This adds a new performance statistic, the geometric mean predictive density (GMPD), which is particularly useful for discrete outcomes because there, the GMPD is a geometric mean of probabilities and hence bounded by zero and one. As explained in the documentation for argument stats in the ?summary.vsel help, the SE of the GMPD is derived using the delta method ($SE{GMPD} = SE{MLPD} \cdot GMPD$; sorry for the bad math formatting: GitHub does not seem to support \text{} or \mathrm{}). The confidence interval (CI) for the GMPD is calculated by exponentiating the CI bounds of the MLPD CI. This is also what is done in a similar case in Stata, for example (see here).
This adds a new performance statistic, the geometric mean predictive density (GMPD), which is particularly useful for discrete outcomes because there, the GMPD is a geometric mean of probabilities and hence bounded by zero and one. As explained in the documentation for argument
stats
in the?summary.vsel
help, the SE of the GMPD is derived using the delta method ($SE{GMPD} = SE{MLPD} \cdot GMPD$; sorry for the bad math formatting: GitHub does not seem to support\text{}
or\mathrm{}
). The confidence interval (CI) for the GMPD is calculated by exponentiating the CI bounds of the MLPD CI. This is also what is done in a similar case in Stata, for example (see here).