margins.des {catregs}

Creates a design matrix of idealized data for illustrating model predictions.

Description

Create a data frame of idealized data for making model predictions/predicted margins that will be used with margins.dat for generating/plotting model predictions. Given a model object (generalized linear model or generalized linear mixed model), a grid of independent variable values, and a list of any variables (factor variables in particular) to exclude from the design matrix, the function returns the design matrix as a data.frame object. All covariates are set to their means in the data used to estimate the model object. If there are factors in the model, they need to be excluded using the "excl" option.

Supported models include OLS via the lm function, logistic and Poisson regression via the glm function, negative binomial regression via MASS::glm.nb, ordinal logistic regression via MASS::polr, multinomial logistic regression via nnet::multinom, partial proportional odds models via vgam::vglm, linear mixed effects models with or without serial correlation specified via nlme::lme, mixed effects logistic regression via lme4::glmer, mixed effects Poisson regression via lme4::glmer, mixed effects negative binomial regression via lme4::glmer.nb, and mixed effects ordinal logistic regression via ordinal::clmm. Zero-inflated and hurdle models are also supported via pscl::zeroinfl and pscl::hurdle, respectively.

For multinomial regression model, as estimated via the nnet package, you need to provide the data used in the nnet function that defined the model. For partial proportional odds models, as estimated via the vgam package, you need to specify an ordinal model via MASS::polr and provide that model to the margins.des function (the data for the model are not part of a vgam object.) For mixed effects regression models, as estimated by lme4/lmerTest, you need to provide the data used in the glmer or lmer function that defined the model. For mixed effects ordinal logistic regression models, as estimated via the ordinal package, the outcome variable in the regression model (i.e., the clmm function) needs to be named "dv."

Usage

margins.des(mod,ivs,excl="nonE",data)

Arguments

mod

A model object. The model should be regression model for limited dependent variables, such as a logistic regression. Specifically, supported models include lm, glm, polr, multinom, vgam, zeroinf (pscl), amd hurdle (pscl). vgam is supported for partial proportional odds models, not models for count outcomes. zerotrunc is only supported with bootstrapped inference, and may take a while.

ivs

This should be an 'expand.grid' statement including all desired variables and their corresponding levels in the design matrix.

excl

If you want to exclude covariates from the design matrix, you can list them here. This is designed to exclude factor variables from the design matrix, as they do not have means, but may be useful in other specialized cases. Default is "nonE," corresponding to excluding none of the variables.

data

If the model is a multinomial model, you also need to provide the data. This is because nnet objects do not include the relevant information for computing the means of covariates.

Value

design

Returns a data.frame containing the design matrix for model predictions.

Author(s)

David Melamed

Examples

data("Mize19AH")
m1 <- glm(alcB ~woman*parrole + age + race2 +
race3 + race4 + income + ed1 + ed2 + ed3 +
ed4,family="binomial",data=Mize19AH)
des1 <- margins.des(m1,expand.grid(parrole=1,woman=1))
des1
des2<-margins.des(m1,expand.grid(woman=c(0,1),parrole=c(0,1)))
des2
des3 <- margins.des(m1,expand.grid(age=seq(20,75,5),parrole=c(0,1)))
des3