This article proposes a new class of rating scale models, which merges advantages and overcomes shortcomings of the traditional linear and ordered latent regression models. Both parametric and semi-parametric estimation is considered. The insights of an empirical application to satisfaction data are threefold. First, the methods are easily implementable in standard statistical software. Second, the non-linear model allows for exible marginal effects, and predicted means respect the boundaries of the dependent variable. Third, average marginal effects are similar to ordinary least squares estimates.
rating variables non-linear least squares quasi-maximum likelihood semiparametric least squares subjective well-being