Dealing with Missing Values in Data

Ji Kaiser
Faculty of Civil Engineering, Czech Technical University
Czech Republic
[email protected]
Abstract: Many existing industrial and research data sets contain missing values due to various
reasons, such as manual data entry procedures, equipment errors and incorrect measurements.
Problems associated with missing values are loss of efficiency, complications in handling and
analyzing the data and bias resulting from differences between missing and complete data. The
important factor for selection of approach to missing values is missing data mechanism. There are
various strategies for dealing with missing values. Some analytical methods have their own approach
to handle missing values. Data set reduction is another option. Finally missing values problem can be
handled by missing values imputation. This paper presents simple methods for missing values
imputation like using most common value, mean or median, closest fit approach and methods based
on data mining algorithms like k-nearest neighbor, neural networks and association rules, discusses
their usability and presents issues with their applicability on examples.
Key words: missing data imputation, missing values, missing values imputation, missing data
mechanisms, closest fit, k-nearest neighbor, neural networks, association rules, data mining

1.

Introduction

Many existing industrial and research data sets contain missing values. Data sets contain missing
values due to various reasons, such as manual data entry procedures, equipment errors and incorrect
measurements. It is usual to find missing data in most of the information sources used.
Missing values usually appears as NULL values in database or as empty cells in spreadsheet table.
Some flat-file formats use various symbols for missing values e.g. arff files uses ? symbol for
missing values. These forms of missing values can be easily detected. However missing values can
also appears as outliers or wrong data (i.e. out of boundaries). These data must be removed before
intended analysis, and are much harder to find out.
The paper presents an overview missing values problem and strategies for dealing with missing
values in data. Main part of the paper is dedicated to missing values imputation methods, discusses
their usability and presents issues with their applicability on examples.

2.

The Missing Values Problem

Missing value is a value that we intended to obtain during data collection (interview, measurement,
observation) but we didnt because of various reasons. Missing values can appear because
respondent did not answer all questions in questionnaire, during manual data entry process, incorrect
measurement, faulty experiment, some data are censored or anonymous and many others.
Luengo, J. (2011) introduced three problems associated with missing values as
- loss of efficiency,
- complications in handling and analyzing the data,
- bias resulting from differences between missing and complete data.
Loss of efficiency is caused by time-consuming process of dealing with missing values. This is closely
connected to the second problem - complications in handling and analyzing the data. Complications in
handling and analyzing data lie in fact that most methods and algorithms are usually unable to deal
with missing values and missing values problem must be solved prior to analyses during data
preparation phase. The other problem - bias resulting from differences between missing and complete
data lies in fact that imputed values are not exactly the same as known values of completed data set
and should not be handled the same way. The same problem occurs also if data set is reduced and
some cases (rows of data set) are removed or ignored.

42

JOURNAL OF SYSTEMS INTEGRATION 2014/1

DEALING WITH MISSING VALUES IN DATA

3.

Characteristics of Missing Data Mechanisms

The missing data mechanism is usually classified as missing completely at random, missing at random
or not missing at random (Lakshminarayan, K., Harp S. A. & Samad, T. 1999), (Horton, N. J. &
Kleinman, K. P. 2007). Unfortunately identification of missing data mechanism is not always easy.

3.1

Missing Completely at Random (MCAR)

The missing data mechanism is considered as missing completely at random (MCAR), when the
probability of a record having a missing value for an attribute does not depend on either the observed
data or the missing data. MCAR is sometimes called uniform non-response. An example of a MCAR
mechanism would be that a laboratory sample is dropped, so the resulting observation is missing.
Data which is missing due to structural reasons cannot be regarded as MCAR (London School of
Hygiene and Tropical Medicine, 2013).

