Management Accounting Research 25 (2014) 6–29

Contents lists available at ScienceDirect

Management Accounting Research

journal homepage: www.elsevier.com/locate/mar

Relating alternative forms of contingency fit to the

appropriate methods to test them
Michael Burkert a,∗ , Antonio Davila b , Kandarp Mehta b , Daniel Oyon a
a
HEC Lausanne, Faculty of Business and Economics, University of Lausanne, 1015 Lausanne, Switzerland
b
IESE Business School, University of Navarra, Avenida Pearson, 21, 08034 Barcelona, Spain

a r t i c l e i n f o a b s t r a c t

Article history: At the core of contingency theory, a major theory in management accounting, is the concept
Received 8 November 2011 of fit. We critically discuss forms of fit as presented in overview articles from the manage-
Received in revised form 15 July 2013
ment accounting field, highlighting forms of fit that have not appeared in prior overview
Accepted 30 July 2013
articles (matching fit with hetero-performance on the fit line and/or asymmetric effects
of mis-fit on performance). We also address some confusing arguments in the literature
Keywords:
concerning the moderation form of fit and what has been referred as the mediation form
Contingency theory
Matching form of fit of fit. In a second step, we reevaluate the appropriateness of statistical techniques used
Moderation form of fit to test sub-forms of fit, highlighting the difficulties in differentiating conclusively between
Mediation form of fit them. Specifically, we present polynomial regression analysis (PRA) in conjunction with the
Polynomial regression analysis response surface methodology (RSM) as a powerful methodological alternative and discuss
Covariance-based structural equation its ability to differentiate between the sub-forms of fit. We also discuss the strengths and
modeling weaknesses of structural equation modeling (SEM) to test for forms of fit.
Monte Carlo simulation
© 2013 Elsevier Ltd. All rights reserved.

1. Introduction Moers (1999, 2003) made a first important step addressing

Contingency theory is a major theory in management
accounting research (Chenhall, 2003; Fisher, 1995; Franco-
Santos et al., 2012). Much of the contingency work in
management accounting evaluates the impact of states of
misfit on performance based on the matching, moderation,
or mediation form of fit (Gerdin and Greve, 2004, 2008).
The challenge for researchers is to theoretically derive and
properly test the predicted form of fit. Failure to accept a
true form or reject a false form of fit severely impedes the
advancement of theory-consistent management account-
ing knowledge (Luft and Shields, 2003). Hartmann and
∗ Corresponding author. Tel.: +41 798627835.
E-mail addresses: Michael.Burkert@unil.ch (M. Burkert),
ADavila@iese.edu (A. Davila), Kmehta@iese.edu (K. Mehta),
Daniel.Oyon@unil.ch (D. Oyon).
issues related to the mismatch between verbal statements
of hypotheses in management accounting research and
the use of moderated regression analysis (MRA). More
recently, Gerdin and Greve (2008) discussed the suitability
of different techniques to test contingency theory’s forms
of fit, calling for hypothesizing and testing more specific
sub-forms of fit. Building on their research, we address
several remaining ambiguities regarding how to conduct
contingency research. These ambiguities are related to both
specific sub-forms of fit and the power of the methods
to test them. Our aim is to reinforce the theoretical and
methodological ground of future contingency-driven stud-
ies in management accounting.
Adequate theoretical progress relies on empirical meth-
ods that are able to discriminate between different forms of
fit so that they allow “(. . .) coming up with the most likely
story by eliminating alternative explanations” (Hartmann
and Moers, 2003, p. 808). Our review of recommendations

