conting (version 1.7)

bcct: Bayesian Analysis of Complete Contingency Tables

Description

These functions implement a Bayesian analysis of complete contingency tables. This is accomplished using an MCMC algorithm where the null moves are performed using a Metropolis-Hastings algorithm and the between models moves are performed using a reversible jump algorithm.

bcct should be used initially, and bcctu should be used to do additional MCMC iterations, if required.

bcctsubset and bcctsubsetu operate on a subset of models.

Usage

bcct(formula, data, n.sample, prior = "SBH", start.formula = NULL, 
start.beta = NULL, start.sig = NULL, save = 0, name = NULL, null.move.prob=0.5, 
a = 0.001, b = 0.001, progress = FALSE)
bcctu(object, n.sample, save = NULL, name = NULL, progress = FALSE)
bcctsubset(subsetformula, data, n.sample, prior = "SBH", start.formula = NULL, 
start.beta = NULL, start.sig = NULL, save = 0, name = NULL, null.move.prob=0.5, 
a = 0.001, b = 0.001, progress = FALSE)
bcctsubsetu(object, n.sample, save = NULL, name = NULL, progress = FALSE)

Arguments

formula

An object of class "formula": a symbolic description of the maximal model.

subsetformula

A list with elements of class "formula": symbolic description of the subset of models. The first element in the list should be the maximal model.

object

An object of class "bcct" produced as a previous call to bcct or bcctu.

data

An object of class "data.frame" (or "table") containing the variables in the model. If the model variables are not found in data, the variables are taken from environment(formula), typically the environment from which bcct is called.

n.sample

A numeric scalar giving the number of (additional, in the case of bcctu) MCMC iterations to perform.

prior

An optional argument giving the prior to be used in the analysis. It can be one of c("UIP","SBH"), where "UIP" = unit information prior; and "SBH" = Sabanes-Bove & Held prior. The default value is "SBH".

start.formula

An optional argument giving an object of class "formula": a symbolic description of the starting model in the MCMC algorithm. If NULL (the default) the starting model will be the maximal model.

start.beta

An optional argument giving the starting values of the log-linear parameters for the MCMC algorithm. It should be a vector of the same length as the number of log-linear parameters in the starting model implied by the argument start.formula. If NULL (the default) the starting value will be the posterior mode under the maximal model.

start.sig

An optional argument giving the starting value of sigma^2 (under the Sabanes-Bove & Held prior) for the MCMC algorithm when the argument of prior is "SBH". If NULL (the default) the starting value will be one.

save

An optional argument for saving the MCMC output mid-algorithm.

For bcct and bcctsubset, if positive, the function will save the MCMC output to external text files every save iterations. If zero (the default), the function will not save the MCMC output to external files.

For bcctu and bcctsubsetu, if non-NULL, the function will save the MCMC output to external text files every save iterations. If NULL (the default), it will inherit the value of save from the previous call to bcct or bcctu.

name

An optional argument giving a prefix to the file name of the external files saved if the argument save is positive. For bcct, a value of NULL means the external files will not have a prefix. For bcctu, a value of NULL, means the prefix will be inherited from the previous call to bcct or bcctu.

null.move.prob

An optional scalar argument giving the probability of performing a null move in the reversible jump algorithm, i.e. proposing a move to the current model. The default value is 0.5.

a

The shape hyperparameter of the Sabanes-Bove & Held prior, see Overstall & King (2014). The default value is 0.001. A value of a = -1 gives the Gelman prior (Gelman, 2006), i.e. a uniform prior on the standard deviation.

b

The scale hyperparameter of the Sabanes-Bove & Held prior, see Overstall & King (2014). The default value is 0.001. A value of b = 0 gives the Gelman prior (Gelman, 2006), i.e. a uniform prior on the standard deviation.

progress

Logical argument. If TRUE, then a progress bar will be displayed. The default value is FALSE.

Value

The functions will return an object of class "bcct" which is a list with the following components:

BETA

An n.sample by p matrix containing the sampled values of the log-linear parameters, where p is the number of log-linear parameters in the maximal model. For elements of this matrix which correspond to a log-linear parameter which is not present for the current model a zero is returned.

MODEL

A vector of length n.sample giving the sampled model indicators in hexadecimal format.

SIG

A vector of length n.sample giving the sampled values for sigma^2 under the Sabanes-Bove & Held prior. If the unit information prior is used then the components of this vector will be one.

rj_acc

A binary vector of the same length as the number of reversible jump moves attempted. A 0 indicates that the proposal was rejected, and a 1 that the proposal was accepted.

mh_acc

A binary vector of the same length as the number of Metropolis-Hastings moves attempted. A 0 indicates that the proposal was rejected, and a 1 that the proposal was accepted.

priornum

A numeric scalar indicating which prior was used: 1 = "UIP", 2 = "SBH".

maximal.mod

An object of class "glm" giving the fit of the maximal model.

