UPG: Efficient MCMC Samplers for Bayesian probit regression and...

In UPG: Efficient Bayesian Algorithms for Binary and Categorical Data Regression Models

Efficient MCMC Samplers for Bayesian probit regression and various logistic regression models

Description

UPG estimates Bayesian regression models for binary or categorical outcomes using samplers based on marginal data augmentation.

Usage

UPG(y,
    X,
    model,
    Ni          = NULL,
    baseline    = NULL,
    draws       = 1000,
    burnin      = 1000,
    A0          = 4,
    B0          = 4,
    d0          = 2.5,
    D0          = 1.5,
    G0          = 100,
    verbose     = TRUE,
    gamma.boost = TRUE,
    delta.boost = TRUE,
    beta.start  = NULL)

Arguments

y	a binary vector for probit and logit models. A character, factor or numeric vector for multinomial logit models. A numerical vector of the number of successes for the binomial model.
X	a matrix of explanatory variables including an intercept in the first column. Rows are individuals, columns are variables.
model	indicates the model to be estimated. 'probit' for the probit model, 'logit' for the logit model, 'mnl' for the multinomial logit model or 'binomial' for the binomial logit model.
Ni	a vector containing the number of trials when estimating a binomial logit model.
baseline	a string that can be used to change the baseline category in MNL models. Default baseline is the most commonly observed category.
draws	number of saved Gibbs sampler iterations. Default is 1000 for illustration purposes, you should use more when estimating a model (e.g. 10,000).
burnin	number of burned Gibbs sampler iterations. Default is 1000 for illustration purposes, you should use more when estimating a model (e.g. 2,000).
A0	prior variance for the intercept, 4 is the default.
B0	prior variance for the coefficients, 4 is the default.
d0	prior shape for working parameter delta, 2.5 is the default.
D0	prior rate for working parameter delta, 1.5 is the default.
G0	prior variance for working parameter gamma, 100 is the default.
verbose	logical variable indicating whether progress should be printed during estimation.
gamma.boost	logical variable indicating whether location-based parameter expansion boosting should be used.
delta.boost	logical variable indicating whether scale-based parameter expansion boosting should be used.
beta.start	provides starting values for beta (e.g. for use within Gibbs sampler). Baseline coefficients need to be zero for multinomial model.

Value

Depending on the estimated model, UPG() returns a UPG.Probit, UPG.Logit, UPG.MNL or UPG.Binomial object.

Author(s)

Gregor Zens

See Also

summary.UPG.Probit to summarize a UPG.Probit object and to create tables. predict.UPG.Logit to predict probabilities using a UPG.Logit object. plot.UPG.MNL to plot a UPG.MNL object.

Examples

# load package
library(UPG)

# estimate a probit model using example data
# warning: use more burn-ins, burnin = 100 is just used for demonstration purposes
data(lfp)
y = lfp[,1]
X = lfp[,-1]
results.probit = UPG(y = y, X = X, model = "probit", burnin = 100)

# estimate a logit model using example data
# warning: use more burn-ins, burnin = 100 is just used for demonstration purposes
data(lfp)
y = lfp[,1]
X = lfp[,-1]
results.logit = UPG(y = y, X = X, model = "logit", burnin = 100)

# estimate a MNL model using example data
# warning: use more burn-ins, burnin = 100 is just used for demonstration purposes
data(program)
y = program[,1]
X = program[,-1]
results.mnl = UPG(y = y, X = X, model = "mnl", burnin = 100)

# estimate a binomial logit model using example data
# warning: use more burn-ins, burnin = 100 is just used for demonstration purposes
data(titanic)
y  = titanic[,1]
Ni = titanic[,2]
X  = titanic[,-c(1,2)]
results.binomial = UPG(y = y, X = X, Ni = Ni, model = "binomial", burnin = 100)

UPG documentation built on Nov. 4, 2023, 5:06 p.m.