Data Quality Assessment
Leo L. Pipino, Yang W. Lee, and Richard Y. Wang
How good is a company’s data quality? Answering this question requires usable data
quality metrics. Currently, most data quality measures are developed on an ad hoc
basis to solve specific problems [6, 8], and fundamental principles necessary for devel-
oping usable metrics in practice are lacking. In this article, we describe principles that
can help organizations develop usable data quality metrics.
Studies have confirmed data quality is a multi-dimensional concept [1, 2, 6, 9,
10, 12]. Companies must deal with both the subjective perceptions of the individu-
als involved with the data, and the objective measurements based on the data set in
question. Subjective data quality assessments reflect the needs and experiences of
stakeholders: the collectors, custodians, and consumers of data products [2, 11]. If
stakeholders assess the quality of data as poor, their behavior will be influenced by this
assessment. One can use a questionnaire to measure stakeholder perceptions of data
quality dimensions. Many healthcare, finance, and consumer product companies
have used one such questionnaire, developed to assess data quality dimensions listed
in Table 1 [7]. A major U.S. bank that administered the questionnaire found custo-
dians (mostly MIS professionals) view their data as highly timely, but consumers dis-
agree; and data consumers view data as difficult to manipulate for their business
purposes, but custodians disagree [4, 6]. A follow-up investigation into the root
causes of differing assessments provided valuable insight on areas needing improve-
ment.
Objective assessments can be task-independent or task-dependent. Task-inde-
pendent metrics reflect states of the data without the contextual knowledge of the
application, and can be applied to any data set, regardless of the tasks at hand. Task-
dependent metrics, which include the organization’s business rules, company and
government regulations, and constraints provided by the database administrator, are
developed in specific application contexts.
COMMUNICATIONS OF THE ACM April 2002/Vol. 45, No. 4ve
211
Leo L. Pipino (Leo_Pipino@uml.edu) is professor of MIS in the College of Management at the University of
Massachusetts Lowell.
Yang W. Lee (y.wlee@neu.edu) is an assistant professor in the College of Business Administration at
Northeastern University in Boston, MA.
Richard Y. Yang (rwang@bu.edu) is an associate professor at Boston University and Co-director of the Total
Data Quality Management (TDQM) program at MIT Sloan School of Management in Cambridge, MA.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies
are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy
otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.
© 2002 ACM
In this article, we describe the subjective and objective assessments of data qual-
ity, and present three functional forms for developing objective data quality metrics.
We present an approach that combines the subjective and objective assessments of
data quality, and illustrate how it has been used in practice. Data and information are
often used synonymously. In practice, managers differentiate information from data
intuitively, and describe information as data that has been processed. Unless specified
otherwise, this paper will use
data interchangeably with
information.
212 April 2002/Vol. 45, No. 4ve
COMMUNICATIONS OF THE ACM
Table 1. Data quality dimensions.
Functional Forms
When performing objective assessments, companies should follow a set of principles
to develop metrics specific to their needs. Three pervasive functional forms are sim-
ple ratio, min or max operation, and weighted average. Refinements of these func-
tional forms, such as addition of sensitivity parameters, can be easily incorporated.
Often, the most difficult task is precisely defining a dimension, or the aspect of a
dimension that relates to the company’s specific application. Formulating the metric
is straightforward once this task is complete.
Simple Ratio. The simple ratio measures the ratio of desired outcomes to total
outcomes. Since most people measure exceptions, however, a preferred form is the
number of undesirable outcomes divided by total outcomes subtracted from 1. This
simple ratio adheres to the convention that 1 represents the most desirable and 0 the
least desirable score [1, 2, 6, 9]. Although a ratio illustrating undesirable outcomes
gives the same information as one illustrating desirable outcomes, our experience sug-
gests managers prefer the ratio showing positive outcomes, since this form is useful
for longitudinal comparisons illustrating trends of continuous improvement. Many
traditional data quality metrics, such as
free-of-error,
completeness, and
consistency take
this form. Other dimensions that can be evaluated using this form include
concise rep-
resentation,
relevancy, and
ease of manipulation.