1044-5005/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.mar.2013.07.008
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
and practices in management accounting reveals a number
of unsettled issues related to the theorizing and testing of
contingency hypotheses. After addressing the sub-forms of
fit, we discuss the appropriateness of various methods to
conclusively test for them.
First, we compare different classifications of forms of
contingency fit. We find that some forms of fit that have
been discussed and tested in management accounting liter-
ature do not belong to contingency theory (mediation form
of fit) while others have not been addressed (contingency
theory’s matching form of fit with hetero-performance
on the fit-line (Donaldson, 2001) and/or with asymmet-
ric effects of mis-fit on performance (Klaas and Donaldson,
2009)).
Second, we evaluate the main methods discussed in the
literature for testing contingency theory predictions based
on matching and moderation forms of fit. We build on the
work of Gerdin and Greve (2008), who discuss the appro-
priateness of various methods to test sub-forms of fit but
take a more critical view. We find specifically that the tra-
ditional approaches used to test matching forms of fit are
at high risk of reaching incorrect conclusions. We explain
why these methods may either not be powerful enough to
detect existing matching forms of fit or may even lead to
erroneous acceptance of such a form of fit when the null
hypothesis or another form of fit is reflected in the dataset
instead. We propose that extending moderated regression
analysis (MRA) to polynomial regression analysis (PRA) and
using response-surface methodology (RSM) offers a power-
ful alternative to testing for matching form of fit hypotheses
(Edwards, 2007). We point to the increasing relevance of
this approach in the organizational behavior field, where
its use has been the basis for considerable theory progress.
Third, we address misconceptions regarding how to
best test and interpret moderation forms of fit. We chal-
lenge the view that the moderator (contingency) variable
and the independent variable (e.g., management control
systems) should not be conceptually related (and hence
correlated). In this regard, previous studies have discussed
the possibility that a path model may instead be the cor-
rect alternative (Duh et al., 2006; Gerdin, 2005a; Gerdin
and Greve, 2004; Hartmann and Moers, 2003; Shields and
Shields, 1998). Referring to such arguments, Gerdin and
Greve (2004) argue that if the moderator (contingency)
and the independent variables are conceptually related
and hence significantly correlated, a statistically signifi-
cant moderation effect should be ignored and a mediation
form of fit model should be tested instead. In contrast, we
argue that contingency theory implies that, in the long
run, the contingency variable is associated with the inde-
pendent variable (so-called selection forces) (Donaldson,
2001; Meilich, 2006). Therefore, these variables are likely
to be conceptually related and hence significantly corre-
lated. We outline that statistically significant moderation
effects should be interpreted and that any path model with
performance as the dependent variable is outside the scope
of contingency theory.
To corroborate our argument, we perform a Monte Carlo
simulation. We find that if the moderation form of fit is
ignored (because moderator and independent variable are
related) and a mediation form of fit model is tested instead,
7
true moderation forms of fit will be rejected in up to 100%
of cases, and false mediation form of fit models will erro-
neously be accepted.
In this context, we also address Hartmann and Moers
(2003, p. 808) legitimate conjecture that there is a risk
associated with accepting a spurious moderation effect
if the moderator and independent variable are strongly
related. However, when extending moderated regression
analysis (MRA) to PRA, thereby adequately controlling for
non-linear relationships, this risk is substantially reduced,
allowing the moderation form of fit to be interpreted
(MacCallum and Mar, 1995). We also address miscon-
ceptions in the management accounting field about how
to probe for interaction effects in order to differentiate
between sub-forms of moderation fit.
Fourth, we provide recommendations on the possi-
bilities of covariance-based structural equation modeling
(SEM) to test contingency-based hypotheses and address
issues related to avoiding the risk of using these techniques
inappropriately (Henri, 2007). Particularly in the context
of testing moderating effects and quadratic effects, these
advanced statistical techniques can decrease the number
of type I and type II errors if applied correctly. Conversely,
they increase the risk of such errors considerably if applied
inappropriately. We contribute to the contingency litera-
ture by making researchers in the management accounting
field aware of when to use (or not use) a specific SEM
approach to test for matching and moderation forms of fit.
The paper proceeds as follows: Section 2 provides an
overview of different forms of contingency fit. Section 3
re-evaluates the main methods of testing fit discussed by
Gerdin and Greve (2008) and extends the discussion to
residual analysis and PRA. Section 4 provides an overview
of SEM techniques and discusses when to use them to test
contingency hypotheses. Section 5 summarizes the find-
ings and concludes.
2. Forms of contingency fit and their implications
for performance
In the 1960s, contingency theory emerged as an impor-
tant organizational theory that views the organization as
an open system for which no general optimal structure
exists. Instead, external and internal context factors such
as size, the company’s technology or the competitive envi-
ronment determine the optimal design of an organizational
structure (see Schoonhoven, 1981, for an early discus-
sion). However, contingency theory is properly viewed
more as a meta-theory or a general idea than as a conven-
tional theory with a precise set of interrelated propositions
(Schoonhoven, 1981). Applying the idea that no general
optimal structure exists in the management accounting
context, researchers have argued that there is no optimal
design of a management control system (MCS)1 ; instead,
contingency factors such as size, technology, environmen-
tal uncertainty or strategy determine the specific optimal
1
We follow Chenhall’s (2003, p. 129) definition of MCS as encompassing
management accounting systems (e.g., cost accounting systems) but also
including controls such as personal or clan controls.
8 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29

Table 1
Forms of contingency fit: Abbreviated summary of various typologies.

