r\0)
\^ ^
Ihm
miswey
HD28
.M414
MIT LIBRARIES
3 9080 00932 7534
A Process View of Data Quality
Henry Kon
Jacob Lee
Richard
WP #3766
Wang
March 1993
PROFIT #93-05
From Information Technology
"PROFIT" Research Initiative
Sloan School of Management
Productivity
Massachusetts Institute of Technology
Cambridge,
MA 02139 USA
(617)253-8584
Fax: (617)258-7579
Copyright Massachusetts
Institute
of Technology 1993. The research described
herein has been supported (in whole or in part) by the Productivity From Information
Technology (PROFIT) Research Initiative at MIT. This copy is for the exclusive use of
PROFIT
sponsor firms.
Productivity
From Information Technology
(PROFIT)
The Productivity From Information Technology (PROFIT) Initiative was established
on October 23, 1992 by MIT President Charles Vest and Provost Mark Wrighton "to
study the use of information technology in both the private and public sectors and
to enhance productivity in areas ranging from finance to transportation, and from
manufacturing to telecommunications." At the time of its inception, PROFIT took
over the Composite Information Systems Laboratory and Handwritten Character
Recognition Laboratory. These two laboratories are
lated to context mediation
now
and imaging respectively.
involved in research re-^^cr^^- ,.,^.~.
OF TECHNOLOGY
MAY 2
3 1995
LIBRARIES
PROFFT has undertaken joint efforts with a number of research centers,
laboratories, and programs at MIT, and the results of these efforts are documented
in Discussion Papers published by PROFIT and/or the collaborating MIT entity.
In addition,
Correspondence can be addressed
to:
The "PROFIT"
Room
Initiative
E5 3-3 10,
MIT
50 Memorial Drive
Cambridge, MA 02142-1247
Tel: (617) 253-8584
Fax: (617) 258-7579
E-Mail: profit@mit.edu
ABSTRACT
We
posit that the term data quality,
though used in a variety of research and practitioner
and defined. To improve data quality, we must
bound and define the concept of data quality. In the past, researchers have tended to take a productoriented view of data quality. Though necessary, this view is insufficient for three reasons. First,
data quality defects in general, are difficult to detect by simple inspection of the data product.
Second, definitions of data quality dimensions and defects, while useful intuitively, tend to be
ambiguous and interdependent. Third, in line with a cornerstone of TQM philosophy, emphasis
should be placed on process management to improve product quality.
contexts, has been inadequately conceptualized
The
objective of this paper
is to characterize the concept of data quality from a process
formal process model of an information system (IS) is developed which offers precise
process constructs for characterizing data quality. With these constructs, we rigorously define the
key dimensions of data quality. The analysis also provides a framework for examining the causes of
data quality problems Finally, facilitated by the exactness of the model, an analysis is presented
of the interdependencies among the various data quality dimensions.
persfjective.
A
.
INTRODUCTION
1.
Total Quality
Management techniques aim
to
drive defect levels in business processes to a
minimum based on continuous improvement and customer focus
these processes, however, has been
[Johnson, Leitch,
need
clearly a
We
&
measured defective
[Ishikawa, 1985).
postulate that
we
and continual
that supports
in the 10-75% range for a variety of applications
Neter, 1981; Laudon, 1986; Morey, 1982; O'Neill
for coordinated, concerted
The data
effort to
&
Vizine-Goetz, 1988).
There
address the data quality issue.
can neither measure nor manage what
we
cannot define, and
this is the
motivation for a precise explication of the concept of data quality. In order to relate our work
research,
we
Indeed, quality in
conformance to customer requirements [Juran
Data quality, when interpreted as
its
& Gryna,
most general sense,
1988
1991]. Included in the scales are
IS success, has
operationalized based on
been variously operationalized into scales
IS (Davis, 1989;
Moore
&
Benbesat,
such items as "makes job easier" and "clear and understandable". These
however, do not define (nor are intended
.
is
p.2.2].
such as the "perceived usefulness" and "perceived ease of use" of an
the data
to pjast
consider data quality broadly to involve the satisfaction of information systenw users with
the information they receive.
scales,
is
to define) objective
Such objective definitions would be required
Measuring the quality of a data
and nveasurable
to facilitate the
characteristics of
management
of data quabty.
set requires specific characteristics of the data as a basis for
measxirement
Other researchers have developed scales
specific technical characteristics of the data
for user information satisfaction,
such as accuracy and timeliness.
intuitive definitions of these terms are provided.
