JSS Journal of Statistical Software

December 2011, Volume 45, Issue 2. http://www.jstatsoft.org/

Multiple Imputation with Diagnostics (mi) in R:

Opening Windows into the Black Box

Yu-Sung Su Andrew Gelman

Tsinghua University Columbia University

Jennifer Hill Masanao Yajima

New York University University of California, Los Angeles

Abstract
Our mi package in R has several features that allow the user to get inside the impu-
tation process and evaluate the reasonableness of the resulting models and imputations.
These features include: choice of predictors, models, and transformations for chained im-
putation models; standard and binned residual plots for checking the fit of the conditional
distributions used for imputation; and plots for comparing the distributions of observed
and imputed data. In addition, we use Bayesian models and weakly informative prior
distributions to construct more stable estimates of imputation models. Our goal is to
have a demonstration package that (a) avoids many of the practical problems that arise
with existing multivariate imputation programs, and (b) demonstrates state-of-the-art di-
agnostics that can be applied more generally and can be incorporated into the software
of others.

Keywords: multiple imputation, model diagnostics, chained equations, weakly informative

prior, mi, R.

1. Introduction
The general statistical theory and framework for managing missing information has been well
developed since Rubin (1987) published his pioneering treatment of multiple imputation meth-
ods for nonresponse in surveys. Several software packages have been developed to implement
these methods to deal with incomplete datasets. However, each of these imputation packages
is, to a certain degree, a black box, and the user must trust the imputation procedure without
much control over what goes into it and without much understanding of what comes out.
2 mi: Multiple Imputation with Diagnostics in R

Model checking and other diagnostics are generally an important part of any statistical pro-
cedure. Examining the implications of imputations is particularly important because of the
inherent tension of multiple imputation: that the model used for the imputations is not in
general the same as the model used for the analysis (Meng 1994; Fay 1996; Robins and Wang
2000). We have created an open-ended, open source mi package, not only to solve these
imputation problems, but also to develop and implement new ideas in modeling and model
checking.
Our mi package in R (R Development Core Team 2011) has several features that allow the user
to get inside the imputation process and evaluate the reasonableness of the resulting model
and imputations. These features include: choice of predictors, models, and transformations
for chained imputation models; standard and binned residual plots for checking the fit of the
conditional distributions used for imputation; and plots for comparing the distributions of
observed and imputed data. mi uses an algorithm known as a chained equation approach
(Buuren and Groothuis-Oudshoorn 2011; Raghunathan, Lepkowski, Van Hoewyk, and Solen-
berger 2001); the user specifies the conditional distribution of each variable with missing
values conditioned on other variables in the data, and the imputation algorithm sequentially
iterates through the variables to impute the missing values using the specified models.
We omit a description of the theoretical background of multiple imputation since this material
is available from many other sources (e.g., Little and Rubin (2002); Gelman and Hill (2007,
chapter 25)). Rather, the major goal is to demonstrate the flexible way in which users
can perform multiple imputation with mi and to introduce functions for diagnostics after
imputation. The paper proceeds as follows: In Section 2, we provide an overview of steps to
perform sensible multiple imputation. In Section 3, we demonstrate some novel features and
functions of mi that address some imputation problems that have been neglected by other
software. These features include: (1) Bayesian regression models to address problems with
separation; (2) imputation steps that deal with semi-continuous data; (3) modeling strategies
that handle issues of perfect correlation and structural correlation; (4) functions that check the
convergence of the imputations; and (5) plotting functions that visually check the imputation
models. In Section 4, we demonstrate how to apply these functions using an example of a study
of people living with HIV in New York City (Messeri, Lee, Abramson, Aidala, Chiasson, and
Jessop 2003). In Section 5, we discuss future plans for our mi package. The package is available
from the Comprehensive R Archive Network at http://CRAN.R-project.org/package=mi.

2. Basic setup
The procedure to obtain sensible multiply imputed datasets approach requires four steps:
setup, imputation, analysis, and validation. Each step is divided into substeps as follows:

1. Setup.
ˆ Display of missing data patterns.
ˆ Identifying structural problems in the data and preprocessing.
ˆ Specifying the conditional models.
2. Imputation.
ˆ Iterative imputation based on the conditional model.
 Journal of Statistical Software 3

ˆ Checking the fit of conditional models and checking to see if the imputed values
are reasonable.
ˆ Checking the convergence of the procedure.

3. Analysis.

ˆ Obtaining completed data.

ˆ Pooling the complete case analysis on multiply imputed datasets.

4. Validation.

ˆ Sensitivity analysis.
ˆ Cross validation.
ˆ Compatibility check.

At first glance, it may seem more complicated to conduct multiple imputation using mi
compared to other available imputation software. However this is because we outline steps
that other packages have traditionally ignored. mi is designed for both novice and experienced
users. For the novice users, mi has a step-by-step interactive interface where users choose
options from the given multiple choices and a graphical user interface (GUI) where users click
buttons (see Section 4). For more experienced users, mi has simple commands that users can
use to conduct a multiple imputation. This section simply describes the core functions. In
Section 4, we will demonstrate how users can easily implement these imputation steps using
mi via an example.
The implementation of the mi package is straightforward. The core function is a generic
function mi(object, ...) which executes one of three methods depending on whether the
input is a data.frame, or of S4 class mi.preprocessed, or mi. The mi.preprocessed class
defines the output return by mi.preprocess() when it recodes special variables in a dataset
(see Section 2 and Section 3). The mi class defines the output returned by mi() when it
finishes a multiple imputation with a dataset. The usages of the S4 methods for signature
data.frame, mi.preprocessed, and mi are described respectively below:

mi(object, info, n.imp = 3, n.iter = 30,

R.hat = 1.1, max.minutes = 20, rand.imp.method = "bootstrap",
preprocess = TRUE, run.past.convergence = FALSE,
seed = NA, check.coef.convergence = FALSE,
add.noise = noise.control())

mi(object, n.imp = 3, n.iter = 30,

R.hat = 1.1, max.minutes = 20, rand.imp.method = "bootstrap",
run.past.convergence = FALSE,
seed = NA, check.coef.convergence = FALSE,
add.noise = noise.control())

mi(object, info, n.iter = 30, R.hat = 1.1,

max.minutes = 20, rand.imp.method = "bootstrap",
run.past.convergence = FALSE, seed = NA)
4 mi: Multiple Imputation with Diagnostics in R

ˆ object: A data frame or an mi object that contains an incomplete dataset. mi only

recognizes the special value NA as the missing data.
ˆ info: The mi.info matrix returned by mi.info(). This matrix contains information
about the data (e.g., number of cells that are missing in a variable, variable types, etc.),
and some parameters that control the imputation procedure (see Section 2).
ˆ n.imp: The number of independent imputation chains with different sets of starting
values randomly drawn from the observed ones (see rand.imp.method). The default
is 3 chains. A minimum of 2 chains are required in order to conduct the Gelman and
Rubin convergence diagnostic (Gelman and Rubin 1992; Gelman, Carlin, Stern, and
Rubin 2004).
ˆ n.iter: The maximum number of imputation iterations. The default is 30 iterations.

ˆ R.hat: The value of the R̂ statistic used as a convergence criterion. The default is 1.1
(Gelman and Rubin 1992; Gelman, Carlin, Stern, and Rubin 2004).
ˆ max.minutes: The maximum number of minutes to operate the whole imputation pro-
cess. The default is 20 minutes.
ˆ rand.imp.method: The method used for random imputing starting values of the missing
values. Currently, mi() only implements the bootstrap method: missing values are
filled in with values that are randomly sampled from the observed data.
ˆ preprocess: Default is TRUE. mi() will transform the variables that are not of standard
distribution. These types of variable are nonnegative, positive-continuous, and
proportion. The transformed variables then can be modeled using linear regression
(see Section 3 for details).
ˆ run.past.convergence: Default is FALSE, meaning mi() stops if the imputation is
converged. If the value is TRUE, mi() will run until the values of either n.iter or
max.minutes are reached even if the imputation is converged.
ˆ seed: The random number seed. The default is NA.

