Institut f. Statistik u.

Wahrscheinlichkeitstheorie
1040 Wien, Wiedner Hauptstr. 8-10/107
AUSTRIA
http://www.statistik.tuwien.ac.at
Visualization of missing values using
the R-package VIM
M. Templ and P. Filzmoser
Forschungsbericht CS-2008-1
Mai 2008
Kontakt: P.Filzmoser@tuwien.ac.at
Visualization of Missing Values using the R-Package VIM
Matthias Templ
1,2
and Peter Filzmoser
1
1
Department of Statistics and Probability Theory, Vienna University of Technology, 1040
Vienna, Austria
2
Statistics Austria, 1110 Vienna, Austria
Abstract: This paper introduces new tools for the visualization of missing values. The
tools can be used for exploring the data and the structure of the missing values. Depending
on this structure, the tools can be helpful for identifying the mechanism generating the
missings. This knowledge is important for selecting an appropriate imputation method to
reliably estimate the missing values.
The visualization tools are implemented in the R-package VIM (visualization and
imputation of missing values). A graphical user interface allows an easy handling of the
plot methods. The developed tool can be used for data from ofcial statistics, but also for
data from various other elds. Special attention is given to data with spatial coordinates,
and in that case the missing data information can be displayed with maps.
1 Introduction
Data often contain missing values, and the reasons are manifold. Missing values occur
when measurements fail, or in case of non-respondents in surveys, when analysis results
get lost, or when measurements do not fulll some prior knowledge (edits) of the data,
i.e. are implausible. Examples for missing values in the natural sciences are broken
measurement units for measurements of ground water quality or temperature, lost soil
samples in geochemistry, or soil samples which must be re-analyzed but where the soil
samples are exhausted. Examples for missing values in ofcial statistics are respondents
who deny information about their income, small companies which do not report their
turnover, values which do not fulll pre-dened editing rules, etc.
Missing values are often of great interest and they must be replaced by meaningful
values. Moreover, most of the standard statistical methods can only be applied to complete
data, and deleting whole columns or rows of data matrices where missing values appear
would result in a loss of important available information. The estimation of missing values
is known under the name imputation (Little and Rubin, 1987). In order to be able to
choose a proper imputation method one must be aware of the missing data mechanism(s).
The quality of the imputed values depends on the imputation itself and on the imputation
method used.
Although there exists a comprehensive literature on the estimation of missing values
(see, e.g., Little and Rubin, 1987; Schafer, 1997), the visualization of missings is treated
only in a few papers (e.g. Eaton et al., 2005; Cook and Swayne, 2007). This is also
reected in the statistical software. Visualization tools for missing values are rarely or not
implemented in SAS, SPSS, or STATA, and only few plots are available in R and GGOBI
(see, e.g., Cook and Swayne, 2007).
1
In this paper we introduce several graphical presentations to visualize missing values.
We will demonstrate the usefulness of the plots with real data (Section 2). With an ap-
propriate visualization of the missings it can also be possible to detect the missing values
mechanism (see e.g. in Figure 8) which is important for the selection of an appropriate
imputation method.
All visualization tools shown in the following are implemented in the R-package VIM
(Visualization and Imputation of Missing Values) which has been written by the rst
author of this paper. A screen-shot of the implementation and few comments about the
usage of the package are given in Section 3.
1.1 Missing Values Mechanisms
There are three important cases to distinguish that are the responsible generating processes
behind the missing values (see Rubin, 1976; Little and Rubin, 1987; Schafer, 1997). The
missing values are Missing Completely At Random (MCAR) if the distribution of miss-
ingness neither depends on the observed part X
obs
nor on the missing part X
miss
. Thus
the probability of missingness is given by
P(X
miss
|X) = P(X
miss
)
with the complete data X = (X
obs
, X
miss
). In other words, the missing data mechanism
does not depend on the variable of interest, nor on any other variable which is observed
in the data set.
When the distribution of missingness depends on the observed part X
obs
, the missing
values are said to be Missing At Random (MAR), and the probability of missingness is
P(X
miss
|X) = P(X
miss
|X
obs
). (1)
Hence the distribution of missingness depends not on the missing part X
miss
.
When Equation (1) is violated and the patterns of missingness are in some way re-
lated to the outcome variables, i.e. the probability of missingness depends on X
miss
, the
missing data are said to be Missing Not At Random (MNAR). This relates to the equation
P(X
miss
|X) = P(X
miss
|(X
obs
, X
miss
)).
Hence, the missings can not be fully explained by the observed part of the data.
A practical example for the missing values mechanism which is adequate for the data
used in this paper is given by Little and Rubin (1987). Considering two variables age and
income the data are MCAR if the probability of missing is the same for all individuals,
regardless of their age or income. If the probability that income is missing varies accord-
ing to the age of the respondents (e.g. more missings for higher age) but does not vary
according to the income of respondents with the same age (e.g. for a considered age class
the distribution of the missings show a random pattern), then the missings in variable in-
come are MAR. However, if the probability that income is recorded varies according to
income for those with the same age (e.g. more missings for high income than for low
income in certain age classes), then the missings in variable income are MNAR (Little
and Rubin, 1987). Naturally, MNAR is difcult to be detected (see Section 1.2).
2
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
GG
GG
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
x
y
Complete data
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
GG
GG
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
