MultiVarMI (version 1.0)

MI: Bayesian Multiple Imputation for Multivariate Data

Description

This function implements the multiple imputation framework as described in Demirtas (2017) "A multiple imputation framework for massive multivariate data of different variable types: A Monte-Carlo technique."

Usage

MI(dat, var.types, m)

Arguments

dat

A data frame containing multivariate data with missing values.

var.types

The variable type corresponding to each column in dat, taking values of "NCT" for continuous data, "O" for ordinal or binary data, or "C" for count data.

m

The number of stochastic simulations in which the missing values are replaced.

Value

A list containing m imputed data sets.

References

Demirtas, H. and Hedeker, D. (2011). A practical way for computing approximate lower and upper correlation bounds. The American Statistician, 65(2), 104-109.

Demirtas, H., Hedeker, D., and Mermelstein, R. J. (2012). Simulation of massive public health data by power polynomials. Statistics in Medicine, 31(27), 3337-3346.

Demirtas, H. (2016). A note on the relationship between the phi coefficient and the tetrachoric correlation under nonnormal underlying distributions. The American Statistician, 70(2), 143-148.

Demirtas, H. and Hedeker, D. (2016). Computing the point-biserial correlation under any underlying continuous distribution. Communications in Statistics-Simulation and Computation, 45(8), 2744-2751.

Demirtas, H., Ahmadian, R., Atis, S., Can, F.E., and Ercan, I. (2016). A nonnormal look at polychoric correlations: modeling the change in correlations before and after discretization. Computational Statistics, 31(4), 1385-1401.

Demirtas, H. (2017). A multiple imputation framework for massive multivariate data of different variable types: A Monte-Carlo technique. Monte-Carlo Simulation-Based Statistical Modeling, edited by Ding-Geng (Din) Chen and John Dean Chen, Springer, 143-162.

Ferrari, P.A. and Barbiero, A. (2012). Simulating ordinal data. Multivariate Behavioral Research, 47(4), 566-589.

Fleishman A.I. (1978). A method for simulating non-normal distributions. Psychometrika, 43(4), 521-532.

Vale, C.D. and Maurelli, V.A. (1983). Simulating multivariate nonnormal distributions. Psychometrika, 48(3), 465-471.

See Also

MVN.corr, MVN.dat, trMVN.dat

Examples

# NOT RUN {
library(PoisBinOrdNonNor)
set.seed(1234)
n<-1e5
lambdas<-list(1, 3) #2 count variables
mps<-list(c(.2, .8), c(.6, 0, .3, .1)) #1 binary variable, 1 ordinal variable with skip pattern
moms<-list(c(-1, 1, 0, 1), c(0, 3, 0, 2)) #2 continuous variables

################################################
#Generate Poisson, Ordinal, and Continuous Data#
################################################
#get intermediate correlation matrix
cmat.star <- find.cor.mat.star(cor.mat = .8 * diag(6) + .2, #all pairwise correlations set to 0.2
                               no.pois = length(lambdas),
                               no.ord = length(mps),
                               no.nonn = length(moms),
                               pois.list = lambdas,
                               ord.list = mps,
                               nonn.list = moms)

#generate dataset
mydata <- genPBONN(n,
                   no.pois = length(lambdas),
                   no.ord = length(mps),
                   no.nonn = length(moms),
                   cmat.star = cmat.star,
                   pois.list = lambdas,
                   ord.list = mps,
                   nonn.list = moms)

cor(mydata)
apply(mydata, 2, mean)

#Make 10 percent of each variable missing completely at random
mydata<-apply(mydata, 2, function(x) {
  x[sample(1:n, size=n*0.1)]<-NA
  return(x) 
  }
)


#Create 5 imputed datasets
mydata<-data.frame(mydata)
mymidata<-MI(dat=mydata,
             var.types=c('C', 'C', 'O', 'O', 'NCT', 'NCT'),
             m=5)

#get the means of each variable for the m imputed datasets  
do.call(rbind, lapply(mymidata, function(x) apply(x, 2, mean)))

#get m correlation matrices of for the m imputed dataset
lapply(mymidata, function(x) cor(x))

#Look at the second imputed dataset
head(mymidata$dataset2)

##run a linear model on each dataset and extract coefficients
mycoef<-lapply(mymidata, function(x) {
  fit<-lm(X6~., data=data.frame(x))
  fit.coef<-coef(fit)
  return(fit.coef)
})

do.call(rbind, mycoef)
# }