28 Journal of Organizational and End User Computing, 21(3), 24-36, July-September 2009
Copyright © 2009, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global
is prohibited.
rate, did not actually report a cell error rate of
6.9% on completed spreadsheets. Rather, they
reported that auditors had “issues” concerning
6.9% of the cells on the initial review of mod-
els that they subsequently discussed with the
developers. The authors provided no details on
the denition of an “issue” or its relation to an
actual error. Moreover, the initial model version
that was reviewed was revised an average of six
times before the model was complete. Thus the
estimate of a 6.9% error rate applies only to is-
sues raised by auditors on initial model versions,
not to errors in completed spreadsheets. If, as
seems warranted, we exclude this study from
the calculation of the cell error rate, the average
cell error rate decreases from 5.2% to 1.3% on
a base of only 13 audited spreadsheets.
Our review of the literature on spreadsheet
errors draws attention to several shortcomings.
First, no generally accepted classication of
errors exists. Second, the classications that
do exist are of more theoretical than practical
value. Third, existing estimates of error rates
are based on extremely limited data. Fourth, no
studies of errors have fully revealed the sources
of spreadsheets tested, how errors were dened,
and the auditing methods that were used.
RESEARCH DESIGN
Our research into spreadsheet errors is predicat-
ed on several guiding principles or constraints.
First, our interest is in errors in completed,
operational spreadsheets, not errors made in
a laboratory setting or errors made during the
development of a spreadsheet. A second prin-
ciple is that we conduct the audit using only the
information in the spreadsheet itself, without
relying on the developer. A third principle is
that we use an explicit auditing protocol that
any moderately experienced user of Excel can
master.
These principles limit our study in certain
ways. When we work without access to the
spreadsheet developers, we are not able to check
our understanding of a model with an expert. In
practice, this means that at times we will accept a
suspicious formula as correct because we cannot
be sure that it is incorrect. It also means that we
cannot hope to uncover errors in formulating
the underlying problem or errors in interpret-
ing spreadsheet results. Detecting these types
of errors requires a much more time-intensive
longitudinal study of how spreadsheets are
used in the broader context of problem solving.
Offsetting these limitations is the fact that we
can audit a much larger volume of spreadsheets
using our protocol than otherwise.
Our procedures almost certainly lead to an
underestimate of the actual error rates in the
spreadsheets we audit. We may occasionally
classify a cell incorrectly as an error, but we
have been conservative in limiting our deni-
tion of errors to cells for which we can have
a high degree of condence that our judgment
is correct. On the other hand, there are entire
classes of errors that we cannot hope to iden-
tify with the procedures used in this study. For
example, most (but not all) errors in input data
are beyond our scope.
Sample Spreadsheets
Our sample of 50 spreadsheets came from a
wide variety of sources. Some were obtained
during site visits to companies. We carried
out site visits at two consulting companies, a
bank, a college, a large energy company, and
a state government agency. We also obtained
spreadsheets from a variety of organizations
through the alumni and faculty networks at
the Tuck School of Business. No single source
contributed more than ve spreadsheets to the
sample of 50 analyzed here.
We also obtained spreadsheets from vari-
ous Web sites. Several software companies post
spreadsheets on their Web site, either to illustrate
how to use their software or to showcase the
results that practitioners have had with their
software. The Web site of Decisioneering Inc.
(http://www.decisioneering.com), makers of the
Crystal Ball add-in for Excel, is typical. This
site lists hundreds of sample models used in
industries ranging from aerospace to utilities.
Each of these spreadsheets has been contributed