This paper develops and compares two alternative approaches to accommodate scale heterogeneity (also referred to as heteroskedasticity) in latent class models. Our modelling approach compares two different representations of heteroskedasticity, respectively associating the heterogeneity in scale factor with respondent's characteristics (i.e. observed scale heterogeneity) or deriving it probabilistically (i.e. unobserved scale heterogeneity). The results reveal a number of benefits associated with this type of approach, particularly when heterosckedasticity can be linked to observed characteristics of the respondent. Our data comes from a discrete choice experiment eliciting recreational users preferences for farmland walking trails in Ireland