x
y
Chosen mechanism: MNAR
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
GG
G
G
G
G
GG
GG
G
G
G
G
G
G
G
G
G
G
G
G
G
G
x
y
User knowledge
Figure 1: Simulated bivariate data set. LEFT: Complete data. MIDDLE: Red points are
chosen as missings in y depending on the value of y (MNAR). RIGHT: Information is
only available for x-values in practice.
Appropriate visualization tools for missing values should be helpful for distinguishing
between the three missing values mechanisms. However, there are some limitations that
will be described in the following.
1.2 Limitations for the Detection of the Missing Values Mechanism
In practice it is often difcult to detect the missing values mechanism exactly, because this
would require the knowledge of the missing values themselves (Little and Rubin, 1987).
In the following we want to give a simple example in order to show the limitations for the
detection of the missing values mechanism.
Figure 1 shows a highly correlated bivariate data set. From the complete data (graphic
on the left side of the gure) some observations are marked as missings in y (graphic in the
middle of the gure) depending on the value of y (the higher y the higher the probability
of missingness). So, the missing values mechanism is constructed as MNAR. In practice,
however, we only know the x-part of the observations with missing values in y (graphic
on the right of the gure) and can only observe that for increasing x-values the amount of
missingness also increases. Therefore, we will assume a MAR situation knowing that this
could also be a MNAR situation, i.e. we cannot distinguish between MAR and MNAR.
We know from our construction that after taking the relationship between x and the miss-
ing data pattern into account, the probability of missingness in y still depends on y (see
the graphic in the middle of Figure 1 where for approximately equal x-values only high
y-values were set to be missing). However, in real world situations we will determine
a MAR situation for this example because of the high correlation between x and y (see,
e.g., Little and Rubin, 1987). This results in good estimation by using well-established
imputation methods for MAR situations (see, e.g., Dempster et al., 1977).
3
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
x
y
Complete data
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G G
G
G
G G
G
G
G
G
G
G
G G
x
y
Chosen mechanism: MNAR
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
G
x
y
User knowledge
Figure 2: Simulated bivariate data set. LEFT: Complete data. MIDDLE: Red points are
chosen as missings in y depending on the value of y (MNAR). RIGHT: Information is
only available for x-values in practice.
In Figure 2 we see a similar picture as in Figure 1 but with uncorrelated variables.
Some y-values are again marked as missing depending on the value of y (MNAR). In
practice, however, we will detect a MCAR situation because the missingness seems to be
completely independent from the data values. On the other hand, we know that this could
also be a MNAR situation, and so we can not distinguish between MCAR and MNAR.
We can summarize that, for a bivariate data set where one variable contains missings
which are generated with the MNAR mechanism, we can have the following situations:
(a) The variables are correlated: MNAR would be classied as MAR, and thus we only
detect MCAR and MAR.
(b) The variables are uncorrelated: MNAR would be classied as MCAR, and thus we
only detect MCAR and MAR.
Multivariate data with missing values in several variables can make it even more compli-
cated to distinguish between the missing values mechanisms. The situation can become
even worse in case of outliers, inhomogeneous data or very skewed data distributions.
2 Visualization Methods for Missing Values
The visualization tools proposed in this section do not rely on any statistical model as-
sumptions. The tools are made available as R-package VIM, and through a graphical user
interface they are easy to handle.
We illustrate the visualization tools on a subset of the European Survey of Income and
Living Conditions (EU-SILC) from Statistics Austria from 2004, see Table 1. This very
famous and complex data set is mainly used for measuring poverty and for monitoring
4
Table 1: Explanation of variables used from the EU-SILC data set.
name meaning
py010n employee cash or near cash income
py035n contributions to individual private pension plans
py050n cash benets or losses from self-employment
py070n values of goods produced by own-consumption
py080n pension from individual private plans
py090n unemployment benets
py100n old-age benets
py110n survivor benets
py120n sickness benets
py130n disability benets
py140n education-related allowances
pek n net income
pek g gross income
P001000 employment situation
r007000 occupation
P033000 years of employment
P029000 hours worked per week, miscellaneous employment
P014000 profession
bundesld region
age age
sex sex
the Lisbon 2010 strategy of the European Union. This data set includes a high amount of
missings which were imputed with model based imputation methods (Statistics Austria,
2006). Since a high amount of missings are not MCAR one has to think about which
variables should be included for imputation. The proposed visualization tools are helpful
for this decision.
2.1 Aggregation Plot
It is often of interest, how many missing values are contained in the single variables.
Even more interesting, there can be certain combinations of variables with a high number
of missing values. Figure 3 shows this information. The plot on the left hand side shows a
bar for each considered variable, and the bar height corresponds to the number of missing
values in the variable. The aggregation plot on the right hand side shows for the same
variables (horizontal axis) all different combinations that are present in the observations
with missing and non-missing values (vertical axis). The color red indicates missingness,
the color blue represents available data. The numbers to the right are the frequencies of
observations to the corresponding combinations. For example, the top row represents a
combination where the rst three variables of an observation have missing values and
5
p
y
0
1
0
n
p
y
0
3
5
n
p
y
0
5
0
n
p
y
0
7
0
n
p
y
0
8
0
n
p
y
0
9
0
n
p
y
1
0
0
n
p
y
1
1
0
n
p
y
1
2
0
n
p
y
1
3
0
n
p
y
1
4
0
n
p
e
k
_
n
N
u
m
b
e
r

