conting (version 1.7)

bict: Bayesian Analysis of Incomplete Contingency Tables

Description

These functions implement a Bayesian analysis of incomplete contingency tables. This is accomplished using a data augmentation MCMC algorithm where the null moves are performed using the Metropolis-Hastings algorithm and the between models moves are performed using the reversible jump algorithm. This function can also accomodate cases where one of the sources observes a mixture of individuals from target and non-target populations. This results in the some of the cell counts being censored.

bict should be used initially, and bictu should be used to do additional MCMC iterations, if needed.

Usage

bict(formula, data, n.sample, prior = "SBH", cens = NULL, start.formula = NULL, 
start.beta = NULL, start.sig = NULL, start.y0 = NULL, save = 0, name = NULL, 
null.move.prob=0.5, a = 0.001, b = 0.001, progress = FALSE)
bictu(object, n.sample, save = NULL, name = NULL, progress = FALSE)

Arguments

formula

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

object

An object of class "bict" produced as a previous call to bict or bictu.

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 bict is called.

n.sample

A numeric scalar giving the number of MCMC iterations to peform.

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".

cens

A numeric vector indicating the row numbers of the data.frame in data which correspond to the censored cells. This can be found using the function find_cens.

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.

start.y0

An optional argument giving the starting values of the missing and censored cell counts. This should have the same length as the number of missing and censored cell counts.

save

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

For bict, 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 bictu, 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 bict or bictu.

name

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

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 "bict" 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.

Y0

An n.sample by k matrix giving the sampled values of the missing and censored cell counts, where k is the total number of missing and censored cell counts.

missing1

A vector of the same length as the number of missing cell counts giving the row numbers of the data.frame in data (or the elements of the variables) which correspond to the missing cell counts.

missing2

A vector of the same length as the number of censored cell counts giving the row numbers of the data.frame in data (or the elements of the variables) which correspond to the censored cell counts.

missing_details

The rows of the data.frame in data (or the elements of the variables) which correspond to the missing cell counts.

censored_details

The rows of the data.frame in data (or the elements of the variables) which correspond to the censored cell counts.

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.

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 the orthogonal projections method for GLMs proposed by Forster et al (2012). For details on these methods applied to log-linear models through the data-augmentation algorithm see Overstall & King (2014), and the references therein. For details on the censored approach see Overstall et al (2014).

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/

Overstall, A.M., King, R., Bird, S.M., Hutchinson, S.J. & Hay, G. (2014) Incomplete contingency tables with censored cells with application to estimating the number of people who inject drugs in Scotland. Statistics in Medicine, 33 (9), 1564--1579.

See Also

bict.fit, spina, ScotPWID.

Examples

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

data(spina)
## Load the spina data

test1<-bict(formula=y~(S1 + S2 + S3 + eth)^2,data=spina,n.sample=50, prior="UIP")
## Let the maximal model be the model with two-way interactions. Starting from the 
## posterior mode of the model with two-way interactions, do 50 iterations under the 
## unit information prior.

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

test1

#Number of cells in table = 24
#
#Maximal model =
#y ~ (S1 + S2 + S3 + eth)^2
#
#Number of log-linear parameters in maximal model = 15 
#
#Number of MCMC iterations = 100 
#
#Computer time for MCMC = 00:00:01 
#
#Prior distribution for log-linear parameters = UIP 
#
#Number of missing cells = 3 
#
#Number of censored cells = 0

summary(test1)
## Summarise the result. Will get:

#Posterior summary statistics of log-linear parameters:
#            post_prob post_mean post_var lower_lim upper_lim
#(Intercept)         1    1.0427 0.033967    0.6498    1.4213
#S11                 1   -0.3159 0.015785   -0.4477   -0.1203
#S21                 1    0.8030 0.018797    0.6127    1.1865
#S31                 1    0.7951 0.003890    0.6703    0.8818
#eth1                1    2.8502 0.033455    2.4075    3.1764
#eth2                1    0.1435 0.072437   -0.4084    0.5048
#S21:S31             1   -0.4725 0.002416   -0.5555   -0.3928
#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 0.36 ~S1 + S2 + S3 + eth + S2:S3                                           
#2 0.19 ~S1 + S2 + S3 + eth + S2:S3 + S2:eth                                  
#3 0.12 ~S1 + S2 + S3 + eth + S1:eth + S2:S3                                  
#4 0.12 ~S1 + S2 + S3 + eth + S1:S2 + S1:S3 + S1:eth + S2:S3 + S2:eth + S3:eth
#5 0.10 ~S1 + S2 + S3 + eth + S1:S3 + S1:eth + S2:S3                          
#6 0.06 ~S1 + S2 + S3 + eth + S1:S3 + S1:eth + S2:S3 + S2:eth                 
#
#Total number of models visited =  8 
#
#Posterior mean of total population size = 726.75 
#95 % highest posterior density interval for total population size = ( 706 758 ) 
#
#Under the X2 statistic 
#
#Summary statistics for T_pred 
#   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#  8.329  15.190  20.040  22.550  24.180 105.200 
#
#Summary statistics for T_obs 
#   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#  5.329  18.270  22.580  21.290  24.110  37.940 
#
#Bayesian p-value =  0.45


# }