3.2

Missing at Random (MAR)

The missing data mechanism is considered as missing at random (MAR), when the probability of
a record having a missing value for an attribute could depend on the observed data, but not on the
value of the missing data itself. Data which is incomplete only due to structural reasons are MAR.
Horton, N. J. & Kleinman, K. P. (2007) describe MAR mechanism as states that the missingness
depends only on observed quantities, which may include outcomes and predictors (in which case the
missingness is sometimes labeled covariate dependent missingness (CDM)).
A special case of MAR is uniform non-response within classes. For example (London School of
Hygiene and Tropical Medicine, 2013), suppose we seek to collect data on income and property tax
band. Typically, those with higher incomes may be less willing to reveal them. If we have everyone's
property tax band and given property tax band non-response to the income question is random, then
the income data is missing at random. The reason (or mechanism) for it being missing depends on
property band. Given property band missingness does not depend on income itself.

3.3

Not Missing at Random (NMAR)

The missing data mechanism is considered as not missing at random (NMAR), when the probability of
a record having a missing value for an attribute could depend on the value of the attribute. Missing
data mechanism that is considered as NMAR is non-ignorable. This can be solved by going back to
the source of data and obtaining more information about the mechanism or obtain complete data set.
Unfortunately it is very rare to know the appropriate model for the missing data mechanism.
Examples NMAR missing data mechanism (Lakshminarayan, K., Harp S. A. & Samad, T. 1999)
include a sensor not detecting temperatures below a certain threshold, people not filling in yearly
income in surveys if the income exceeds a certain value.
Some sources, e.g. (Horton, N. J. & Kleinman, K. P. 2007), called this mechanism missing not at
random (MNAR).

4.

Strategies for Dealing with Missing Data

There are three main strategies for dealing with missing data. The simplest solution for the missing
values imputation problem is the reduction of the data set and elimination of all samples with missing
values (Kantardzic, M. 2003). Another solution is to treat missing values as special values. Finally
missing values problem can be handled by various missing values imputation methods. Unfortunately
missing values imputation methods are suitable only for missing values caused by missing completely
at random (MCAR) and some of them for missing at random (MAR) mechanism. If missing values are
caused by not missing at random mechanism (NMAR) it must be handled by going back to the source
of data and obtaining more information or the appropriate model for the missing data mechanism have
to be taken into account.

4.1

Using Missing Values Policy of Used Analytical Method

There is no need to use a special method for dealing missing values if method that is used for data
analysis has its own policy for handling missing values. Decision rules extraction methods may

JOURNAL OF SYSTEMS INTEGRATION 2014/1

43

JI KAISER

consider attributes with missing values as irrelevant (Luengo, J. 2011). Association rules extraction
methods may ignore rows with missing values (conservative approach) or handle missing values as
they are supporting the rule (optimistic approach) or are in contrary with the rule (secured approach)
as described by Berka, P. (2003).

4.2

Reducing the Data Set

The simplest solution for the missing values imputation problem is the reduction of the data set and
elimination of all missing values. This can be done by elimination of samples (rows) with missing
values (Kantardzic, M. 2003) or elimination of attributes (columns) with missing values
(Lakshminarayan, K., Harp S. A. & Samad, T. 1999). Both approaches can be combined. Elimination
of all samples is also known as complete case analysis.
Elimination of all samples is possible only when large data sets are available, and missing values
occur only in a small percentage of samples and when analysis of the complete examples will not lead
to serious bias during the inference. Elimination of attributes with missing values during analysis is not
possible solution if we are interested in making inferences about these attributes. Both approaches are
wasteful procedures since they usually decrease the information content of the data.

4.3

Treating Missing Attribute Values as Special Values.

This method deals with the unknown attribute values using a totally different approach. Rather than
trying to find some known attribute value as its value, we treat missing value itself as a new value for
the attributes that contain missing values and treat it in the same way as other values (GrzymalaBusse J. W., Hu M. 2001). Instead of storing value of the attribute we store the information that the
value is missing. This approach assumes that we handle these values as they dont influence future
analyses.

