JM Imputation of clustered data with mixed variable types

Share:

Description

Impute a clustered dataset with mixed data types as outcome. A joint multivariate model for partially observed data is assumed and imputations are generated through the use of a Gibbs sampler where the covariance matrix is updated with a Metropolis-Hastings step. Fully observed categorical covariates may be considered as covariates as well, but they have to be included as dummy variables.

Usage

1
2
3
4
jomo1ranmix(Y.con, Y.cat, Y.numcat, X=NULL, Z=NULL, clus, 
beta.start=NULL, u.start=NULL, l1cov.start=NULL, l2cov.start=NULL, 
l1cov.prior=NULL, l2cov.prior=NULL, nburn=1000, nbetween=1000, nimp=5, 
output=1, out.iter=10) 

Arguments

Y.con

A data frame, or matrix, with continuous responses of the joint imputation model. Rows correspond to different observations, while columns are different variables. Missing values are coded as NA.

Y.cat

A data frame, or matrix, with categorical (or binary) responses of the joint imputation model. Rows correspond to different observations, while columns are different variables. Categories must be integer numbers from 1 to N. Missing values are coded as NA.

Y.numcat

A vector with the number of categories in each categorical (or binary) variable.

X

A data frame, or matrix, with covariates of the joint imputation model. Rows correspond to different observations, while columns are different variables. Missing values are not allowed in these variables. In case we want an intercept, a column of 1 is needed. The default is a column of 1.

Z

A data frame, or matrix, for covariates associated to random effects in the joint imputation model. Rows correspond to different observations, while columns are different variables. Missing values are not allowed in these variables. In case we want an intercept, a column of 1 is needed. The default is a column of 1.

clus

A data frame, or matrix, containing the cluster indicator for each observation.

beta.start

Starting value for beta, the vector(s) of fixed effects. Rows index different covariates and columns index different outcomes. For each n-category variable we define n-1 latent normals. The default is a matrix of zeros.

u.start

A matrix where different rows are the starting values within each cluster for the random effects estimates u. The default is a matrix of zeros.

l1cov.start

Starting value for the covariance matrix. Dimension of this square matrix is equal to the number of outcomes (continuous plus latent normals) in the imputation model. The default is the identity matrix.

l2cov.start

Starting value for the level 2 covariance matrix. Dimension of this square matrix is equal to the number of outcomes (continuous plus latent normals) in the imputation model times the number of random effects. The default is an identity matrix.

l1cov.prior

Scale matrix for the inverse-Wishart prior for the covariance matrix. The default is the identity matrix.

l2cov.prior

Scale matrix for the inverse-Wishart prior for the level 2 covariance matrix. The default is the identity matrix.

nburn

Number of burn in iterations. Default is 1000.

nbetween

Number of iterations between two successive imputations. Default is 1000.

nimp

Number of Imputations. Default is 5.

output

When set to any value different from 1 (default), no output is shown on screen at the end of the process.

out.iter

When set to K, every K iterations a message "Iteration number N*K completed" is printed on screen. Default is 10.

Details

TThe Gibbs sampler algorithm used is described in detail in Chapter 9 of Carpenter and Kenward (2013). Regarding the choice of the priors, a flat prior is considered for beta and for the covariance matrix. A Metropolis Hastings step is implemented to update the covariance matrix, as described in the book. Binary or continuous covariates in the imputation model may be considered without any problem, but when considering a categorical covariate it has to be included with dummy variables (binary indicators) only.

Value

On screen, the posterior mean of the fixed effects estimates and of the covariance matrix are shown. The only argument returned is the imputed dataset in long format. Column "Imputation" indexes the imputations. Imputation number 0 are the original data.

References

Carpenter J.R., Kenward M.G., (2013), Multiple Imputation and its Application. Chapter 9, Wiley, ISBN: 978-0-470-74052-1.

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
#First of all we load and attach the data:

data(cldata)
attach(cldata)

#Then we define the inputs:
# nimp, nburn and nbetween are smaller than they should. This is
#just because of CRAN policies on the examples.

Y.con=data.frame(measure,age)
Y.cat=data.frame(social)
Y.numcat=matrix(4,1,1)
X=data.frame(rep(1,1000),sex)
Z<-data.frame(rep(1,1000))
clus<-data.frame(city)
beta.start<-matrix(0,2,5)
u.start<-matrix(0,10,5)
l1cov.start<-diag(1,5)
l2cov.start<-diag(1,5)
l1cov.prior=diag(1,5);
l2cov.prior=diag(1,5);
nburn=as.integer(100);
nbetween=as.integer(100);
nimp=as.integer(5);

#Then we can run the sampler:

imp<-jomo1ranmix(Y.con, Y.cat, Y.numcat, X,Z,clus,beta.start,u.start,l1cov.start, 
          l2cov.start,l1cov.prior,l2cov.prior,nburn,nbetween,nimp)

cat("Original value was missing (",imp[4,1],"), imputed value:", imp[1004,1])

# We run our substantive model on the 5 imputed datasets:

estimates<-matrix(0,5,5)
ses<-matrix(0,5,5)
for (i in 1:5) {
  dat<-imp[imp$Imputation==i,]
  fit<-lm(measure~age+sex+factor(social)+factor(clus),data=dat)
  estimates[i,1:5]<-coef(summary(fit))[2:6,1]
  ses[i,1:5]<-coef(summary(fit))[2:6,2]
}

# And finally we aggregate results with Rubin's rules, using BaBooN package:

#library("BaBooN")
#MI.inference(estimates[,1], ses[,1]^2)
#MI.inference(estimates[,2], ses[,2]^2)
#MI.inference(estimates[,3], ses[,3]^2)
#MI.inference(estimates[,4], ses[,4]^2)
#MI.inference(estimates[,5], ses[,5]^2)

Want to suggest features or report bugs for rdrr.io? Use the GitHub issue tracker.