Surveys aimed at conducting choice modelling analyses routinely include questions about perceptions and attitudes. While researchers have occasionally included responses to such questions in their models in a deterministic fashion, it is well known that this can lead to endogeneity bias. This has led to a growing popularity for latent variable approaches that jointly model the response to stated choice tasks and attitudinal questions. Such hybrid frameworks have proved a fruitful approach for explaining differences across individual respondents in sensitivities. At the same time, a separate stream of research has started to openly question the nature of the taste heterogeneity retrieved in Mixed Logit analyses, showing that at least part of the heterogeneity retrieved in such models is in fact scale heterogeneity, i.e. variation in absolute rather than relative sensitivities. In the present paper, we combine these two approaches. Specifically, we hypothesise that differences across respondents in survey engagement, understanding or attention result in differences in response quality, expressed as scale heterogeneity. We model this through a random scale approach that interacts with a latent variable model. Here, we find clear evidence of a link between this latent variable, model scale, and the response to various questions about survey realism and complexity. The resulting model is able to better represent the observed choices, while also leading to noticeable differences in the retrieved heterogeneity patterns.