ˆ check.coef.convergence: Default is FALSE. mi() only checks the convergence of the

means and standard deviations of the imputed values. If the value is TRUE, mi() also
checks the convergence of the coefficients of imputation models.
ˆ add.noise: A list of parameters for controlling the process of adding noise to mi() via
noise.control(). This is to fill in the missing values with values that are randomly
sampled from the observed ones. This step addresses the problem of collinearity that
impedes appropriate imputation of missing data (see Section 3 for details).

mi() is a wrapper of several key components: the imputation information matrix, variable
types and imputation models.

2.1. Imputation information matrix

mi.info() produces a matrix of imputation information that is needed to impute the missing
data. After the information is extracted from a dataset, users can still alter the default model
 Journal of Statistical Software 5

specifications that are automatically created using this imputation information. Such a matrix
of imputation information allows the users to have control over the imputation process. It
contains the following information:

ˆ name: The names of variables in the dataset.

ˆ imp.order: A vector that records the order of each variable in the iterative imputation
process. If such a variable is missing for all the observations (see all.missing), is an
identification variable (see is.ID), and is not included (see include), the imp.order
slot will record an NA.

ˆ nmis: A vector that records the number of data points that are missing in each variable.

ˆ type: A vector that contains the information of the variable types which are determined
by typecast() (see Section 2).

ˆ var.class: A vector that records the classes of the input variables.

ˆ level: A list of the levels of the input variables.

ˆ include: A vector of indicators that decide whether or not (Yes/No) to include a specific
variable in an imputation process. If include is No, the variable will not show up either
as a predictor or as a variable to be imputed.

ˆ is.ID: A vector of indicators that determine whether or not (Yes/No) a specific variable
is an identification (ID) variable. If a variable is detected as an ID variable, it will not
be included in the imputation process; thus the include slot records a No value. ID
variables are usually not problematic as dependent variables, since in most of the cases,
they have no missing values. But when they are included in a model as predictors, they
induce an unwanted order effect of the data into the model (unless the data is a repeated
measure study and ID variables are treated as categorical variables). However, because
ID variables are hard to detect, users should carefully check to see if all such variables
have been detected.

ˆ all.missing: A vector of indictors that identify whether or not (Yes/No) a variable is

missing for all the observation. If the value is TRUE, such a variable will be excluded
in the imputation process because it is not possible to impute sensible values. The
include slot records a No value if all.missing is TRUE.

ˆ missing.index: A vector that stores the index number of the missing units in a variable.

ˆ collinear: A vector of indicators that shows whether or not (Yes/No) a variable is

perfectly collinear with another variable. mi.info() uses cor() to compute the Pearson
coefficients, a measure of the correlation (linear dependence) between two variables, of
the data. If the Pearson coefficients are larger than 0.99999 (arbitrary chosen as the
default in mi), mi.info() sets the value of the collinear slot to TRUE. Such a variable
will be excluded in the imputation process (thus the include slot records a No value) if
and only if these two variables have the same missing data pattern, meaning that they
are missing in the same units.
6 mi: Multiple Imputation with Diagnostics in R

ˆ imp.formula: A vector of formulas that records the imputation formulas used in the
imputation models. This formula represents the linear specification of the appropriate
model. For instance, for a binary variable this formula represents the linear function of
covariates that is set equal to the logit of the expectation of the response variable.

ˆ determ.pred: The name of the corresponding correlated variable. This slot is NULL if
there is no corresponding correlated variable as identified by cor().

ˆ params: A list of parameters to pass on to the imputation models.

ˆ other: Other options. This is currently not used.

Users can alter the mi.info matrix using update(). For instance, if we have a variable x1
in a dataset, and we do not want to include it in the imputation process, we can update the
include slot of the mi.info matrix by:

R> info

names include order number.mis all.mis type collinear

1 x1 Yes 1 40 No continuous No
2 x2 Yes 2 13 No continuous No
...

R> info <- update(info, "include", list("x1" = FALSE))

R> info

names include order number.mis all.mis type collinear

1 x1 No NA 40 No continuous No
2 x2 Yes 1 13 No continuous No
...

2.2. Variable types

mi handles eleven variable types. Within mi(), mi.info() uses typecast() to automatically
identify eight different variable types; mi.preprocess() specifies the log-continuous type
via transformation (see Section 3). The variable types that are not automatically identified
by typecast() are count and predictive-mean-matching type.1 These two types must be
user-specified via update(). typecast() identifies variable types using the rules depicted in
Figure 1.
The rules, which typecast() uses to identify each variable type, are listed as follows:

1. fixed: Any variable that contains a single unique value.

2. binary: Any variable that contains two unique values.

1
predictive-mean-matching is not really a variable type. We include it here to invoke mi() to fit a model
using the predictive mean matching method (see Section 2)
 Journal of Statistical Software 7

3. ordered-categorical: Any variable that has the ordered attribute that is determined
by is.ordered() in R. Or any numerical variable that has 3 to 5 unique values.

4. unordered-categorical: Any factor or character variable that is determined by

is.character() or is.factor() in R. This type of variable often is not saved as a
character or factor variable. Additionally, typecast() will identify a numeric unordered
categorical variable as a continuous or an ordered categorical variable if users are not
vigilant about identifying it. An unordered categorical variable, once is identified by
typecast() or is specified by user, it is going to be imputed with a multinomial log-
linear model (see Section 2) but is included in the imputation models of other variables
as a factorized predictor (i.e., R will split this variable into the respective indicator
variables).

5. proportion: Any numerical variable that has its values fall between 0 and 1, not
including 0 and 1.

6. positive-continuous: Any numerical variable that is always positive and has more
than 5 unique values.

7. nonnegative: Any numerical variable that is always nonnegative and has more than
5 unique values.

8. continuous: Any numerical variable that is modeled as continuous without transfor-

mation.

9. log-continuous: log-scaled continuous variable, specified by mi.preprocess() (see

Section 3).

10. count: A user-specified variable type.

11. predictive-mean-matching: An user-specified variable type (see Section 2).

Once the variable type is determined by typecast(), the type information will be stored in
the mi.info matrix. Nonetheless, users can alter this default judgment. For instance, if riot
is the number of riots in a specific year, its values are very likely to fall between 0 and any
positive integer. Hence, typecast() is going to identify riot as either ordered-categorical,
positive-continuous or nonnegative type, depending on number of unique values it has
and whether or not its values contains 0. You can alter this judgment by updating the type
slot in a mi.info matrix as:

R> info

names include order number.mis all.mis type correlated

1 riot Yes 1 23 No nonnegative No
...

R> info <- update(info, "type", list("riot" = "count"))

R> info
8 mi: Multiple Imputation with Diagnostics in R

L==1
fixed
L==2
binary
is.character(x)
unordered−categorical
is.factor(x)
unordered−categorical L=length(unique(x))
x is.ordered(x)
ordered−categorical
is.numeric(x)
continuous
all(0<x<1)
proportion
2<L<=5
ordered−categorical
binary
L>5 & all(x>=0) mi.preprocess()
nonnegative
log−continuous
L>5 & all(x>0)
positive−continuous log−continuous

Figure 1: Illustration of the rules of typecast() to identify and classify different variable
types. mi currently handles eleven variable types: fixed, binary, unordered-categorical,
ordered-categorical, proportion, positive-continuous, nonnegative, continuous,
log-continuous, count, and predictive-mean-matching. But mi can only automatically
identify the first eight variable types. mi.preprocess() specifies log-continuous variable
type via transformation. Users have to specify count and predictive-mean-matching vari-
able types manually via update() or interactively via mi.iteractive().

names include order number.mis all.mis type correlated

1 riot Yes 1 23 No count No
...