Drazin and Van de Ven (1985) Gerdin and Greve (2004) Gerdin and Greve (2008) Donaldson (2001) and Klaas
and Donaldson (2009)

1. Selection fit 1. Congruence fit 1. Congruence fit 1. Managerial choice

2. Interaction fit No distinction
between matching and moderation
form of fit
2. Moderation form of fit
2.1 Multiplicative form of fit 2.1 Moderation form of fit
Distinction between general
interaction, symmetrical
interaction and cross-over
interaction

2.2 Matching form of fit 2.2 Matching form of fit:

Distinction between general
interaction, symmetrical
interaction and cross-over
interaction
– 2.3 Mediation form of fit
2.2.1 Traditional matching
form of fit with
iso-performance on the fit line
2.2.2 Neo-contingency’s
matching form of fit with
hetero-performance on the fit
line
2.2.3 Matching form of fit with
asymmetric effects of misfit on
performance
–

3. Systems fit – – 3. Multifit

– 4. Configuration fit: Distinction Not considered to be part of

between configuration contingency theory
contingency fit and
configuration congruence fit

design (Chenhall, 2003). It has been argued that, if MCS
deviates from this optimal design (for example, if too few or
too many performance indicators are provided), companies
suffer from lower performance because they are in a state of
‘mis-fit’. The MCS variable often captures different aspects
of MCS (for example, formal versus informal), higher versus
lower sophisticated versions of Balanced Scorecard-like
performance measurement systems (Lee and Yang, 2011)
or higher or lower measurement diversity (Ittner et al.,
2003).
2.1. Forms of contingency fit
Different forms of fit have been discussed in both the
general management literature (Drazin and Van de Ven,
1985; Schoonhoven, 1981; see also Southwood, 1978) and
the management accounting field (for example, Gerdin and
Greve, 2004, 2008). Chenhall and Chapman (2006) note
that the proliferation of different typologies of contingency
fit has led to some confusion. Table 1 contrasts the classi-
fications of Gerdin and Greve (2004, 2008) with those of
Donaldson (2001) and Drazin and Van de Ven (1985).
The first category, which is often referred to as the selec-
tion form of fit (Drazin and Van de Ven, 1985), implies that
companies have fully adapted their MCS to the demands of
their context (Hartmann, 2005).2 Because companies are
in equilibrium, no differences in performance are expected.
Several studies following this approach in the management
accounting field have investigated the context factors that
determine the design of MCS (for example, Chenhall and
Morris, 1986).
Contingency researchers argue, however, that reaching
this equilibrium takes time and many companies are in
states of misfit as long as they have not yet fully adapted
their MCS to the context factors. Companies move on a con-
tinuum between states of fit and misfit (Donaldson, 2001).
The fact that different companies are located in different
places on this continuum allows for the analysis of the per-
formance implications of MCS-contingency combinations.
The second category entails the effects of MCS-
contingency combinations on performance. Donaldson
(2001) distinguishes between two specific forms of fit
within this category: (1) matching form of fit and (2) mod-
eration form of fit.3 Gerdin and Greve (2004) distinguish a
third form: the mediation form of fit.
Drazin and Van de Ven (1985) and Donaldson (2001)
identify a third category of fit named respectively systems fit
and multi-fit. The level of fit is the result of the combination
of several MCS and contingency variables. The category of
systems fit assumes that fit is the result of various combina-
tions of MCS characteristics (for example, looking at control
systems as a package) and several contingency variables
(for example, size and environmental uncertainty). The
multiplicity of variables can be restricted to either one MCS
variable with several contingency variables or vice versa.
Both alternatives lead to an additive model. Donaldson
(2001, p. 200) indicates: “This additive model applies both
to combining multiple fits of an organizational structure

