This dissertation carries out a series of Monte Carlo simulations seeking the

implications for welfare estimates from three research practices commonly implemented in empirical applications of mixed logit and latent class logit.

Chapter 3 compares welfare measures across conditional logit, mixed logit, and

latent class logit. The practice of comparing welfare estimates is widely used in the field. However, this chapter shows comparisons of welfare estimates seem unable to

provide reliable information about the differences in welfare estimates that result

from controlling for unobserved heterogeneity. The reason is that estimates from

mixed logit and latent class logit are inherently inecient and inaccurate.

Researchers tend to use their own judgement to select the number of classes of

a latent class logit. Chapter 4 studies the reliability of welfare estimates obtained under two scenarios for which an empirical researcher using his/her judgement would

arguably choose less classes than the true number of classes. Results show that

models with a number of classes smaller than the true number tend to yield down-

ward biased and inaccurate estimates. The latent class logit with the true number

of classes always yield unbiased estimates but their accuracy may be worse than

models with the smaller number of classes.

Studies implementing discrete choice experiments commonly obtain estimates

of preference parameters from latent class logit models. This practice, however,

implies a mismatch: discrete choice experiments are designed under the assumption

of homogeneity in preferences, and latent class logit search for heterogeneous preferences. Chapter 5 studies whether welfare estimates are robust to this mismatch.

This chapter checks whether the number of choice tasks impact the reliability of

welfare estimates. The findings show welfare estimates are unbiased regardless the

number of choice tasks, and their accuracy increases with the number of choice tasks.

However, some of the welfare estimates are inefficient to the point that cannot be

statistically distinguished from zero, regardless the number of choice tasks.

Implications from these findings for the empirical literature are discussed.