o
f

M
i
s
s
i
n
g
s
0
1
0
0
2
0
0
3
0
0
4
0
0
C
o
m
b
i
n
a
t
i
o
n
s
3827
8
13
2
2
53
1
5
1
23
1
2
2
11
119
2
1
48
1
9
1
425
2
2
1
3
4
1
6
1
1
11
32
1
4
p
y
0
1
0
n
p
y
0
3
5
n
p
y
0
5
0
n
p
y
0
7
0
n
p
y
0
8
0
n
p
y
0
9
0
n
p
y
1
0
0
n
p
y
1
1
0
n
p
y
1
2
0
n
p
y
1
3
0
n
p
y
1
4
0
n
p
e
k
_
n
Figure 3: Number of missing values for a subsample of the EU-SILC data from Statistics
Austria. LEFT: Barplots for the number of missing values in each variable. RIGHT:
Aggregation plot showing all combinations of missing (red) and non-missing (blue) parts
in the observations, and the corresponding frequencies.
the remaining variables have no missings. This combination appears 4 times in the data.
Another example: The last row reects the 3827 observations that have no missing values.
As a result we observe an exceptionally high number of missings for variable py010n
(employee cash or near cash income). The combination with variable py035n (contribu-
tions to individual private pension plans) still has 32 missing values.
2.2 Matrix Plot
The matrix plot visualizes all cells of the data matrix by horizontal lines. Red lines are
drawn for missing values, and a grey scale is used for the available data values. For the
grey scale the data are rst scaled to the interval [0,1] for each variable by subtracting the
mean and dividing by the range. Then small values are assigned a light grey, high values
a dark grey and values equal to 0 are colored in white. Additionally, the observations can
be sorted by the magnitude of a selected variable.
Figure 4 shows a matrix plot of a subset of the EU-SILC data where the sorting was
done for variable pek n (net income). It can be seen that the higher the net income, the
more missing values in the variables py010y (employee cash or near cash income), and
py050n (employees income). Thus, the missing data mechanism was detected to be MAR
for these two variables which should be considered when applying imputation methods
on these variables.
6
0
1
0
0
0
2
0
0
0
3
0
0
0
4
0
0
0
I
n
d
e
x
P
0
0
1
0
0
r
0
0
7
0
0
p
y
0
1
0
n
p
y
0
3
5
n
p
y
0
5
0
n
p
y
0
9
0
n
p
y
1
0
0
n
p
e
k
_
n
r
e
g
i
o
n
a
g
e
Figure 4: Matrix plot of a subset of the EU-SILC data which are sorted by variable pek n.
2.3 Spineplot with Missings
When plotting a histogram of a variable one could show the amount of missings of a
second variable by splitting each bin into two parts. This is shown in Figure 5 (left), where
the histogram represents the frequencies of the variable age and the bins (age classes) are
split according to the numbers of missing (red) and observed (blue) values of variable it
py010n (employee cash or near cash income) for the corresponding age class.
Another way of presenting this information can be done by the spineplot (see, e.g.,
Hummel, 1996). Figure 5 (right) shows such a spineplot for the same two variables as
used in Figure 5 (left). The horizontal axis is scaled according to relative frequencies
of the age classes. The vertical axis shows the proportion of missing (red) and observed
(blue) values of variable py010n (employee cash or near cash income) for each age class,
and thus they sum up to 1. Since the area of the bin for each age interval reects the
number of values, it is now possible to compare the areas for the missing proportions
among the different age classes. For example, strictly decreasing red areas for increasing
