Contrast Coding: A Refinement of ANOVA in Behavioral Analysis

Author(s): Frank A. Buckless and Sue Pickard Ravenscroft

Source: The Accounting Review , Oct., 1990, Vol. 65, No. 4 (Oct., 1990), pp. 933-945
Published by: American Accounting Association

Stable URL: https://www.jstor.org/stable/247659

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

American Accounting Association is collaborating with JSTOR to digitize, preserve and extend
access to The Accounting Review

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
THE ACCOUNTING REVIEW
Vol. 65, No. 4
October 1990
pp. 933-945

Contrast Coding: A Refinement of

ANOVA in Behavioral Analysis
Frank A. Buckless
North Carolina State University
Sue Pickard Ravenscroft
Eastern Michigan University

SYNOPSIS: As behavioral accounting research has progressed, research

designs have become more complex. New topic areas are initially investi-
gated using single-factor designs, which may result in significant effects
and a substantial amount of explained variation. In subsequent investi-
gations, explained variation is increased by employing factorial designs
and/or multiple-level explanatory variables.
Traditionally, ANOVA is the statistical analysis approach employed to
test research hypotheses. A limitation of ANOVA is that it only detects sig-
nificant differences among cell means, but does not indicate the functional
form of the relationship among cell means. We propose that researchers
employ contrast coding-a refinement of ANOVA-to test research
hypotheses. Contrast coding requires researchers to specify a priori the
functional form of the relationship among cell means.
This article demonstrates that contrast coding provides greater
statistical power than the conventional ANOVA without increasing Type I
error rates. Examples from recent accounting literature illustrate when the
use of contrast coding is most advantageous. The examples include a
factorial design and a single-factor, multiple-level experiment. Formulas for
calculating the net benefit of employing contrast coding and the signifi-
cance of the net benefit are presented. The examples and analysis support
the use of contrast coding with certain research designs. The primary
benefit of the a priori specification required by contrast coding is greater
statistical power.

Key Words: Contrast coding, ANOVA, Behavioral research, Statistical

power.

Submitted May 1989.

Revisions received November 1989 and April 1990.
Accepted May 1990.

933

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
 934 The Accounting Review, October 1990

TN a discussion of the increasing num

interactions among explanatory variables, Keppel (1982, 180) says that "if behavior
is complexly determined," we will need factorial experiments that allow explora-
tion of interactions. In a like manner, researchers in behavioral aspects of accounting
would benefit by applying the tools available to analyze interactions.
Recent articles concerning cognitive distinctions between expert and novice audi-
tors (e.g., Butt 1988), auditor performance (Kaplan and Reckers 1984), and contract
effects on performance (Chow et al. 1988) provide evidence that interactions among
explanatory variables are replacing main effects as the focus of research interest.
The purpose of this note is to illustrate the increased statistical power of contrast
coding over the conventional ANOVA when multiple-level factors are involved or when
certain types of interactive relationships among factors are hypothesized. This in-
creased power both decreases the likelihood of incorrectly finding nonhypothesized
effects and strengthens the probability of finding hypothesized effects when they
actually exist. Furthermore, contrast coding provides this advantage without increas-
ing the likelihood of Type I error, the rejection of a null hypothesis when the null is
true. The use of contrast coding presumes that researchers develop theoretical formula-
tions for their hypotheses.
The remainder of the note is organized as follows. First, a discussion of ANOVA
and contrast coding is presented and illustrated. Next, recent accounting research that
focuses on interactive relationships is briefly described, followed by two extended ex-
amples of the use of contrast coding. The examples are a 2 x 2 factorial experiment and
an experiment with one multiple-level factor. The final section includes the conclu-
sions of this note.

I. Discussion of ANOVA and Contrast Coding

A limitation of ANOVA is that it indicates whether cell means are different from
one another, but does not specify the pattern of relationships among cell means (i.e., the
source of the differences among means). Furthermore, ANOVA spreads the variance
among the multiple patterns being tested, thereby altering the likelihood of finding sig-
nificance for any of the patterns. Using the main and interaction effects of ANOVA to
explain the pattern of relationships among cell means is appropriate only when the pat-
tern hypothesized is testable by the conventional ANOVA and no variable has more
than two levels.
For any set of G sample means being compared, many possible sets of G -1 orthog-
onal contrasts exist. Consider the case of the simple 2 x 2 ANOVA, with two levels (high
and low) for each factor. The default weights assigned by ANOVA for testing the main
and interaction effects are shown in table 1. ANOVA requires that the weights for each
contrast sum to zero and that the sum of the product for each pair of contrasts also sum
to zero. The first constraint ensures estimation of population parameters. The combina-
tion of the first and second constraints ensures orthogonality of the contrasts. Keppel
(1982, 112) provides a simple formula for computing the sum of squares for ANOVA
using cell sizes, cell means, and default weights. The sum of squares formula is as
follows:,

