Understand customer decision making: Discrete choice models in marketing
In our blog article on Marketing Mix Modeling, we introduced you to the benefits of data-driven analysis and statistical modeling of your marketing mix. Marketing Mix Modeling uses multivariate econometric time series models to identify the strengths and weaknesses of a marketing program and to determine the right mix of promotional activities to maximize return on investment.
In this article, we would like to discuss another class of models, which are used in marketing analysis and modeling, so-called discrete choice models. These models allow you to generate a better understanding of customer decision making process, and can be used to make more accurate predictions on purchase decisions. In addition, they enable the evaluation of how customers perceive advertising offers, product messages or brand strategies and how they think about new products or improved product features. They can also evaluate sensitivities to price changes or promotions, and analyse demand changes. If you are looking to add new products to your portfolio, discrete choice models can help identify product features that have the potential to increase demand for the products.
Discrete choice models belong to the class of choice models, a family of statistical techniques that allow to determine an optimal combination of (marketing) explanatory variables. They make use of the fact that consumers make decisions when they decide for or against the purchase of products or services. Information about these decisions is used to derive model-based insights about customer preferences and the effectiveness of marketing. Marketing influencing variables include all variables that can be changed to increase the effectiveness of marketing and advertising (for example: price, packaging, brand positioning, communication, advertising). Related methods in this context are conjoint or tradeoff analyses, as well as advanced variants that can account for more complex decisions. We will go into technical and methodological details in the second part of our article series on discrete choice models.
Discrete choice models are used to model choice decisions (for example, buying/not buying a product, using a service, choosing a brand) from among alternatives. The different alternatives of each decision are characterized by different (product-specific) attributes that play a different role for customers in their purchase decision. For example, one relevant attribute is price. A discrete choice model allows to identify the influence of price changes on the customers' demand for the product alternatives. Furthermore, a discrete choice model allows to identify the influence of marketing instruments. Available marketing tools vary by industry, but typically include pricing, advertising, trade promotions, consumer promotions such as coupons and sweepstakes, in-store merchandising, or longer-term decisions such as assortment variations. Discrete choice models determine the influence of these variables on the choice decision. In addition, discrete choice models allow customer-specific information such as age or gender to be taken into account in the modeling and thus allow customer- or group-specific analyses to be performed.
Consumer choices and preferences are studied in choice experiments. Let's assume that the goal is to optimize a product design. Usually, a number of different sizes, formats, labels, inscriptions and prices can be considered. To evaluate which combination of options is best, stated preference surveys can be conducted. In this classical approach, potential customers are asked which product from a range of options they prefer and whether they would buy the preferred product or not. Nowadays, this is increasingly done in online surveys. Stated-preference surveys confront respondents with hypothetical markets in which they eypress preferences. In addition, there are surveys of actual (rather than hypothetical) preferences (so-called revealed-preference surveys). Here, the aim is to reveal the preferences of consumers who have already made a purchase decision retrospectively.
In the age of Big Data, the availability of data on choice situations increases rapidly, even without explicit surveys, and can be analyzed using discrete choice models. In particular, sales data can be used to draw conclusions about preferences as well as the effects of marketing if the existing range of products is evaluated at the same time, representing the available alternative products to a purchase. The marketing recommendations derived from discrete choice modeling can be tested with A/B tests, which can be carried out afterwards.
From the field of discrete choice models, the multinomial logistic regression model (short: MNL model) serves as the standard procedure. It is often the starting point for modeling choice decisions due to its attractive properties in terms of rigorous embedding in economic theory. A MNL model allows to generate the following information:
- Average utility of alternatives for consumers.
- Insight into brand loyalty.
- Sensitivities of users to price changes (willingness to pay) or marketing mix. This is expressed in terms of elasticities or cross-elasticities (i.e., determining the effects of percentage changes on the dependent variables).
- Model-based predictions for changes in covariates or introduction of new alternatives.
We will discuss advanced variants, all of which are relevant to practice, in the second part of our series on discrete choice models.
A well known example of the usefulness of discrete choice models in marketing is the study of the cracker dataset by Jain et al. 1994. The dataset contains information on 136 consumers who have a choice between four types of cookies (three well-established brands and one private-label brand). Each consumer makes between 14 and 77 purchases for which the available alternatives are known. For each purchase, information is available on the price as well as on different marketing activities. Since these variables can be influenced in the context of marketing, it is possible to take the results of discrete choice modeling into account for future marketing and thus optimize it. Information about brand loyalty can also be extracted from the data, since the buying behavior of customers has been observed over time. By estimating a simple MNL model, benefits, sensitivities, and loyalty can be quantified and predictions for future buying behavior based on changes in marketing efforts can be made.
More details can be found in the article by Elshiewy et al. 2017, which provides an overview of different approaches to modeling the data. The article highlights specific uses of discrete choice modeling in marketing. In addition, the authors provide the R code of their modeling. We reproduced several of the modeling approaches presented by Elshiewy et al. 2017, using the statistical framework Stan. Key findings of the statistical analysis are:
- Detached from price, advertising, and brand loyalty, two of the three brands generate on average higher utility for customers than the private-label brand. The benefits of the third brand, on the other hand, do not differ significantly from those of the private-label.
- Customers obtain higher utility from lower prices. Decreasing prices thus increase the purchasing probability.
- Use of marketing instruments increases the purchase probability of the brands on average. It can be determined which of the advertising measures used increases the choice probability the most.
- Consumers tend to decide for brands that they also decided for in their previous decision, so on average they are brand-loyal.
- Consumers are less price-sensitive towards the private-label brand than toward other brands (analysis of elasticities). In addition, an analysis of the so-called cross-elasticities shows how the probability of choosing a brand changes when the price of another brand is reduced.
The above example illustrates the usefulness of discrete choice models for understanding complex buying behavior. Furthermore, the approach allows to quantify the effectiveness of advertising measures and to make model-based predictions. In discrete choice modeling, the MNL model is considered as a benchmark approach, but it has a number of limitations. Therefore, more advanced methods should be used in practice. For example, a specific property of the model (the so-called IIA assumption) leads to the fact that the cross-elasticities (see point 5.) are constant between alternatives. As a consequence, price reductions of one brand increase the purchase probability of the other brands proportionally. In practice, this assumption is implausible if some brands are more similar to each other than others (for example, two luxury brands are more similar to each other than a luxury brand and a sports brand). Moreover, an MNL model is not able to account for unobserved effects in repeated choice decisions or between customers. This results in the problem that heterogeneity between customers is often not yet sufficiently accounted for. We address extensions of the model that are relevant to practice in the second part of our series on discrete choice models.
- Jain, D. C., Vilcassim, N. J., and Chintagunta, P. K. (1994). A random-coefficients logit brand-choice model applied to panel data. Journal of Business & Economic Statistics, 12(3), 317–328.
- Elshiewy, O., Guhl, D. and Boztug, Y. (2017): Multinomial Logit Models in marketing - From Fundamentals to State-of-the-Art. Marketing ZFP - Journal of Research and Marketing, 39(3), 32–55.