3
Drazin and Van de Ven (1985) did not distinguish matching and mod-
eration form of fit but label both interaction form of fit instead. Moderation
2
It has also been referred as managerial choice (Donaldson, 2001) or form of fit has also been labeled as multiplicative form of fit (Gerdin and
congruence fit (Gerdin & Greve, 2004). Greve, 2004).
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
Fig. 1. Traditional contingency theory’s matching form of fit with iso-performance on the fit-line. Panel A plots the relationship between the MCS and
performance for different values of the contingency variable in a three dimensional plot. The bold line on the top of the surface represents the line of perfect
fit, i.e. it is the line with all matches that produce the highest performance (Panel B adapted from Donaldson (2001, p. 211)).
variable to various contingencies and to combining mul-
tiple fits of various organizational structure variables to a
contingency.”4
Finally, Gerdin and Greve (2004) identify a fourth cat-
egory named configuration fit or gestalts (Venkatraman,
1989).5 This category drops the idea of fit as a continuum
where companies move gradually between states of fit and
misfit and assumes that only a few archetypes of specific
configurations exist. Within this framework, companies do
not make small movements on a continuum from states of
misfit to states of fit but rather “quantum jumps” from one
configuration to another. In the management accounting
context, this idea implies that only a few configurations of
specific control packages are feasible, each leading to the
same performance. The literature has opposing views on
whether this configuration fit is part of contingency theory,
and there are arguments for both views.6 We do not make a
judgment as to whether or not configuration fit belongs to
contingency theory, but it is important to be aware of the
theoretical differences between multi-fit and configuration
fit.
Most studies in management accounting consider fit
to be a continuum and focus on matching, moderation or
mediation forms of fit. Only a few studies apply the con-
figuration form of fit (Chenhall and Langfield-Smith, 1998).
4
Donaldson (2001; pp. 196–200) explains in more detail why his multi-
fit and Drazin and Van de Ven’s (1985) system fit refer to the same form
of fit, resulting in the above-described additive model.
5
Gerdin and Greve (2004) separate configuration contingency that is
comparable to multi-fit, but they emphasize what they call configuration
congruence, which is associated with the idea of “quantum jumps” or only
a few feasible configurations.
6
Donaldson (2001, pp. 141–152) provides a detailed argument as to
why configuration fit may not be part of contingency theory and highlights
the weak empirical support for its inclusion. Briefly, if several config-
urations lead to optimal performance, a choice exists between several
equally good configurations given the same level of contingencies. Having
the choice between several equally performing structures (MCS) given a
certain level of the contingency (rather than only one as implied by contin-
gency theory) is called equifinality (Donaldson, 2001, p. 143). According
to Donaldson (2001), the fact that only a few configurations are possi-
ble (“quantum jumps”) defeats the core idea of contingency theory of
organizations moving between states of fit and mis-ft.
9
Therefore, we restrict our discussion to the first three forms
of fit.
2.2. Matching, moderation, and mediation forms of fit
Drazin and Van de Ven (1985) propose two main
methods for testing interaction fit–namely, moderated
regression analysis (MRA) and the deviation score tech-
nique. While the former technique requires introducing
an interaction term in the regression equation, the lat-
ter implies calculating a deviation score from an optimal
match between structure and contingency. It was by some
understood that both techniques could be used to test inter-
action fit and that these two methods only represented
two different statistical techniques. Drazin and Van de Ven
(1985, p. 519) note: “The deviation-score approach and the
interaction approach are similar only to the extent that they
attempt to model the same underlying bivariate fit. Statisti-
cally, however they are quite different” (emphasis added).
This understanding has evolved over time, however; in par-
ticular, Hartmann and Moers (1999) and Donaldson (2001)
highlight that these two approaches not only differ statis-
tically but also represent two theoretically different forms
of fit: matching and moderation.
2.2.1. Matching form of fit
The classical form of fit is the matching form of fit
with iso-performance on the fit-line (Donaldson, 2001;
