Handling unlabeled discrete choice data

Abstract

Introduction Before we continue to look at some of the richer sets of behavioral outputs from the basic MNL model, we want to make an important diversion. Discrete choice data may come in one of many forms. Aside from revealed preference (RP) and stated preference (SP) data (see Chapter 6), discrete choice data may be further categorized as being either labeled or unlabeled in nature. In labeled choice data, the names of alternatives have substantive meaning to the respondent beyond their relative order of appearance in a survey (e.g., the alternatives might be labeled Dr House, Dr Cameron, Dr Foreman, Dr Chase). In unlabeled choice data, the names of the alternatives convey only the relative order of their appearance within each survey task, (e.g., drug A, drug B, drug C). Aside from affecting what outputs can appropriately be derived for the study (e.g., elasticities have no substantive meaning in unlabeled experiments), from the perspective of the overall study this decision is important, as it might directly impact upon the type and number of parameters that can or will be estimated as part of the study. As we show below, typically, unlabeled experiments will involve the estimation of generic parameters only, whereas labeled experiments may involve the estimation of alternative-specific and/or generic parameter estimates, hence potentially resulting in more parameter estimates than with an identical, though unlabeled, experiment