age would indicate a MARsituation for variable py050n because the proportion of missing
values in py050n would then also strictly decrease. Another typical MAR situation could
be found when using variable py100n (old-age benets) and age. The higher the age the
higher the probability of missing values in py100n. In this spineplot, however, one can
only see that for age classes > 60 years the proportion of missing values of the employees
income becomes very small, which is logical.
2.4 Scatterplots with Missing Data Information in the Margins
Additionally to a standard scatterplot that shows the data values for a pair of variables,
information about missing values can be shown. Figure 6 uses the variables pek n (net
income) and py130n (unemployability) of the EU-SILC data and shows the scatterplot.
In addition, boxplots for missing (red) and available (blue) data are shown in the outer
7
F
r
e
q
u
e
n
c
y
20 40 60 80
0
1
0
0
2
0
0
3
0
0
4
0
0
5
0
0
F
r
e
q
u
e
n
c
y
20 40 60 80
0
1
0
0
2
0
0
3
0
0
4
0
0
5
0
0
57 42 54 53 71 66 52 55 33 6 2 0 1 2
age
missing in py010n
observed in py010n
missing in py010n
observed in py010n
20 30 40 50 60 70 80
age
15 20 25 30 35 40 45 50 55 60 65 75 85
m
i
s
s
.

i
n

p
y
0
1
0
n
o
b
s
e
r
v
e
d

i
n

p
y
0
1
0
n
0
.
0
0
.
2
0
.
4
0
.
6
0
.
8
1
.
0
Figure 5: Histogram (left) and Spineplot (right) of age with color coding of the amount
of missing and available data in variable py010n (employee cash or near cash income).
plot margins. For example, along the horizontal axis the two parallel boxplots both rep-
resent the variable net income, but the red boxplot is for those values of the net income,
where no values for unemployability are available, and the blue boxplot for net income
values, where also information for unemployability is available. A comparison of the two
boxplots can indicate the missing data mechanism. In this example the net income for
persons that provided no information for unemployability seems to be higher than that
where information is available.
The plot additionally shows a univariate scatterplot of the values that are missing in
the second variable (red points), and the frequency of these values by a number (lower
left). The number in the lower left corner (here 0) is the number of observations that are
missing in both variables.
This kind of bivariate scatterplot can easily be extended to a scatterplot matrix rep-
resenting all combinations of a set of variables by such scatterplots. Another extension
already included in the R-package VIM is that the information about missingness of more
than one variable is represented, allowing for more than two-dimensional relations.
2.5 Parallel Coordinate Plots with Missings
Parallel coordinate plots show each observation of the scaled data (usually interval scaled
in [0,1]) by a line, where the variables are presented by parallel axes (Wegman, 1990).
Similar to previous plots, the information of missingness of another variable can be color
coded. Figure 7 shows a parallel coordinate plot of a subset of the EU-SILC data where
the lines color refers to observations which are missing (red) of available (blue) for vari-
able py050n (employees income). Missingness in py050n is related with several of the
8
G G G G G G G G G G G GG G G
0
15
0
pek_n (net income), transformed
p
y
1
3
0
n