At
For example, a research subject
the importance of information accuracy, but the term itself
is left
which do include
best,
however, only
may be
undefined [Baroudi
asked to rate
&
Orlikowski,
Alternatively, intuitive definitions
1988).
[Bailey
output information
&
may
be provided such
system and data are important
to users.
It is
means
a
The measures themselves are
and general foundation
the correctness of the
satisfaction research deals with data quality issues, its focus
the objective assessment of data quality, but rather in understanding
itself.
-
Pearson, 1983 p.5411-
Although user information
not of the data
accuracy
as;
for data quality definition
is
what overall
characteristics of a
of measuring characteristics of the data users
subjective
developed.
the user or organizational context [Morey, 1982; Ballou
broken
timeliness.
definition
of
down
&
Pazer, 1985;
Huh,
et al., 1990).
less so
is that,
although a data<entered view
and scope of the dimensions are unsatisfying. For example, accuracy
agreement with an identified source" (Huh,
certain variable are recorded" [Ballou
k
et
al.,
is
there
and
taken, the
"all
values for a
While these definitions may be
among them, and
is
defined as "a measure
1990 p.560], or completeness as
Pazer, 1985 p.l531.
they are imprecise, there are interdependencies exist
is
on
Data quality
into various dimensions such as accuracy, completeness, consistency,
The problem with these approaches
and
and highly aggregated, and no rigorous
Other researchers working on data quality have focused more on the data product and
typically
not
is
is
no
intuifive,
theoretical basis
offered for them.
In
summary,
a product view of data quality focuses
satisfaction of the information
on
characteristics of the data product or
systems users with the information system. Such a view
is intuitive,
on
as
we
tend to think of product quality in terms of product characteristics and user requirements. However,
we
believe that the product-oriented approach to defining data quality
is
inherently lacking in three
respects:
(1)
Data quality defects, unlike those of
nuny
standard manufactured products, are difficult
observe or detect. In product manufacturing, each manufactured item
may be compared
to a
to
standard
product specification to determine conformance. In data nnanufacturing, each piece of non-redundant
data
is
unique. There
prenuse of
this
paper
is
no standard product.
is that
the
way
How do we know if the data represents the 'true' value? A
to detect defects is to focus
on the process of data creation and
manipulation. Thus process quality beconoes a surrogate for data quality
(2)
Product-oriented efforts to derive a mininul and orthogonal set of data quality dimensions
show unconvincing
results. Objections
cannot be answered satisfactorily.
causes.
.
such as
"Why
this set of
dimensions?" and "Are there any others?"
Various quality dimensions are interdependent, having conunon
For example, timeliness and completeness are related, as data which
slow in arriving, results both in untinnely data elements and incomplete data
while useful for general understanding, do not
management.
facilitate the clear
is
generated too late or
sets.
is
Simple definitions,
communication needed
for qiiality
(3)
as per the quality experts
Finally,
Deming (Deming,
and Taguchi (Taguchi,
1982)
causes of quality problems in general are inherent in the production process.
that causes of defects in the production process should be detected
analyze and improve data quality,
we must
A
key
TQM
1979],
philosophy
and eliminated. Thus,
in
order
is
to
look at production process quality, and not simply at product
quality.
We
approach
have a poor grounding as
will
be more
data into output data.
receiver
and
model
fruitful.
A
to the
process
is
believe that a process-oriented
the set of inter-related activities that transforms input
This paper defines and formalizes the data production process using a source-
of an information system
(IS).
We
posit that the constructs
developed are fundamental
sufficiently precise to serve as a platform for characterizing the concept of data quality. Thus, data
quality problems can be defined in terms of
LL
component defects in
objectives of this paper are: (1) to develop a set of formal process constructs which constitute
model of an
dimensions of data quality and
The contribution of the
to
(3) the
use of these constructs
Mario Bunge
in
p>aper
is
the development of a theoretically
define five (non-orthogonal)
The
grounded process model of an
model developed draws from
IS
the
Senuntics [Bunge, 1974) and Ontology [Bunge, 1977]. The process approach has
implications for systems design and of)eration.
data defects does not suggest
We adopt a
to
provide an analysis of their interdependencies.
information system and conceptualization of data quality.
of
and analysis of orthogonal
IS as a data delivery vehicle, (2) the identification
components of process non<onformance, and
work
the data production process.
