A Generalized Two-Part Model for Fractional Response Variables with Excess Zeros

The fractional probit (or fractional logit) model is used when the outcome variable is a fractional response variable, i.e. a variable taking a value between zero and one. In case of excess zeros, the fractional probit model might not be the optimal modeling device since this model does not predict zeros. As a solution, the two-part model has been proposed, which assumes different processes for having a (non-)zero outcome and, conditionally on having a non-zero outcome, the actual outcome. However, the two-part model assumes independence of these processes. This paper proposes a generalization of the two-part model which allows for dependence of these processes and which also nests the two-part model as a special case. A simulation study indicates that the proposed estimator performs well in finite samples. Two empirical examples illustrate that the model proposed in this paper improves upon the fractional probit and two-part model in terms of model fit and also leads to different marginal effects.