(
d
i
s
a
b
i
l
i
t
y

b
e
n
e
f
i
t
s
)
,

t
r
a
n
s
f
o
r
m
e
d
7 8 9 10 11
7
8
9
1
0
Figure 6: Scatterplot of net income (pek n) and unemployability (py130n) with missings
data information in the plot margins.
presented variables. In Figure 7 we can see that a high portion of small values in P033000
(years of employment) implies missings in py050n. Additionally, missings in py050n oc-
cur only for employees which have more than one employment (in this case the values
were set to 0 in variable P029000). Furthermore, Figure 7 shows that the amount of
missings depends on the actual values for the variables P001000 (different employment
situation) and bundesld (region). Finally, for variable pek g (gross income) missing values
only occur in a certain range.
2.6 Parallel Boxplots for Missing Values
Boxplots were already used in Figure 6 for comparing the values of a variable that are
either available or missing in a second variable. This can be extended for a comparison
of the information of missingness of several variables. Figure 8 shows the values of
variable pek n (net income) in form of a boxplot (left). The other boxplots shown in the
gure also refer to the values of pek n, but they are grouped according to missingness
(red) or non-missingness (blue) of each observation in another variable. Some of the
boxplots show a clear dependence between the magnitude of the values of pek n and the
presence of missing values. For example, missing values in variable py080n (annuity)
appear especially for high values of the net income. This indicates a clear MAR situation
for missings in py080n.
9
sex log(pek_g) P033000 P029000 P014000 P001000 age bundesld
observed in py050n
missed in py050n
Figure 7: Parallel coordinate plot for a subset of the EU-SILC data. The color indicates
missing values in variable py050n (employees income).

5
0
5
1
0

5
0
5
1
0
l
o
g
(

p
e
k
_
n

)
o
b
s
.

i
n

p
y
0
1
0
n
o
b
s
.

i
n

p
y
0
3
5
n
o
b
s
.

i
n

p
y
0
5
0
n
o
b
s
.

i
n

p
y
0
7
0
n
o
b
s
.

i
n

p
y
0
8
0
n
o
b
s
.

i
n

p
y
0
9
0
n
o
b
s
.

i
n

p
y
1
0
0
n
o
b
s
.

i
n

p
y
1
1
0
n
o
b
s
.

i
n

p
y
1
2
0
n
o
b
s
.

i
n

p
y
1
3
0
n
o
b
s
.

i
n

p
y
1
4
0
n
m
i
s
s
.

i
n

p
y
0
1
0
n
m
i
s
s
.

i
n

p
y
0
3
5
n
m
i
s
s
.

i
n

p
y
0
5
0
n
m
i
s
s
.

i
n

p
y
0
7
0
n
m
i
s
s
.

i
n

p
y
0
8
0
n
m
i
s
s
.

i
n

p
y
0
9
0
n
m
i
s
s
.

i
n

p
y
1
0
0
n
m
i
s
s
.

i
n

p
y
1
1
0
n
m
i
s
s
.

i
n

p
y
1
2
0
n
m
i
s
s
.

i
n

p
y
1
3
0
n
m
i
s
s
.

