binomlogit (version 1.2)

IndivdRUMIndMH: Data-augmented independence Metropolis-Hastings sampling for the binary logit model

Description

IndivdRUMIndMH simulates the posterior distribution of the regression coefficients of a binary logit model and returns the MCMC draws. The sampling procedure is based on an algorithm using data augmentation, where the regression coefficients are estimated by rewriting the binary logit model as a latent variable model called difference random utility model (dRUM). This dRUM representation of the binary logit model is equivalent to the individual dRUM representation of the binomial logit model. The logistic distribution of the error in the dRUM representation is approximated by a normal distribution with same mean and variance. The posterior of this approximate regression model is then used as independence proposal density for the regression coefficients.

Usage

IndivdRUMIndMH(y, X, sim = 12000, burn = 2000, acc = 0, b0, B0,  start, verbose = 500)

Arguments

y

logical or coercible to integer with values 0 and 1 (Bernoulli data).

X

a design matrix of predictors.

sim

number of MCMC draws including burn-in. The default value is 12000 draws.

burn

number of MCMC draws discarded as burn-in. Default is a burn-in of 2000 draws.

acc

number of MCMC draws at the beginning of the burn-in phase where each proposed parameter vector is accepted with probability 1 rather than according to the MH acceptance rule. Choose a small number acc > 0 if the sampler gets stuck at the starting values, otherwise acc is set to 0.

b0

an optional vector of length dims = ncol(X) containing the prior mean. Otherwise a vector of zeros is used.

B0

an optional dims x dims prior variance-covariance matrix. Otherwise a diagonal matrix with all diagonal elements equal to 10 is used.

start

an optional vector of length dims = ncol(X) containing the starting values for the regression parameters. Otherwise a vector of zeros is used.

verbose

an optional non-negative integer indicating that in each verbose-th iteration step a status report is printed (default: verbose = 500). If 0, no output is generated during MCMC sampling.

Value

The output is a list object of class c("binomlogitIndiv","binomlogitMH", "binomlogit") containing

beta

a dims x sim matrix of sampled regression coefficients from the posterior distribution

sim

the argument sim

burn

the argument burn

acc

the argument acc

dims

number of covariates (dims = ncol(X))

N

number of binary observations

b0

the argument b0

B0

the argument B0

duration

a numeric value indicating the total time (in secs) used for the function call

duration_wBI

a numeric value indicating the time (in secs) used for the sim-burn MCMC draws after burn-in

rate

acceptance rate based on the sim-burn MCMC draws after burn-in

To display the output use print, summary and plot. The print method prints the number of observations and covariates entered in the routine, the total number of MCMC draws (including burn-in), the number of draws discarded as burn-in, the runtime used for the whole algorithm and for the sim-burn MCMC draws after burn-in and the acceptance rate. The summary method additionally returns the length of the acceptance phase during burn-in, the used prior parameters b0 and B0 and the posterior mean for the regression coefficients without burn-in. The plot method plots the MCMC draws and their acf for each regression coefficient, both without burn-in.

Details

For details concerning the algorithm see the papers by Fussl, Fruehwirth-Schnatter and Fruehwirth (2013) and Fruehwirth-Schnatter and Fruehwirth (2010).

References

Agnes Fussl, Sylvia Fruehwirth-Schnatter and Rudolf Fruehwirth (2013), "Efficient MCMC for Binomial Logit Models". ACM Transactions on Modeling and Computer Simulation 23, 1, Article 3, 21 pages.

Sylvia Fruehwirth-Schnatter and Rudolf Fruehwirth (2010), "Data augmentation and MCMC for binary and multinomial logit models." In Statistical Modelling and Regression Structures - Festschrift in Honour of Ludwig Fahrmeir, T. Kneib and G. Tutz, Eds. Physica-Verlag, Heidelberg, pp. 111-132.

See Also

dRUMIndMH, dRUMAuxMix, dRUMHAM

Examples

## Data-augmented independence Metropolis-Hastings sampling for the
## binary logit model

## load caesarean birth data
data(caesarean_binary)
y <- as.numeric(caesarean_binary[,1])                                                 
X <- as.matrix(caesarean_binary[,-1])

## start independence MH sampler
indivMH1 <- IndivdRUMIndMH(y,X)

print(indivMH1)
summary(indivMH1)
plot(indivMH1)

## Not run: 
# ## load simulated data set
# data(simul_binary)
# y <- as.numeric(simul_binary[,1])                                           
# X <- as.matrix(simul_binary[,-1])
# 
# ## use a small acc>0 (e.g. acc=50), otherwise the sampler gets stuck at
# ## the starting values
# indivMH2 <- IndivdRUMIndMH(y,X,acc=50)
# 
# print(indivMH2)
# summary(indivMH2)
# plot(indivMH2)
# ## End(Not run)