Paper Qbjectivg and Stratggy
The
a
We
concept of data quality.
From a product view
alone, even a robust classification of
how system design and operation might be enhanced
three step strategy.
First,
we
to
reduce problenre.
define an ideal IS as one that delivers data to the end
user via a data production process that conforms to user requirements. The data production process nr\aps
some aspects
of interest in the world into a symbolic representation for use
necessary conditions for this
mapping
to
conform
to user requirements.
by a data
Second,
user.
We identify
we decompose
the
production process into orthogonal components. Thirdly, these process components form the basis for the
characterization of data quality dimensions and defects through conformance
and non conformance
to
user requirements.
In Section 2
Section 3
we
constructs.
we
present an overall description and high level fonnalization of the nvodel.
In
present the data production process in greater detail, including the detailed process
In section 4
we develop
interdependencies. In Section 5
the definitions of data quality dimensions
we conclude
the paper
and discuss its
implications.
and analyze
their
THE roEAL INFORMATION SYSTEM
2.
Wand and Weber
postulated generally that
"information systems are built to provide
information that otherwise would have required the effort of observing or predicting some reality with
which we are concerned. From
perceived reality.
someone.
"
We
It
is
1:
An IS
is
a
by
we
world system as p)erceived by
means of providing a mapping from aspects
that data product defects are
literature in general,
some
a representation of
stating the following premise.
representation of those aspects via
We posit
of a real
is
1990).
this postulate
Premise
an information system
human-created representation
a
[Wand & Weber,
adopt
this point of view,
some human
due
to data
of the
world into a symbolic
perception.
production process defects
.
As per
operationalize quality as conformance to user requirements.
the quality
We
thus
operationalize data quality as coriformance of the data production process to user requirements. In this
respect, data quality is a binary condition [Pall, 1987].
We
do not aim here
to characterize levels of
data quality. Either the process confonns (has quality) or does not conform (does not have quality) with
respect to user requirements. This leads to the next premise:
Premise
Data quality
2:
is
conformance of the data production process
to a single user
requirement.
Figure
on behalf of a
(Arrow
A),
1
data user.
and
our model.
illustrates
We see
from the figure
This observation (perception)
is
results in a perceived reality as construed
that a data originator observes the
the measurement of the world of interest
by the
encoded (data entry) into the IS via son>e data input device (Arrow
Premise
3:
The
world
originator records states of the world as
originator.
B).
Next, this f)erception
is
We assunne
opposed
to
changes in
states.
For example, the originator would record inventory levels as opposed to changes in inventory.
This encoding results in the symbolic representation of the originator's perception as data. Lastly, the
user decodes (interprets) the symbols (Arrow C).
perceived
reality
+ A
e
e
&
input
storage
device
display
device
data production
timeO
proceed
to formalize
[Bunge, 1977]. The world
via attributes
is
our notion of an ideal
made up
of things.
which are defined and ascribed
A
of thing? with resp>ect to reference frames.
timeX
An inionnation system model
Figiire 1:
We
user
The
by
IS
first
stating
some
ontological principles
characteristics of things of the
to the things
by the observer.
We
world are construed
measure the attributes
reference franrte nuiy include elements such as the time of
measurement and the condition of the observer,
e.g.,
the observer's velocity
when measuring
the speed of
another object. The reference frame, in general, would provide information about any variable factors
believed to affect the outcome value of die measuren^nt.
We
are interested in a set of r things in the world
interested in a set of
q
attributes (ak
I
kal,2,...,q).
(xi
I
i
=
l,...,r).
For each attribute there
frames from which to measure it The reference frames chosen up to time
Each
an aspect of the world that
some form
component
xi
we
are
may be one or more reference
tl is
a set {mj
I
j=l,2,...,n}.
triple
0)ijk=<xi, mj,
is
For each thing
of measurement.
is
of interest
Knowledge
and has an associated value
of
vjiit
and
its
of the data originator's perceived reality.
perceived reality
is
given in Section
3.1.
ak>
vijic
This value
associated triple <xi, mj,
A
is
obtained via
ay> constitutes a unit
more formal discussion
of the concept of
For a thing xi and reference frame mi,
let
Aij=|<xi,
mj,ak>
I
k=l,2...q)
and
Hi=U
Aij
;=i
where Hi
the history of the thing xj over
is
all
the reference frames of interest.
