A CUB Model Strategy to Select Anchoring Vignettes (303627)

Keywords: anchoring vignettes, CUB modeling, feeling, individual heterogeneity, ordered variables, response scales, uncertainty

Anchoring vignettes are additional questions (to be answered by respondents) introduced in the literature as a tool to identify and correct the systematic differences in the use of response scales within countries or socio-economic groups when respondents evaluate themselves. Answers are usually provided on an ordinal scale.

There is no benchmarks on the number of vignettes to propose. First, vignettes should be designed in order to provide discriminatory power, that is they should be equally spaced through the distribution of the self-evaluations. Then, a fewer or larger number of vignettes is related to a less or more heterogeneity in the population under investigation. A good strategy is to ask more vignettes during the pretest and then analyze the stability of estimates in the parametric statistical model that exploits vignette data.

CUB (Combination of discrete Uniform and -shifted- Binomial distributions) is a new class of statistical models, where the response is modeled as the mixture of two latent components, one related to individual feeling towards the item and another to the uncertainty in the response process.

In this work we aim at exploiting the CUB modeling in supporting the administration of survey vignettes. The idea is to specify and estimate standard or extended versions of CUB models for screening the vignette questions, in order to highlight those questions with potential problems of understanding.

For the selection of vignettes, several CUB models (in which vignettes are considered as covariates) were implemented for both components of feeling and uncertainty, respectively. After a backward selection, the best obtained model stresses the vignettes that are significant for explaining either the uncertainty or the feeling component: according to our proposed strategy, these vignettes are selected for the main survey (for instance, after a pretest) or recommended for the analysis involving vignette data.