4.4

Replace Missing Value with Mean

This method replaces each missing value with mean of the attribute (Kantardzic, M. 2003). The mean
is calculated based on all known values of the attribute. This method is usable only for numeric
attributes and is usually combined with replacing missing values with most common attribute value for
symbolic attributes.

4.5

Replace Missing Value with Mean for the Given Class

This method is similar to the previous one. The difference is in that the mean is not calculated from all
known values of the attribute, but only attributes values belonging to given class are used. This
approach is possible only for classification problems where samples are classified in advance
(Kantardzic, M. 2003) or there is a possibility to create the classes.

4.6

Replace Missing Value with Median for the Given Class

Since the mean is affected by the presence of outliers it seems natural to use the median instead just
to assure robustness. In this case the missing data for a given attribute is replaced by the median of
all known values of that attribute in the class where the instance with the missing value belongs
(Acua, E. & Rodriguez, C. 2004). This method is usable only for numeric attributes and requires
existence of classes or possibility to create classes as previous method

4.7

Replace Missing Value with Most Common Attribute Value

This method simply uses most common attribute value for missing value imputation (Grzymala-Busse
J. W., Hu M. 2001). The most common value of all values of the attribute is used. This method is
usable only for symbolic attributes and is usually combined with replacing missing values with missing
values imputation using mean for numeric attributes.

4.8

Concept Most Common Attribute Value

This method is similar to the previous one. The concept most common attribute value method is
a restriction of the previous method to the concept (Grzymala-Busse J. W., Hu M. 2001). This method

44

JOURNAL OF SYSTEMS INTEGRATION 2014/1

DEALING WITH MISSING VALUES IN DATA

uses most common value of the attribute but uses cases belonging to the given class or concept
instead of using global most common value. This method is usable for symbolic attributes and requires
existence of classes or possibility to create classes.

4.9

Closest Fit

The closest fit algorithm (Grzymala-Busse, J. W., Goodwin, L. K., Grzymala-Busse, W. J. & Zheng X.
2005) for missing attribute values is based on replacing a missing attribute value with an existing value
of the same attribute from another case that resembles as much as possible the case with missing
attribute values. The main idea is based on searching through the set of all cases, considered as
vectors of attribute values, for a case that is the most similar to the given case with missing attribute
values. E.g. if 5 attributes of given case is known and 6th is being searched, the 6th attribute value is
taken from the case which has other 5 attributes values most similar to the given case. The proximity
measure between two cases x and y is the Manhattan distance between x and y, i.e.,
n

distance ( x, y )

distance ( xi , y i ) ,

(1)

i 1

Where

where r is the difference between the maximum and minimum of the known values.
Problem with using only one closest case may lied in replacing missing value by an outlier. The
problem is illustrated by following example (Tab. 1):
Tab. 1: Closest fit example
Property table
2

floor area [m ] rental price [CZK]

45

10 000

52

13 000

52

11 000

54

55

18 000

62

13 500

62

12 000

Missing value imputation using only one closest case leads to replace missing value with value 18 000
2
2
(from the case with floor area 55 m ), but there are properties with even larger floor area (62 m ) with
lower price.
This problem may be partially solved by using more than one closest case.
Another problem is in assumption of same distance between all values of symbolic attributes. Some
attributes can be measured on ordinal scale. E.g. the C grade is closer to B than to A, but it is not
possible to specify exactly the distance between these values. Described method uses Manhattan
distance but there are several other ways of measuring distance including Euclidean distance and
many others.

