Professional Documents
Culture Documents
1996
in
Joline Momson
Department of MIS
University of Wisconsin, Eau Claire
Eau Claire, WI 54702
Diane Janvrin
Department of Accounting
University of Iowa
Iowa City, IA 52242
Abstract
End-user developed applications introduce many
control risks into organizations. Literature suggests that
influencing factors include developer experience, design
approach, application type, problem complexity, time
pressure, and presence or absence of reviav procedures.
This research explores the impacts of diffeTent design
approaches through two field experiments evaluating the
use versus non-use of a structured design methodology
when developing complex spreadsheets. Our results
indicated that subjects using the methodology showed a
significant reduction in the number of linking errors,
i.e., mistakes in creating links between values that must
pow from one area of the spreadsheet to another or from
one worksheet to another in a common workbook. We
also observed that factors such as gender, application
influenced
expertise, and Workgroup configuration
spreadsheet error rates as well.
directions.
2. Background
Cotterman
& Kumar
[9] classify
end-user
computing
risks
according
to
three
major
dimensions:
end-user
development,
end-user
operation,
and end-user control of information
systems.
To our knowledge,
no work
has
specifically identified and attempted to empirically
validate factors impacting
outcomes in end-user
developed applications.
Literature suggests that
potential
factors include
developer
attributes,
development
approach, developer
configuration,
problem/process
characteristics
and application
domain. The following sections explore these areas
and identify measures for predicting outcomes.
1. Introduction
During the past decade, the use of end-user
especially microcomputer
developed applications,
However,
spreadsheets, has grown exponentially.
when identifying
and evaluating
risks within
developed
computer
end-user
organizations,
Research has
applications
are often overlooked.
revealed that as many as 44% of all end-user
spreadsheets contain at least one error [5]. Such
errors can have catastrophic results. For example,
a Florida based construction company lost $254,000
by relying on an incorrect spreadsheet analysis for
project costing [34].
Previous research suggests that errors in enduser developed applications may be influenced by
several factors. The purpose of this research is to
of different
design
investigate
the impacts
error
rates and
approaches
on spreadsheet
Developer Attributes
Cotterman and I&mar [9] define an end-user as
unit or person who has an
my 0% anizational
interaction
with the information
system
as a
producer or consumer of information.
Our work
focuses on end-users who are both producers and
consumers of information.
Several studies have
indicated that end-users are diverse, and that this
diversity
leads to a need for differentiated
education, training, support, software tools and risk
assessment (e.g., [14,19, 301).
Harrison
& Rainer [17] assert that specific
attributes
include
computer
gender,
age,
experience, computer confidence, math anxiety, and
cognitive style. Curtis et. al.s field studies on
346
1060-3425/96 $5.00 0 1996 IEEE
1996
Halverson
[28]
note
that
the
increased
communication
and coordination
costs of working
in groups may outweigh the benefits of reduced
error rates.
Problem/Process Characteristics
Several
the
difficulties
studies
identify
versus non-complex
associated
with
complex
programming
problems (e.g., [13, 321). Studies on
the effect of time pressure
on programming
activities indicate that due to time pressure, enduser developers may not spend sufficient time on
problem definition
and diagnosis [2, 61. These
researchers also noted an absence of formal review
procedures
within
end-user
validation
and
developed applications.
Development Approach
activities
for
end-user
developed
Design
applications tend to be unstructured and often nonSalchenberger [33] notes that endl-users
existent.
tend to develop and implement
models without
analysis
or
requirements
first
performing
developing a system design. Brown & Gould [5]
observed that their participants
spent little time
Pl arming before launching into spreadsheet coding.
Ronen, Palley and Lucas [31] suggest that s#everal
attributes of spreadsheets (i.e., non-professional
developers, shorter development life cycles, ease of
modification) tend to preclude a formal sprt?adsheet
analysis and design process.