Schoonhoven, 1981). The relationship between the MCS
variable and performance is curvilinear (i.e., inverted U-
shaped) so that there is only one optimal MCS given a
certain level of the contingency variable (Fig. 1, Panel A).
The classical matching form of fit predicts several optimal
MCS-contingency combinations (matches), each producing
the same level of performance. This assumption of iso-
performance is reflected with the zero values on the fit-line
(Fig. 1, Panel B).7 Furthermore, any deviation from this fit
7
It is important to note that iso-performance and equifinality are
distinct concepts. Equifinality means that various configurations (con-
trol packages) lead to high performance. Thus, “(. . .) equifinality carries
10 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
Fig. 2. Neo-contingency’s matching form of fit with hetero-performance on the fit-line. Panel A plots the relationship between the MCS and performance
for different values of the contingency variable. Optimal matches along the fit-line imply increasing levels of performance (hetero-performance). Panel
B reports the level of performance for combinations of six values of the MCS variable and six values of the contingency variable (Panel B adapted from
Donaldson (2001, p. 283)).
line in either direction causes the same negative effects on
performance, so that both “under-fit” (too few MCS, e.g.,
too few performance indicators being available) and “over-
fit” (too much MCS, e.g., too many performance indicators
being provided) are equally harmful (Meilich, 2006, p. 167).
Referring to neo-contingency theory, Donaldson (2001)
explains a second sub-form of matching fit. Neo-
contingency theory addresses the question of how
companies transition to higher performance as they change
their levels of contingencies (for example by increasing
their size or moving from low technology to high technol-
ogy; Donaldson, 2001). It represents an extension of the
classical contingency theory’s matching form of fit, which
did not address the effect of changes in a company’s con-
tingency variables. For example, if all optimal matches
between MCS and contingency (such as size) produce
the same level of performance, why would small startup
companies that rely, in the first stage of the company’s life-
cycle, on a merely informal approach to control centered
around the company’s founder have an interest in invest-
ing resources in formalizing MCS (Davila and Foster, 2005,
2007; Greiner, 1998; Cardinal et al., 2004)? The concept of
matching forms of fit with hetero-performance on the fit-
line may offer an explanation, as it suggests that changes in
contingencies (such as size or technology) can move orga-
nizations to higher performance. However, the first effect
of a change in the level of a contingency variable is for
companies to move into misfit, requiring them to adapt
their MCS to the new level of the contingency variable and
achieve higher performance. This extension to contingency
theory replaces the assumption of iso-performance on the
fit-line with the assumption of hetero-performance (Fig. 2)
(Donaldson, 2001, p. 283, see also Donaldson, 2006).
connotations of choice, in that if there is more than one way to achieve
high performance, then an organization can choose between them”
Donaldson (2001, p. 194). Iso-performance, on the other hand, means that
the level of performance of different structure/contingency combinations
is the same.
However, it still implies a matching form of fit, as the
basic relationship between the MCS variable and perfor-
mance remains curvilinear, as illustrated by Fig. 2. The
aspect of hetero-performance on the fit line has nei-
ther received much attention in overview articles on
contingency theory nor been explicitly theorized upon
by researchers from the management accounting field.
Whether theory allows researchers to expect iso- or hetero-
performance largely depends on the specific performance
variables examined. In the example above, a performance
variable like return on sales may increase once the start-up
company reaches the second stage of its life cycle and starts
generating profits, but self-reported measures of perfor-
mance (for example, “how would you rate the performance
of your company compared to similar companies”) or sales
growth in percent may decrease or be unaffected by firm
growth.
A third sub-form of matching fit that has recently been
proposed predicts either iso- or hetero-performance on
the fit-line, but performance decreases more in the case of
“under-fit” compared to “over-fit” (Klaas and Donaldson,
2009). Hence, in contrast to the two forms of matching fit
discussed above, this form predicts an asymmetric effect