SSmodel s[=(cJ(ABJ](
Etci)2

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
 Buckless and Ravenscroft-Contrast Coding 935

Table 1
Weights Assigned to Treatment Cells by ANOVA

Interaction of
A B Factor A Factor B AxB

Low Low -1 -1 1
Low High -1 1 -1
High Low 1 -1 -1
High High 1 1 1

0 0 0

where:

SS = sum
s = numb
c=weights assigned to the cell means,
= =cell identifier, and
AB = cell means.

The disordinal interaction represented in figure 1 is used to illustrate the ANOVA

sum of squares calculations. The ANOVA sum of squares calculation for the interaction
effect is presented in table 2. Notice that the total between-group sum of squares is allo-
cated to the interaction effect.
Although the conventional 2 x 2 ANOVA is appropriate for detecting effects with
the pattern of means shown in figure 1, ANOVA is less powerful as a statistical tool for
testing ordinal interactions2 of the type illustrated in figure 2 (Bobko 1986; Hunter
1982a) when no main effects are hypothesized. Hypothesized patterns of means fre-
quently involve nonsymmetric patterns (Bobko 1986; Staw 1976) such that the effect is
due entirely to the difference of one experimental cell from the other three cells, which
are approximately equal.
The pattern of means in figure 2 shows a significant effect occurring in only one of
the four cells. The ANOVA sum of squares calculation for the ordinal interaction in
figure 2 is presented in table 3, panel A. Although main effects for factors A and B

The derivation of this formula following the standard regression approach is:

Sum of squares regression = (EXy)2 _ [sI~c(AB) fl2 - s[E(c)AB]2

x2 sE(c)2 C)
where:
x = weight assigned to each observation, ANOVA assignes either + 1 or - 1,
y =raw score (dependent measure),
c =weight assigned to each cell mean,
s = number of subjects per cell, and
AB= cell means.

2 A mathematical definition of an interaction requires that the coefficients sum to zero within each ro
within each column of the design. Social scientists broadly define interactions as effects that differ depending
upon the levels of several variables combined (see Keppel 1982, 178). We are using this broader definition of an
interaction in this note.

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
936 The Accounting Review, October 1990

Figure 1
Graph of Disordinal Interaction

5+

a 4

3+

ol,
0
Low High

Factor A
5 Factor B - Low + Factor B - High

Panel A. Disordinal Interaction:

FACTOR A
Low High

F
A Low 5 3
C
T
0
R

B High 3 5

appear in table 3, the cru

B are both high. The misleading main effects of table 3 occur because of the default
weights assigned by ANOVA for testing main and interaction effects (see table 1). The
conventional ANOVA spreads the between-group sum of squares among the main and

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
 Buckless and Ravenscroft-Contrast Coding 937

Table 2
ANOVA Sum of Squares Calculation for Figure 1

Sum of Squares for A Main Effect:

s[(-1)5 + (-1)3 + (1)3 + (1)5]2 -0

- _sO
(-1)2 + (1)2 + (1)2 + (1)2 4
Sum of Squares for B Main Effect:

s[(-1)5 + (1)3 + (1)3 + (1)5]2 - _sO

(-1)2 + (1)2 + (1)2 + (1)2 4

Sum of Squares for A x B Interaction Effect:

s[(5)3 + (1)3 + (1)3 + (1)5]2 = s16=

(1)2 + 1)2 + (1)2 + (1)2 4

Table 3
Sum of Squares Calculation for Figure 2

Panel A. ANOVA Sum of Squares Calculation:

Sum of Squares for A Main Effect:

s[(-1)3 + (1)3 + (1)3 + (1)5]2 -s4 =s1

(_1)2 + (1)2 + (1)2 + (1)2 4
Sum of Squares for B Main Effect:

s[(-1)3 + (1)3 + (1)3 + (1)5]2 = s4=

(-1)2 + (1)2 + (1)2 + (1)2 4

Sum of Squares for A x B Interaction Effect:

s[(1)3 + (1)3 + (-1)3 + (1)5]2 -s4= sl

(1)2 + (1)2 + (1)2 + (1)2 4

Panel B. Contrast Coding Sum of Squares Calculation:

