Heterogeneity in Ordered Choice Models: A Review With Applications to Self‐assessed Health

Abstract

Abstract. Discrete variables that have an inherent sense of ordering across outcomes are commonly found in large data sets available to many economists, and are often the focus of research. However, assumptions underlying the standard ordered probit (which is usually used to analyse such variables) are not always justified by the data. This study provides a review of the ways in which the ordered probit might be extended to account for additional heterogeneity. Differing from other reviews in scope, application and relevance in economic settings, a series of issues pertaining to choices of variables, and the economic assumptions underlying each model are discussed in the context of measuring the underlying health of respondents. The models are applied to a wave of the household, income and labour dynamics in Australia survey, in order to check the appropriateness of such assumptions in an applied context.

The monetary subjective health evaluation for commuting long distances in Chile: A latent class analysis

“…The average of SAH for individuals in Chile is 3.68, indicating that most of Chileans consider their health to be good. In comparison to other studies that use SAH with the same scale as Chile, this average is higher to the mean score of 3.36 for Australians (Greene et al, ) and lower to the score of 4.04 for those commuters in the UK (Künn‐Nelen, ).…”

“…To quantify these health costs, individuals' self‐assessed health (SAH) status is used as our key dependent variable, which is highly correlated with mortality and morbidity and has been shown to be a good indicator of the underlying health (Idler & Benyamini, ; McCallum, Shadbolt, & Wang, ; Miilunpalo, Vuori, Oja, Pasanen, & Urponen, ; Wu et al, ). Furthermore, since it is costly and difficult to have objective health measures, SAH questions have been designed in socio‐economic surveys to capture all aspects of health in a subjective way (Greene, Harris, Hollingsworth, & Weterings, ; Mu, ).…”

Threshold Specification and Parameter Identification in the Generalized Ordinal Probit Model

“…Greene et al . () claim that in the GOP of Terza (), when the regressors of the latent index and the thresholds are the same, variation in the threshold coefficients across outcome categories is not allowed. Expression shows this assertion to be untrue.…”

“…In one of the earlier issues of this journal, Greene et al . () state that the generalized ordinal probit model (GOP) of Terza () does not allow for variation in threshold coefficients across alternatives when the threshold covariates coincide with those of the latent regression. This is incorrect.…”

“…Greene et al . () claim that when the above model allows the covariates to ‘…influence only the first threshold…’ and forces ‘…the other thresholds to be constant…’. In fact, no such restriction is imposed, regardless of the specification of and .…”