gllm (version 0.38)

emgllm: Generalized log-linear modelling by EM and iterative proportional fitting

Description

Fits log-linear models for incomplete contingency tables, including some latent class models, via an EM approach.

Usage

emgllm(y,s,X,maxit=1000,tol=0.00001)

Value

A list with components:

deviance

the final model deviance (-2 log likelihood)

observed.values

the observed counts in y

fitted.values

the expected counts under the fitted model

full.table

the expected counts for the full (unobserved) table.

Arguments

y

is the observed contingency table.

s

is a vector of indices, one for each cell of the full (unobserved) contingency table, representing the appropriate cell of y

X

is the design matrix, or a formula.

maxit

is the number of EM iterations.

tol

is the convergence criterion for the LR criterion.

Details

The generalized log-linear model allows for modelling of incomplete contingency tables, that is tables where one or more dimensions have been collapsed over. These include situations where imprecise measures have been calibrated using a "perfect" gold standard, and the true association between imperfectly measured variables is to be estimated; where data is missing for a subsample of the population; latent variable models where latent variables are "errorless" functions of observed variables - eg ML gene frequency estimation from counts of observed phenotypes; specialised measurement models eg where observed counts are mixtures due to perfect measures and error prone measures; standard latent class analysis; symmetry and quasi-symmetry models for square tables.

The general framework underlying these models is summarised by Espeland (1986), and Espeland & Hui (1987), and is originally due to Thompson & Baker (1981). An observed contingency table \(y\), which will be treated as a vector, is modelled as arising from an underlying complete table \(z\), where observed count \(y_j\) is the sum of a number of elements of \(z\), such that each \(z_i\) contributes to no more than one \(y_j\). Therefore one can write \(y=F'z\), where \(F\) is made up of orthogonal columns of ones and zeros.

We then specify a loglinear model for \(z\), so that \(log(E(z))=X'b\), where \(X\) is a design matrix, and \(b\) a vector of loglinear parameters. The loglinear model for \(z\) and thus \(y\), can be fitted via an iterative proportional fitting algorithm for \(b\) and \(z\), with an EM fitting for \(y\), \(z\) and \(b\) (Haber 1984).

The emgllm function is a wrapper for C code implementing the approach in Haber (1984).

References

Espeland MA (1986). A general class of models for discrete multivariate data. Commun. Statist.-Simula 15:405-424.

Espeland MA, Hui SL (1987). A general approach to analyzing epidemiologic data that contains misclassification errors. Biometrics 43:1001-1012.

Haber M (1984). AS207: Fitting a general log-linear model. Appl Statist 33:358-362.

Thompson R, Baker RJ (1981). Composite link functions in generalized linear models. Appl Stat 30: 125-131.

Examples

#
# latent class analysis: two latent classes
#
# Data matrix 2x2x2x2x2 table of responses to five binary items
#
y<-c( 3,   6,   2,   11,   1,   1,   3,    4,
      1,   8,   0,   16,   0,   3,   2,   15,
     10,  29,  14,   81,   3,  28,  15,   80,
     16,  56,  21,  173,  11,  61,  28,  298)
#
# Scatter matrix: full table is 2x2x2x2x2x2
#
s<-  c(1:32,1:32)
#
# Design matrix: x is the latent variable (2 levels), 
# a-e are the observed variables
#
i<-rep(1,64)
x<-as.integer(gl(2,32,64))-1
a<-as.integer(gl(2,16,64))-1
b<-as.integer(gl(2,8 ,64))-1
c<-as.integer(gl(2,4 ,64))-1
d<-as.integer(gl(2,2 ,64))-1
e<-as.integer(gl(2,1 ,64))-1
X<-cbind(i,x,a,b,c,d,e,x*cbind(a,b,c,d,e))
colnames(X)<-c("Int","X","A","B","C","D","E","AX","BX","CX","DX","EX")
res<-emgllm(y,s,X, tol=0.01)
res
#
# Obtain standard errors for parameter estimates
#
summary(scoregllm(y,s,X,as.array(res$full.table)))