Sum of Squares for Model Contrast:

s[(-1)3 + (1)3 + 1)3 + (3)5]2 s36

( - 1)2 + (1)2 + (1)2 + (3)2 12 s3

interaction effects. However, reliance on the main effects "provides on

often misleading picture of what has happened" (Kirk 1982, 358) and inc
lihood of misinterpreting the results (Keppel 1982, 179; Hays 1981, 365)
We propose that if researchers expect patterns of means such as tho
figure 2, or employ research designs with more than two treatment lev
rely on contrast coding instead of conventional ANOVA. Contrast coding requires a
priori reasoning and explicit hypotheses, which enable researchers to select appropri-
ate weights.
Contrast weights for the relationship in figure 2 can be determined by subtracting

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
938 The Accounting Review, October 1990

Figure 2
Graph of Ordinal Interaction

5 -

a) 4

0
Low High

Factor A
l Factor B - Low + Factor B - High

Panel A. Ordinal Interaction:

FACTOR A
Low High

F
A Low 3 3
C
T
0
R

B High 3 5

the overall expected mean fo

priate set of weights for the
ANOVA, the contrast weights assigned to the experimental cells should sum to zero.3

3Hays (1981, 641-45) provides an excellent discussion on the selection of suitable weights.

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
 Buckless and Ravenscroft-Contrast Coding 939

Table 4
Weights Assigned to Treatment Cells by Contrast Coding

A B Contrast

Low Low -1
Low High -1
High Low -1
High High 3

Except for the difference in weights, the ANOVA SSmodel in equation (1) can also
be used to calculate the sum of squares for contrast coding. The SSmodel for figure 2
using contrast coding is presented in table 3, panel B. Contrast coding allocates the
total between-group sum of squares entirely to the model contrast.4 A comparison of the
expected sum of squares for ANOVA and contrast coding shows the F-value for the
latter to have an expected value three times as large as that for any of the ANOVA
effects.5

II. Empirical Examples

Research Involving Interactions

In this section we briefly review examples from recent accounting literature pre-
dicting interactions and provide a numerical example of how contrast coding can be
employed.
Blocher et al. (1986) predicted a disordinal interaction tested by the conventional
2x2 ANOVA. They hypothesized that decision accuracy would be affected inter-
actively by task complexity and mode of information presentation. Their prediction
was that accuracy would be higher with graphic information in a low complexity task,
higher with tabular information in a more complex task. Given this hypothesized dis-
ordinal interaction, ANOVA is the recommended analysis.
Butt (1988) explored frequency encoding by experienced and novice auditors using
two modes of presentation of information and two tasks. Using a similar paradigm,
Bonner (1990) measured differences in cue weighting and cue selection between expe-
rienced and novice auditors in two risk assessment tasks. Performance was expected to
differ between the two subject groups on one task but not on the other. Similarly,
Marchant (1989) hypothesized an interaction between audit expertise and the ability to
generate hypotheses using analogies. The hypotheses in these three studies were sta-
tistically tested and moderately supported by using ANOVA and additional t-tests.

4To rely on an interactive effect using ANOVA, the predicted pattern of means would imply that both main
effects as well as the A x B interaction must be significant. The tests for these effects are lower in power, how-
ever, than the single degree of freedom test using contrast coding.
5 This implies that contrast coding has higher power than ANOVA. If the expected treatment cell means
were 3, 3, 3, 3 (no effect) instead of 3, 3, 3, 5, the expected F-value would be zero for both contrast coding and
ANOVA. This implies that the alpha risk is identical for both methods.

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
 940 The Accounting Review, October 1990