2.3. Imputation models

By default, mi chooses the conditional models via type.model(), a function that determines
which imputation models to use based on the variable types determined by typecast(). Ta-
ble 1 lists the default regression models corresponding to variable types. mi.fixed() just
copies the values from the observed one. mi.categorical() uses multinom() (multino-
mial log-linear model, Venables and Ripley 2002) to impute unordered categorical variables.
mi uses mi.continuous() to impute positive-continuous, nonnegative, and proportion
variable types.
mi.continuous(), mi.binary(), mi.count(), and mi.polr() fit Bayesian version of the
generalized linear models (bayesglm() and bayespolr(), see Gelman, Jakulin, Pittau, and
Su 2008) of arm (Gelman, Su, Yajima, Hill, Pittau, Kerman, and Zheng 2011). The Bayesian
version of the generalized linear model that we use is different from the classical generalized
linear model in that it adds a Student-t prior on the regression coefficients. Gelman, Jakulin,
Pittau, and Su (2008) propose a new prior distribution for classical (nonhierarchical) logistic
regression models, constructed by first scaling all nonbinary variables to have mean 0 and
standard deviation 0.5, and then placing an independent Student-t prior distribution on the
 Journal of Statistical Software 9

Variable types mi regression functions

binary mi.binary
continuous mi.continuous
count mi.count
fixed mi.fixed
log-continuous mi.continuous
nonnegative mi.continuous
ordered-categorical mi.polr
unordered-categorical mi.categorical
positive-continuous mi.continuous
proportion mi.continuous
predictive-mean-matching mi.pmm

Table 1: List of mi regression functions, corresponding to variable types.

coefficients. As a default choice, they recommend the Cauchy distribution with center 0 and
scale 2.5. At the present time, mi() does not allow users to alter the priors.
While this is a reasonably wide selection of models, it is still possible that none provide
completely adequate fit for the data which could lead to improper imputation. Achieving
appropriate fit can be particularly challenging when constraints exist in the data, when het-
eroskedasticity exists in a model, when multiple modes exist or a variety of other modeling
challenges. To address this issue, mi offers an option to impute these data using the predictive
mean matching method (Rubin 1987; Heitjan and Little 1991; Schenker and Taylor 1996).
The predictive mean matching method (mi.pmm()) works in the following way. For each ob-
servation with a missing value on a given variable, we find the observation (from among those
with observed values on that variable) with the closest predictive mean for that variable.
The observed value from this “match” is then used as the imputed value. Currently, these
predictions are obtained from bayesglm(). However, we plan to extend the predictive mean
matching options in the future. This method can be problematic when rates of missingness
are high or when the missing values fall outside the range of the observed data. We are
continuing to develop more flexible imputation models to better address the issue of creating
appropriate imputations.

3. Novel features
Our mi has some novel features that address some open issues in multiple imputation.

3.1. Bayesian models to address problems with separation

Logistic regression, and more generally, discrete data models, commonly suffer from the prob-
lem of separation. This problem occurs whenever the outcome variable is perfectly predicted
by a predictor or a linear combination of the predictors. This can happen even with a modest
number of predictors, particularly if the proportion of “successes” in the response variable is
relatively close to 0 or 1. The risk of separation typically increases as the number of pre-
dictors increases. However, multiple imputation is generally strengthened by including many
variables, which can help to impute more precisely and also may help to satisfy the miss-
10 mi: Multiple Imputation with Diagnostics in R

mi functions Bayesian functions

mi.continuous() bayesglm() with gaussian family
mi.binary() bayesglm() with binomial family (default uses logit link)
mi.count() bayesglm() with quasi-poisson family (overdispersed poisson)
mi.polr() bayespolr()

Table 2: Lists of Bayesian generalized linear models used in mi regression functions.

ing at random assumption. When imputing large-scale surveys to create public-use multiply
imputed datasets, several hundred variables might need to be imputed. And it is unclear
how or if we should start discarding subsets of variables from some or all of the conditional
models. Separation problems can cause the chained equation algorithms to either fail or
impute unreasonable values. Moreover, even without perfect separation, imposing priors on
regression coefficients improves imputation in cases of near-separation and/or collinearity or
near-collinearity of predictors (Gelman, Jakulin, Pittau, and Su 2008).
To address problems with separation, we have augmented our mi to allow for Bayesian version
of the generalized linear models with Student-t prior distributions on regression coefficients
(the default prior uses t distribution with center 0, degree of freedom 1 and scale 2.5). The
models, as implemented in the functions bayesglm() and bayespolr(), automatically handle
separation (Gelman, Jakulin, Pittau, and Su 2008). The corresponding imputation models
are listed in Table 2.2

3.2. Imputing semi-continuous data with transformation

Semi-continuous data (positive-continuous, nonnegative and proportion variable types
in mi) are typically not modeled in a reasonable way in other imputation software. The
difficulty comes from the fact that these kinds of data have bounds or truncations and are
not of standard distributions. Our algorithm models these data using transformations via
mi.preprocess(). These transformations are automatically performed in mi() by setting
the option preprocess = TRUE (which is the default).
For the nonnegative variable type, mi.preprocess() creates two ancillary variables. One
is an indicator for which values of the nonnegative variable are greater than 0. The other
ancillary variable takes the log of the original variable for any value that is greater than 0. For
the positive-continuous variable type, mi.preprocess() takes the log of the variable. For
the proportion variable
  type, mi.preprocess() does a logit transformation on the variable
x
as logit (x) = log 1−x .
Figure 2 illustrates this transformation process. Users can transform the data back to the orig-
inal scale using mi.postprocess(). This is implemented automatically in mi.completed()
(see Section 4). For the positive-continuous variable (x1), it is going to be transformed
back as x1c = exp(x1.mi.log) × x1.mi.ind; for the nonnegative variable (x2), x2 c =
−1
exp(x2.mi.log); and for the proportion variable (x3), x3 c = logit (x3.mi.logit).
After the transformation, mi() uses mi.continuous() (Gaussian linear model) to impute
these log transformed and logit transformed variables. Nonetheless, if the transformation is
2
We are working on Bayesian version of multinomial models for unordered categorical variables. mi() now
uses mi.categorical(), which uses multinom() (multinomial log-linear model, Venables and Ripley 2002), to
handle with unordered categorical variables.
 Journal of Statistical Software 11

Original Dataset Transformed Dataset

x1 x2 x3 x1.mi.log x1.mi.ind x2.mi.log x3.mi.logit
0.00 6.00 NA NA 0.00 1.79 NA
NA 5.00 0.18 NA NA 1.61 −1.49
6.00 10.00 0.54 1.79 1.00 2.30 0.15
19.00 NA 0.51 mi.preprocess() 2.94 1.00 NA 0.04
0.00 NA 0.43 NA 0.00 NA −0.27
5.00 NA 0.98 1.61 1.00 NA 4.00
NA 11.00 0.26 NA NA 2.40 −1.06
10.00 NA 0.82 2.30 1.00 NA 1.54
18.00 5.00 0.16 2.89 1.00 1.61 −1.65
0.00 14.00 NA NA 0.00 2.64 NA
… … … … … … …

Figure 2: Illustration of the way in which mi.preprocess() transforms the data of

nonnegative (x1), positive-continuous (x2) and proportion (x3) variable types.

not enough to achieve a well-fitting model, we recommend that users utilize the predictive
mean matching method as a work-around.

3.3. Imputing data with collinearity

mi can deal with two types of data with collinearity. One type is the perfect correlation of
two variables (e.g., x1 = 10x2 − 5). The other type is for data with additive constraints across
several variables (e.g., x1 = x2 + x3 + x4 + x5 + 10).

Perfect correlation
In real datasets, a variable may appear multiple times or with different scale. For example,
GDP per capita and GDP per capita in thousand dollars could both be in a dataset. For
these variables, if the missingness pattern of these two variables is the same, mi() will include
only one of the duplicated variables in the iterative imputation process. To impute data
with such a perfect correlation, mi() will firstly identify such a pair of correlated variables
via mi.info() and exclude one of them from the imputation process. If the absolute value
of a Pearson coefficient of two variables is larger or equal to an arbitrary chosen threshold
of 0.9999, these two variables are treated as correlated. Then mi() will use mi.copy() to
impute the missing values of the excluded variable by duplicating values from the correlated
variable. If for some reasons, the duplicated variables do not have the same missingness
pattern (missing the same units), mi() keeps them in the iterative imputation process and
utilizes the technique described in the next section (Section 3).

