predict.bayesQR {bayesQR}

Calculate predicted probabilities for binary quantile regression

Description

predict.bayesQR is an S3 method that calculates predicted probabilities for binary quantile regression, both with or without adaptive lasso.

Usage

  ## S3 method for class 'bayesQR'
predict(object, X, burnin=0, ...)

Arguments

object

an output object of the bayesQR function, S3 class bayesQR, with binary dependent variable, with or without adaptive lasso. At leas 9 estimated quantiles are required.

X

matrix of predictors (should be of the same type as the variables used to estimate the model).

burnin

the number of burnin draws that should be discared (default=0, meaning all draws are included).

...

additional parameters passed to the generic predict function.

Details

predict.bayesQR is an S3 method that calculates the predicted probabilities based on a matrix of predictors X and an object containing the parameter estimates of a sequence of binary quantile regressions. The rationale behind the approach is described in Kordas (2006) and applied in Migueis, V.L., Benoit, D.F. and Van den Poel, D. (2012). Note that the more quantiles are estimated, the more fine-grained the predicted probabilities will be.

Value

A vector containing the predicted probabilities.

Author(s)

Dries F. Benoit

References

Kordas, G. (2006). Binary regression quantiles, Journal of Applied Econometrics, 21(3), 387-407.

Migueis, V.L., Benoit, D.F. and Van den Poel, D. (2012). Enhanced decision support in credit scoring using Bayesian binary quantile regression, Journal of the Operational Research Society, (in press).

Examples

# Simulate data from binary regression model
set.seed(1234)
n <- 200
X <- matrix(runif(n=n*2, min=-5, max=5),ncol=2)
ystar <- cbind(1,X)%*% c(1,1.5,-.5) + rnorm(n=n, mean=0, sd=abs(2*X[,1]))
y <- as.numeric(ystar>0)

# Estimate a sequence of binary quantile regression models
# NOTE: to limit execution time of the example, ndraw is set
#       to a very low value. Set value to 4000 for a better
#       approximation of the posterior distirubtion.
out <- bayesQR(y ~ X, quantile=seq(.1,.9,.1), ndraw=400)

# Calculate predicted probabilities
pred <- predict(object=out, X=cbind(1,X), burnin=20)

# Make histogram of predicted probabilities 
hist(pred,breaks=10)

# Calculate Percentage Correclty Classified (PCC)
mean(y==as.numeric(pred>.5))

# Compare with logit model
mylogit <- glm(y ~ X, family=binomial(logit))

# Make histogram of predicted probabilities 
hist(mylogit$fit,breaks=10)

# Calculate Percentage Correclty Classified (PCC)
mean(y==as.numeric(mylogit$fit>.5))