Their results could be presented more powerfully and efficiently, however, if contrast
coding were used.6 An illustration from another discipline of the use of contrast coding
to examine the effect of expertise is Bettman and Sujan (1987).
In another study focused on interactions, Kaplan and Reckers (1984) examined the
effect of the audit client's environment upon auditors' risk assessment regarding a par-
ticular accounting cycle. They asked two groups of subjects (audit seniors and audit
managers) to assess the likelihood of error in the accounts receivable system twice, first
after reading a general client scenario and then again after reading additional materials
regarding the client's internal controls over inventory. No statistically significant dif-
ferences were found between the two groups using ANCOVA. Again, the results could
be tested more powerfully and efficiently using contrast coding."
The role of interactive effects in human information processing continues to be an
important and unresolved question. Experts in many fields report that they use cues in
a configural (interactive) fashion. However, most results from research on human infor-
mation processing indicate that such processing is linear or can be adequately repre-
sented by linear models. Brown and Solomon (1989, 10) argue that configural infor-
mation processing has not been detected for several reasons. They cite Yntema and
Torgerson (1961), who showed "that even when information processing strategies are
known to be configural, ANOVA models will attribute most of the explained variance to
main effects rather than ordinal interactions." The discussion in the previous section of
this note demonstrates why that effect occurs. Brown and Solomon (1989, 6) urge re-
searchers to "have expectations concerning specific cue interactions, including the
nature (form) of such interactions, and to employ research designs capable of effec-
tively detecting expected interactions."
The audit expertise paradigm (Libby 1987) is used to illustrate the employment of
contrast coding. The audit expertise paradigm predicts an ordinal interaction of the
type illustrated in figure 2. Factor A represents experience: novices designated by low
and experts by high. Factor B represents the type of task: generic tasks designated as
low and domain-specific ones as high. The default weights assigned by ANOVA for
testing the main and interaction effects are shown in table 1. Unless a researcher spec-
ifies alternative weights, ANOVA automatically uses the default weights.
Hypothetical data consistent with the hypothesized effect in figure 2 are presented
in table 5. These hypothetical data are used to illustrate the calculation of the ANOVA
sum of squares. The cell size was seven. In this example, an equal number of subjects
per cell is assumed, but appropriate adjustments for unequal cell sizes can be made (see
Keppel 1982, 347-48). The calculations of the ANOVA sum of squares corresponding to
the main and interaction effects are presented in table 6, panel A. The ANOVA sum-
mary results are presented in table 7, panel A.
The statistical analysis of the data in figure 2 should also include a test of the expec-
tation that the other three cells are statistically equivalent, since the entire predicted

6 We were unable to recalculate results using contrast coding, as the cited articles did not provide data
necessary to perform the computations.
I Had Kaplan and Reckers (1984) hypothesized that audit seniors' risk assessments would be high in the
low control consciousness condition and that the other three experimental cells would be approximately equal
(based on the representativeness heuristic), the resulting F-statistic would have been 3.53, p = 0.07. This is in
contrast to the reported results of F=0.708 and p=0.40.

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
 Buckless and Ravenscroft-Contrast Coding 941

Table 5
Cell Means of Numerical Example

Experience

Task Novice Expert

Generic 2.71 2.71 2.71

Domain Specific 2.43 4.71 3.57

2.57 3.71 3.14

Table 6
Sum of Squares for Numerical Example

Panel A. ANOVA Sum of Squares Calculations:

Experience Main Effect:

7f(- 1)2.71 + (- 1)2.43 + (1)2.71 + (1)4.7112 36.40

= = 9.10
(1J2 +(1J2 +(1)2 +(1J2 4

Task Main Effect:

7[(- 1)2.71 + (1)2.43 + (- 1)2.71 + (1)4.7112 20.72

= = 5.18
(1)2 + (1)2 + (1J2 + (1)2 4

Experience by Task Interaction Effect:

7[(1)2.71 + (- 1)2.43 + (-1)2.71 + (1)4.7 112 36.39

_= =9.10
(1) +(-1) +(1)2 +(1)2 4

Panel B. Contrast Coding Sum of Squares Calculations:

Model Contrast:

7[(- 1)2.71 + (- 1)2.43 + (1)2.71 + (3)4.7112 276

1)2 + )2(-12(-12_1+(3)2
+ 12 += = 12
23.01
Between-Group Residual Sum of Squares:
23.38 - 23.01 = .37

effect is the difference of one cell from the other three cells. The equality of
ing three cells is tested by performing a semiomnibus F-test8 on the remain
plained between-group residual.9 The numerator for the semiomnibus F-test

8 A semiomnibus F-test has greater than one degree of freedom in the numerator.
' A total of three orthogonal contrasts can be constructed for the current example. According to Cohen and
Cohen (1975, 197), the second and third orthogonal contrasts are necessary to explain all the variance of the
data. Winer (1971, 175) states that "in practice the comparisons that are constructed are those having some
meaning in terms of the experimental variables." This suggests that only comparisons of theoretical interest
should be tested. Use of the semiomnibus F-test has the advantage of testing precisely the theoretically pre-
dicted equalities, rather than testing relationships solely because they comprise a set of complete and orthog-
onal contrasts.