Additive constraints
General additive constraints can cause problems with the chained equation algorithm. How-
ever, they can be difficult to identify when we have missing data. Moreover, even if we identify
such dependencies, if the variables in this grouping do not all have the same missing data
pattern the solution is typically not obvious. This problem can exist either in deterministic
situations (such as inclusion of all the items in a scale as well as the total scale score) or
situations included in the model that by chance a subset of variables end up being func-
tionally dependent on each other (i.e., each is a linear combination of the rest). Figure 3
displays a hypothetical example of a dataset with additive constraints. In this example, we
have the number of female students (female), the number of male students (male) and the
12 mi: Multiple Imputation with Diagnostics in R

Structured Correlated Dataset

class total female male
1 66 NA 35
2 NA 27 23
3 76 37 NA
4 NA 31 NA
5 51 24 NA
6 73 39 34
7 NA NA 39
8 NA NA NA
9 46 26 NA
10 59 NA NA
… … … …

Figure 3: Illustration of a dataset with additive constraints. In this dataset, total = female+
male.

total number of students (total) in a given class. Hence, total = female + male. Such a
problem could easily be dealt with if an investigator spots such a problem beforehand and
takes out one of the variables. However, this problem could easily remain undetected, if such
a problem of additive constraints exists across many variables without logically related and
explicit variable names.
Computational issues can arise when more than one variable is missing. Take class 4 in
Figure 3 for an example. mi() will first randomly assign a value for male in the class 4, say
20. Then mi() imputes total by regressing it on male and female. In this case, the value
would be 51. Using the imputed value for total, the next imputation of mi() for male will
be 20. Using this imputed value for male, the next imputed value for total will be 51. This
situation will continue to repeat like this over and over. The result of this problem is that
mi() will not be able to explore the entire response surface of the imputed variables.
To deal with this problem, we introduce an artificial set of prior distributions on the missing
data into the iterative imputation process. The purpose is to create noise that breaks the
deterministic structure to force mi() to explore more of the response surface of the imputed
variables. At the same time, because the priors originate from the observed data, they also
ensure that the imputed values do not deviate too far from the observation. mi() currently
offers two options via noise.control(), each of which temporarily adds prior information to
the model fits.

ˆ Reshuffling noise: By default, mi() adds noise to the iterative imputation process by
randomly imputing missing data for a given variable from the empirical marginal distri-
bution. In other words, mi() imputes the missing data by sampling from the observed
data. In every iteration, mi() decides whether or not to impute the missing data from
the marginal distribution based on a random bernoulli variable q with a probability
p = Ks , where s is the number of imputation iteration and K is specified by the user
(the default is 1). If q = 1, mi() imputes the missing data with values from the observed
data. Otherwise, it imputes missing data with values from the conditional models.
The influence of the noise gradually declines because p gradually decreases to the zero
as the number of iterations increases. This means mi() eventually imputes the missing
 Journal of Statistical Software 13

data only from the conditional models. Hence, depending on the size of K (default is
1), users can control how much power they want the noise to insert into the iterative
imputation process. The standard usage is as below:
R> IMP <- mi(data, add.noise = noise.control(method = "reshuffling",
+ K = 1))
ˆ Fading empirical noise. In each iteration, mi() augments the data by pct.aug = 10
percent (default) of the completed data by randomly adding new data that are drawn
from the observed data. Hence the noise is added to all variables. Thus if a completed
dataset has 250 data points, mi() will augment such a dataset with 25 new data points
from the observed data of the complete case. The standard usage is as below:
R> IMP <- mi(data, add.noise = noise.control(method = "fading",
+ pct.aug = 10))

By default, mi() uses the reshuffling noise. If users have faith that their data have neither of
the two correlation problems, they can choose not to add noise into the imputation process by
specifying mi(data, add.noise = FALSE, ...). By default, if any of the two methods of
adding noise is used, mi() will run 20 more iterations (controlled by post.run.iter, default
is 20) without adding any noise to mitigate the influence of the noise. To change the default
number of iterations from 20 to 30, user simply runs:
R> IMP <- mi(data, add.noise = noise.control(post.run.iter = 30))
If post.run.iter is set to zero, mi() will not run any additional iteration to mitigate the in-
fluence of the noise. This is not recommended, though, users can still run additional iterations
later by doing the following:
R> IMP <- mi(IMP)

3.4. Checking the convergence of the imputations

Our mi offers two ways to check the convergence of the multiple imputation procedure. By
default, mi() monitors the mixing of each variable by the variance of its mean and standard
deviation within and between different chains of the imputation. If the R̂ statistic is smaller
than 1.1, (i.e., the difference of the within and between variance is trivial), the imputation is
considered converged (Gelman, Carlin, Stern, and Rubin 2004). Additionally, by specifying
mi(data, check.coef.convergence = TRUE, ...), users can check the convergence of the
parameters of the conditional models.

3.5. Model checking and other diagnostic using graphics

Model checking and other diagnostics are generally an important part of any statistical pro-
cedure. This is particularly important to imputation because the model used for imputation
in general is not the same as the model used for the analysis. Yet, there is a noticeable dearth
of such checks in the multiple imputation world. Thus imputations are, for the most part,
a black box. This likely stems from uncertainty regarding how to assess the appropriateness
of imputed values that are proxies for data points that are by definition unknown. Our mi
addresses this problem with three strategies.
14 mi: Multiple Imputation with Diagnostics in R

ˆ Imputations are typically generated using models, such as regressions or multivariate

distributions, which are fit to observed data. Thus the fit of these models to the observed
data can be checked (Gelman, Mechelen, Verbeke, Heitjan, and Meulders 2005) using
standard graphical diagnostics.

ˆ Imputations can be checked using a standard of reasonability: the differences between

observed and missing values, and the distribution of the completed data as a whole, can
be checked to see whether they make sense in the context of the problem being studied
(Abayomi, Gelman, and Levy 2008).

So far, mi only implements the first two solutions with various plotting functions. We demon-
strate the usages of these functions in Section 4.

4. Example
In this Section, we demonstrate some basic steps of mi with an example.

4.1. A study of HIV-positive people in New York City

The CHAIN dataset included in mi is a subset of the Community Health Advisory and Informa-
tion Network (CHAIN) study. This study is a longitudinal cohort study of people living with
HIV in New York City and is conducted by Columbia University School of Public Health
(Messeri, Lee, Abramson, Aidala, Chiasson, and Jessop 2003). The CHAIN cohort was re-
cruited in 1994 from a large number of medical care and social service agencies serving HIV
in New York City. Cohort members were interviewed up to 8 times through 2002. A total of
532 CHAIN participants completed at least one interview at either the 6th, 7th or 8th rounds
of interview, and 508, 444, 388 interviews were completed respectively at rounds 6, 7 and 8
(CHAIN 2009). For simplicity, our analysis here discards the time aspect of the dataset and
use only the 6th round of the survey. The dataset has 532 observations and has the following
8 variables:

ˆ h39b.W1: Log of self reported viral load level, 0 represents undetectable level.

ˆ age.W1: The respondent’s age at time of interview.

ˆ c28.W1: The respondent’s family annual income. Values range from under USD 5, 000
to USD 70, 000 or over.

ˆ pcs.W1: A continuous scale of physical health with a theoretical range between 0 and
100 (better health is associated with higher scale values).

ˆ mcs37.W1: A dichotomous measure of poor mental health: 0 = No, 1 = Yes.

ˆ b05.W1: Ordered interval for the CD4 count (the indicator of how much damage HIV
has caused to the immune system).

ˆ haartadhere.W1: A three-level-ordered variable: 0 = not currently taking highly active

antiretroviral therapy (HAART), 1 = taking HAART nonadherent, 2 = taking HAART
adherent.
 Journal of Statistical Software 15

To use the data, users must first load the mi package:3

R> library("mi")

Loading required package: MASS

Loading required package: nnet
Loading required package: car
...
Loading required package: arm
Loading required package: Matrix
Loading required package: lattice
...
Loading required package: lme4
...
Loading required package: abind

mi (Version 0.09-13, built: 2011-2-15)

Then load the CHAIN dataset in the memory:

R> data("CHAIN")

4.2. Setup
The first thing to do is to set up the imputation. As with most statistical procedures, one
must start with some preliminary analysis to avoid trivial problems. When that is completed,
two key steps must be done: choosing the conditional models and specifying the models.
Preliminary analysis is crucial in an iterative procedure such as multiple imputation that uses
the chain equation algorithm. Users do not want simple mistakes that arise in the early stages
to ruin the end result after a long iteration. In a small dataset, this may not be a serious
issue, but for a large dataset, this may be costly. There are problems which mi automatically
detects. But there are problems that are not possible to detect automatically by mi. For
those problems that are difficult to detect, our mi will raise flags so that user can keep them
in mind.

