When TTO data from all respondents are pooled and jointly modelled one obtains a model for the whole sample, averaging out individual differences between respondents. In such a model it is assumed that all respondents behave similarly in valuing the EQ-5D health states. Fundamentally this means that it is assumed that all variability in health state observations is due to random error and that all respondents have the same utility model. However, this is not the case in reality. For instance, we see in the data of many (if not all) valuation studies that some respondents don’t consider any health state to be worse than dead while others do. This is a clear indication that different respondents have different preferences and therefore a different utility model.
We hypothesize that there are two fundamental aspects that are the main drivers of the difference in preferences of respondents. The first is trading behaviour: people use the TTO scale in a fundamentally different way. The second is domain behaviour: people differ in their opinion regarding the relative importance of the 5 EQ-5D domains. By grouping the respondents according to their preference structure and analysing these groups separately, we effectively remove the structural variability due to differences in preferences within each group.
The first objective of the study is to obtain the groups of preference structure and to use these to accurately simulate the observed data obtained with an EQ-VT study. The second objective is to approach the modelling of EQ-VT data as a mix of models rather than as a single model, thereby enhancing both the accuracy and the transparency of the modelling exercise. This paper presents the first results of the study, namely those for trading behaviour.