JOURNAL OF SYSTEMS INTEGRATION 2014/1

45

JI KAISER

4.10 Missing Values Imputation Using k-Nearest Neighbor

The closest fit method looks for the most similar case with known attribute value. K nearest neighbor
usually looks for more than one (k) similar cases with known values of the attribute, whose value is
being searched. Nearest neighbor methods are usually used to classify objects. To classify a new
object, with input vector y, we simply examine the k closest data set points to y and assign the object
to the class that has the majority of points among these k (Hand, D., Mannila, H. & Smyth, P. 2001).
The k-nearest neighbor algorithm does not create explicit models (Batista, G.E.A.P.A. & Monard, M.C.
2003). There are several ways for using known values of k-nearest neighbor. One possibility for
numeric attributes is to impute a weighted mean of the k nearest-neighbors attribute values (The
MathWorks, Inc. 2013). The weights are inversely proportional to the distances from the neighboring
columns.
The main question is selection of k parameter. Another question is handling symbolic attributes.
There are several metrics that can be used for measuring the distance (The MathWorks, Inc. 2013). It
is possible to use Euclidean distance, Manhattan distance and many others. HUANG, Ch. & LEE, H.
(2004) introduced k-nearest neighbor missing values imputation method that uses grey relational
analysis to determine the relationships among a referential observation and compared observations by
calculating the grey relational coefficient and the grey relational grade.
Some implementations of the k-nearest neighbor method for missing values imputation (Improved
outcomes software 2004) are able to compute the k nearest neighbors only on complete datasets so
missing values have to be filled in with an initial approximation.

4.11 Missing Values Imputation Using Neural Networks

Neural networks constitute a class of predictive modeling system that works by iterative parameter
adjustment (Chen, Z. 2001). The network structure, also called topology or architecture, includes the
neural framework (number of neurons, number of layers, neuron model type, etc.) and the
interconnection structure. Single-layer network has only an input and output layer. In a multilayer
network, one or more hidden layers are inserted between the input and the output layer.
Gupta, A. & Lam, M. (1998) introduced following procedure for reconstruction of missing values using
multilayered networks and backpropagation algorithm:
Step 1 Collect all training cases without any missing value and call them the complete set.
Step 2 Collect all training and test cases with at least one missing value and call them the
incomplete set.
Step 3 For each pattern of missing values, construct a multi-layered network with the number
of input nodes in the input layer equal to the number of non-missing attributes, and the
number of output nodes in the output layer equal to the number of missing attributes.
Each input node is used to accept one non-missing attribute, and each output node to
represent one missing attribute.
Step 4 Use the complete set and the backpropagation algorithm to train each network
constructed in step 3. Since the complete set does not have missing values, different
patterns of input-output pairs can be obtained from the complete set to satisfy the
input-output requirements for different networks from step 3. As the output of
a network is between 0 and 1, data have to be converted to values between 0 and 1
for this reconstruction procedure.
Step 5 Use the trained networks from step 4 to calculate the missing values in the incomplete
set.
The main question is strategy of neural networks construction. There are two main approaches. The
first one is to start with a simple network and add extra nodes to the network until such addition does
not improve the network performance. The second approach is to start with a large network and delete
nodes from the network as long as the deletion does not deteriorate the network performance
(commonly referred to as the node-pruning method). Fahlman S. E. & Lebiere, Ch. (1991) proposed
a network construction method called cascade-correlation under the first approach. A node-pruning
method, relevance assessment, was proposed by Mozer, M. C. & Smolensky, P. (1989).

46

JOURNAL OF SYSTEMS INTEGRATION 2014/1

DEALING WITH MISSING VALUES IN DATA

Another question is handling symbolic attributes, because step 4 of the procedure requires data
conversion to values between 0 and 1.

4.12 Missing Values Imputation Using Association Rules

An association rule is a simple probabilistic statement about the co-occurrence of certain events. For
binary variables association rule takes the following form (Hand, D. J., Manilla, H. & Smyth, P. 2001):

(2)

where A, B, and C are variables and

...