Display of missing data patterns

Users can get the glimpse of the data by looking at the missingness pattern.

R> missing.pattern.plot(CHAIN, clustered = FALSE)

Or simply type:
3
The printout of the loaded information shows that mi depends upon on several R packages, including MASS
and nnet (Venables and Ripley 2002), car (Fox and Weisberg 2011), arm (Gelman et al. 2011), Matrix (Bates
and Maechler 2011), lme4 (Bates, Maechler, and Bolker 2011), R2WinBUGS (Sturtz, Ligges, and Gelman
2005), coda (Plummer, Best, Cowles, and Vines 2006) and abind (Plate and Heiberger 2011). In the last line
of the loaded information, R prints out the version number of mi. Users are welcome to report bugs or make
suggestions to us with the attached version number.
16 mi: Multiple Imputation with Diagnostics in R

CHAIN CHAIN

haartadhere.W1 haartadhere.W1
b05.W1 mcs37.W1
mcs37.W1 pcs.W1
pcs.W1 age.W1
c28.W1 c28.W1
age.W1 b05.W1
h39b.W1 h39b.W1

Index Ordered by number of missing items

Figure 4: Missingness pattern plot. The observed values are plotted with blue and the missing
values are in red. In the right panel, variables and cases are ordered by proportion missing
and clustered by similar missingness pattern.

R> mp.plot(CHAIN, clustered = FALSE)

Figure 4 (left) shows the data matrix with observed values in blue and missing values in red.
At this point it is difficult to detect anything, but we also have the option to order them by
the missing data rates (the x-axis is the data index and the y-axis is the variable index) and
to cluster them by the missingness pattern.

R> mp.plot(CHAIN, y.order = TRUE, x.order = TRUE, clustered = TRUE)

Figure 4 (right) reveals that h39b.W1 variable has the highest missing rate followed by b05.W1
and c28.W1. The other variables have missingness only for the observations that have missing
data for all of the variables. When there exist observations with all variables missing, the user
should either remove these observations (since they will not add any new information to the
imputation procedure) or include more variables (with observed values for these observations)
from the full dataset if these exist.

Identifying structural problems in the data and preprocessing

Before starting to impute the missing data, we can review some basic data information, which
will feed into the imputation process, and determine whether or not we want to adjust this
information.

R> info <- mi.info(CHAIN)

R> info

names include order number.mis all.mis type collinear

1 h39b.W1 Yes 1 179 No nonnegative No
2 age.W1 Yes 2 24 No positive-continuous No
3 c28.W1 Yes 3 38 No positive-continuous No
4 pcs.W1 Yes 4 24 No positive-continuous No
5 mcs37.W1 Yes 5 24 No binary No
6 b05.W1 Yes 6 63 No ordered-categorical No
7 haartadhere.W1 Yes 7 24 No ordered-categorical No
 Journal of Statistical Software 17

By default, mi.info() prints out seven out of the fourteen categories of the mi.info matrix
(see Section 2). We can see from this output that h39b.W1, age.W1, c28.W1, and pcs.W1
are variable types that need special treatment (see Section 3 for data preprocessing and
transformation).
So to address the variable types that mi.info() identifies as requiring special treatment,
mi preprocesses the data via mi.preprocess() and may change the default judgement re-
turned by typecase() (see Section 2 and Figure 2). mi.preprocess returns an S4 object
mi.preprocess that stores the transformed data in the slot data and the new mi.info object
in the slot mi.info.

R> CHAIN.new <- mi.preprocess(CHAIN)

R> attr(CHAIN.new, "mi.info")

names include order number.mis all.mis type collinear

1 h39b.W1.mi.log Yes 1 367 No log-continuous No
2 age.W1.mi.log Yes 2 24 No log-continuous No
3 c28.W1.mi.log Yes 3 38 No log-continuous No
4 pcs.W1.mi.log Yes 4 24 No log-continuous No
5 mcs37.W1 Yes 5 24 No binary No
6 b05.W1 Yes 6 63 No ordered-categorical No
7 haartadhere.W1 Yes 7 24 No ordered-categorical No
8 h39b.W1.mi.ind Yes 8 179 No binary No

The new information matrix shows that h39b.W1, age.W1, c28.W1, and pcs.W1 have been
transformed into new variables with different scales and types. Also, the transformed variables
have been attached with new suffixes.

Specifying the conditional models

mi() chooses the conditional models based on the variable types that are determined by
typecast() (see Section 2). By changing the variable types, mi() will choose different con-
ditional models to fit the altered variables. For example, you can change the type of h39b.W1
from nonnegative to continuous as:

R> info <- mi.info(CHAIN)

R> info

names include order number.mis all.mis type collinear

1 h39b.W1 Yes 1 179 No nonnegative No
...

R> info.upd <- update(info, "type", list("h39b.W1" = "continuous"))

R> info.upd

names include order number.mis all.mis type correlated

1 h39b.W1 Yes 1 179 No continuous No
...

By default, mi() assumes linearity between the outcomes and additive predictors.
18 mi: Multiple Imputation with Diagnostics in R

R> info$imp.formula

h39b.W1
"h39b.W1 ~ age.W1 + c28.W1 + pcs.W1 + mcs37.W1 + b05.W1 + haartadhere.W1"
age.W1
"age.W1 ~ h39b.W1 + c28.W1 + pcs.W1 + mcs37.W1 + b05.W1 + haartadhere.W1"
c28.W1
"c28.W1 ~ h39b.W1 + age.W1 + pcs.W1 + mcs37.W1 + b05.W1 + haartadhere.W1"
pcs.W1
"pcs.W1 ~ h39b.W1 + age.W1 + c28.W1 + mcs37.W1 + b05.W1 + haartadhere.W1"
mcs37.W1
"mcs37.W1 ~ h39b.W1 + age.W1 + c28.W1 + pcs.W1 + b05.W1 + haartadhere.W1"
b05.W1
"b05.W1 ~ h39b.W1 + age.W1 + c28.W1 + pcs.W1 + mcs37.W1 + haartadhere.W1"
haartadhere.W1
"haartadhere.W1 ~ h39b.W1 + age.W1 + c28.W1 + pcs.W1 + mcs37.W1 + b05.W1"

If you want to change the fitted formulas by adding interactions or add squared terms, you
can alter the imp.formula slot of the mi.info matrix via update() or interactively via
mi.interactive():

R> info.upd <- update(info, "imp.formula", list("h39b.W1" =

+ "h39b.W1 ~ age.W1^2 + c28.W1*pcs.W1 + mcs37.W1 +
+ b05.W1 + haartadhere.W1"))
R> info.upd$imp.formula["h39b.W1"]

h39b.W1
"h39b.W1 ~ age.W1^2 + c28.W1*pcs.W1 + mcs37.W1 + b05.W1 + haartadhere.W1"

4.3. Imputation
Once everything has been setup correctly, actual imputation is simple. However, there are
still a few things users should check: the fit of the conditional models and convergence of the
imputation algorithm. Diagnostic tools are integrated as parts of mi(), but decisions about
how to use the diagnostic information must be made by the users. We will provide general
guidelines here.

Iterative imputation based on the conditional model

mi() imputes the missing values based on the conditional models. That is it will draw values
(that is, stochastically sample) from the fitted conditional distributions. As demonstrated in
the previous sections, you can modify the mi.info object and pass it into mi() to alter these
model settings . If no mi.info object is passed into mi(), mi() will call mi.info() internally
and use the default setting. Although this is not recommended, we have made the defaults
as reasonable as possible.