R=U
Then
Hi
1=1
is
the set of triples that are of interest as of time
with reference fran>es
up
to
The knowledge of
This
may be
time
function (Arrow C).
where
t2
D, where
t|.
is
The temporal
Each
we are
are defined for specific times
ti
cell
and
its
triple Oijk
and each
data element constitutes an
and wff s form part of a symbolic language
in the next section.
therefore be represented
by
the binary relation g
wff s delivered as output
interested in reference frames
g expresses the actual production process
asf?ects of the
then encoded into a symbolic form.
corresponds to a
cell
fimction, decoding function
the set of atomic
Recall that
Specifically,
wff.
D
is
These wff s are then interpreted by the end-user via a decoding"
The encoding
be formally presented
^
originator,
represents a value vjjk-
The data production process can
to the set
by the data
R, as perceived
cell
the set of aspects of the world of interest
is
ti
atomic well-formed formula (wff).
will
Then R
a relational table for example. In this case, each
data element within the
which
ti-
-
up
to a data
set
consumer as of time
R
t2,
to t\.
how
aspects of the world are
production process are inherent in the specification of
and
from the
mapped
R and
to
D, as they
t2.
Definition: Data quality
is
defined as the conformance of g to user requirenvents.
We refer to g as gp when g conforms to user requirements.
Each
triple in
R must be imiqudy represented as
data in D. This leads to the proposition:
Proposition: For quality data,
gp must be a
bijective function (one-to-one
and onto)
aspects of things to the appropriate wff s for the user's symbolic language.
Formally,
that
maps
gp:
The
precise form of
(e.g. relational
gp must be determined by user requirements because various symbolic
or hierarchical)
may
In the last section
abstraction. In this section
we
we
representations
above requirement.
satisfy the
THE DATA PRODUCTION PROCESS
3.
terms of
R-»D
represented the production of data as the relation
delve into greater engineering detail
-
g, a
high level
specifying the production process in
specific steps.
its
Our treatment
Weber, 1988]
of
an
IS is similar in
in the sense that
it
is
an
IS
methodology and
spirit to
Wand and Weber,
e.g.,
(Wand
modeling formalism based on Bunge [Bunge, 1974, Bunge,
&
1977].
However, where
Wand and Weber
production
including measurement, encoding, and decoding of data between the data originator
itself,
and the data
focus on the design and analysis of systems,
we
focus on data
user.
In the following sub-sections
we decompose the production process g.
Measurement and Perceived Reality
3.1.
Each
triple
(Ojjic
R
e
is
assigned a value vji^ e
V
the data originator via a function:
X:R->V
This
is
analogous
state function iterates
to the state function
over a domain which
proposed by Bunge [Bunge, 1977
is
The function % on the other hand,
thing.
p.l251.
However, Bunge's
a set of reference frames for an attribute of a particular
iterates
over R, which
is
preferable for the current
presentation.
The data
originator's
knowledge of
(Dijk consists
not only of the value vijk but also the
appropriately,
we
a set of atomic statements which represents the originator's knowledge of R. For example,
we
More
associated attribute ak of the thing xi as perceived via a reference frame nw.
define the measurement function (Arrow
A in Fig.
1) as:
F:R-»S
where S
is
may be
interested in the attribute "temperature" in the city of
interest.
The temperature
in
Rome
is
to
"Rome" which
is
be n>easured by an alcohol thermometer
becomes the reference frame and the mode of measurement, the thermometer,
function
F.
the particular thing of
Let the value of the temperature
on a
particular
day t be
ZS'F.
Then
daily.
is
Thus, the date
represented by the
F (Rome,
where
s is the
thermometer
We
is
t,
temperature) =
s
Rome on day
atomic statement "The temperature in
note also that different measurement procedures are appropriate for different attributes.
argument of
What
F.
In the temperature example,
this point, the difference
F represents measurement
between an
depends on what
is
fixed
and what
interested the "temperature at
the attribute
is
via a
thermometer.
is
and a reference frame needs
attribute
included as part of attribute and what
is
the particular attribute in the
If
"weight"
F would represent a weighing machine.
attribute of interest, then a
At
by an alcohol
28°F.
The particular form of the measurement represented by F depends on
were the
as measured
t
is
included in a reference frame?
variable across instances of measurement.
8KX)am on Valentine's Day 1993"
at
If,
to
be highlighted.
This distinction
various locations within
"temperature at 8:00am on Valentine's Day 1993" with
M
we
for example,
corresponding
are
Then
Italy.
to the set of
various locations.
