Package BaBooN

July 2, 2014
Version 0.1-6
Date 2011-02-24
Title Bayesian Bootstrap Predictive Mean Matching - Multiple and single imputation for discrete data
Author Florian Meinfelder <[email protected]>
Maintainer Florian Meinfelder <[email protected]>
Depends R (>= 2.12.0), MASS, nnet
Description The package contains two variants of Bayesian Bootstrap
Predictive Mean Matching to multiply impute missing data. The
rst variant is a variable-by-variable imputation combining
sequential regression and Predictive Mean Matching (PMM) that
has been extended for unordered categorical data. The Bayesian
Bootstrap allows for generating approximately proper multiple
imputations. The second variant is also based on PMM, but the
focus is on imputing several variables at the same time. The
suggestion is to use this variant, if the missing-data pattern
resembles a data fusion situation, or any other
missing-by-design pattern, where several variables have
identical missing-data patterns. Both variants can be run as
'single imputation' versions, in case the analysis objective is of a purely descriptive nature.
License GPL (>= 2)
URL http://www.r-project.org
Repository CRAN
Date/Publication 2011-03-26 15:43:18
NeedsCompilation no
1
2 BBPMM
R topics documented:
BBPMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
BBPMM.row . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
MI.inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
rowimpPrep . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
summary.imp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
summary.impprep . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Index 14
BBPMM (Multiple) Imputation through Bayesian Bootstrap Predictive Mean
Matching (BBPMM)
Description
BBPMM performs single and multiple imputation (MI) of mixed-scale variables using a chained
equations approach and (Bayesian Bootstrap) Predictive Mean Matching.
Usage
BBPMM(data, M=10, nIter=10, outfile=NULL, ignore=NULL,
vartype=NULL, stepwise=TRUE, maxit=3, maxPerc = 0.98, verbose=TRUE, setSeed,...)
Arguments
data A partially incomplete data frame or matrix.
M Number of multiple imputations. If M=1, no Bayesian Bootstrap step is carried
out. Default=10.
nIter Number of iterations of the chained equations algorithm before the data set is
stored as an imputed data set. Default=10.
outfile A character string that species the path and le name for the imputed data sets.
If outfile=NULL (default), no data set is stored
ignore A character or numerical vector that species either column positions or variable
names that are to be excluded from the imputation model and process, e.g. an
ID variable. If ignore=NULL (default), all variables in data are used in the
imputation model.
vartype A character vector that ags the class of each variable in data (without the vari-
ables dened by the ignore argument), with either M for metric-scale or C
for categorical. The default (NULL) takes over the classes of data. Overrul-
ing these classes can sometimes make sense: e.g. an ordinal-scale variable is
originally classied as factor, but treating it as metric-scale variable within the
imputation process might still be a better choice (considering the robust proper-
ties of PMM to model misspecication).
stepwise Performs backwise selection for each imputation model based on the Schwarz
(Bayes) Information criterion. Default=TRUE.
BBPMM 3
maxit Imported argument from the nnet package that species the maximum number
of iterations for the multinomial logit model estimation. Default=3.
maxPerc The maximum percentage the mode category of a variable is allowed to have in
order to try regular imputation. If a variable is approximately Dirac distributed,
i.e. if it has (almost) no variance, imputation is carried out by simple hot deck
imputation. Default = 0.98.
verbose The algorithm prints information on imputation and iteration numbers. De-
fault=TRUE.
setSeed Optional argument to x the pseudo-random number generator in order to allow
for reproducable results.
... Further arguments passed to or from other functions.
Details
BBPMM is based on a chained equations approach that is using a Bayesian Bootstrap approach and
Predictive Mean Matching variants for metric-scale, binary, and multi-categorical variables to gen-
erate multiple imputations. In order to emulate a monotone missing-data pattern as well as possible,
variables are sorted by rate of missingness (in ascending order). If no complete variables exist, the
least incomplete variable is imputed via hot-deck. The starting solution then builds the imputation
model using the observed values of a particular y variable, and the corresponding observed or al-
ready imputed values of the x variables (i.e. all variables with fewer missing values than y). Due
to the PMM element in the algorithm, auto-correlation of subsequent iterations is virtually zero.
Therefore, a burn-in period is not required, and there is no need to administer high values (> 20)
to nIter either.
If M=1, no Bayesian Bootstrap step is carried out for the chained equations. Note that in this
case the algorithm is still unlikely to converge to a stable solution, because of the Predictive Mean
Matching step.