R> CHAIN.new <- mi.preprocess(CHAIN)

R> IMP <- mi(CHAIN.new)
 Journal of Statistical Software 19

Beginning Multiple Imputation ( Fri Jul 02 10:54:34 2010 ):

Iteration 1
Chain 1 : h39b.W1.mi.log* age.W1.mi.log* c28.W1.mi.log* pcs.W1.mi.log* mcs37.W1* b05.W1* haartadhere.W1* h39b.W1.mi.ind*
Chain 2 : h39b.W1.mi.log* age.W1.mi.log* c28.W1.mi.log* pcs.W1.mi.log* mcs37.W1* b05.W1* haartadhere.W1* h39b.W1.mi.ind*
Chain 3 : h39b.W1.mi.log* age.W1.mi.log* c28.W1.mi.log* pcs.W1.mi.log* mcs37.W1* b05.W1* haartadhere.W1* h39b.W1.mi.ind*
Iteration 2
Chain 1 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log* mcs37.W1* b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 2 : h39b.W1.mi.log age.W1.mi.log* c28.W1.mi.log* pcs.W1.mi.log* mcs37.W1 b05.W1 haartadhere.W1* h39b.W1.mi.ind
Chain 3 : h39b.W1.mi.log age.W1.mi.log* c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind*
...
Iteration 30
Chain 1 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1* b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 2 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 3 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1* b05.W1 haartadhere.W1 h39b.W1.mi.ind
Reached the maximum iteration, mi did not converge ( Wed Feb 09 16:49:30 2011 )
Run 20 more iterations to mitigate the influence of the noise...
Beginning Multiple Imputation ( Wed Feb 09 16:49:30 2011 ):
Iteration 1
Chain 1 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 2 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 3 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Iteration 2
Chain 1 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 2 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 3 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
...
Iteration 20
Chain 1 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 2 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 3 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
mi converged ( Wed Feb 09 16:50:06 2011 )

By default, mi() will perform reshuffling, and run 20 more iterations after the first mi() is
finished (add.noise = noise.control(K = 1, post.run.iter = 20)). The star symbols
attached to the variable names indicate that reshuffling is being implemented for that variable
in that iteration. Before doing mi, if users want to use the mi() to transform the data, they
can transform the data with their own design or use mi.preprocess() to do the job. Here,
we use mi.preprocess() to transform the data before we starting the imputation process.
There are other options to specify number of iterations (n.iter), how long mi() should run
(max.minutes), whether or not mi should continue when it converged (run.past.convergence),
etc (see Section 2 or type ?mi in the R console for details).

Checking the fit of conditional models and imputed values

If the fit of our imputation models to the observed data is poor, it is unlikely that we will
impute reasonable values for the missing values even our data are truly missing at random.
We can check this fit through the binned residuals plot for the observed data and through
overlaid histograms comparing observed and imputed data. Moreover, if the missing at ran-
dom assumption is not appropriate we may be able to detect that by comparing imputed
values to observed values based on what we know about the science of the phenomenon being
measured by the variables in our dataset. mi provides three different plots to visually inspect
the fit of the conditional models.
Imputation may take some time to run, depending on the size of the data. Thus we suggest
that users may want to check the fit of the conditional models (Gelman, Mechelen, Verbeke,
Heitjan, and Meulders 2005; Abayomi, Gelman, and Levy 2008) by plotting the mi object
(Figure 5) after a reasonable number of iterations rather than waiting for convergence. Fur-
thermore, if working with a large sample size, it may also be helpful to diagnose the imputation
procedure for a random sample from the full sample.
R> plot(IMP)
The first plot displays histograms of the observed (in blue color), the imputed (in red color)
and the completed (observed plus imputed, in black color) values. The second is a binned
20 mi: Multiple Imputation with Diagnostics in R

Figure 5: Diagnostic plots for checking the fit of the conditional imputation models. To save
page space, only four out of the eight plots for each variable are displayed. The blue color is
for the observed value and the red color is for the imputed one. By default, these values are
plotted against an index number. Plotting against a variable that contains more information
is a strongly recommended alternative. Fitted lowess curves are also plotted for the observed
data. A small amount of random noise (jittering) is added to the points so that they do not
fall on top of each other.
 Journal of Statistical Software 21

3 chains, each with 59 iterations (first 0 discarded)

80% interval for each chain R−hat medians and 80% intervals
0 2 4 1 1.5 2+ 2.25

mean(h39b.W1.mi.log) 2.2
●
●

mean(h39b.W1.mi.log) ●
●

mean(age.W1.mi.log) ● 2.15

mean(c28.W1.mi.log) ●

3.735
mean(pcs.W1.mi.log) ●
●
●
●

mean(age.W1.mi.log) 3.73
mean(mcs37.W1) ●

3.725
mean(b05.W1) ●

mean(haartadhere.W1) ●
1.06
1.05
mean(h39b.W1.mi.ind) ● mean(c28.W1.mi.log) 1.04 ●
●
●

1.03
sd(h39b.W1.mi.log) ●
1.02

sd(age.W1.mi.log) ●

3.755
sd(c28.W1.mi.log) ●

mean(pcs.W1.mi.log) 3.75 ●
●
●

sd(pcs.W1.mi.log) ●

3.745
sd(mcs37.W1) ●

sd(b05.W1) ● 0.285
0.28
sd(haartadhere.W1) ●
mean(mcs37.W1) 0.275 ●

0.27 ●

sd(h39b.W1.mi.ind) ●
0.265

0 2 4 1 1.5 2+ 3.65

mean(b05.W1) 3.6 ●
●

3.55

1.88
1.87
mean(haartadhere.W1) ●

1.86 ●
●

1.85

0.52
0.5 ●

mean(h39b.W1.mi.ind) ●
●

0.48
0.46

0.22
0.21
sd(h39b.W1.mi.log) 0.2 ●
●
●

0.19
0.18

0.202
0.2
sd(age.W1.mi.log) ●

0.198 ●
●

0.196

0.63

sd(c28.W1.mi.log) 0.62 ●
●
●

0.61

0.315
0.31
sd(pcs.W1.mi.log) ●
●

0.305
0.3

0.45

sd(mcs37.W1) 0.445
●
●

0.44

1.38
1.36
sd(b05.W1) ●
●

1.34
1.32

0.875
0.87
sd(haartadhere.W1) ●
●

0.865
0.86

0.5005 ●
●

sd(h39b.W1.mi.ind) 0.5

0.4995

Figure 6: Plot of the summary of the mean and the standard deviation of each variable for
the different chains of imputations. All the R̂ statistics are smaller than 1.1, indicating the
imputation is converged.
22 mi: Multiple Imputation with Diagnostics in R

residual plot that plots the average of residuals in bins against the expected values (Gelman,
Goegebeur, Tuerlinckx, and Van Mechelen 2000). The third is a bivariate scatterplot that
plots the observed or imputed values for each observation against the predicted values, overlaid
with fitted lowess curves of the observed data (Cleveland 1979).
Figure 5 displays the selected variables using these four diagnostic plots. The histograms
show that the imputed values are all within reasonable ranges and do not differ much from
the observed values. The binned residual plots show that there is room for improvement on the
imputation models of h39b.W1.mi.log and pcs.W1.mi.log as there are a number of residuals
that fall outside of the 95% error bounds (the dotted lines with light color). Each point in
a binned residual plot is the average of the points that fall in a given “bin” (interval of the
variable on the x-axis) from a standard residual point. The bivariate scatterplot demonstrates
that there is no significant difference between observed and imputed values.
If users discover a problem when accessing these plots and want to alter the model specification
to fix it, they can fix the mi.info object via update() or they can perform transformations
to their variables before feeding the dataset into the mi program.
Checking the convergence of the procedure
Checking the convergence of multiple imputation is still an open research question. By de-
fault, mi() checks the mean and standard deviation of each variable for different chains. It
considers the imputation to have converged when the R̂ < 1.1 for all the parameters (Gelman,
Carlin, Stern, and Rubin 2004). There is a R.hat option in mi() that allows users to adopt
more stringent rule on checking convergence using the R̂ statistics (mi(CHAIN, R.hat = 1).
Users can also check the convergence of parameters of each conditional model by specify-
ing mi(CHAIN, check.coef.convergence = TRUE). Figure 6 shows that the R̂ value of each
variable is smaller than 1.1 indicate that the imputation is converged.
R> plot(as.bugs.array(IMP@mcmc))
If the imputation is not converged or if conditional models seem to be fit reasonably but
users may still want to run the imputation longer, users can do this by feeding the previous
returned mi object into mi(). Imputation will continue from where it left off. If the previous
mi() object is converged, you have to specify run.past.convergence = TRUE to force mi()
to run for more iterations.
R> IMP <- mi(IMP, run.past.convergence = TRUE, n.iter = 10)
Beginning Multiple Imputation ( Wed Feb 09 16:51:21 2011 ):
Iteration 21
Chain 1 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 2 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 3 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Iteration 22
Chain 1 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 2 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 3 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
...
Iteration 38
Chain 1 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 2 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
Chain 3 : h39b.W1.mi.log age.W1.mi.log c28.W1.mi.log pcs.W1.mi.log mcs37.W1 b05.W1 haartadhere.W1 h39b.W1.mi.ind
mi converged ( Wed Feb 09 16:52:13 2011 )

