Imputation of multivariate panel or cluster data
Gibbs sampler for the multivariate linear mixed model with incomplete data described by Schafer (1997). This function will typically be used to produce multiple imputations of missing data values in multivariate panel data or clustered data. The underlying model is
yi = Xi%*%beta + Zi%*%bi + ei, i=1,...,m,
where
yi = (ni x r) matrix of incomplete multivariate data for subject or cluster i;
Xi = (ni x p) matrix of covariates;
Zi = (ni x q) matrix of covariates;
beta = (p x r) matrix of coefficients common to the population (fixed effects);
bi = (q x r) matrix of coefficients specific to subject or cluster i (random effects); and
ei = (ni x r) matrix of residual errors.
The matrix bi, when stacked into a single column, is assumed to be normally distributed with mean zero and unstructured covariance matrix psi, and the rows of ei are assumed to be independently normal with mean zero and unstructured covariance matrix sigma. Missing values may appear in yi in any pattern.
In most applications of this model, the first columns of Xi and Zi will be constant (one) and Zi will contain a subset of the columns of Xi.
pan(y, subj, pred, xcol, zcol, prior, seed, iter=1, start)
y |
matrix of responses. This is simply the individual yi matrices stacked upon one another. Each column of y corresponds to a response variable. Each row of y corresponds to a single subject-occasion, or to a single subject within a cluster. Missing values (NA) may occur in any pattern. |
subj |
vector of length nrow(y) giving the subject (or cluster) indicators i for the rows of y. For example, suppose that y is in fact rbind(y1,y2,y3,y4) where nrow(y1)=2, nrow(y2)=3, nrow(y3)=2, and nrow(y4)=7. Then subj should be c(1,1,2,2,2,3,3,4,4,4,4,4,4,4). |
pred |
matrix of covariates used to predict y. This should have the same number of rows as y. The first column will typically be constant (one), and the remaining columns correspond to other variables appearing in Xi and Zi. |
xcol |
vector of integers indicating which columns of pred will be used in Xi. That is, pred[,xcol] is the Xi matrices (stacked upon one another). |
zcol |
vector of integers indicating which columns of pred will be used in Zi. That is, pred[,zcol] is the Zi matrices (stacked upon one another). |
prior |
a list with four components (whose names are a, Binv, c, and Dinv, respectively) specifying the hyperparameters of the prior distributions for psi and sigma. For information on how to specify and interpret these hyperparameters, see Schafer (1997) and the example below. Note: This is a slight departure from the notation in Schafer (1997), where a and Binv were denoted by "nu1" and "Lambdainv1", and c and Dinv were "nu2" and "Lambdainv2". |
seed |
integer seed for initializing pan()'s internal random number generator. This argument should be a positive integer. |
iter |
total number of iterations or cycles of the Gibbs sampler to be carried out. |
start |
optional list of quantities to specify the initial state of the Gibbs sampler. This list has the same form as "last" (described below), one of the components returned by pan(). This argument allows the Gibbs sampler to be restarted from the final state of a previous run. If "start" is omitted then pan() chooses its own initial state. |
The Gibbs sampler algorithm used in pan() is described in detail by Schafer (1997).
A list containing the following components. Note that when you are using pan() to produce multiple imputations, you will be primarily interested in the component "y" which contains the imputed data; the arrays "beta", "sigma", and "psi" will be used primarily for diagnostics (e.g. time-series plots) to assess the convergence behavior of the Gibbs sampler.