Almost
all published
spreadsheet
(design
approaches agree on one principle:
the importance
of keeping spreadsheet work areas small in order to
make errors easier to find, limit the damaging
effects of errors, and make spreadsheets easier to
understand and maintain.
Researchers have noted
that structured design methodologies
increase the
accuracy and reliability of information systems (e.g.,
[241). Salchenberger [33] proposes the use of
structured
design methodologies
in end-user
development
and discusses their effectiveness in
two situations. Cheney, Mann & Amoroso [17] also
design
use of
a structured
the
support
methodology; they found that the more structured
the tasks being performed by the end-user, the more
likely the success of the application.
Ronen, Palley
and Lucas [31] describe a structured spreadsheet
design approach based on modifying
conventional
data Bow diagram symbols to create spreadsheet
Several practitioner-directed
flow
diagrams.
articles encourage the use of structured
design
spreadsheet
planning
for
and
approaches
development.
Application
Domain
Research on end-user development
generally
falls into one of two streams: general application
spreadsheet
development.
development
and
(Notable exceptions include Frank, Shamir & Briggs
[15] and Batra [3]). Par&o and Halverson [28]
appropriately
summarize the state of spreadsheet
research as follows:
ln the past, spreadsheeting has been the Rodney
Although
Dangerfield of computer applications.
everyone agrees that it is extremely important, fm
researchers have studied it in any depth. Given its
potential for harm as well as for good, this situation
needs to change.
Psychologists Hendry and Green [20] assert that
spreadsheets
have
the
ability
to
simplify
programming
by suppressing
the complexities
associated with conceptually simple operations like
adding a list of numbers or handling input and
output.
As a result, users tend to ignore
spreadsheet weaknesses such as the potential for
experienced
when
errors
or the difficulties
debugging or reusing spreadsheets.
Outcomes
Salchenberger [33] suggests that factors to use in
of end-user
developed
judging
the quality
use,
applications
include
reliability,
ease of
maintainability,
and auditability.
Reliability can bc
defined as the absence of logical and mathematical
errors in a model.
One of the earliest studies on
spreadsheet outcomes (conducted by Brown and
Gould [5] in 1987), found that of 27 spreadsheets
created by experienced
spreadsheet users, 41%
contained user-generated errors. In another study,
Developer Configuration
Par&o Rr Halverson [28] and Nardi & Miller [26]
both note that the accuracy of end-user developed
applications may be increased by having multiple
collaborate.
However,
Panko
&
developers
347
Developer Attributes
/ *Age
Gender
* Computer Confidence
* Domam Expertise
Application Expertise
* Development Expertise
- Math Anxiety
- Cognitive Style
- Training
* Experience
\.
-x
Application
I
Type
i
End User-Developed
Application Outcomes
-i
Reliability
Ease of Use
Maintainability
Auditability
cost
-+
_+sSS--I
/
l-----l
* Problem Complexity
l
ProblcmiProcess
characteristics
,r
- Existence of
Kevlew Procedures
Development Approach
1996
defines maintainability
as the ability to change or
update the system as needs change, and auditability
as the ability to determine the source of data values
in the model. Few researchers have evaluated endfor
user
developed
applications
either
maintainability
or auditability.
We propose the
inclusion of cost (usually measured in terms of
development time) as another important end-user
development outcome.
Based on prior theoretical and empirical work in
this area, we have developed an initial research
framework
(Figure 1) that summ arizes end-user
development risk factors and their relationships to
potential outcomes. We now focus on determining
the impacts
of
development
approach
on
The particular
spreadsheet reliability
and cost.
structured design methodology used in our studies
is now described.
Framework
Figure
348
External
2. Basic
Modeling
Symbols
1996
Homebuilders
Income
Statement
Market
pgi--pshaE
\_,/
Figure 3. Example
Manufacturer
Design
for Small
4. Evaluation
Home Fixture
between
developer
confidence
and improved
outcomes was noted by Harrison & Rainer [17].