4.4. Analysis
One of the nice features of multiply imputed data is that we can conduct analyses as if the data
were complete. Results from an analysis performed on each dataset must then be combined
in a sensible way, for instance by using formulas proposed by Rubin (1987).
 Journal of Statistical Software 23

Obtaining completed datasets

If the users prefer to perform separate data analyses for each dataset by themselves, they
can easily extract the completed datasets from the mi object via mi.completed(). This will
return a list that contains multiple datasets.

R> IMP.dat.all <- mi.completed(IMP)

They can extract just one dataset from a specific chain of imputations via mi.data.frame().

R> IMP.dat <- mi.data.frame(IMP, m = 1)

mi also offers an option to output these multiply imputed datasets into files. Currently, mi
only supports three data formats: csv, dta, and table. The default output data format is
csv.

R> write.mi(IMP)

The output files shall be stored under the working directory. The file names will be
midata1.csv, midata2.csv, midata3.csv, . . ., and so on.

Pooling the complete case analysis on multiply imputed datasets

mi() facilitates the analysis process by providing functions that perform these separate anal-
yses and then combine the separate estimates into one estimate and standard error. mi
currently offers seven regression functions: lm.mi(), glm.mi(), polr.mi(), bayesglm.mi(),
bayespolr.mi(), lmer.mi(), and glmer.mi().

R> fit <- lm.mi(h39b.W1 ~ age.W1 + c28.W1 + pcs.W1 + mcs37.W1 +

+ b05.W1 + haartadhere.W1, IMP)
R> display(fit)

=======================================
Separate Estimates for each Imputation
=======================================

** Chain 1 **
lm(formula = formula, data = mi.data[[i]])
coef.est coef.se
(Intercept) 14.78 1.46
age.W1 -0.06 0.02
c28.W1 -0.44 0.10
pcs.W1 -0.04 0.02
mcs37.W1 1.20 0.43
b05.W1 -0.99 0.14
haartadhere.W1 -1.52 0.22
---
n = 532, k = 7
residual sd = 4.36, R-Squared = 0.24
24 mi: Multiple Imputation with Diagnostics in R

** Chain 2 **
lm(formula = formula, data = mi.data[[i]])
coef.est coef.se
(Intercept) 14.11 1.47
age.W1 -0.07 0.02
c28.W1 -0.33 0.10
pcs.W1 -0.04 0.02
mcs37.W1 1.54 0.44
b05.W1 -1.03 0.15
haartadhere.W1 -0.88 0.22
---
n = 532, k = 7
residual sd = 4.37, R-Squared = 0.20

** Chain 3 **
lm(formula = formula, data = mi.data[[i]])
coef.est coef.se
(Intercept) 15.93 1.47
age.W1 -0.11 0.02
c28.W1 -0.22 0.09
pcs.W1 -0.03 0.02
mcs37.W1 1.21 0.43
b05.W1 -1.11 0.15
haartadhere.W1 -1.03 0.22
---
n = 532, k = 7
residual sd = 4.38, R-Squared = 0.21

=======================================
Pooled Estimates
=======================================
lm.mi(formula = h39b.W1 ~ age.W1 + c28.W1 + pcs.W1 + mcs37.W1 +
b05.W1 + haartadhere.W1, mi.object = IMP)
coef.est coef.se
(Intercept) 14.94 1.81
age.W1 -0.08 0.04
c28.W1 -0.33 0.16
pcs.W1 -0.03 0.02
mcs37.W1 1.32 0.49
b05.W1 -1.05 0.16
haartadhere.W1 -1.14 0.45
---

Nonetheless, if users are running regression models that are not supported by mi, mi offers a
function mi.pooled() that allows them to pool the regression estimates together with different
chains. The following code demonstrates that by feeding mi.pooled() a list of coefficients
 Journal of Statistical Software 25

and standard deviations of these coefficients, mi.pooled() will produce the pooled estimates
for the users.

R> m <- m(IMP)

R> coefs <- sds <- vector("list", m)
R> for(i in 1:m) {
+ dat <- mi.data.frame(IMP, i)
+ fit <- lm(h39b.W1 ~ c28.W1, data = dat, subset = mcs37.W1 == 1)
+ coefs[[i]] <- coef(fit)
+ sds[[i]] <- se.coef(fit)
+ }
R> mi.pooled(coefs, sds)

$coefficients
(Intercept) c28.W1
6.4413698 -0.2381554

$se
(Intercept) c28.W1
1.2055315 0.2931557

4.5. Validation
The validation step is still under construction and is not included in the current version.
However, we present some ideas of the ways in which users can validate their results obtained
from mi.

Sensitivity analysis
Multiple imputation is based on many assumptions about the conditional models. Thus it is
natural to test how sensitive imputed values are to these assumptions. Since mi is extremely
flexible about the model specification and also provides diagnostics to examine model fit,
the user can check the sensitivity of the results from pooled analyses to changes in model
specification within the subset of models that appear to be equally plausible based on the
graphical diagnostics.

Cross validation
We can use cross-validation to explore robustness to violations of our structural assumptions.
For instance, if we want to test the assumption of missing at random, after obtaining the
completed dataset (original data plus imputed data) using mi, we can create missing values
on these imputed datasets based on a missing at random missing data mechanism that mimics
patterns of missingness actually found in the data and re-impute the missing data (Gelman,
King, and Liu 1998). By comparing the imputed dataset before and after this test, we can
assess how our procedure is likely to fair under this particular assumption of missing at
random. Similar checks could be performed to assess performance of the procedure under the
assumption of data not missing at random.
26 mi: Multiple Imputation with Diagnostics in R

Compatibility check
We will use graphical diagnostics to assess the extent to which our conditional models fail to
provide consistent information about the underlying joint distribution of the data and explore
the extent to which such incompatibility might impact our results.

4.6. Interactive interface

mi has an interactive program where users do not have to type commands to perform multiple
imputation. By calling mi.iteractive() and giving it the data to be imputed, it will walk
the users through all the necessary settings and steps as discussed in the previous sections.

R> data("CHAIN")
R> IMP <- mi.interactive("CHAIN")

-----------------------------------
creating information matrix:
-----------------------------------

done

-----------------------------------
Would you like to:
-----------------------------------

1: look at current setting

2: proceed to mi with current setting
3: change current setting

Selection:

Additionally, migui (Lee and Su 2010) offers a graphical user interface (GUI) of mi where
users can do multiple imputation by clicking buttons. To call up the GUI, simply type the
followings in the R console:

R> library("migui")
R> migui()

5. Conclusions and future plans

The major goal of mi is to make multiple imputation transparent and easy to use for the
users. Here are four characteristics of the package that we believe are particularly valuable.

1. Graphical diagnostics of imputation models and convergence of the imputation process.

2. Use of Bayesian versions of regression models to handle the issue of separation.

 Journal of Statistical Software 27

Figure 7: The screenshot of the graphical user interface of mi.

3. Imputation model specification is similar to the way in which you would fit a regression
model in R.

4. Automatical detection of problematic characteristics of data such as high levels of

collinearity across variables followed by either a fix or an alert to the user. In par-
ticular, mi adds noise into the imputation process to solve the problem of additive
constraints.