of misfit on performance. This sub-form of matching fit
captures whether “(. . .) under-fit is worse than over-fit.”
(Naman and Slevin, 1993, p. 146; Tushman and Nadler,
1978). The conceptual argument is that, “(. . .) [f]rom a pro-
cess viewpoint, under-fit might represent ‘critical’ factors
and over-fit merely ‘slack’ wastage.” (Naman and Slevin,
1993, p. 146). Fig. 3 depicts a conjectured example in the
management accounting context. Given a certain strategy,
having too few performance measures is detrimental to
performance, but having too many leads to only small per-
formance decreases. Thus, overinvestment in performance
measures is not as harmful as underinvestment, because
the additional cost for gathering and processing this extra
information does not severely affect performance. This rea-
soning results in a matching form of fit with asymmetric
effects of misfit on performance.
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
Fig. 3. Matching form of fit with hetero-performance on the fit-line but with asymmetric effects of misfit on performance.
2.2.2. Moderation form of fit
In sharp contrast to the matching form of fit, the mod-
eration form of fit implies that the relationship between
the MCS variable and performance is linear and varies at
different levels of the contingency variable (Panel A, C and
D, Fig. 4).8 Conceptually, this means that the contingency
determines the effect of the MCS variable on performance.
Some of the early contingency work had argued that the
MCS effect on performance is negative for one level of the
contingency and positive for another (see Schoonhoven’s
(1981, p. 353–354) discussion of Galbraith, 1973). Figure 4,
Panel D visualizes such a non-monotonic interaction where
the relationship between MCS and performance is posi-
tive for one value of the contingency variable but negative
for another, with the lines crossing each other within the
range of the data. In a recent study, Zatzick et al. (2012)
provide evidence for such a form of fit. They find that the
effect of total quality management (TQM) on performance
is positive at low levels but negative at high levels of dif-
ferentiation strategy.
Fig. 4 Panel C visualizes monotonic interaction effects
depicting a situation where the relationship between MCS
and performance is always positive but where the lines
cross over within the range of data.
8
In some cases, the terms interaction and moderation are used to refer
to a matching form of fit as well. Matching and moderation forms of fit
are different both from a theoretical perspective and an empirical design
perspective. Interaction form of fit was used to refer to both before theory
distinguished them (Drazin & Van de Ven, 1985). Because interaction is
often associated with an interaction term characteristic of moderation
form of fit, they are often used interchangeably.
11
Fig. 4 Panels A and B depict another sub-form
of moderation fit which entails a linear monotonic
interaction. The relationship between MCS and perfor-
mance is always positive, but the effect of the MCS
variable depends on the level of the contingency variable
(Donaldson, 2001, p. 191).9
It is relevant to highlight that there is no true ‘fit-line’
for moderation forms of fit, as the relationship between
MCS and performance is linear and not inverted U-shaped
(Donaldson, 2001; p. 190). For instance, finding support
for a monotonic moderation form fit (Fig. 4, Panel A)
suggests at first glance that best performance is always
achieved when the value of the MCS variable is highest.
One might therefore question whether moderation forms
of fit belong to contingency theory, because there is no opti-
mal fit between MCS and the contingency variables. This
is particularly true for monotonic moderation forms of fit
that indicate that more MCS (for example, many perfor-
mance indicators provided) is always better. We suggest
that the discussed moderation forms of fit still fall under
the umbrella of this theory for two reasons.
First, moderation forms of fit share with matching
forms of fit the broader idea that the effect of MCS on
performance depends on the context and that the context
9
Moderation may refer to the strength (mainly tested by subgroup
correlation analysis) or the form of the relationship between variables
(Venkatraman, 1989). It is as Hartmann and Moers (1999, p. 297) outline
that contingency hypotheses are more of form type because researchers
are mainly interested in the effect of the moderator variable on the slope
of the regression line. This form type of moderation effect is also the main
type of effect of interest to researchers in social sciences. We therefore
focus on the form type.
12 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29

