ACD (version 1.5.3)

loglinML: Fitting Log-Linear Models via Maximum Likelihood

Description

loglinML fits log-linear models by ML (maximum likelihood). For complete data, it is based on a object of the class readCatdata. For missing data, it is based on a object of the class satMarML (under MAR or MCAR). Depending on the formulation (freedom equations or constraints) and on the model type (ordinary or generalized log-linear model), different arguments must be informed.

Usage

loglinML(obj, A, X, U, XL, UL, start, maxit=100, trace=0, epsilon1=1e-6, 
	epsilon2=1e-6, zeroN, digits)

Arguments

obj

object of the class readCatdata (for complete data) or satMarML (for missing data).

A

a matrix that specifies the log-linear functions of the probabilities to be modeled; by default, it is diag(S) %x% cbind(diag(R-1),rep(-1,R-1)), which generates logits with the last (R) category as baseline.

X

a model specification matrix for the freedom equation formulation of the ordinary log-linear model.

U

a matrix for the constraint formulation of the ordinary log-linear specification.

XL

a model specification matrix for the freedom equation formulation of the generalized log-linear model.

UL

a matrix for the constraint formulation of the generalized log-linear model.

start

by default, the function uses the proportions of the complete data as starting values in the iterative process, but the current argument allows the user to inform an alternative starting value for the parameters of the model if the freedom equation formulation is considered and the matrix A is modeling S*(R-1) log-linear functions; a vector with these values must be informed.

maxit

the maximum number of iterations (the default is 100).

trace

the alternatives are: 0 for no printing (default), 1 for showing only the value of the likelihood ratio statistics at each iteration of the iterative process, and 2 for including also the parameter estimates at each iteraction.

epsilon1

the convergence criterion of the iterative process is attained if the absolute difference of the values of the likelihood ratio statistic in successive iterations is less than the value defined in epsilon1, 1e-6 by default.

epsilon2

the convergence criterion of the iterative process is attained if the absolute differences of the values of estimates for all parameters of the marginal probabilities of categorization in consecutive iterations are less than the value defined in epsilon2, 1e-6 by default.

zeroN

values used to replace null frequencies in the denominator of the Neyman statistic; by default, the function replaces the values by 1/(R*nst), where nst is the sample size of the missingness pattern associated to the corresponding subpopulation; the user may indicate alternative values in a matrix with S rows and an additional column relatively to the number of columns of Rp; the first column relates to the completely categorized "missingness" patterns, and the remaining columns to the other missingness patterns as they appear in Rp; the values must be non-negative and less or equal to 0.5.

digits

integer value indicating the number of decimal places to round results when shown by print and summary; this argument works also when specified directly in both generic functions; default value is 4.

Value

An object of the class loglinML is a list containing most of the components of the argument obj as well as the following components:

thetaH

vector of ML estimates for all product-multinomial probabilities under the log-linear model for the marginal probabilities of categorization and, in the case of missing data, under an assumption of an ignorable missingness mechanism.

VthetaH

corresponding estimated covariance matrix.

beta

vector of ML estimates for the parameters of the log-linear model (only for freedom equation formulation).

Vbeta

corresponding estimated covariance matrix (only for the freedom equation formulation).

Fu

observed log-linear functions, without model constraints.

VFu

corresponding estimated covariance matrix.

FH

ML estimates for the log-linear functions under the fitted model.

VFH

corresponding estimated covariance matrix.

QvH

likelihood ratio statistic for testing the goodness of fit of the log-linear model (for missing data, conditional on the assumed missingness mechanism).

QpH

Pearson statistic for testing the goodness of fit of the log-linear model (for missing data, conditional on the assumed missingness mechanism).

QnH

Neyman statistic for testing the goodness of fit of the log-linear model (for missing data, conditional on the assumed missingness mechanism).

QwH

Wald statistic for testing the goodness of fit of the log-linear model (for missing data, conditional on the assumed missingness mechanism).

glH

degrees of freedom for testing the goodness of fit of the log-linear model (for missing data, conditional on the assumed missingness mechanism).

QvHMCAR

likelihood ratio statistic for the conditional test of both the log-linear model and MCAR given a MAR assumption (for missing data only).

QpHMCAR

Pearson statistic for the conditional test of both the log-linear model and MCAR given a MAR assumption (for missing data only).

QnHMCAR

Neyman statistic for the conditional test of both the log-linear model and MCAR given a MAR assumption (for missing data only).

glHMCAR

degrees of freedom for the conditional test of both the log-linear model and MCAR given a MAR assumption (for missing data only).

ystH

for complete data, it has the ML estimates for the frequencies under the log-linear model; for missing data, it has the ML estimates for the augmented frequencies under both the log-linear model and the assumed missingness mechanism.