The
free-of-error dimension represents data correctness. If one is counting the data
units in error, the metric is defined as the number of data units in error divided by
the total number of data units subtracted from 1. In practice, determining what con-
stitutes a data unit and what is an error requires a set of clearly defined criteria. For
example, the degree of precision must be specified. It is possible for an incorrect char-
acter in a text string to be tolerable in one circumstance but not in another.
The
completeness dimension can be viewed from many perspectives, leading to
different metrics. At the most abstract level, one can define the concept of schema
completeness, which is the degree to which entities and attributes are not missing
from the schema. At the data level, one can define column completeness as a func-
tion of the missing values in a column of a table. This measurement corresponds to
Codd’s column integrity [3], which assesses missing values. A third type is called pop-
ulation completeness. If a column should contain at least one occurrence of all 50
states, for example, but it only contains 43 states, then we have population incom-
pleteness. Each of the three types (schema completeness, column completeness, and
population completeness) can be measured by taking the ratio of the number of
incomplete items to the total number of items and subtracting from 1.
The
consistency dimension can also be viewed from a number of perspectives, one
being consistency of the same (redundant) data values across tables. Codd’s Referen-
tial Integrity constraint is an instantiation of this type of consistency. As with the pre-
viously discussed dimensions, a metric measuring consistency is the ratio of violations
of a specific consistency type to the total number of consistency checks subtracted
from one.
Min or Max Operation. To handle dimensions that require the aggregation of
multiple data quality indicators (variables), the minimum or maximum operation can
be applied. One computes the minimum (or maximum) value from among the nor-
malized values of the individual data quality indicators. The min operator is conser-
vative in that it assigns to the dimension an aggregate value no higher than the value
of its weakest data quality indicator (evaluated and normalized to between 0 and 1).
COMMUNICATIONS OF THE ACM April 2002/Vol. 45, No. 4ve
213
The maximum operation is used if a liberal interpretation is warranted. The individ-
ual variables may be measured using a simple ratio. Two interesting examples of
dimensions that can make use of the min operator are
believability and
appropriate
amount of data. The max operator proves useful in more complex metrics applicable
to the dimensions of
timeliness and
accessibility.
Believability is the extent to which data is regarded as true and credible. Among
other factors, it may reflect an individual’s assessment of the credibility of the data
source, comparison to a commonly accepted standard, and previous experience. Each
of these variables is rated on a scale from 0 to 1, and overall believability is then
assigned as the minimum value of the three. Assume the believability of the data
source is rated as 0.6; believability against a common standard is 0.8; and believabil-
ity based on experience is 0.7. The overall believability rating is then 0.6 (the lowest
number). As indicated earlier, this is a conservative assessment. An alternative is to
compute the believability as a weighted average of the individual components.
A working definition of the
appropriate amount of data should reflect the data
quantity being neither too little nor too much. A general metric that embeds this
tradeoff is the minimum of two simple ratios: the ratio of the number of data units
provided to the number of data units needed, and the ratio of the number of data
units needed to the number of data units provided.
Timeliness reflects how up-to-date the data is with respect to the task it’s used for.
A general metric to measure timeliness has been proposed by Ballou et al., who sug-
214 April 2002/Vol. 45, No. 4ve
COMMUNICATIONS OF THE ACM
Figure 1. Dimensional data quality assessment across roles.
gest timeliness be measured as the maximum of one of two terms: 0 and one minus
the ratio of currency to volatility [2]. Here, currency is defined as the age plus the
delivery time minus the input time. Volatility refers to the length of time data remains
valid; delivery time refers to when data is delivered to the user; input time refers to
when data is received by the system; and age refers to the age of the data when first
received by the system.
An exponent can be used as a sensitivity factor, with the max value raised to this
exponent. The value of the exponent is task-dependent and reflects the analyst’s judg-
ment. For example, suppose the timeliness rating without using the sensitivity factor
(equivalent to a sensitivity factor of 1) is 0.81. Using a sensitivity factor of 2 would
then yield a timeliness rating of 0.64 (higher sensitivity factor reflects fact that the
data becomes less timely faster) and 0.9 when sensitivity factor is 0.5 (lower sensitiv-
ity factor reflects fact that the data loses timeliness at a lower rate).