beta |
array of dimension c(length(xcol),ncol(y),iter) = (p x r x number of Gibbs cycles) containing the simulated values of beta from all cycles. That is, beta[,,T] is the (p x r) matrix of simulated fixed effects at cycle T. |
sigma |
array of dimension c(ncol(y),ncol(y),iter) = (r x r x number of Gibbs cycles) containing the simulated values of sigma from all cycles. That is, sigma[,,T] is the simulated version of the model's sigma at cycle T. |
psi |
array of dimension c(length(zcol)*ncol(y), length(zcol)*ncol(y), iter) = (q*r x q*r x number of Gibbs cycles) containing the simulated values of psi from all cycles. That is, psi[,,T] is the simulated version of the model's psi at cycle T. |
y |
matrix of imputed data from the final cycle of the Gibbs sampler. Identical to the input argument y except that the missing values (NA) have been replaced by imputed values. If "iter" has been set large enough (which can be determined by examining time-series plots, etc. of "beta", "sigma", and "psi") then this is a proper draw from the posterior predictive distribution of the complete data. |
last |
a list of four components characterizing the final state of the Gibbs sampler. The four components are: "beta", "sigma", "psi", and "y", which are the simulated values of the corresponding model quantities from the final cycle of Gibbs. This information is already contained in the other components returned by pan(); we are providing this list merely as a convenience, to allow the user to start future runs of the Gibbs sampler at this state. |
This function assumes that the rows of y (and thus the rows of subj and pred) have been sorted by subject number. That is, we assume that subj=sort(subj), y=y[order(subj),], and pred=pred[order(subj),]. If the matrix y is created by stacking yi, i=1,...,m then this will automatically be the case.
Schafer JL (1997) Imputation of missing covariates under a multivariate linear mixed model. Technical report 97-04, Dept. of Statistics, The Pennsylvania State University,
Schafer JL (2001). Multiple imputation with PAN. Chapter 12, pp357-77. of New Methods for the Analysis of Change. Edited by Collins LM, Sayer AG. American Psychological Association, Washington DC.
Schafer JL, Yucel RM (2002). Computational strategies for multivariate linear mixed-effects models with missing values. Journal of Computational and Graphical Statistics. 11:437-457
########################################################################
# This example is somewhat atypical because the data consist of a
# single response variable (change in heart rate) measured repeatedly;
# most uses of pan() will involve r > 1 response variables. If we had
# r response variables rather than one, the only difference would be
# that the vector y below would become a matrix with r columns, one
# for each response variable. The dimensions of Sigma (the residual
# covariance matrix for the response) and Psi (the covariance matrix
# for the random effects) would also change to (r x r) and (r*q x r*q),
# respectively, where q is the number of random coefficients in the
# model (in this case q=1 because we have only random intercepts). The
# new dimensions for Sigma and Psi will be reflected in the prior
# distribution, as Dinv and Binv become (r x r) and (r*q x r*q).
#
# The pred matrix has the same number of rows as y, the number of
# subject-occasions. Each column of Xi and Zi must be represented in
# pred. Because Zi is merely the first column of Xi, we do not need to
# enter that column twice. So pred is simply the matrix Xi, stacked
# upon itself nine times.
#
data(marijuana)
attach(marijuana)
pred <- with(marijuana,cbind(int,dummy1,dummy2,dummy3,dummy4,dummy5))
#
# Now we must tell pan that all six columns of pred are to be used in
# Xi, but only the first column of pred appears in Zi.
#
xcol <- 1:6
zcol <- 1
########################################################################
# The model specification is now complete. The only task that remains
# is to specify the prior distributions for the covariance matrices
# Sigma and Psi.
#
# Recall that the dimension of Sigma is (r x r) where r
# is the number of response variables (in this case, r=1). The prior
# distribution for Sigma is inverted Wishart with hyperparameters a
# (scalar) and Binv (r x r), where a is the imaginary degrees of freedom
# and Binv/a is the prior guesstimate of Sigma. The value of a must be
# greater than or equal to r. The "least informative" prior possible
# would have a=r, so here we will take a=1. As a prior guesstimate of
# Sigma we will use the (r x r) identity matrix, so Binv = 1*1 = 1.
#
# By similar reasoning we choose the prior distribution for Psi. The
# dimension of Psi is (r*q x r*q) where q is the number of random
# effects in the model (i.e. the length of zcol, which in this case is
# one). The hyperparameters for Psi are c and Dinv, where c is the
# imaginary degrees of freedom (which must be greater than or equal to
# r*q) and Dinv/c is the prior guesstimate of Psi. We will take c=1
# and Dinv=1*1 = 1.
#
# The prior is specified as a list with four components named a, Binv,
# c, and Dinv, respectively.
#
prior <- list(a=1,Binv=1,c=1,Dinv=1)