Different types of moderation forms of fit

Panel A: Panel B
Monotonic interaction

High Chigh High 5 10 15 20 25

Cmiddle 4 8 12 16 20
Performance

MCS
Clow 3 6 9 12 15

2 4 6 8 10

Low Low 1 2 3 4 5

Low High Lo w High

MCS Contingency

Panel C Panel D
Monotonic and cross-over interaction Symmetrical and cross-over interaction

Chigh Chigh
High High
Performance
Performance

Clow

Clow
Low
Low

Low High Low High

MCS MCS

Fig. 4. Different types of moderation forms of fit. Panel A plots the relationship between the MCS variable and performance for three different values of the
contingency variable (Clow , Cmiddle and Chigh ). Panel B reports the level of performance for combinations of five values of the MCS variable and five values
of the contingency variable. Panel C plots monotonic interaction with cross-over between the MCS variable and performance for two different values of
the contingency variable (Clow and Chigh ) with cross-over interaction. Panel D plots a non-monotonic interaction with cross-over between the MCS variable
and performance for two different values of the contingency variable (Clow and Chigh ).

leads companies to adapt their MCS. It is worth noting
that most researchers seem to follow this argument, as the
vast majority of contingency studies in both the general
management field and management accounting have
tested moderation rather than matching forms of fit.
Second, from a conceptual point of view, there must be
a point at which performance suffers if the company over-
invests in MCS at the expense of other relevant activities.
Thus, finding a monotonic moderation form of fit instead
of a matching form of fit could be interpreted to mean that
companies are smart enough not to overinvest in MCS (for
example too much measurement diversity) so that no such
data points exist in the sample or even in reality.
Alternatively, such data range restrictions may be
caused when variables are categorical by nature (for
example a variable like use of Balanced Scorecards mea-
sured categorically; e.g., (1) no Balanced Scorecard, (2)
simple Balanced Scorecard, or (3) fully developed Bal-
anced Scorecard) or are bounded by the way they are
operationalized (for example, interdependence being
bounded at zero).
Moreover, having both independent variables (MCS
and contingency variable) measured dichotomously would
lead to two optimal matches and two states of mis-fit only.
In such a case, while it is possible to differentiate between
iso- and hetero-performance, matching or moderation can-
not be statistically separated because curvilinearity cannot
be tested.
Given such possible data range restrictions, it is there-
fore useful for researchers to discuss whether a moderation
form of fit might represent a truncated version of a match-
ing form of fit.
Theorizing on the specific sub-forms of moderation fit
– as called for by Schoonhoven (1981), Luft and Shields
(2003) and Gerdin and Greve (2008) – is therefore as rele-
vant as discussing the issue of data range restrictions (i.e.,
truncated forms of matching fit). Moreover, it is important
to use empirical methods that differentiate conclusively
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
Path model (so-called mediation form of fit) versus moderation form of fit
Path Model
Contingency MCS
Moderation form of fit
Contingency
MCS
Fig. 5. Path model (so-called mediation form of fit) versus moderation form of fit. The upper half depicts a path model, while the lower half depicts a
moderation form of fit model (Hartmann, 2005).
between the sub-forms of matching and moderation forms
of fit.10
2.2.3. Mediation form of fit
A final form of fit highlighted in Gerdin and Greve’s
(2004) classification is the mediation form of fit. In this form
of fit, the independent variable affects an endogenous vari-
able that then affects performance; hence, it is different
from a moderation form of fit, as illustrated in Fig. 5.
This form of fit has also been used in various empirical
studies (for example, Chong and Chong, 1997; Duh et al.,
2006; Gerdin, 2005a). The models tested in those studies
neither argued theoretically nor tested empirically for true
mediation. For true mediation, it is not sufficient to show
that effect a and effect b (see Fig. 6) are both significant
but that the mediation variable mediates the effect of the
independent variable on the dependent variable (if there
is full mediation, the effect size is effect a multiplied by
effect b, and this must be shown to be significant). Labeling
such path models as mediation is problematic as long as
there is no explicit test for mediation reported. Moreover,
these models do not represent contingency theory mod-
els, as its key concept of fit is missing (Hartmann, 2005).
Neither true mediation models nor these path models
(mediation form of fit models) belong to contingency the-
ory. True mediation models require certain theoretical as
10
Though we agree that it is important to be precise about the theory
predicting a specific sub-form of fit, we argue that Gerdin and Greve’s
(2008) classification adds unnecessary complexity. Gerdin and Greve
(2008) refer to “general interaction” for an interaction term being sig-
nificant without having specified (or predicted) the particular type of
interaction. We do not believe that theorizing general interaction is use-
ful; instead, the researcher should directly argue what specific sub-form
of fit is expected.
13
Performance
Performance
well as statistical conditions. First, on theoretical grounds,
the independent variable is a cause of the outcome vari-
able (Jacobucci, 2008). Therefore, the first step to test for
mediation in the classic four-step procedure (Baron and
Kenny, 1986) is to empirically show a direct link between
