Dummy and effects coding variables in discrete choice analysis

Abstract

Discrete choice models typically incorporate product/service attributes, many of which are categorical. Researchers code these attributes in one of two ways: dummy coding and effects coding. Whereas previous studies favor effects coding citing that it resolves confounding between attributes, our analysis demonstrates that such confounding does not exist in either method, even when a choice model contains alternative specific constants. Furthermore, we show that because of the lack of understanding of the equivalence between the two coding methods, a sizeable number of previously published articles have misinterpreted effects coded results. The misinterpretation generates conflicting preference ordering and renders t‐statistics, marginal willingness to pay, as well as consumer surplus/compensating variation estimates invalid. We show that severe misinterpretation occurs for any categorical attribute that contains more than two discrete levels. The frequency of two‐level attributes used in discrete choice analyses may have led some past studies to overlook this error. Given its equivalence and lower likelihood of misinterpretation, we recommend dummy coding.

Would You Buy Plant-Based Beef Patties? A Survey on Product Attribute Preference and Willingness to Pay among Consumers in Liaoning Province, China

“…Compared with effect coding, the welfare estimation results of dummy coding are easier to interpret and less prone to welfare estimation errors [ 39 ]. In this paper, dummy coding is used to assign values to the product attributes of plant-based beef patties.…”

Purchasing habits, age effects and Chinese consumers' willingness to pay for chilled pork: Evidence from a random Nth‐price auction experiment

“…Future research may benefit from incorporating welfare evaluations within this domain, as delineated by Small and Rosen (1981) and Hu et al. (2022).…”

Which factors drive the choice of the French‐speaking Quebec population towards a COVID‐19 vaccination programme: A discrete‐choice experiment

“…To add realism to the exercise, we included an opt‐out alternative (i.e., respondents were given the possibility not to choose a vaccine) 49,50 . Each level was dummy‐coded: the level took the value of $$1$$ when it was described in the scenario and $$0$$ otherwise (see Hu et al 51 for a discussion about dummy and effects coding). Considering the panel nature of the data, the alternative specific constant (ASC) was invariant across individuals and choice tasks and coded as a triad of (1, 1, 0) (alternative A, alternative B and opt out).…”

“…To add realism to the exercise, we included an opt-out alternative (i.e., respondents were given the possibility not to choose a vaccine). 49,50 Each level was dummy-coded: the level took the value of 1 when it was described in the scenario and 0 otherwise (see Hu et al 51 for a discussion about dummy and effects coding).…”