(3)

i.e., the conditional probability that C = 1 given that A = 1 and B = 1. The conditional probability p is
referred to as the confidence of the rule, and p (A = 1, B = 1, C = 1) is referred to as the support.
The support can be used as a constraint of minimum count of cases supporting association rule. The
If part of the rule is often called antecedent and the then part is often called consequent (Larose D.
T. 2005).
There are several ways of generating association rules, e.g.: apriori algorithm (Berka, P. 2003), (Hand,
D. J., Manilla, H. & Smyth, P. 2001), (Kantardzic M. 2003), GUHA method (Berka, P. 2003) and other
algorithms derived from apriori algorithm (Burita L. et al., 2012) or (Hipp, J., Gntzer, U. & G.
Nakhaeizadeh, G. 2000).
The input for missing values imputation is incomplete data set. Algorithms for association rules
generation are usually unable to handle missing values. Some association rules extraction methods
may ignore rows with missing values (conservative approach) or handle missing values as they are
supporting the rule (optimistic approach) or are in contrary with the rule (secured approach) as
described by Berka, P. (2003).
Another possibility is getting complete data set for association rules generation. First way to get
complete data set for association rules generation is reducing the data set.by elimination of cases
and/or attributes as described in one of previous chapters. Another way to get complete data set for
association rules generation is to handle missing values as special values. Using special values to
obtain complete data set for association rules can be recommended because association rules with
special values can be easily ignored. Association rules cant be often generated directly from numeric
attributes. Missing values imputation may be directly used only for symbolic attributes.
Algorithms for association rules generation usually generates rules based on setting of minimum
support, minimum confidence and maximum number of rules parameters. Most analyses, that use
association rules methods, use only few association rules. For missing values imputation usually many
more rules are required because not all association rules are suitable for missing values imputation.
Useable association rule for missing values imputation has the consequent containing value of
attribute whose value is being searched and the antecedent correspond to values of other given case
attributes. If complete data set for association rules generation was obtained by replacing missing
values by special values association rules with these special values in consequent must be omitted. It
is possible to use only association rules with consequent length equal to 1 and impute missing values
one by one in cases with more than one missing values. Another possibility is using association rules
with consequent length equal to the count of missing values in given case.
Support and confidence of the rule are two criteria that should be maximized during making decision
about using the association rule for missing value imputation. For missing value imputation can be
used the rule with maximum confidence along with required support. It is also possible to ignore
support of the rule and use only confidence. Setting required min. support of the rule highly depends
on given data set.
There is a possibility to combine association rules approach with most common attribute value
method. At first the most common attribute value method can be used if it is not possible to find
suitable association rule with required support for missing value imputation. At second the most
common attribute value method can be used if there is no suitable association rule with required
support and confidence not lower then relative frequency of occurrence of most common attribute

JOURNAL OF SYSTEMS INTEGRATION 2014/1

47

JI KAISER

value. It is also possible to use only few association rules to improve missing values imputation
accuracy using the most common attribute value.
Following tables and figures show how missing values imputation accuracy may be influenced by
setting of requirement of minimum support of association rules. Analyses were made using 5
categorical attributes from STULONG study data set. Missing values were imputed using combination
of association rules and most common attribute value method. There were 5% of missing values
generated. Missing values were generated under MCAR conditions.
Following table and figure show example where minimum of incorrectly imputed values is reached with
ignored support parameter (minimum required support was set to 0 counts).
Tab. 2: Incorrectly imputed values using combination of association rules and most common
st
attribute values with different requirement of minimum association rules support (1 example)
Requirement of min. association rules support

Incorrectly imputed values

[count of cases]

[count]

[%]

<0;52>

538

20,3865

(52;65>

541

20,5002

(65;78>

547

20,7275

(78;233>

548

20,7654

(233; 2323>

546

20,6897

(2323; )

591

22,3948

Fig. 1: Count of incorrectly imputed values using combination of association rules and most
common attribute values with different requirement of minimum association rules support
st
(1 example)
Next table and figure show example where minimum of incorrectly imputed values is reached with
minimum required support set within the interval (60; 86> counts).

