household purchasing power. The results in Table 2 showed that there were, in general significant differences in customer loyalty for households with different characteristics such as income, education, age, children and employment status. These results indicate that while there are reasonably loyal customers, they are difficult to identify from their demographic characteristics.


The multinomial logit model has increasingly been used to estimate models of brand choice for packaged goods where panel data are widely available. Although the impact of customer loyalties, product promotions, and consumer purchasing strategies on consumer brand choice have been widely investigated in these models, little research has been conducted to analyze the effects of promotions and purchasing strategies on customer loyalty. The results of the current study provide some evidence that customer loyalties as defined by Guadagni and Little (1983), were dependent on promotions and purchasing strategies. The effects of the various store promotions and consumer purchasing strategies on customer loyalties, however were mixed and more studies should be conducted to provide a better understanding of these relationships about brand choice. In addition since store promotions, purchasing strategies and loyalties are all included as independent variables in most models, caution must be exercised in interpreting the results due to the presence of multicollinearity.

Future research may want to consider alternative analytical techniques, which can address some of the methodological limitations of multinomial logit. Overwhelmingly, logistic regression models or some variation has been the modelling technique of choice by most researchers who have examined the issue of household brand choice behavior. Yet logistic regression models are often ineffective in estimating the effects of covariates such as household characteristics and marketing mix variables; leading to unstable and inaccurate estimate's of causal effects and choice probabilities. Several aspects of Partial Least Squares a structural modeling technique makes it ideal for addressing issues of multicollinearity, heterogeneity across households, nonlinear relationships, and interactions between causal factors.

Another interesting area of research is the issue of how to model cross-sectional heterogeneity. One common approach is to incorporate household characteristics in the empirical model. In the logit model, however, this method becomes cumbersome because household characteristics do not vary across the choice alternatives and have to be incorporated as alternative specific variables. The customer loyalty variables proposed by Guadagni and Little (1983) is widely accepted, and used in logit models as measures that capture much of the cross-sectional heterogeneity. Our results, provide some evidence that these measures were not affected by most household characteristics implying that the two approaches may be independent measures.


Aaker, David, (1991) Managing Brand Equity: Capitalizing on the Value of a Brand Name, The Free Press, New York.

Ben Akiva, M. and S. R. Lerman (1985),Discrete Choice Analysis: Theory and Application to Travel Demand, Cambridge, MA, MIT Press.

Blattberg, R. and Neslin S. (1990), Sales Promotion: Concepts, Methods, and Strategies, Prentice-Hall Inc., Englewood Cliffs New Jersey.

Guadagni P. and J. Little (1983)," A Logit Model of Brand Choice Calibrated on Scanner Data", Marketing Science, Vol. 2. no.3 (Summer) pp.203-238.

Gupta, S. (1988), "Impact of Sales Promotion When, What, and How Much to Buy," Journal of Marketing Research, Vol. 25, no.4 (November) pp. 342-355.

Krishanmurthi L., and S.P. Raj (1988)" A Model of Brand Choice and Purchase Quantity Price Sensitivities" Marketing Science, Vol. 6. No. 1 (Winter)pp.1-20.

Lattin, J. (1988)," "The Impact of Store Brands on the Nature of Manufacturer's Trade Deals and Retail Price Promotion" working paper, Stanford University.

Lattin, J. and Bucklin, R. (1989)," Reference Effects of Price and Promotion on Brand Choice Behavior" Journal of Marketing Research, 26, (August), 299-310.

McFadden, D. (1973) "Conditional Logit Analysis of Qualitative Choice Behavior" in P Zarembka (ed.), Frontiers in Econometrics, New York, Academics Press, pp.105-142.

Neslin, S. C. Henderson and J. Quelch (1985) "Consumer Promotions and the Acceleration of Product Purchases" Marketing Science, 4, (Spring), pp. 147-165

Ortmayer, G. J. Lattin, and D. Montgomery (1991) " Individual Differences in Response to Consumer Promotions" International Journal of Research in Marketing, 8,pp.169-188

Tellis, G. ( 1988)," Advertising Exposure, Loyalty, and Brand Purchase" Journal of Marketing Research, 25, (May), pp.134-144.



Richard Tay, Nanyang Technological University
William J. Qualls, MIT-Sloan School of Management


AP - Asia Pacific Advances in Consumer Research Volume 1 | 1994

Share Proceeding

Featured papers

See More


L7. The Joy of Shopping: Reconciling Mixed Effects of Positive Emotions on Shopping Behavior

Kelley Gullo, Duke University, USA
Duncan Simester, Massachusetts Institute of Technology, USA
Gavan Fitzsimons, Duke University, USA

Read More


J13. The Good and Bad of a Delayed Indulgence: Effects on Self-Perceptions and Purchase Satisfaction

Argiro Kliamenakis, Concordia University, Canada
Kamila Sobol, Concordia University, Canada

Read More


M7. The Mixed Effects of Nostalgia on Consumer Switching Behavior

Zhongqiang (Tak) Huang, University of Hong Kong
Xun (Irene) Huang, Nanyang Technological University, Singapore
Yuwei Jiang, Hong Kong Polytechic University

Read More

Engage with Us

Becoming an Association for Consumer Research member is simple. Membership in ACR is relatively inexpensive, but brings significant benefits to its members.