Multinomial regression and its application in discrete choice experiments

This master’s thesis is focused on econometric models for discrete choice analysis. Main goal of this thesis is to introduce the topic of discrete choice experiments and it's econometric aspects in the Czech republic. Theory summarizes the main model, which is the conditional logit model, as well as appropriate extensions. For labelled discrete choice experiments, suitable models are nested logit, heteroscedastic logit and probit model. A mixed logit model is suggested as an alternative for unlabelled choice experiment. Econometric theory is accompanied by some crucial specifics for modelling discrete choice experiments, like specifying the constant for various choice experiments or the concept of effects coding. The use of presented models is demonstrated on unlabelled choice experiment focused on flood protection. Each respondent was asked to select a preferred land in nine different scenarios. Each choice task had three possibilities: land A, land B or none. The key focus was on the various plane, point and linear aspects of land. Plane measures include strip crop rotation, grain field and corn field. Considered point measures are ponds and dry polder. Linear measures were field margin and grassing along the stream. Several models are introduced to highlight different possibilities in specifying the models and their differences. All calculations are conducted in R. Models suggest that people prefer strip crop rotation to corn and grain field, ponds to dry polders and field margins to grassing along the stream. Preferences are not entirely homogenous throughout the respondents. The most appropriate model is a mixed logit model with jointly normal attributes and log-normally distributed (negative) cost coefficient.