48

JOURNAL OF SYSTEMS INTEGRATION 2014/1

DEALING WITH MISSING VALUES IN DATA

Tab. 3: Incorrectly imputed values using combination of association rules and most common
nd
attribute values with different requirement of minimum association rules support (2 example)
Requirement of min. association rules support

Incorrectly imputed values

[count of cases]

[count]

[%]

<0; 2>

568

21,5233

(2; 3>

569

21,5612

(3; 38>

570

21,5991

(38; 60>

568

21,5233

(60; 86>

566

21,4475

(86; 2289>

567

21,4854

(2289; )

647

24,5169

Fig. 2: Count of incorrectly imputed values using combination of association rules and most
common attribute values with different requirement of minimum association rules support
nd
(2 example)

5.

Conclusion

Selection of missing values imputation method highly depends on given data set, structure of
attributes and missing data mechanism. Missing data mechanism is a key factor to decide if missing
values can be imputed using some of described methods. Missing data mechanism can be considered
as missing completely at random, missing at random or not missing at random. If missing data
mechanism is considered as not missing at random, imputation can not be done without knowledge of
this mechanism. Unfortunately missing data mechanism is usually unknown.
Some analytical methods have their own mechanism for dealing with missing data so missing values
imputation methods should be used only if necessary. It is also possible to use data set reduction by
eliminating all missing values. This can be done by eliminating cases (rows) or/and attributes
(columns) with missing values but this approach usually decrease the information content of the data.
Most often used missing values imputation methods are simple solutions like imputation using mean or
most common value of given attribute. These methods dont consider dependencies among attributes.

JOURNAL OF SYSTEMS INTEGRATION 2014/1

49

JI KAISER

Another possibility for missing values imputation is using described methods that are based on data
mining methods like k-nearest neighbor, neural networks or association rules. These methods are
more complicated and often dont represent exact procedure for imputation of missing values. Results
of missing values imputation may vary based on setting of various parameters as shown on examples
of missing values imputation using association rules.
Selection of missing values imputation method must be also done with consideration of structure of
given dataset attributes. Some methods more suits for numeric attributes and some for symbolic
attributes. Methods can be often combined.

Acknowledgements
The study (STULONG) was realized at the 2nd Department of Medicine, 1st Faculty of Medicine of
Charles University and Charles University Hospital, U nemocnice 2, Prague 2 (head. Prof. M.
Aschermann, MD, SDr, FESC), under the supervision of Prof. F. Boudk, MD, ScD, with collaboration
of M. Tomekov, MD, PhD and Ass. Prof. J. Bultas, MD, PhD. The data were transferred to the
electronic form by the European Centre of Medical Informatics, Statistics and Epidemiology of Charles
University and Academy of Sciences (head. Prof. RNDr. J. Zvrov, DrSc). The data resource is on
the web pages http://euromise.vse.cz/challenge2003. At present time the data analysis is supported
by the grant of the Ministry of Education CR Nr LN 00B 107.