A similarly constructed metric can be used to measure
accessibility, a dimension
reflecting ease of data attainability. The metric emphasizes the time aspect of accessi-
bility and is defined as the maximum value of two terms: 0 or one minus the time
interval from request by user to delivery to user divided by the time interval from
request by user to the point at which data is no longer useful. Again, a sensitivity fac-
tor in the form of an exponent can be included.
If data is delivered just prior to when it is no longer useful, the data may be of
some use, but will not be as useful as if it were delivered much earlier than the cut-
off. This metric trades off the time interval over which the user needs data against the
time it takes to deliver data. Here, the time to obtain data increases until the ratio
goes negative, at which time the accessibility is rated as zero (maximum of the two
terms).
In other applications, one can also define accessibility based on the structure and
relationship of the data paths and path lengths. As always, if time, structure, and path
lengths all are considered important, then individual metrics for each can be devel-
oped and an overall measure using the min operator can be defined.
Weighted Average. For the multivariate case, an alternative to the min operator
is a weighted average of variables. If a company has a good understanding of the
importance of each variable to the overall evaluation of a dimension, for example,
then a weighted average of the variables is appropriate. To insure the rating is nor-
malized, each weighting factor should be between zero and one, and the weighting
factors should add to one. Regarding the believability example mentioned earlier, if
the company can specify the degree of importance of each of the variables to the over-
all believability measure, the weighted average may be an appropriate form to use.
Assessments in Practice
To use the subjective and objective metrics to improve organizational data quality
requires three steps (see Figure 2):
• Performing subjective and objective data quality assessments;
• Comparing the results of the assessments, identifying discrepancies, and deter-
mining root causes of discrepancies; and
• Determining and taking necessary actions for improvement.
COMMUNICATIONS OF THE ACM April 2002/Vol. 45, No. 4ve
215
To begin the analysis, the subjective and objective assessments of a specific
dimension are compared. The outcome of the analysis will fall into one of four quad-
rants (see Figure 3). The goal is to achieve a data quality state that falls into Quad-
rant IV. If the analysis indicates Quadrants I, II, or III, the company must investigate
the root causes and take corrective actions. The corrective action will be different for
each case, as we illustrate using the experiences of two companies.
Global Consumer Goods, Inc., (GCG), a leading global consumer goods com-
pany, has made extensive use of the assessments [4]. At GCG, results of subjective
assessments across different groups indicated that consistency and completeness were
two major concerns. When these assessments were compared to objective assessments
of data being migrated to GCG’s global data warehouse, the objective measures cor-
roborated the subjective assessment (Quadrant I). This agreement led to a corporate-
wide initiative to improve data consistency and completeness. Among the
measurements used was a metric measuring column integrity of the transaction
tables. Prior to populating their global data warehouse, GCG performed systematic
null checks on all the columns of its detailed transaction files. GCG conducted col-
umn integrity analysis using a software tool called Integrity Analyzer [5] to detect
missing values, which indicated the database state did not reflect the real-world state
and any statistical analysis would be useless. Although GCG could simply have meas-
ured consistency and completeness on an ad hoc basis, performing the measurements
based on the approach presented here enabled GCG to continually monitor both
objective measures and user assessments, thereby institutionalizing its data quality
improvement program.
216 April 2002/Vol. 45, No. 4ve
COMMUNICATIONS OF THE ACM
Figure 2. Data quality assessments in practice.
A leading data product manufacturing company, Data Product Manufacturing,
Inc., (DPM), which provides data products to clients in the financial and consumer
goods industries, among others, illustrates the issue of conflicting assessments. Unlike
GCG, DPM found discrepancies between the subjective and objective assessments in
its data quality initiative. DPM’s objective assessment indicated its data products were
of high quality, but its client’s assessments (subjective assessments) indicated a lack of
confidence in the data products in terms of believability, timeliness, and free of error
(Quadrant III). Further analysis revealed the clients’ subjective assessments were
based on the historical reputation of the data quality. DPM proceeded to implement
a data quality assurance program that included training programs for effective use of
data. They also incorporated the results of the objective assessments in an overall
report that outlined the complexities of client deliverables.