If
on
"temperature in
the
first
we may be
the other hand,
Rome" corresponds
case, the date
and time are
interested in the "temperature in
domain
at various times,
then
M corresponds to the set of various times.
to the attribute while
fixed for the
Rome"
of interest (and therefore incorporated as part
of the attribute definition) while locations vary (and therefore incorporated as reference frames).
the second, the location
is
fixed while the times of
In
measurement vary.
statements that forms the originator's knowledge or pjerceived
We provide below a more detailed example to be used
Note
In
that S is the set of
reality.
for the rennainder of the paper.
Example:
We use a
frozen-foods company's inventory system as the domain of interest. The things of interest will
be n cold-storage warehouses
attributes of interest:
xi, 1
"Warehouse
^
i
id",
^
n,
owned by
"number
company.
this
A
given warehouse will have 3
of frozen dinners at begiiming of day" as
procedure pi and "temperature at beginning of day" as measured by procedure p2 which
QTY
(a2)
and
TEMP (33)
The reference frame
v2
(in units of
is
respectively.
the date
t
QTY
on which
is
equal to the
number
the measurenwnts are taken.
one) and V3 (in "¥) correspond to ID,
QTY and TEMP
corresponding statements are given by
where
of dinners physically
F(xi, ai,
t)
=
F(xi, 32,
t)
= S2
F(xi, as,
t)
= S3
SI
The values vi
measured by
we call ID (aj),
on the
shelves.
(3-digit string),
respectively for
day
t.
The
si="the ID of warehouse
S2="the
is
pi
v] at the beginning of the
QTY of frozen dinners
in
warehouse
i
day
t"
beginning of day
at the
t
as
measured by procedure
V2 and
"
S3="TEMP
in
warehouse
summary, thus
In
is
i
are instantiated via
far
i
at the
we have
measurement
beginning of day
t
as measured by procedure p2
defined thing s with their attributes and reference frames which
into atomic statements. In the next section
the encoding of statements into data,
V3"
is
and the
interpretation of the data
by
we develop
a formalism for
the user.
Symbolic Language
3.2.
The knowledge
Bunge [Bunge,
1974).
of the data originator
encoded as wffs of a symbolic language defined by
is
We introduce the encoding function
E represents the mapping used by the originator to
E which
is
a
component
of a symbolic language.
map statements in S to corresponding wffs in
D.
E:S^D
The decoding function
E"^
(Arrow
C in
Figure
1) is
used by the user
to interpret
wff back to the staten^ents
they signify.
We
note that an atomic wff not only contains data that represents the value
additional information about the reference frame and attribute definitions, as well as
attained.
Often,
when
the term "data quality" is used,
data elements alone, or whether
it
on a
In addition to the data values, the user nuiy
understand the data. Each of these
names)
modem
is
is
not clear whether
it
how
screen.
they were
Consider
in the cells represent attribute values.
need column names, table names, and other information
may have some
to
type of quality associated with them.
databases, this extra information
referred to as metadata
The data
but also
refers to the quality of
refers to the quality of this additional information as well.
the display of a relational database table
In
it
vijk,
(e.g.,
besides the data elements and their column
[McCarthy, 1984J, which
may
be implemented via schema
information, data dictionaries, semantically rich data model constructs, or systems documentation,
depending on the boundary of the concept of the data and
IS.
regarding the intrinsic n>eaning of the data [Collett, Huhns,
k
Metadata
may
Shen, 1991; Siegel
well as indicators of the quality of the data value as an electronic artifact
Wang, Kon,
We
the time of
&
&
Madnick, 1991], as
k
[Wang
Madnick, 1990;
Madnick, 19931.
ackiwwiedge the
(wssibility of inaccuracies in
metadata as well as in data. For example,
measurement may have been recorded wrongly. Thus the possible need
metadata as well. This recursiveness in metadata
k Madnick,
include both ii\formation
1991;
Wang, Reddy,
is
for qualification of
a universal problem in metadata modeling [Siegel
k Kon, 1993]. We assume only one level of metadata.
We
can
now more
completely characterize g as a composite function.
measurement function F and the encoding function
E.
This
is
reflected in Fig. 2.
It
consists of the
(1), (2)
and
We now
define R.
(3)
provide an example
(1)
THINGS: The
(2)
ATTRIBUTES:
and °F
for
January
1,
based on the example started
in Section 3.1:
things of interest are corporate warehouses.