Value
impdata The imputed data set, if M=1, or a list containing M imputed data sets.
misOverview The percentage of missing values per incomplete variable.
indMatrix A matrix with the same dimensions as data minus ignore containing ags for
missing values.
M Number of (multiple) imputations.
nIter Number of iterations between two imputations.
References
Koller-Meinfelder, F. (2009) Analysis of Incomplete Survey Data Multiple Imputation Via Bayesian
Bootstrap Predictive Mean Matching, doctoral thesis.
See Also
BBPMM.row
4 BBPMM.row
Examples
### sample data set with non-normal variables
n <- 50
x1 <- round(runif(n,0.5,3.5))
x2 <- as.factor(c(rep(1,10),rep(2,25),rep(3,15)))
x3 <- round(rnorm(n,0,3))
y1 <- round(x1-0.25*(x2==2)+0.5*x3+rnorm(n,0,1))
y1 <- ifelse(y1<1,1,y1)
y1 <- as.factor(ifelse(y1>4,5,y1))
y2 <- x1+rnorm(n,0,0.5)
y3 <- round(x3+rnorm(n,0,2))
data1 <- as.data.frame(cbind(x1,x2,x3,y1,y2,y3))
misrow1 <- sample(n,20)
misrow2 <- sample(n,15)
misrow3 <- sample(n,10)
is.na(data1[misrow1, 4]) <- TRUE
is.na(data1[misrow2, 5]) <- TRUE
is.na(data1[misrow2, 6]) <- TRUE
### imputation
imputed.data <- BBPMM(data1, nIter=5, M=5)
BBPMM.row (Multiple) Imputation of variable vectors
Description
BBPMM.row performs single and multiple imputation (MI) of metric scale variable vectors. For
MI, parameter draws from a posterior distribution are replaced by a Bayesian Bootstrap step. Im-
putations are generated using Predictive Mean Matching (PMM) as described in Little (1988).
Usage
BBPMM.row(misDataPat, blockImp=length(misDataPat$blocks),
M=10, outfile=NULL, manWeights=NULL, stepwise=TRUE, verbose=TRUE,
tol=0.25, setSeed=NULL, ...)
Arguments
misDataPat An object created by rowimpPrep that contains information on all identied
missing-data patterns.
blockImp A scalar or vector containing the number(s) of the block(s) considered for im-
putation. Per default only the last block is imputed.
M Number of multiple imputations. If M=1, no Bayesian Bootstrap step is carried
out.
outfile A character string that species the path and le name for the imputed data sets.
If outfile=NULL (default), no data set is stored.
BBPMM.row 5
manWeights Optional argument containing manual (non-negative) weights for the PMM step.
manWeights can either be a list containing a vector for each missingness pattern,
or just a vector, if only one missingness pattern/block exists. In either case, the
number of elements in the vector(s) must match the number of variables in the
corresponding block. Note that the higher the weight the higher the importance
of a good match for the corresponding variables predictive means.
stepwise Performs backwise selection for each imputation model based on the Schwarz
(Bayes) Information criterion. Default=TRUE.
verbose The algorithm prints information on weighting matrices and imputation num-
bers. Default=TRUE.
tol Imported argument from function qr that species the tolerance level for linear
dependencies among the complete variables and defaults to 0.25.
setSeed Optional argument to x the pseudo-random number generator in order to allow
for reproducable results.
... Further arguments passed to or from other functions.
Details
The simultaneous imputation of several variables is useful for missing-by-design patterns, such
as data fusion or split questionnaire designs. The predictive means of the imputation variables
are weighted by the inverse of the covariance matrix of the residuals from the regression of these
variables on the complete variables. The intuitive idea behind is that distances between predictive
means should be punished more severely, if the particular variable can be explained well by the
(completely observed) imputation model variables. Through partialization and subsequent usage of
the residuals the weight matrix is transformed into a diagonal matrix. The calculated weights can
be adjusted by manual weights. Since the weight matrix is a Mahalanobis type of distance matrix,
the weights are in the denominator and therefore the lower the weight, the higher the inuence.