Companies like GCG and DPM that assess subjective and objective data quality
go a long way toward answering the question posed at the beginning of this article:
How good is my company’s data quality? Such assessments also help answer other
questions posed by practitioners: How does my data quality compare with others in
my industry? Is there a single aggregate data quality measure? If dimensional data
quality metrics are developed and assessment data is collected and analyzed over time
across an industry, that industry can eventually adopt a set of data quality metrics as
a de facto standard, or benchmark performance measure. In the long term, different
benchmarks and aggregate performance measures can be established across industries.
In practice, companies wish to develop a single aggregate measure of their data
quality—an index of data quality. A single-valued, aggregate data quality measure
would be subject to all the deficiencies associated with widely used indexes like the
Dow Jones Industrial Average and the Consumer Price Index. Many of the variables
and the weights would be subjective. Issues that arise when combining values associ-
ated with different scale types (ordinal, interval, and ratio) further complicate mat-
ters. But if the assumptions and limitations are understood and the index is
interpreted accordingly, such a measure could help companies assess data quality sta-
tus. From the practitioner’s viewpoint, such an index could help to succinctly com-
COMMUNICATIONS OF THE ACM April 2002/Vol. 45, No. 4ve
217
Figure 3. Subjective and objective assessments.
municate the state of data quality to senior management and provide comparative
assessments over time.
Conclusion
Experience suggests a “one size fits all” set of metrics is not a solution. Rather, assess-
ing data quality is an on-going effort that requires awareness of the fundamental prin-
ciples underlying the development of subjective and objective data quality metrics. In
this article, we have presented subjective and objective assessments of data quality, as
well as simple ratio, min or max operators, and weighted average—three functional
forms that can help in developing data quality metrics in practice. Based on these
functional forms, we have developed illustrative metrics for important data quality
dimensions. Finally, we have presented an approach that combines the subjective and
objective assessments of data quality, and demonstrated how the approach can be
used effectively in practice.
References
1. Ballou, D.P. and Pazer, H.L. Modeling data and process quality in multi-input, multi-out-
put information systems.
Management Science 31, 2, (1985), 150–162.
2. Ballou, D.P., Wang, R.Y., Pazer, H. and Tayi, G.K. Modeling information manufacturing
systems to determine information product quality.
Management Science 44, 4 (1998),
462–484.
3. Codd, E.F., Relational database: a practical foundation for productivity, the 1981 ACM
Turing Award Lecture.
Commun. ACM 25, 2 (1982), 109–117.
4. CRG, Information Quality Assessment (IQA) Software Tool. Cambridge Research Group,
Cambridge, MA, 1997.
5. CRG, Integrity Analyzer: A Software Tool for Total Data Quality Management. Cambridge
Research Group, Cambridge, MA, 1997.
6. Huang, K.,Lee, Y., and Wang, R.
Quality Information and Knowledge. Prentice Hall, Upper
Saddle River: N.J. 1999.
7. Kahn, B.K., Strong, D.M., and Wang, R.Y. Information Quality Benchmarks: Product and
Service Performance.
Commun. ACM, (2002).
8. Laudon, K.C. Data quality and due process in large interorganizational record systems.
Commun. ACM 29,1 (1986), 4–11.
9. Redman, T.C., ed. Data
Quality for the Information Age. Artech House: Boston, MA., 1996.
10. Wand, Y. and Wang, R.Y. Anchoring data quality dimensions in ontological foundations.
Commun. ACM 39,11 (1996), 86–95.
11. Wang, R.Y. A product perspective on total data quality management.
Commun.ACM 41,
2 (1998), 58–65.
12. Wang, R.Y. and Strong, D.M. Beyond accuracy: what data quality means to data con-
sumers.
Journal of Management Information Systems 12, 4 (1996), 5–34.
218 April 2002/Vol. 45, No. 4ve
COMMUNICATIONS OF THE ACM