ID,
QTY and TEMP
QTY and TEMP
REraRENCE FRAME:
(3)
spjecification
are the attributes of interest, using units of single dinners
respectively.
Information
is
to
be generated once
at the
beginning of each day from
1993.
(4)
MEASUREMENT procedures pi and p2 for QTY and TEMP respectively.
(5)
LANGUAGE:
Decoding
is
Encoding should be done using the following notation: 4
a reverse
mapping
of the encoding through
digits, left justified.
which the user interprets the
numerical data.
INSTANTIATION DELAYS:
(6)
Values are
to
be available on the
IS display within three
hours
of measurement.
In the following section
we
use the IS model to analyze and define several fundamental-
dimensions of data quality.
DEFINING THE DIMENSIONS OF DATA QUAUnf
4.
We
embodied
have shown
in g, the
In this section
requirements.
we
mapping from R to D.
look at
how
depends on the quality of the data production process as
We have also decomposed g into its various process components.
the data production process
fails in
temns of non<onfonnance to user
We define the specific components of non-conformance by which data quality is lost
The reasons
cycle (SDLC). First,
IS.
that data quality
for
iwn-conformance
may correspond
to
any of several stages
in a
systems design
life
an analysis and design stage transforms user requirements into a design model of the
Second, an implementation stage converts the design specification into an operational system.
identify here a third
component not
identified in the
SDLC
We
perspective: the data production stage
Deviations nnay exist between the intended operation and the actual operation of the
IS.
We
.
define
data quality over the ag gregate tramformation from user requirements through actual data as delivered
and
interpreted.
In Table
1
below, each item describes a component of non-conformance of the process constructs.
11
Process Construct
Thing
Accuracy
signifies (e.g.,
if
point concept which requires each wff to properly correspond to the triple
is a
there are
two dinners,
observed and the symbol
'two' is
'2'
is
entered).
It
o) it
characterizes
the relation from a single aspect of the world to a single wff.
gp((o) = ga(w) for
In terms of the process constructs, a wff
is
€ R.
all co
accurate
if,
and only
if
Ep(Fp(xi, mj, ak)) = Ea(Fa(xi, mj, a^))
i.e.
the aggregate result of
Note
offset
measurement and encoding
is
that this condition accounts for the case
according to P.
where measurement
errors
and encoding
errors
each other.
Completeness
4.1^.
Completeness
is
a set-based concept that requires that
all o)
e
Rp are measured and mapped
onto
the co-domain:
co-domain (gp) = co-domain
In terms of process constructs, a data set
Rp =
Da
is
complete
means
that all the aspects of the
For example,
specifies only
only
Rp
is
violated
if,
and only
if
Ra/«ind
Ep(Fp((o))=Ea(Fa((o)) for
This
(ga).
if
P
world of
specifies all
warehouses in California or
if
P
all co
e
interest are
Rp
measured and encoded
company warehouses
specifies the attributes
accurately.
as things of interest and
QTY and TEMP and A
A
specifies
QTY - these are incompleteness problems - one in things and one in attributes.
A.13.
Relevance
Relevance
is
a point concept
It
requires that each wff in
Da be
the result of a
mapping from an
aspect of the world of interest Rp.
g-lp(di) €
4.1.4.
Rp
for all di e
Da
Timeliness
Timeliness requires that the time delay between measurement and the instantiation of wff s
at
or below the
maximum
specified.
13
is
te-td
(te
and
td
were defined
4.1.5.
in Section 3.3
^ DImax
each
^or
and DImax
co
specification item 6 in Secfion 3.4.)
irv
Interpretability
map
In general, interpretability requires that the user
per
e Rg
wff back to the statements they signify as
This requires that the user be able to interpret the data element as well as various aspects of the
P.
metadata.
For example, the data element
number
three.
number
of frozen dinners).
The
itself
definition of the attribute
reliability of the data, for
must be interpreted so
must also be known or
when
the data
symbol
interpreted, (e.g.,
In addition, reference frame information
example, based on
that the
may
means
the
QTY means
the
'3'
provide indicators of the
was measured.
Thus, for data to be interpretable with respect to the user
for di €
Da, di
is
Ep-1 (di) = Ea"^
To be
nrtay
specific,
we
distinguish
not be well formed. This necessitates the
The key
interpreted
We
to interpretability
4^
is
due
is,
(di) for di
Da
Da-
first
of the
is
e
Da
the actual data
set, for
which some elements
two requirements.
given a wff, whether accurate or not,
can be properly
it
by a user of the language.
posit that these five dinoensions are sufficient to characterize data of perfect quality with
respect to a given user specification.
