The PLM Procedure

Scoring Data Sets for Zero-Inflated Models

The PLM procedure can score new observations for zero-inflated models with the SCORE statement. If you specify the ILINK option, the computed statistics are for estimated counts.

In the following formula, $\mb{x}$ is the design row for covariates that correspond to the the Poisson or negative binomial component, $\hat{\bbeta }$ is the column vector of the fitted regression parameters; $\mb{z}$ is the design row for covariates that correspond to the zero inflation component, $\hat{\bgamma }$ is the column vector of the fitted regression parameters; g and g$^{-1}$ are the link and inverse link functions for the Poisson or negative binomial component, $g_ z$ and $g^{-1}_ z$ are the link and inverse link functions for the zero inflation component; $\Phi $ is the standard normal cumulative distribution function and $\alpha $ is the nominal significance level. Let

\[  \begin{array}{lclcl} \mu &  = &  g^{-1}(\eta ) &  = &  g^{-1}(\mb{x}\hat{\bbeta }) \\ \mu _ z &  = &  g^{-1}_ z(\eta _ z) &  = &  g^{-1}_ z(\mb{z}\hat{\bgamma }) \\ v_1 &  = &  (\frac{\mathrm{d}\mu }{\mathrm{d}\eta })^2 \mb{x}\hat{\mb{V}}_{\bbeta ,\bbeta }\mb{x}’\\ v_2 &  = &  (\frac{\mathrm{d}\mu _ z}{\mathrm{d}\eta _ z})^2 \mb{z}\hat{\mb{V}}_{\bgamma ,\bgamma }\mb{z}’\\ v_{12} &  = &  -(\frac{\mathrm{d}\mu }{\mathrm{d}\eta }) (\frac{\mathrm{d}\mu _ z}{\mathrm{d}\eta _ z}) \mb{x}\hat{\mb{V}}_{\bbeta ,\bgamma }\mb{z}’\\ \end{array}  \]

The formula for statistics in the SCORE statement for zero-inflated models are listed as follows.

\[  \begin{array}{lclcl} \mr{PZERO} &  = &  \omega &  = &  \mu _ z \\ \mr{PRED/ILINK} &  = &  p_ c &  = &  \mu (1-\omega )\\ \mr{STD/ILINK} &  = &  s_ c &  = &  \sqrt {p_ c^2v_2+(1-\omega )^2v_1+ v_1v_2+2p_ c(1-\omega )v_{12}+v^2_{12}} \\ \mr{UCLM/ILINK} &  = &  u_ c &  = &  p_ c\exp (\Phi ^{-1}(1-\alpha /2) s_ c/p_ c) \\ \mr{LCLM/ILINK} &  = &  l_ c &  = &  p_ c/\exp (\Phi ^{-1}(1-\alpha /2) s_ c/p_ c) \\ \mr{PRED} &  = &  p_ l &  = &  g(p_ c) \\ \mr{STD} &  = &  s_ l &  = &  g(s_ c)/(\frac{\mathrm{d}\mu }{\mathrm{d}\eta }) \\ \mr{UCLM} &  = &  u_ l &  = &  g(u_ c) \\ \mr{LCLM} &  = &  l_ l &  = &  g(l_ c) \\ \end{array}  \]