Dear reader,
We are (trying to) model the determinents of firms receiving fines (panel data set), and the height of the fine(s). We observe, however, that many firms do not get any fines. As such, if our dependent variable is a fine dummy (yes/no fine in particular year), or our dependent variable is log(total fine amount), the results are vastly different. The latter omits those firm-year observations for which firms receive no fines.
We therefore have reason to suspect that the propensity of firms to receive a firm is determined by other reasons than the (total) fine height of a firm in a particular year.
I therefore investigated a hurdle model, as explained here:
https://www.stata.com/features/overview/hurdle-models/
However, an assumption this model makes is that the selection model (yes/no fine in a particular year) error terms are not correlated with the outcome equations (fine height) error terms. An alternative would be to use a copula hurdle model, where the dependence in error terms between selection and outcome equations are modelled. I am wondering though if that would be necessary in this case, and whether Stata offers perhaps a better alternative for this case.
Thanks for reading,
We are (trying to) model the determinents of firms receiving fines (panel data set), and the height of the fine(s). We observe, however, that many firms do not get any fines. As such, if our dependent variable is a fine dummy (yes/no fine in particular year), or our dependent variable is log(total fine amount), the results are vastly different. The latter omits those firm-year observations for which firms receive no fines.
We therefore have reason to suspect that the propensity of firms to receive a firm is determined by other reasons than the (total) fine height of a firm in a particular year.
I therefore investigated a hurdle model, as explained here:
https://www.stata.com/features/overview/hurdle-models/
However, an assumption this model makes is that the selection model (yes/no fine in a particular year) error terms are not correlated with the outcome equations (fine height) error terms. An alternative would be to use a copula hurdle model, where the dependence in error terms between selection and outcome equations are modelled. I am wondering though if that would be necessary in this case, and whether Stata offers perhaps a better alternative for this case.
Thanks for reading,