This
Dp from
a wff, and
to the
We make no clain\s as to
interdependendes that
exist
among
necessity (minimality) of the dinnensions.
the dimensions, which
we discuss next.
Interdependendea
The above aiulysis demonstrates the complexity underlying the concept
Intuitive or simple definitions of data quality
may be
inadequate to operationalize data quality
measures. Sinularly, attempts to find orthogonal dimensions of data quality
discuss below, interdependendes exist
constructs.
may be
misguided. As
we
the dimensions, even given the exactness of our model
The discussion on interdependendes
illustrate several
will not
is
presented to
to reflect the current state of
our dynamic
be a formal one, but rather
obvious interdependendes.
Tinneliness
world.
amoi^
of data quality.
and accuracy: As data
ages,
it
where the
data
is
In applicatioiu
timeliness of data delivery
is
latest
may cease
taken to represent the current sute of the world,
related to data accuracy.
number may beconne inaccurate with time.
14
For example, the value of a person's phone
Timeliness and completeness
mentioned
or
is
slow
earlier, timeliness
in arriving^ results
is
unavailable.
If
we
consider timeliness in terms of instantiation delay then, as
and completeness are
related concepts.
Data which
is
generated too
both in untimely data values and (tempwrarily) incomplete data
Completeness and accuracy
desired wff
:
A
:
wff which
The required data
is
inaccurate
is
late,
sets.
unusable by the user, and thus the
for practical purposes, not in the database for the user,
is,
causmg incompleteness.
Accuracy and interpretability
in
metadata
is
:
As we include metadata
interpretable as to
what the data
is.
is
our definition of the data, inaccuracy
For example,
related to the interpretability of data.
height as of a certain date, then the date
in
metadata
if
a wff
were
express a child's
to
which makes
for the height value
it
Thus, inaccuracy in the date would affect the interpretability of
the height.
Thus we see
that interdependencies exist
DISCUSSION AND CONCLUSION
5.
A model of an
information system has been presented in which data process quality was treated
as a surrogate for data product quality.
constructs.
Using the
among dimensions often though of as independent.
nrtodel
we
The model contains an
explicit
and orthogonal
set of
process
are able to develop five data quality dimension definitions: accuracy,
completeness, relevance, timeliness, and interpretability. They are shown, under the explicitly stated
assumptions of the model,
We
have seen
to inter-relate in various ways.
that data quality is not a simple concept.
minimal and orthogonal
set of
data quality dimensior\s
single dimensions nor their interdependencies
world and an infonnation system are required
intuitive definitions of data,
applicability to solving
quality data.
have
is
trivial definitions.
IS,
and
A
this paper.
for detailed analysis. Efforts in the past
to the
some
Neither
well-defined model of the
measurement, and the information system
problems related
that there exists
brought into question by
may be
which appeal
to
limited in their
development of information systems
that deliver
Data quality improvements must come through assisting and infonning those who
manage, design, imi^ement, and operate infonnation systems. In
of the
The notion
its
this respect, a process oriented
impacts on data quality, are more informative than
This paper has
shown
static
analyas
data product dimensions.
that interp reta bility, accuracy, timeliness, relevance,
and completeness
have complex constructs underlying them. More rigorous definitions of data, schfema infonnation, user
understanding of data, generation of data, and time are necessary to define rigorously the components of
data quality.
Even the
relatively simple source-receiver
requires a thorough analysis for
We
-
its
model
of an IS developed in this paper
implications to data quality.
have demonstrated a certain arbitrariness
in past conceptualizations of data characteristics
considering words such as timeliness and con\pleteness as indeperulent dimensions of data quality
15
when
in fact there are strict relationships
between them.
It is
clear
from
data quality requires a strong model of the process by which the data
components and stages of data
creation.
this effort that operationalizing
and
solid definitions for
Data quality characteristics can be traced
to characteristics of
the information system delivery vehicle that creates
This research has denK)nstrated the
is
created
it.
utility of
using fundamental concepts, such as those
provided by the process model, as a foundation for data quality measurement. The constructs are donriain
and data model independent. The primitives developed here provide an unambiguous vocabulary
further
work
in developing data quality measures,
systems.
16
and may
assist in
for
designing quality information
References
[1]
Bailey,
E.
J-
&
Pearson,
S.
