Using conditioning on observed choices to retrieve individual-specific attribute processing strategies

Abstract

With the growing reliance on Stated Choice (SC) data, researchers are increasingly interested in understanding how respondents process the information presented to them in such surveys. Specifically, it has been argued that some respondents may simplify the choice tasks by consistently ignoring one or more of the attributes describing the alternatives, and direct questions put to respondents after the completion of SC surveys support this hypothesis. However, in the general context of issues with response quality in SC data, there are certainly grounds for questioning the reliability of stated attribute processing strategies. In this paper, we take a different approach by attempting to infer attribute processing strategies through the analysis of respondent-specific coefficient distributions obtained through conditioning on observed choices. Our results suggest that a share of respondents do indeed ignore a subset of explanatory variables. However, there is also some evidence that the inferred attribute processing strategies are not necessarily consistent with the stated attribute processing strategies. Additionally, there is some evidence that respondents who claim to have ignored a certain attribute may simply have assigned it lesser importance. The results produced by the inferring approach not only lead to slightly better fit but also more consistent results.

Behavioral implications of providing real incentives in stated choice experiments

“…However, as opposed to our hypothesis 6 stated above, it has been argued in the literature, that when merely asked to state non-attendance there is no incentive to provide a truthful answer, regardless of the setting being hypothetical or real (e.g. Hess and Hensher, 2010;Alemu et al, 2013;Carlsson et al, 2010). Instead it is suggested to rely on non-attendance being inferred from the model.…”

“…Based on these conditional distributions which assume a normal distributed random parameter for each individual for each attribute, i.e. As noted by Hess and Hensher (2010) and Scarpa et al (2012), a value of 2 for κ nk would seem like a reasonable threshold value above which non-attendance observations can be expected since normal distributions with spread-to-mean ratios above 2 are over-dispersed 7 . In other words, when κ nk >2 the noise-to-signal ratio can be considered high, and it may thus be inferred that individual n has ignored attribute k. The share of individuals ignoring a given attribute can then be obtained by aggregating over the individuals.…”

“…Following Campbell and Lorimer (2009) and Hess and Hensher (2010), in order to assess the accuracy of the stated attribute non-attendance we thus estimate a model with two states for each attribute, depending on whether or not the attribute is considered or ignored. In other terms, we essentially estimate two beta coefficients for each attribute; one for respondents who said they considered the attribute and one for respondents who said they ignored it.…”

“…In that case it may be argued that it is better to infer non-attendance from analytical models rather than relying on respondent statements. To continue our investigation in this direction, we follow the approach to inferring non-attendance suggested by Hess and Hensher (2010) that was also recently applied in Scarpa et al (2013). This approach to inferring attribute non-attendance is based on the coefficient of variation of individual-specific posterior means.…”

Choice modelling with search and sort data from an interactive choice experiment

“…Rose et al, 2005). There is evidence to suggest that attribute attendance as stated by the respondent is not reliable, and that sounder methodologies include a stochastic treatment of attribute attendance (Hensher et al, 2007) or the use of other model outputs such as conditional parameter estimates (Hess and Hensher, 2010). Cloaking tools might not cover all attribute nonattendance, as revealed attributes might still be ignored, but attributes that are not revealed can be definitively considered as not attended to, and removed from the utility expressions of a choice model.…”