As with many other software packages, mi is continually being augmented and improved.
One caution with the current incarnation is that mi may take some time to converge with big
datasets with a high rate of missingness across many variables. We are currently investigating
approaches to increase the computational efficiency of the algorithm.
Another future direction includes expanding the functionality of mi to allow for imputation of
time-series cross-sectional data, hierarchical or clustered data. Currently, it is only possible
to include group or time indicators as predictors in the imputation process to capture the
group-specific or time-specific aspect of missingness patterns. We would like to use multilevel
models or approximations to these models to these types of data (Gelman and Hill 2007).
Finally, as discussed in Section 4, we will incorporate tools and functions to perform sensitivity
analysis, cross-validation and compatibility checks for mi.

Afterward (December 1, 2011)

The bulk of this article was written in 2009 after the call for papers for this special issue. Since
then, our software and others’ have advanced in many ways. We are delighted to see that many
of the novel features of mi described in our paper have since been independently developed
and incorporated into other multiple imputation software. Our goal was and continues to be
to use mi as an open-source vehicle for disseminating ideas for flexible, understandable, and
28 mi: Multiple Imputation with Diagnostics in R

checkable missing data imputation. In particular, we are looking forward to seeing one of
the important features of mi – model checking and cross validation after imputation – to be
incorporated in other software, just as we will continue to take advantage of developments
elsewhere in improving our programs. The synergy available from open-source software should
be a general benefit.

Acknowledgments
We thank Peter Messeri for the CHAIN example, Maria Grazia Pittau for helpful discussions
and her early works on mi, Benjamin Goodrich for his current efforts on making mi better, the
US Nation Science Foundation, National Institutes of Health, Institute of Education Sciences,
and the Wang Xuelian Foundation for partial support of this research.

References

Abayomi K, Gelman A, Levy M (2008). “Diagnostics for Multivariate Imputations.” Journal

of the Royal Statistical Society C, 57(3), 273–291.

Bates D, Maechler M (2011). Matrix: Sparse and Dense Matrix Classes and Methods.
R package version 1.0-1, URL http://CRAN.R-project.org/package=Matrix.

Bates D, Maechler M, Bolker B (2011). lme4: Linear Mixed-Effects Models Using S4 Classes.
R package version 0.999375-42, URL http://CRAN.R-project.org/package=lme4.

Buuren SV, Groothuis-Oudshoorn K (2011). “mice: Multivariate Imputation by Chained

Equations in R.” Journal of Statistical Software, 45(3), 1–67. URL http://www.
jstatsoft.org/v45/i03/.

CHAIN (2009). “NY HIV Data CHAIN.” URL http://www.nyhiv.org/data_chain.html.

Cleveland WS (1979). “Robust Locally Weighted Regression and Smoothing Scatterplots.”

Journal of the American Statistical Association, 74(368), 829–836.

Fay RE (1996). “Alternative Paradigms for the Analysis of Imputed Survey Data.” Journal
of the American Statistical Association, 91(434), 490–498.

Fox J, Weisberg S (2011). car: Companion to Applied Regression. R package version 2.0-11,
URL http://CRAN.R-project.org/package=car.

Gelman A, Carlin JB, Stern HS, Rubin DB (2004). Bayesian Data Analysis. 2nd edition.
Chapman & Hall/CRC, Boca Raton, Fl.

Gelman A, Goegebeur Y, Tuerlinckx F, Van Mechelen I (2000). “Diagnostic Checks for

Discrete Data Regression Models Using Posterior Predictive Simulations.” Journal of the
Royal Statistical Society C, 49(2), 247–268.

Gelman A, Hill J (2007). Data Analysis Using Regression and Multilevel/Hierarchical Models.
Cambridge University Press, UK.
 Journal of Statistical Software 29

Gelman A, Jakulin A, Pittau MG, Su YS (2008). “A Weakly Informative Default Prior

Distribution for Logistic and Other Regression Models.” The Annals of Applied Statistics,
2(4), 1360–1383.
Gelman A, King G, Liu C (1998). “Not Asked and Not Answered: Multiple Imputation for
Multiple Surveys.” Journal of the American Statistical Association, 93(443), 846–857.
Gelman A, Mechelen IV, Verbeke G, Heitjan DF, Meulders M (2005). “Multiple Imputation
for Model Checking: Completed-Data Plots with Missing and Latent Data.” Biometrics,
61(1), 74–85.
Gelman A, Rubin DB (1992). “Inference from Iterative Simulation Using Multiple Sequences.”
Statistical Science, 7(4), 457–472.
Gelman A, Su YS, Yajima M, Hill J, Pittau MG, Kerman J, Zheng T (2011). arm: Data
Analysis Using Regression and Multilevel/Hierarchical Models. R package version 1.4-13,
URL http://CRAN.R-project.org/package=arm.
Heitjan DF, Little RJA (1991). “Multiple Imputation for the Fatal Accident Reporting Sys-
tem.” Journal of the Royal Statistical Society C, 40(1), 13–29.
Lee D, Su YS (2010). migui: Graphical User Interface of the mi Package. R package
version 0.00-08, URL http://CRAN.R-project.org/package=migui.
Little RJA, Rubin DB (2002). Statistical Analysis with Missing Data. 2nd edition. John
Wiley & Sons, Hoboken.
Meng XL (1994). “Multiple-Imputation Inferences with Uncongenial Sources of Input.” Sta-
tistical Science, 9(4), 538–558.
Messeri P, Lee G, Abramson DM, Aidala A, Chiasson MA, Jessop DJ (2003). “Antiretroviral
Therapy and Declining AIDS Mortality in New York City.” Medical Care, 41(4), 512–521.
Plate T, Heiberger R (2011). abind: Combine Multi-Dimensional Arrays. R package ver-
sion 1.3-0, URL http://CRAN.R-project.org/package=abind.
Plummer M, Best N, Cowles K, Vines K (2006). “coda: Convergence Diagnosis and Out-
put Analysis for MCMC.” R News, 6(1), 7–11. URL http://CRAN.R-project.org/doc/
Rnews/.
Raghunathan TE, Lepkowski JM, Van Hoewyk J, Solenberger P (2001). “A Multivariate
Technique for Multiply Imputing Missing Values Using a Sequence of Regression Models.”
Survey Methodology, 27(1), 85–95.
R Development Core Team (2011). R: A Language and Environment for Statistical Computing.
R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URL http:
//www.R-project.org/.
Robins JM, Wang N (2000). “Inference for Imputation Estimators.” Biometrika, 87(1), 113–
124.
Rubin DB (1987). Multiple Imputation for Nonresponse in Surveys. John Wiley & Sons, New
York.
30 mi: Multiple Imputation with Diagnostics in R

Schenker N, Taylor JMG (1996). “Partially Parametric Techniques for Multiple Imputation.”
Computational Statistics & Data Analysis, 22(4), 425–446.

Sturtz S, Ligges U, Gelman A (2005). “R2WinBUGS: A Package for Running WinBUGS

from R.” Journal of Statistical Software, 12(3), 1–16. URL http://www.jstatsoft.org/
v12/i03/.

Venables WN, Ripley BD (2002). Modern Applied Statistics with S. 4th edition. Springer-
Verlag, New York.

Affiliation:
Yu-Sung Su
Department of Political Science
School of Humanities and Social Sciences
Tsinghua University
Room 153 Minzhai, Qinghua Yuan, Haidian District
Beijing, 100084, China
E-mail: suyusung@tsinghua.edu.cn

Andrew Gelman
Department of Statistics
Columbia University
1255 Amsterdam Avenue
New York, NY 10027, United States of America
E-mail: gelman@stat.columbia.edu
URL: http://www.stat.columbia.edu/~gelman/

Jennifer Hill
Department of Humanities and Social Sciences
Steinhardt School of Culture, Education and Human Development
New York University
246 Greene Street
New York, NY 10003, United States of America
E-mail: jennifer.hill@nyu.edu

Masanao Yajima
Department of Statistics
University of California, Los Angeles
9407 Boelter Hall
 Journal of Statistical Software 31

Los Angeles, CA 90095-1554, United States of America

E-mail: yajima@ucla.edu
URL: http://www.stat.ucla.edu/~yajima/

Journal of Statistical Software http://www.jstatsoft.org/

published by the American Statistical Association http://www.amstat.org/
Volume 45, Issue 2 Submitted: 2009-06-15
December 2011 Accepted: 2011-05-30