W.
(1983).
User Satisfaction. Management
Development
of a Tool for
Measuring and Analyzing Computer
Science, 22(5), pp. 530-545.
& Pazer, H. L. (1985). Modeling Data and Process (Quality in Multi-input, Multi-output
Systems.
Information
Management Science, 11(2), pp. 150-162.
[31 Baroudi, j. j. & Orlikowski, W. J. (1988). A Short-Form Measure of User Information Satisfaction: A
Psychometric Evaluation and Notes on Use. Journal of Management Information Systems,
[41 Bunge, M. (1974). Semantics I: Sense and Reference
Boston: D. Reidel Publishing Company.
(31 Bunge, M. (1977). Ontology I: The Furniture of the World
Boston: D. Reidel Publishing Company.
(b| Collett, C, Huhns, M. N., & Shen, W. (1991). Resource Integration Using a Large Knowledge Base in
Carnot. /£££ Computer,
[7] Davis, F. D. (1989). Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information
Technology. Management Information Systems Quarterly, (September), pp. 319-340.
Cambridge: MIT Center for Advanced Enfgineering.
[81 Deming, W. E. (1982). Out of the Crisis
Huh, Y. U., et al. (1990). Data Quality. Information and Software Technology, 12(8), pp. 559-565.
(91
Ishikawa, K. (1985). What is Total Quality Control?-the Japanese Way
[101
Englewood Cliffs, NJ:
Prentice-Hall.
[Ill Johnson, J. R., Leitch, R. A., & Neter, J. (1981). Characteristics of Errors in Accounts Receivable and
Inventory Audits. Accounting Review, 56(April), pp. 270-293.
[121 Juran, J. M. & Gryna, F. M. (1988). Quality Control Handbook (4th ed.). New York: McGraw-Hill
Book Co.
[13] Laudon, K. C. (1986). Data Quality and Due Process in Large Interorganizational Record Systems.
Communications of the ACM, 22(1), pp. 4-11.
McCarthy, J. L. (1984). Scientific Information - Data + Meta-data. U.S. Naval Postgraduate
[141
[21
Bailou, D. P.
.
.
.
.
School, Monterey,
CA.
1984.
Moore, G. C. & Benbesat, I. (1991). Developing an Instrument to Measure the Perceptions of Adopting
an Information Technology Innovation. 2(3), pp. 192-222.
Morey, R. C. (1982). Estimating and improving the Quality of Information in the MIS.
[161
Communications of the ACM, 25(May), pp. 337-342.
(171 O'Neill, E. T. & Vizine-Goetz, D. (1988). (Quality Control in 0\line Databases. In M. E. Williams
(Ed.), Annual Review of Inofrmation, Science, and Technology (pp. 125-156.). Elsevier Publishing
[151
Company.
[181
[191
[20J
[21]
G. A. (1987). Quality Process Management Englewood Qiffs, NJ: PrenHceHall, Inc.
Siegel, M. & Madnick, S. E. (1991). A metadata approach to resolving semantic conflicts.
Pall,
.
Conference. Barceloiu, Spain. 1991.
Taguchi, G. (1979). Introduction to Off-line Quality Control
Control Association.
Wand,
Magaya, Japan: Central Japan
(Quality
& Weber, R. (1988). An Ontological Aruilysis of Some Fundamental Information Systems
The Ninth International Conference on Information Systems, Minneapolis, Minnesota,
Y.
Concepts.
[221
.
VLDB
USA. 1988.
Wand, Y.
&
Weber, R.
(1990).
Mario Bunge's Ontology as a Fonnal Foundation for Infomution
Bunge's Treatise
J. W. Dom (Ed.), Studies on Mario
Systent^ Concepts. In P. Weingartner& G.
Amsterdam; Rodopi.
Wang, R. Y., Kon, H.
B., & Madnick, S. E. (1993). Data Quality Requirements Analysis and
Data Engineering. Vienna, Austria. 1993
[241 Wang, R. Y., Reddy, P., k Kon, H. B. (1993). An Attribute-based Model of Data for Data Quality
Management, to appear in the Journal of Decision Suuport Systems (DSS)
[25] Wang, Y. R. & Madnick, S. E. (1990). A Polygen Model for Heterogeneous Database Systems: The
[231
Modeling
in
Source Tagging Perspective. Brisbane, Australia. 1990. pp. 519-538.
17
MIT LIBRARIES
3 9080 00932 7534
97
Date Due