As this is somewhat counterintuitive, the reciprocal of the manual weights is taken. Therefore,
the higher the manual weight the higher in the inuence of the corresponding variables predictor
on the overall distance. The donor/recipient ID pairlist for each imputation and identied pattern
(block) is stored. In general, weightMatrix, model and pairlist are list objects named M1 to
M<M>, and each in return is a list object named block1 to block<length(blockImp)>. model
contains another list object with lm-objects for all variables in a particular block. Unlike BBPMM
this algorithm is not based on sequential regression. Therefore, imputed variables are conditionally
independent given the completely observed variables (of which at least one must exist).
Value
impdata A list containing M completed data sets.
weightMatrix A list containing weight matrices for all imputations and blocks.
model A list containing the lm-objects for all imputations and blocks.
pairlist A list containing the donor/recipient pairlist data frames for all imputations and
blocks.
indMatrix A matrix with the same dimensions as data containing ags for missing values.
6 MI.inference
References
Little, R.J.A. (1988) Missing-Data Adjustments in Large Surveys, Journal of Business and Eco-
nomic Statistics, Vol.6, No.3,pp. 287-296.
Koller-Meinfelder, F. (2009) Analysis of Incomplete Survey Data Multiple Imputation Via Bayesian
Bootstrap Predictive Mean Matching, doctoral thesis.
See Also
rowimpPrep, BBPMM
Examples
### sample data set with non-normal variables and a single
### missingness pattern
set.seed(1000)
n <- 50
x1 <- round(runif(n,0.5,3.5))
x2 <- as.factor(c(rep(1,10),rep(2,25),rep(3,15)))
x3 <- round(rnorm(n,0,3))
y1 <- round(x1-0.25*(x2==2)+0.5*x3+rnorm(n,0,1))
y1 <- ifelse(y1<1,1,y1)
y1 <- ifelse(y1>4,5,y1)
y2 <- y1+rnorm(n,0,0.5)
y3 <- round(x3+rnorm(n,0,2))
data <- as.data.frame(cbind(x1,x2,x3,y1,y2,y3))
misrow1 <- sample(n,20)
data[misrow1, c(4:6)] <- NA
### preparation step
impblock <- rowimpPrep(data)
### imputation
imputed.data <- BBPMM.row(impblock, M=5)
MI.inference Multiple Imputation inference
Description
MI.inference applies Rubins combining rules to estimated quantities of interest that are based on
multiply imputed data sets. The function requires as input two vectors of length M for the estimate
and its variance.
Usage
MI.inference(thetahat, varhat.thetahat, alpha=0.05)
MI.inference 7
Arguments
thetahat A vector of length M containing estimates of the quantity of interest based on
multiply imputed data sets.
varhat.thetahat
A vector of length M containing the corresponding variances of thetahat.
alpha The signicance level at which lower and upper bound are calculated. DE-
FAULT=0.05
Details
Multiple Imputation (Rubin, 1987) of missing data is a generally accepted way to get correct vari-
ance estimates for a particular quantity of interest in the presence of missing data. MI.inference
estimates the within variance W and between variance B, and combines them to the total variance
T. Based on the output, further analysis gures, such as the fraction of missing information can be
calculated.
Value
MI.Est A scalar containing the MI estimate of the quantity of interest (i.e. an estimator
averaged over all M data sets).
MI.Var The Multiple Imputation variance.
CI.low The lower bound of the MI condence interval.
CI.up The upper bound of the MI condence interval.
BVar The estimated between variance.
WVar The estimated within variance.
References
Rubin, D.B. (1987). Multiple Imputation for Non-Response in Surveys. New York: John Wiley &
Sons, Inc.