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
 942 The Accounting Review, October 1990

Table 7
Sum of Squares Summary Tables for Numerical Example

Sum of Mean
Factor Squares df Square F-Statistic p

Panel A. ANOVA Summary Table:

Experience (E) 9.10 1 9.10 2.39 >.05

Task (T) 5.18 1 5.18 1.35 >.05
ExT 9.10 1 9.10 2.39 >.05

23.38

Error 91.92 24 3.83

Panel B. Contrast Coding Summary Table:

Model contrast 23.01 1 23.01 6.01 <.05

Residual .37 2 .19 0.05 >.05

23.38

Error 91.92 24 3.83

the total between-group sum

remaining ANOVA degrees of freedom. The denominator is equal to the ANOVA error
sum of squares divided by the within-group degrees of freedom.
The calculation of the sum of squares based on the contrast weights in table 4 is pre-
sented in table 6, panel B. The resulting ANOVA summary results are presented in table
7, panel B. 10 The total sum of squares is identical in both analyses; only the allocation of
that total among the contrast differs because of the choice of weights. Contrast coding
allows the determination of weights to be based on a theoretical model.

Research Involving Multiple-Level Factors

The greatest advantage of using contrast coding instead of ANOVA arises when
multiple-level factors are present in an experiment. Chow (1983) is used to illustrate t
employment of contrast coding when there are more than two levels of a manipulate
variable. Chow predicted that tightness of performance standards and type of pay
scheme would both affect performance." In particular, he predicted that performance
(output) under a budget-based pay scheme would exceed performance under either a
fixed or piece-rate pay scheme. The conventional ANOVA tests for any differences
among the three pay schemes, yet Chow very clearly indicated the exact type of differ-
ence he expected.
Table 8 reports the cell means and standard deviations from the data in Chow

10 Two sets of contrast weights orthogonal to the model contrast weights are (-1,-1, 2, 0) and (-1, 1, 0, 0
Using equation (1), the total sum of squares for the two contrasts is 0.37.
" The following analysis is based on Proposition One of Chow's 1983 article.

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
 Buckless and Ravenscroft-Contrast Coding 943

Table 8
Cell Means for Effect of Pay Schemes and
Performance Standards on Performance

Performance Standard

Average Tight

Fixed pay 26.58 (6.1) -1 28.78 (4.7) -1

Budget-based 27.44 (6.5) 2 28.69 (4.7) 2

Piece-rate 29.22 (5.4) -1 34.56 (5.7) -1

Source: Chow 1983.

(1983, 676). Included in the table are the contrast weights we chose based on Chow's
hypothesis. The reported F-statistic for the pay scheme main effect was 7.86 (p < 0.001).
If the contrast weights had been used instead, Chow would have obtained an F-statistic
of 2.86 (p <0.098). This particular application of contrast coding is useful because
it highlights the need to validate the hypothesized model using contrast weights. The
initial results indicated a significant main effect, but it was not the main effect speci-
fied. Visual inspection of the results in table 8 reveals that the highest performance was
attained in the piece-rate condition and not-as was predicted-in the budget-based
condition. A comparison of the piece-rate mean and the budget-based mean reveals a
significant difference in the direction opposite of that hypothesized. As Chow (1983)
noted, the obtained results are contrary to predictions based on agency theory and for
this reason, warrant further investigation. The use of contrast coding can reduce the
likelihood of falsely accepting research hypotheses.
Rosenthal and Rosnow (1985, 86) express the net benefit of employing contrast
coding instead of ANOVA to test hypotheses involving variables with more than two
levels as follows:12

E(r2) - (1/dfl. (2)

For example, the net benefit of employing contrast coding for Chow (1983) would be
0.20, assuming an expected r2 of 0.70 (i.e., 0.70-1/2). Rosenthal and Rosnow's net
benefit formula highlights the direct relationship between the benefit of using contrast
coding and increases in the degrees of freedom. For example, if the expected r2 is 0.70
and there are six levels, the expected benefit of using contrast coding would be 0.50.
That is, the net expected benefit of contrast coding increases from 0.20 to 0.50 with an
increase in degrees of freedom from two to five.
Rosenthal and Rosnow (1985, 89) also provide a test of significance of the expected
benefit in information. The formula requires use of the degrees of freedom for the effect

'2 The expected benefit is the amount of information in the contrast degree of freedom after subtracting the
information in the unselected average degrees of freedom, or the net information per contrast degree of free-
dom (Rosenthal and Rosnow, 1985, 86). The proportion of variance accounted for by the contrast is represented
by r2, commonly referred to as eta squared (iq2) or the squared multiple-correlation coefficient (R2).