IP

A p by p matrix giving the inverse of the prior scale matrix for the maximal model.

eta.hat

A vector of length n (number of cells) giving the posterior mode of the linear predictor under the maximal model.

save

The argument save.

name

The argument name.

null.move.prob

The argument null.move.prob.

time

The total computer time (in seconds) used for the MCMC computations.

a

The argument a.

b

The argument b.

subset.index

Model indicators (in hexidecimal format) of the subset of models.

Details

For identifiability, the parameters are constrained. The conting-package uses sum-to-zero constraints. See Overstall & King (2014), and the references therein, for more details.

The Metropolis-Hastings algorithm employed is the iterated weighted least squares method for generalised linear models (GLMs) proposed by Gamerman (1997). The reversible jump algorithm employed is that orthogonal projections method for GLMs proposed by Forster et al (2012). For details on these methods applied to log-linear models see Overstall & King (2014), and the references therein.

For details on the unit information and Sabanes-Bove & Held priors for generalised linear models see Ntzoufras et al (2003) and Sabanes-Bove & Held (2011), respectively. See Overstall & King (2014), and the references therein, for their application to log-linear models and contingency tables.

References

Sabanes-Bove, D. & Held, L. (2011) Hyper-g priors for generalized linear models. Bayesian Analysis, 6 (3), 387--410.

Forster, J.J., Gill, R.C. & Overstall, A.M. (2012) Reversible jump methods for generalised linear models and generalised linear mixed models. Statistics and Computing, 22 (1), 107--120.

Gamerman, D. (1997) Sampling from the posterior distribution in generalised linear mixed models. Statistics and Computing, 7 (1), 57--68.

Gelman, A. (2006) Prior distributions for variance parameters in hierarchical models(Comment on Article by Browne and Draper). Bayesian Analysis, 1 (3), 515--534.

Nztoufras, I., Dellaportas, P. & Forster, J.J. (2003) Bayesian variable and link determination for generalised linear models. Journal of Statistical Planning and Inference, 111 (1), 165--180.

Overstall, A.M. & King, R. (2014) conting: An R package for Bayesian analysis of complete and incomplete contingency tables. Journal of Statistical Software, 58 (7), 1--27. http://www.jstatsoft.org/v58/i07/

See Also

bcct.fit, AOH, heart.

Examples

# NOT RUN {
set.seed(1)
## Set seed for reproducibility.

data(AOH)
## Load the AOH data

test1<-bcct(formula=y~(alc+hyp+obe)^3,data=AOH,n.sample=50,prior="UIP")
## Let the maximal model be the saturated model. Starting from the 
## posterior mode of the maximal model do 50 iterations under the unit 
## information prior.

test1<-bcctu(object=test1,n.sample=50)
## Do another 50 iterations

test1
## Printing out a bcct object produces this simple summary

#Number of cells in table = 24 
#
#Maximal model =
#y ~ (alc + hyp + obe)^3
#
#Number of log-linear parameters in maximal model = 24 
#
#Number of MCMC iterations = 100 
#
#Computer time for MCMC = 00:00:01 
#
#Prior distribution for log-linear parameters = UIP

summary(test1)
## Printing out a summary produces a bit more:

#Posterior summary statistics of log-linear parameters:
#            post_prob post_mean post_var lower_lim upper_lim
#(Intercept)         1  2.877924 0.002574   2.78778   2.97185
#alc1                1 -0.060274 0.008845  -0.27772   0.06655
#alc2                1 -0.049450 0.006940  -0.20157   0.11786
#alc3                1  0.073111 0.005673  -0.05929   0.20185
#hyp1                1 -0.544988 0.003485  -0.65004  -0.42620
#obe1                1 -0.054672 0.007812  -0.19623   0.12031
#obe2                1  0.007809 0.004127  -0.11024   0.11783
#NB: lower_lim and upper_lim refer to the lower and upper values of the
#95 % highest posterior density intervals, respectively
#
#Posterior model probabilities:
#  prob model_formula                                 
#1 0.45 ~alc + hyp + obe                              
#2 0.30 ~alc + hyp + obe + hyp:obe                    
#3 0.11 ~alc + hyp + obe + alc:hyp + hyp:obe          
#4 0.06 ~alc + hyp + obe + alc:hyp + alc:obe + hyp:obe
#5 0.05 ~alc + hyp + obe + alc:hyp                    
#
#Total number of models visited =  7 
#
#Under the X2 statistic 
#
#Summary statistics for T_pred 
#   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#  11.79   20.16   23.98   24.70   28.77   52.40 
#
#Summary statistics for T_obs 
#   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#   8.18   24.22   31.51   30.12   35.63   42.49 
#
#Bayesian p-value =  0.28

## For more examples see Overstall & King (2014).


# }