Prior research has also noted a positive relationship
between domain and application
expertise and
improved outcomes (e.g., [2, 61).
Structured
systems design approaches
(e.g.
Constantine and Yourdon [S]), have successfully
been used in systems development for over twentyfive years.
This approach can be applied to
spreadsheet
design
by
identifying
different
worksheets and their corresponding
data inflows
and outflows.
Thus, each individual
worksheet
effectively
becomes a black box that takes a
defined set of inputs and translates it into a
prescribed set of outputs. Since the worksheet links
are explicitly
defined in the design, the result
should be a decreased error rate for linking errors.
Study #l
Subjects, Treatment, and Measures.
Sixty-one
upper- and masters-level accounting and business
administration
majors in an accounting information
systems course at a large public university were
assigned to the treatment (structured design) and
control (ad hoc design) groups based upon course
sections. Assignment
of treatments
to specific
sections was done randomly.
All subjects were
given one hour of formal training in spreadsheet
development,
and then instructed
to develop a
single spreadsheet workbook containing a series of
linked worksheets using Microsoft
Excel.
The
assigned
problem
contained
ten
separate
worksheets, with a total of fifty-one linked data
values among the worksheets. A paper template of
Studies
349
ard Deviations
fStudv
#I 1
j
/
I.441
1.38 /
0.50,
0.78:
1.50
1.56
I
;
1
l.QO/
2.491
2.341
0.68;
1.07;
1.14~
2.161
0.99
0.73
0.94
-FiT ._.--.
Eraertise-.-~
-. - -_r*Pretest Design Confidence
t
t
~
0.89:
(
0.981
(I.58
2.11,
2.61 ~ 1.20
1.301
1.821
2.431
1.02
2.68
2.221
1.00
2.41 /
1.19
2.42
1.89
3.721
2.87 j
0.831
1.89
2.301
4.02 1
1.021
0.63 1
2.261
3.841
I.001
0.781
0.671
0.69
1.191
1.97
0.541
0.60 /
1.w
1.83
0.75
1.053'
2.30
1 .OQ
I.231
2.27,
1.05
2 07
4.031
0 98
U.Y41
0.67
2.73 j
2.931
2_40
4.85
9.51%/
TxZy
1996
t
i
N = 66
"p <= "1; **p<
=.05
The MANOVA
analysis results imply that for
subjects working
alone, the use of the design
significant
had
a
impact.
methodology
Unfortunately,
the advantage
of using
the
structured
design methodology
was weakened
Two
when
the subjects were working in pairs.
potential explanations emerge: (1) subjects working
in pairs gained other advantages that offset the
350
Development
Approach
r* Ad hoc design
(I, 2)
l Saucwred
design
(I. 2)
Figure 4. Model
Reliability
-Link Errors (I, 2)
-Conceptual
Enors (2)
-Oversight
Errors (2)
cost
-Development
Time (1, 2)
* Dev. Approach
Usability (2)
for Experimental
1996
Outcomes
l
Studies.
351
1996
disagreed
with
statements
they
agreed
or
the DFD
design
regarding:
(1) whether
methodology
helped
them
develop
a better
spreadsheet; (2) if using the methodology made the
development
effort take longer; and (3) whether
they would use the methodology the next time they
developed a complex spreadsheet.
Study #2
The problem used in Study #2 was slightly more
complex than the one used in Study #l,. with ten
different worksheets and a total of sixty-six linked
cells. All
subjects
were
required
to work
individually
so as to increase sample size and allow
a more isolated investigation
of the impacts of the
design methodology
and developer
attributes.
Error types observed were expanded to include
conceptual and oversight errors as well as linking
errors; a blank template (from Hilton [2I], p. 439443) with only a single check figure was given.
Explicit post-test measures were introduced
to
assess design confidence, coding confidence, and
confidence that the spreadsheet was error-free.