########################################################################
# Now we are ready to run pan(). Let's assume that the pan function
# and the object code have already been loaded into R. First we
# do a preliminary run of 1000 iterations.
#
result <- pan(y,subj,pred,xcol,zcol,prior,seed=13579,iter=1000)
#
# Check the convergence behavior by making time-series plots and acfs
# for the model parameters. Variances will be plotted on a log
# scale. We'll assume that a graphics device has already been opened.
#
plot(1:1000,log(result$sigma[1,1,]),type="l")
acf(log(result$sigma[1,1,]))
plot(1:1000,log(result$psi[1,1,]),type="l")
acf(log(result$psi[1,1,]))
par(mfrow=c(3,2))
for(i in 1:6) plot(1:1000,result$beta[i,1,],type="l")
for(i in 1:6) acf(result$beta[i,1,])
#
# This example appears to converge very rapidly; the only appreciable
# autocorrelations are found in Psi, and even those die down by lag
# 10. With a sample this small we can afford to be cautious, so let's
# impute the missing data m=10 times taking 100 steps between
# imputations. We'll use the current simulated value of y as the first
# imputation, then restart the chain where we left off to produce
# the second through the tenth.
#
y1 <- result$y
result <- pan(y,subj,pred,xcol,zcol,prior,seed=9565,iter=100,start=result$last)
y2 <- result$y
result <- pan(y,subj,pred,xcol,zcol,prior,seed=6047,iter=100,start=result$last)
y3 <- result$y
result <- pan(y,subj,pred,xcol,zcol,prior,seed=3955,iter=100,start=result$last)
y4 <- result$y
result <- pan(y,subj,pred,xcol,zcol,prior,seed=4761,iter=100,start=result$last)
y5 <- result$y
result <- pan(y,subj,pred,xcol,zcol,prior,seed=9188,iter=100,start=result$last)
y6 <- result$y
result <- pan(y,subj,pred,xcol,zcol,prior,seed=9029,iter=100,start=result$last)
y7 <- result$y
result <- pan(y,subj,pred,xcol,zcol,prior,seed=4343,iter=100,start=result$last)
y8 <- result$y
result <- pan(y,subj,pred,xcol,zcol,prior,seed=2372,iter=100,start=result$last)
y9 <- result$y
result <- pan(y,subj,pred,xcol,zcol,prior,seed=7081,iter=100,start=result$last)
y10 <- result$y
########################################################################
# Now we combine the imputation results according to mitools
########################################################################
# First, we build data frames from above,
d1 <- data.frame(y=y1,subj,pred)
d2 <- data.frame(y=y2,subj,pred)
d3 <- data.frame(y=y3,subj,pred)
d4 <- data.frame(y=y4,subj,pred)
d5 <- data.frame(y=y5,subj,pred)
d6 <- data.frame(y=y6,subj,pred)
d7 <- data.frame(y=y7,subj,pred)
d8 <- data.frame(y=y8,subj,pred)
d9 <- data.frame(y=y9,subj,pred)
d10 <- data.frame(y=y10,subj,pred)
# Second, we establish a S3 object as needed for the function MIcombine
# nevertheless we start with an ordinary least squares regression
require(mitools)
d <- imputationList(list(d1,d2,d3,d4,d5,d6,d7,d8,d9,d10))
w <- with(d,lm(y~-1+pred))
MIcombine(w)
# Now, we can turn to lmer as in lme4 package but in this case it is the
# same.
if(require(lme4)) {
w2 <- with(d,lmer(y~-1+pred+(1|subj)))
b <- MIextract(w2,fun=fixef)
Var <- function(obj) unlist(lapply(diag(vcov(obj)),function(m) m))
v <- MIextract(w2,fun=Var)
MIcombine(b,v)
detach(marijuana)
}
### bivariate example
data(bitest)
attach(bitest)
y <- with(bitest,cbind(y1,y2))
subj <- c(clusterid)
pred <- cbind (int, x1, x2, x3)
xcol <- 1:4
zcol <- 1
a <- 2
c <- 2
id2 <- matrix(c(1,0,0,1),ncol=2,nrow=2)
Binv <- a*id2
Dinv <- c*id2
prior <- list(a=a, Binv=Binv, c=c, Dinv=Dinv)
result <- pan(y, subj, pred, xcol, zcol, prior, seed=12345, iter=1000)Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.