Details

loglinML handles both ordinary and generalized log-linear model types either under a freedom equation formulation or under a constraint formulation. X and U are used for ordinary log-linear models, and XL and UL are used for generalized log-linear models. X and XL are used for the freedom equation formulation, and U and UL are used for the constraint formulation. Namely, the 4 ways with which the function allows to specify the model are: log(Theta)=nu+X%*%Beta, U%*%log(Theta)=0, A%*%log(Theta)=XL%*%Beta, UL%*%A%*%log(Theta)=0, where nu are non-estimated parameters included only to satisfy the natural constraints of the product-multinomial distribution and Beta are the parameters to be estimated.

The generic functions print and summary are used to print the results and to obtain a summary thereof.

References

Paulino, C.D. e Singer, J.M. (2006). Analise de dados categorizados (in Portuguese). Sao Paulo: Edgard Blucher.

Poleto, F.Z. (2006). Analise de dados categorizados com omissao (in Portuguese). Dissertacao de mestrado. IME-USP. http://www.poleto.com/missing.html.

Poleto, F.Z., Singer, J.M. e Paulino, C.D. (2007). Analyzing categorical data with complete or missing responses using the Catdata package. Unpublished vignette. http://www.poleto.com/missing.html.

Poleto, F.Z., Singer, J.M. e Paulino, C.D. (2012). A product-multinomial framework for categorical data analysis with missing responses. To appear in Brazilian Journal of Probability and Statistics. http://imstat.org/bjps/papers/BJPS198.pdf.

Singer, J. M., Poleto, F. Z. and Paulino, C. D. (2007). Catdata: software for analysis of categorical data with complete or missing responses. Actas de la XII Reunion Cientifica del Grupo Argentino de Biometria y I Encuentro Argentino-Chileno de Biometria. http://www.poleto.com/SingerPoletoPaulino2007GAB.pdf.

Examples

	#Example 9.1 of Paulino and Singer (2006)
	e91.TF<-c(3,25,32,68)
	e91.catdata<-readCatdata(TF=e91.TF)
	e91.U<-c(1,-1,-1,1)
	
	e91.X<-rbind(c(0,0), c(0,1),
	             c(1,0), c(1,1))

	e91.X2<-rbind(c(0,0,0),
    	          c(0,1,0),
        	      c(1,0,0),
	              c(1,1,1))
	
	e91.loglinml1<-loglinML(e91.catdata,U=e91.U)
	e91.loglinml2<-loglinML(e91.catdata,X=e91.X)
	e91.loglinml3<-loglinML(e91.catdata,X=e91.X2)
	e91.loglinml4<-loglinML(e91.catdata,A=c(1,-1,-1,1),XL=1)

	# Independence ordinary log-linear model, constraint formulation
	e91.loglinml1 
	# Independence ordinary log-linear model, freedom equation formulation
	e91.loglinml2 
	#Saturated ordinary log-linear model, freedom equation formulation
	e91.loglinml3 

	#Saturated generalized log-linear model, freedom equation formulation
	e91.loglinml4 
	
	#95% confidence interval for log-odds ratio and for odds ratio
	round(e91.loglinml4$beta+c(-1,1)*qnorm(0.975)*sqrt(e91.loglinml4$Vbeta),3)
	round(exp(e91.loglinml4$beta),3)
	round(exp(e91.loglinml4$beta+c(-1,1)*qnorm(0.975)*sqrt(e91.loglinml4$Vbeta)),3)

	#Example 1 of Poleto et al (2012)
	smoking.TF<-rbind(c(167,17,19,10,1,3,52,10,11, 176,24,121, 28,10,12),
	                  c(120,22,19, 8,5,1,39,12,12, 103, 3, 80, 31, 8,14))
	
	smoking.Zp<-t(rep(1,2))%x%cbind(diag(3)%x%rep(1,3), rep(1,3)%x%diag(3))
	
	smoking.Rp<-rbind(c(3,3),c(3,3))
	
	smoking.catdata<-readCatdata(TF=smoking.TF,Zp=smoking.Zp,Rp=smoking.Rp)
	smoking.catdata #Proportions of the complete data
	
	smoking.satmarml<-satMarML(smoking.catdata)
	smoking.satmcarml<-satMarML(smoking.catdata,missing="MCAR")
	
	smoking.E<-rbind(c(1,-1,0),c(0,1,-1))
	
	smoking.A<-diag(2)%x%smoking.E%x%smoking.E
	smoking.loglin2.marml<-loglinML(smoking.satmarml,A=smoking.A,XL=rep(1,8))
	smoking.loglin2.mcarml<-loglinML(smoking.satmcarml,A=smoking.A,XL=rep(1,8))