Examples
### example 1
n <- 100
x1 <- round(runif(n,0.5,3.5))
x2 <- round(runif(n,0.5,4.5))
x3 <- runif(n,1,6)
y1 <- round(x1-0.25*x2+0.5*x3+rnorm(n,0,1))
y1 <- ifelse(y1<2,2,y1)
y1 <- as.factor(ifelse(y1>4,5,y1))
y2 <- x3+rnorm(n,0,2)
y3 <- as.factor(ifelse(x2+rnorm(n,0,2)>2,1,0))
mis1 <- sample(100,20)
mis2 <- sample(100,30)
mis3 <- sample(100,25)
data1 <- data.frame("x1"=x1,"x2"=x2,"x3"=x3,"y1"=y1,"y2"=y2,"y3"=y3)
8 MI.inference
is.na(data1$y1[mis1]) <- TRUE
is.na(data1$y2[mis2]) <- TRUE
is.na(data1$y3[mis3]) <- TRUE
imputed.data <- BBPMM(data1, M=10, nIter=10)
MI.m.meany2.hat <- sapply(imputed.data$impdata,FUN=function(x) mean(x$y2))
MI.v.meany2.hat <- sapply(imputed.data$impdata,FUN=function(x) var(x$y2)/length(x$y2))
### MI inference
MI.y2 <- MI.inference(MI.m.meany2.hat, MI.v.meany2.hat,alpha=0.05)
MI.y2$MI.Est
MI.y2$MI.Var
## Not run:
################################################################
### example 2: a small simulation example
### simple additional function to calculate coverages: #
coverage <- function(value, bounds) {
ifelse(min(bounds) <= value && max(bounds) >= value, 1, 0)
}
### value : true value #
### bounds : vector with two elements (upper and #
### lower bound of the CI) #
### sample size
n <- 100
### true value for the mean of y2
m.y2 <- 3.5
y2.cover <- vector(length=n)
set.seed(1000)
### 100 data generations
time1 <- Sys.time()
for (i in 1:100) {
x1 <- round(runif(n,0.5,3.5))
x2 <- round(runif(n,0.5,4.5))
x3 <- runif(n,1,6)
y1 <- round(x1-0.25*x2+0.5*x3+rnorm(n,0,1))
y1 <- ifelse(y1<2,2,y1)
y1 <- as.factor(ifelse(y1>4,5,y1))
y2 <- x3+rnorm(n,0,2)
y3 <- as.factor(ifelse(x2+rnorm(n,0,2)>2,1,0))
mis1 <- sample(n,20)
mis2 <- sample(n,30)
mis3 <- sample(n,25)
data1 <- data.frame("x1"=x1,"x2"=x2,"x3"=x3,"y1"=y1,"y2"=y2,"y3"=y3)
is.na(data1$y1[mis1]) <- TRUE
is.na(data1$y2[mis2]) <- TRUE
is.na(data1$y3[mis3]) <- TRUE
rowimpPrep 9
sim.imp <- BBPMM(data1, M=3, n.iter=2,stepwise=FALSE, verbose=FALSE)
MI.m.meany2.hat <- sapply(sim.imp$impdata,FUN=function(x) mean(x$y2))
MI.v.meany2.hat <- sapply(sim.imp$impdata,FUN=function(x) var(x$y2)/length(x$y2))
### MI inference
MI.y2 <- MI.inference(MI.m.meany2.hat, MI.v.meany2.hat,alpha=0.05)
y2.cover[i] <- coverage(m.y2, c(MI.y2$CI.low,MI.y2$CI.up))
}
time2 <- Sys.time()
difftime(time2, time1, unit="secs")
### coverage estimator (alpha=0.05):
mean(y2.cover)
## End(Not run)
rowimpPrep Missing-data pattern identier
Description
rowimpPrep identies all missingness patterns within an incomplete data set. Running rowimpPrep
is a prerequisite for BBPMM.row.
Usage
rowimpPrep(data, ID=NULL, verbose=TRUE)
Arguments
data Either a data frame or matrix with missing values.
ID A numeric or character string vector indicating the column positions or names
of the ID variable (if two data sets were stacked that have a joint subset of vari-
ables). The rst element refers to the donor ID, the second element refers to the
recipient ID. This disticntion is only of relevance, if the data set is L-shaped,
i.e. if the data contains only one missing-data pattern (where incomplete cases
are recipients). If ID has only one element, The function assumes that the
identier variables of the two data sets are packed into a single variable. De-
fault=NULL is used, if no ID variable is specied.
verbose Prints information on identied missing-data patterns. Default=TRUE.
Details
rowimpPrep identies all patterns, and allows to decide, whether to impute all missing-data patterns
with BBPMM.rowor just some of them. This comes in handy if variables that were assumed to be
completely observed have missing values. These variables are then likely to dene an unexpected
block of their own. Of course, BBPMM.row can be used to impute missing data that are not missing-
by-design as well, but BBPMM would probably be the better option. Note that all variables listed in
10 summary.imp
compNames are used for the imputation model in BBPMM.row, i.e. completely observed variables (ID
variables aside) which are not to be used in the imputation model, have to be removed from the data
set beforehand.
Value
data The original data set minus the ID variable(s).
key The ID variable(s) from the original data set.
blocks A list containing the column positions of all identied missing-data patterns.
blockNames A list containing the variable names corresponding to object blocks.
compNames A character vector containing the variable names of the (completely observed)
imputation model variables.