Bibliography
Acua, E. & Rodriguez, C., 2004: The Treatment of Missing Values and its Effect on Classifier
Accuracy, Classification, Clustering, and Data Mining Applications, pp. 639-647, Available at:
http://link.springer.com/chapter/10.1007%2F978-3-642-17103-1_60 [Accessed 19 September 2013]
Berka, P., 2003: Dobvn znalost z databz 1st ed., Academia
Burita L., Gardavsky P., Vejlupek T. 2012: K-GATE Ontology Driven Knowledge Based System for
Decision Support. Journal of Systeme Integration Vol 3, No 1, pp. 19-31
Chen, Z., 2001: Data mining and uncertain reasoning: an integrated approach. Wiley
Fahlman S. E. & Lebiere, Ch. 1991: The Cascade-Correlation Learning Architecture, Available at:
http://www.ifi.uzh.ch/ailab/teaching/neuralnets2013/fahlman.pdf [Accessed 16 September 2013]
Grzymala-Busse, J. W., Goodwin, L. K., Grzymala-Busse, W. J. & Zheng X., 2005: A Closest Fit
Approach to Missing Attribute Values in Preterm Birth Data, Lecture Notes in Artificial Intelligence vol.
3642, pp. 342351, Available at: http://sci2s.ugr.es/MVDM/pdf/grzymala_busse_goodwin05.pdf
[Accessed 12 September 2013]
Grzymala-Busse J. W., Hu M., 2001: A Comparison of Several Approaches to Missing Attribute Values
in Data Mining, Revised Papers from the Second International Conference on Rough Sets and Current
Trends in Computing, pp. 378-385, Available at:
http://sci2s.ugr.es/keel/pdf/specific/congreso/grzymala_busse_hu01.pdf [Accessed 19 September
2013]
Gupta, A. & Lam, M., 1998: The weight decay backpropagation for generalizations with missing
values, Annals of Operations Research 78, pp. 165-187. Available at:
http://link.springer.com/article/10.1023%2FA%3A1018945915940 [Accessed 16 September 2013]
Hand, D. J., Manilla, H. & Smyth, P. 2001: Principles of Data Mining, A Bradford Book
Horton, N. J. & Kleinman, K. P., 2007: Much Ado About Nothing: A Comparison of Missing Data
Methods and Software to Fit Incomplete Data Regression Models, The American Statistician 61(1).
Available at: http://www.math.smith.edu/~nhorton/muchado.pdf
HUANG, Ch. & LEE, H. 2004: A Grey-Based Nearest Neighbor Approach for Missing Attribute Value
Prediction, Applied Intelligence 20, pp. 239252. Available at:
http://80.link.springer.com.dialog.cvut.cz/content/pdf/10.1023%2FB%3AAPIN.0000021416.41043..pdf
Improved outcomes software, 2004: Nearest Neighbors Missing Value Estimation [Online] (Updated
21 December 2004) Available at:
http://www.improvedoutcomes.com/docs/WebSiteDocs/PreProcessing/Estimation_of_Missing_Values/
Nearest_Neighbors_Missing_Value_Estimation.htm [Accessed 12 September 2013]
Kantardzic M. 2003: Data Mining Concepts, Models, Methods, and Algorithms, IEEE, pp. 165-176.

50

JOURNAL OF SYSTEMS INTEGRATION 2014/1

DEALING WITH MISSING VALUES IN DATA

Lakshminarayan, K., Harp S. A. & Samad, T., 1999: Imputation of Missing Data in Industrial
Databases, Applied Intelligence 11, pp. 259275. Available at:
http://www.ime.unicamp.br/~wanderson/Artigos/imputation_industrial_databases.pdf
London School of Hygiene and Tropical Medicine, 2013: Missingness mechanisms [Online] Available
at: http://missingdata.lshtm.ac.uk/index.php?view=category&id=40%3Amissingnessmechanisms&option=com_content&Itemid=96 [Accessed 2 September 2013]
Luengo, J., 2011: Missing Values in Data Mining [Online] Available at:
http://sci2s.ugr.es/MVDM/index.php [Accessed 2 September 2013]
The MathWorks, Inc., 2013: Impute missing data using nearest-neighbor method [Online] Available at:
http://www.mathworks.com/help/bioinfo/ref/knnimpute.html [Accessed 12 September 2013]
Mozer, M. C. & Smolensky, P., 1989: Skeletonization: A Technique for Trimming the Fat from a
Network via Relevance Assessment. Available at:
http://digitool.library.colostate.edu///exlibris/dtl/d3_1/apache_media/L2V4bGlicmlzL2R0bC9kM18xL2F
wYWNoZV9tZWRpYS8xNjcwMjE=.pdf [Accessed 16 September 2013]

JEL Classification: C45, C82

JOURNAL OF SYSTEMS INTEGRATION 2014/1

51