i
n

p
y
1
4
0
n
l
o
g
(

p
e
k
_
n

)
Figure 8: The values of a specic variable (here pek n net income) are grouped accord-
ing to missingness in another variable and presented in parallel boxplots. This grouping
is done for several variables of the EU-SILC data.
10
5
13
10
11
25
26
30
2
25
Figure 9: The number of missing values in variable py050n (employees income) are com-
puted for the nine political regions of Austria. Regions with a higher amount of missing
values (see the numbers in each region) receive a higher portion of red.
2.7 Plot Missings in Maps
If geographical coordinates are available for a data set, it can be of interest to check
whether missingness of a variable corresponds to spatial patterns in a map. For example,
the observations of a variable of interest could be drawn by growing dots in the map,
reecting the magnitude of the data values, and missingness in a second variable could be
color coded. This allows conclusions about the relation of missingness to the values of
the variable of interest, and to the spatial location.
For the EU-SILC data only the assignment of each observation to the nine political
regions of Austria is available, and not the spatial coordinates. Therefore, we can only
visualize the amount of missingness of a variable in the map. Figure 9 shows the map of
Austria with the nine political regions. The number of missing values of variable py050n
(employees income) is coded according to an rgb color scheme, resulting in a higher
portion of red for regions with a higher number of missing values for the considered
variable. Additionally, the numbers of missing values are shown in the regions.
3 R-Package VIM
All tools shown in Section 2 for visualizing missing values have been implemented in the
R-package VIM. A graphical user interface (GUI) which was developed with the help of
the R-package tcltk (R Development Core Team, 2007) allows an easy handling of the
functions. Figure 10 shows the main menu. For the visualization of the missing values
only the Data and the Visualization menu are important. In the Data menu one can load
data into the R workspace, select the data of interest, load a background map (optional),
and specify coordinates (optional). After selecting a data set a new frame opens (the right
frame of Figure 10) where variables can be selected. In the middle left area of Figure
11
Figure 10: The VIM package GUI menu together with the the active data selection frame,
the visualization workout sheet, the variable selection frame and the combo box for the
variable selection used for bivariate plots.
10 one can see the possible options and methods which can be chosen for visualizing the
active data set.
Besides the selection of variables one can transform and scale the data with various
techniques. The log-transformation or the box-cox transformation (Box and Cox, 1964)
which are common transformations for data from ofcial statistics can be used, but also
various other transformations are implemented, especially for geochemical data. When
selecting bivariate plots in the graphical user interface, a new selection of two variables
can be done simply by clicking on two variables in a combo box.
4 Conclusions
The detection of missing values mechanisms is usually done by statistical tests or models.
Visualization of the missing values can support the test decision, but also reveals more de-
tails about the data structure. Especially statistical requirements for a test can be checked
graphically, and problems like outliers or skewed data distributions can be discovered.
A graphical user interface was developed and implemented in R to produce several
12
possibilities for the visualization of data with missing values. The tools allow to combine
the information of missingness in a variable with other variables. Some plots allow for
an interactive handling, like the selection of the variable used for sorting in the matrix
plot (see Figure 4). The information resulting from the different graphics can be used for
detecting the missing values mechanism, and thus for selecting an appropriate method for
data imputation.
The tool can be used for data fromessentially any eld. If data with spatial coordinates
are available, the missing value information can be presented in maps. It is also possible
to load a background map and show it on the plot. Spatial patterns of missingness can be
very instructive for example in environmental sciences.
The package VIM is available on the comprehensive R archive network (CRAN).
References
G.E.P. Box and D.R. Cox. An analysis of transformations. Journal of the Royal Statistical
Society, pages 211252, 1964.
D. Cook and D.F. Swayne. Interactive and Dynamic Graphics for Data Analysis With R
and GGobi. Springer, New York, 2007. ISBN: 978-0-387-71761-6.
A.P. Dempster, N.M. Laird, and D.B. Rubin. Maximum likelihood for incomplete data
via the EM algorithm (with discussions). Journal of the Royal Statistical Society, 39:
138, 1977.
C. Eaton, C. Plaisant, and T. Drizd. Visualizing missing data: Graph interpretation user
study. In Human-Computer Interaction - INTERACT 2005, Lecture Notes in Computer
Sciences, Springer, pages 861872, 2005. ISBN 978-3-540-28943-2.
J. Hummel. Linked bar charts: Analysing categorical data graphically. Computational
Statistics, 11:2333, 1996.
R.J.A. Little and D.B. Rubin. Statistical analysis with missing data. Wiley, New York,
1987.
R Development Core Team. R: A language and environment for statistical com-
puting. R Foundation for Statistical Computing, Vienna, Austria, 2007. URL
http://www.R-project.org. ISBN 3-900051-07-0.
D.B. Rubin. Inference and missing data. Biometrika, 63:581592, 1976.
J.L. Schafer. Analysis of incomplete multivariate data. Chapman & Hall, 1997.
Statistics Austria. Einkommen, armut und lebensbedingungen 2004, ergebnisse aus eu-
silc 2004, 2006. Vienna, Austria. ISBN 3-902479-59-0.
E.J. Wegman. Hyperdimensional data analysis using parallel coordinates. Journal of the
American Statistical Association, 85:664675, 1990.
13