Examples
### sample data set with non-normal variables and a single
### missingness pattern
set.seed(1000)
n <- 50
x1 <- round(runif(n,0.5,3.5))
x2 <- as.factor(c(rep(1,10),rep(2,25),rep(3,15)))
x3 <- round(rnorm(n,0,3))
y1 <- round(x1-0.25*(x2==2)+0.5*x3+rnorm(n,0,1))
y1 <- ifelse(y1<1,1,y1)
y1 <- ifelse(y1>4,5,y1)
y2 <- y1+rnorm(n,0,0.5)
y3 <- round(x3+rnorm(n,0,2))
data1 <- as.data.frame(cbind(x1,x2,x3,y1,y2,y3))
misrow1 <- sample(n,20)
is.na(data1[misrow1, c(4:6)]) <- TRUE
### preparation step
impblock <- rowimpPrep(data1)
impblock$blockNames
summary.imp Summary methd for objects of class imp
Description
Returns some information about the incomplete data set and the imputation process.
summary.imp 11
Usage
## S3 method for class imp
summary(object,...)
Arguments
object Either with BBPMM or BBPMM.row generated object.
... Arguments to be passed to or from other functions.
Details
Returns information about the percentage of missing data as well as about the imputation variant,
the number of (multiple) imputations and the number of iterations between two imputations.
Author(s)
Florian Meinfelder
See Also
BBPMM, BBPMM.row
Examples
### sample data set with non-normal variables and two different
### missingness patterns
n <- 50
x1 <- round(runif(n,0.5,3.5))
x2 <- as.factor(c(rep(1,10),rep(2,25),rep(3,15)))
x3 <- round(rnorm(n,0,3))
y1 <- round(x1-0.25*(x2==2)+0.5*x3+rnorm(n,0,1))
y1 <- ifelse(y1<1,1,y1)
y1 <- as.factor(ifelse(y1>4,5,y1))
y2 <- x1+rnorm(n,0,0.5)
y3 <- round(x3+rnorm(n,0,2))
data1 <- as.data.frame(cbind(x1,x2,x3,y1,y2,y3))
misrow1 <- sample(n,20)
misrow2 <- sample(n,15)
misrow3 <- sample(n,10)
is.na(data1[misrow1, 4]) <- TRUE
is.na(data1[misrow2, 5]) <- TRUE
is.na(data1[misrow2, 6]) <- TRUE
### imputation
imputed.data <- BBPMM(data1, nIter=5, M=5)
summary(imputed.data)
12 summary.impprep
summary.impprep Summary methd for objects of class impprep
Description
Returns an overview of missing-data patterns, in particular of missing-by-design patterns.
Usage
## S3 method for class impprep
summary(object, nNames=10L,...)
Arguments
object An object generated by rowimpPrep.
nNames Number of variable names per block to be printed (Default = 10).
... Arguments to be passed to or from other functions.
Details
Returns the number of identied missing-data patterns, the rst nNames variable names per block
and the names of the completely observed variables.
Author(s)
Florian Meinfelder
See Also
rowimpPrep
Examples
### sample data set with non-normal variables and a single
### missingness pattern
set.seed(1000)
n <- 50
x1 <- round(runif(n,0.5,3.5))
x2 <- as.factor(c(rep(1,10),rep(2,25),rep(3,15)))
x3 <- round(rnorm(n,0,3))
y1 <- round(x1-0.25*(x2==2)+0.5*x3+rnorm(n,0,1))
y1 <- ifelse(y1<1,1,y1)
y1 <- ifelse(y1>4,5,y1)
y2 <- y1+rnorm(n,0,0.5)
y3 <- round(x3+rnorm(n,0,2))
data1 <- as.data.frame(cbind(x1,x2,x3,y1,y2,y3))
misrow1 <- sample(n,20)
summary.impprep 13
is.na(data1[misrow1, c(4:6)]) <- TRUE
### preparation step
impblock <- rowimpPrep(data1)
summary(impblock)
Index
Topic datagen
BBPMM, 2
BBPMM.row, 4
Topic robust
BBPMM, 2
BBPMM.row, 4
BBPMM, 2, 6, 9, 11
BBPMM.row, 3, 4, 9, 11
MI.inference, 6
rowimpPrep, 6, 9, 12
summary.imp, 10
summary.impprep, 12
14