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
 944 The Accounting Review, October 1990

from which the contrast is taken and the r2. The formula is:

r2(df- 1)/(1 - r2). (3)

This value is distributed as F, with 1 and (df- 1) as the numerator and denominator
degrees of freedom. As Rosenthal and Rosnow point out, this statistic is conservative.

III. Concluding Remarks

Reliance on the default coefficients of ANOVA is valid and beneficial for explor-
atory research. As theories are developed and refined, however, researchers should
develop explicit hypotheses regarding the form of the relationships between indepen-
dent and dependent variables. This article demonstrates the need to specify a priori th
weights to be used in ANOVA analysis. These weights are to be determined by theory
and are not to be left to the arbitrary weighting scheme used in conventional ANOVA
analysis. Determining weights a priori requires greater effort than relying on conven-
tional ANOVA weights, but the potential benefits to researchers are substantial. Greater
precision in hypotheses formulation provides better guidance for future research
efforts.

References
Bettman, J. R., and M. Sujan. 1987. Effects of framing on evaluation of comparable and
noncomparable alternatives by expert and novice consumers. Journal of Consumer
Research 14 (September): 141-54.
Blocher, E., R. P. Moffie, and R. W. Zmud. 1986. Report format and task complexity: Interaction
in risk judgments. Accounting, Organizations and Society 11 (6): 457-70.
Bobko, P. 1986. A solution to some dilemmas when testing hypotheses about ordinal
interactions. Journal of Applied Psychology 71 (2): 323-26.
Bonner, S. 1990. Experience effects in auditing: The role of task-specific knowledge. The
Accounting Review 65 (January): 72-92.
Brown, C. E., and I. Solomon. 1989. Configural information processing in auditing: A
theoretical and empirical analysis. In Audit Judgment Symposium. Los Angeles:
University of Southern California.
Butt, J. L. 1988. Frequency judgments in an auditing-related task. Journal of Accounting
Research 26 (Autumn): 315-30.
Chow, C. W. 1983. The effects of job standard tightness and compensation scheme on
performance: An exploration of linkages. The Accounting Review 58 (October): 667-85.
, J. C. Cooper, and W. S. Waller. 1988. Participative budgeting: Effects of a truth-inducing
pay scheme and information asymmetry on slack and performance. The Accounting
Review 63 (January): 111-22.
Cohen, J., and P. Cohen. 1975. Applied Multiple Regression/Correlation Analysis for the
Behavioral Sciences. New York: John Wiley & Sons.
Hays, W. L. 1981. Statistics. 3d. ed. New York: Holt, Rinehart & Winston.
Hunter, J. E. 1982a. Critique of analysis of variance: Overview. Working paper. Michigan State
University.
. 1982b. Combinatorial treatments: A defect in ANOVA. Working paper. Michigan State
University.
Kaplan, S. S., and P. M. J. Reckers. 1984. An empirical examination of auditors' initial planning
processes. Auditing: A Journal of Practice & Theory 4 (Fall): 1-19.
Keppel, G. 1982. Design and Analysis: A Researcher's Handbook. Englewood Cliffs, NJ:
Prentice-Hall.
Kirk, R. E. 1982. Experimental Design. 2d ed. Belmont, CA: Brooks/Cole.
Libby, R. 1987. Discussion of human judgment: Limitations and opportunities for research. In
1987 Arthur Young Professors' Roundtable. Columbus: The Ohio State University: 99-103.

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms
Buckless and Ravenscroft-Contrast Coding 945

Marchant, G. 1989. Analogical reasoning and hypothesis generation in auditing. The

Accounting Review 64 (July): 500-13.
Rosenthal, R., and R. Rosnow. 1985. Contrast Analysis: Focused Comparisons in the Analysis
of Variance. Cambridge, England: Cambridge University Press.
Staw, B. 1976. Knee-deep in the big muddy: A study of escalating commitment to a chosen
course of action. Organizational Behavior and Human Performance 16 (1): 27-44.
Winer, B. J. 1971. Statistical Principles in Experimental Design. 2d. ed. New York: McGraw-
Hill.
Yntema, D. B., and W. S. Torgerson. 1961. Man-computer cooperation in decisions requiring
common sense. IRE Transactions on Human Factors in Electronics 2 (March): 20-26.

This content downloaded from

202.43.93.79 on Wed, 20 Sep 2023 04:54:30 +00:00
All use subject to https://about.jstor.org/terms