Additionally,
a measure was included to explore
the impact of potential domain expertise:
subjects
who
were accounting
majors
had additional
experience and coursework in preparing financial
reports and other master budgeting topics than the
non-accounting majors.
Subjects in the treatment
group were also asked to rate their post.-treatment
confidence,
and to rate their
DFD development
the development
approach.
satisfaction
with
Specifically, they were asked the degree to which
352
Table 5. Regression
Analysis
Results
1996
(Study #2)
0.57
0.38
1.02
0.24
o:ss
2.72
0.17
0.08
2.52
-0.12
-0.19
0.29
0.58
-0.15
-5.75
0.06
0.00
-5.63
1.03
0.30
0.29
0.30
0.43
0.06
-3.18
-0.07
-0.15
-3.22
-1 .ll
Pre. DFD
0.14
0.16
0.27
0.16
0.04
-0.34
0.47
0.02
-0.23
-0.64
lomein Experience
(l=Other,
reatment Group (l=Control.
Ex~ertiie
2=Acct
2=Treat)
-3.66
Using
the structured
design methodology
significantly
(p<=.Ol) reduced the number of link
errors, but
had no impact on the number of
oversight or conceptual errors. Interestingly, link
errors were by far the most common error made.
Each spreadsheet (for the combined control and
treatment groups) had an average of 9.07 total
It is
errors; of these, 84% were linking errors.
the design
that
using
important
to note
methodology
had no significant impact on total
time spent on the project (aside from the time taken
to learn the methodology initially).
353
Limitations,
1996
inadvertently
that
suggested
developer
configuration
is an important
factor and has
potential
interactions
with
the development
approach.
As mentioned previously, our initial study was
somewhat confounded because the pairs were selfselected.
Our future studies will address the
impacts of collaborating in non-self-selected groups.
Additionally,
our research has focused on absolute
error rates. A limitation
of this approach is the
question of how error rates are correlated to
ultimate outcomes. i.e., the bottom line or poor
decisions resulting from spreadsheet models. Also,
the impact of the structured development approach
on problems of differing complexity needs to be
explored. Another area that needs to be addressed
in the future involves the outcomes associated with
the use of end-user developed spreadsheets by,
other
users
in
terms
of
usability,
end
maintainability,
and audibility.
Additionally,
it will
be interesting to explore the results when end-user
developers know that their spreadsheets will be
reviewed
or audited.
Finally, we have only
considered spreadsheets; we need to determine if
our overall causal model is applicable to other enduser development domains as well.
As
development
increases
in
end-user
organizations
the identification
of strategies to
minimize risks and improve outcomes is becoming
increasingly important.
This research has made an
important contribution both by identifying end-user
development
risk factors and outcomes, and by
developing and empirically
evaluating the impacts
of a structured
design approach for spreadsheet
development.
;;<=.05;pc=.Ol; pc=.OO1 1
5. Conclusions,
Research
References
Adelson, 13. and Soloway, E., The role of
domain experience in software design, IEEE
Trans. on Swftware Eng., SE-11 (Nov. 1985), 13571360.
Alavi,
M. & Weiss, I., Managing
the risks
associated Twith end-user computing, J. MIS 2,3
(Winter 1986), 413426.
for studying human
Batra, D., A framework
error
behavior
in
conceptual
database
modeling, Info. & Management 25 (Sept. 1993),
121-131.
Billings, R.S. and Wroten, S.P., Use of path
analysis
in
industrial/organizational
psychology:
Criticisms
and suggestions, 1.
Applied Psych. 63 (June 1978), 677-688.
and Future
confirnned
Our
experimental
studies
that
developer
attributes,
specifically
gender
and
domain and application
expertise,
significantly
impacted error rates. Additionally,
we have shown
that using the structured
design methodology
significantly improved reliability
for two different
linked spreadsheet applications.
Study #I also
354
6.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
VW.
355