the independent variable and the final dependent variable
(Fig. 6). Next, the relationships between the independent
variable and the mediating variable (step 2) and between
the mediating variable and the dependent variable are
tested (step 3). Finally, the direct effect from step 1 needs
to become significantly smaller for partial mediation or to
disappear entirely for full mediation (step 4). Ideally, step 4
is explicitly tested with the Sobel (1982) test (1982).11 If the
relationships are significant, then the relationship between
the independent and the outcome variable is (partially)
mediated.12
However, true mediation models with performance
as the dependent variable are outside the scope of
contingency theory.13 Contingency variables affect the
relationship between the structure (i.e., MCS) and the
11
Recently, Zhao et al. (2010) discussed the weaknesses of the classical
Baron and Kenny approach, highlighting that a strict adherence to the four
steps can lead to increased type II and type I errors. They recommend using
an alternative “bootstrap” test (Preacher and Hayes, 2004) that directly
tests for the mediating relationship and that has been shown to be more
powerful (see Bisbe and Malagueno (2012) for an application).
12
Psychologists have traditionally tested true mediating relationships
(Jacobucci, 2008). Also, the studies in management accounting that imply
true mediation often rely on psychological theory and test whether
cognitive perception mediates the assumed positive effect of MCS on
managerial performance (for example, Burney et al., 2009; Burney and
Widener, 2007; Hall, 2008, 2011).
13
Mediation models with MCS as the dependent variable can however
be considered to belong to contingency theory as this simply represents
a more advanced form of selection fit.
14 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
Step1: Test of direct effect
Independent variable
Mediation
variable
Independent variable
Step4
Fig. 6. Steps proposed by Baron and Kenny (1986) to test for mediation.
outcome variable (i.e., performance), and do not have a
direct effect on performance.
It should be noted that Gerdin and Greve (2008) no
longer seem to promote the mediation form of fit. They
note (p. 1004): “Accordingly, path analyses cannot explore
the existence of moderated causal relationship(s) predicted
by contingency theory (Donaldson, 2001; Schoonhoven,
1981), i.e. where the relationship between MAS and per-
formance is moderated by context.” In Section 3.2.2, we
show that path models can incorrectly disguise existing
moderation forms of fit.
2.3. Related theories
2.3.1. The concept of complementarities in organizations
The concept of complementarity, which originated in
economics, postulates that the effect of one activity might
be stronger in the presence of a complementary activ-
ity (Milgrom and Roberts, 1990, 1995). In recent years,
researchers from the broader management field have bor-
rowed the concept to study whether the effect of one
management practice or system depends on the imple-
mentation of an additional practice (see Ennen and Richter,
2010, for an overview). The complementarity view is simi-
lar to contingency theory insofar as the effect of practices or
systems in organizations is argued to depend on other prac-
tices or systems. Given the overlaps in variables of interest
(for example, strategy) and the emphasis on the contextu-
ality in which complementarity relationships arise, Ennen
and Richter (2010) see contingency studies as part of the
group of studies that analyze complementarity relation-
ships. Given the recent interest in how control systems
work as a package (Malmi and Brown, 2008), more studies
of this kind can be expected in the future.
2.3.2. Person-organization fit theory
Another related theory is the person-organization (P-
O) fit theory, which originated from the organizational
behavior field (Ostroff and Judge, 2007). P-O-fit studies ana-
lyze bi-variate effects between a person’s characteristics
and his/her organization or environment (Kristof-Brown
et al., 2005; Ostroff and Schulte, 2007) on a person-related
outcome variable, such as job commitment. This theoretical
Dependentvariable
Dependentvariable
perspective builds on the idea that job performance or
commitment is enhanced when a person’s characteristics
(e.g., personal values) match the elements of the organiza-
tion (e.g., organizational culture or job demands).14 Given
the methodological advancements that have been made in
this field, this stream of literature has recently made con-
siderable theoretical progress paying attention to specific
sub-forms of matching fit, as described in Section 2.2.1.
3. Testing bi-variate forms of fit with performance
as the dependent variable
3.1. Contingency theory’s matching form of fit
The common approach in management accounting
research to test matching form of fit models is to com-
pare a predicted optimal level of the MCS variable with
the actual level, using either the deviation score tech-
nique or residual analysis (for example, Gerdin, 2005b;
Said et al., 2003). Two approaches to the deviation score
are often used: in the first, the fit-line is known on
theoretical grounds; in the second, the fit-line is not
known but estimated empirically (Chenhall and Chapman,
2006). Residual analysis, on the other hand, relies on
predicting the optimal MCS level given the level of the
contingency variables; deviations from this predicted value
are used to test for negative effects of misfit on per-
formance. The following two sub-sections discuss the
traditional methods to test matching forms of fit. Section
3.1.3 presents polynomial regression analysis as a powerful
alternative.
3.1.1. The deviation score approach
The first version of the deviation score techniques
requires using a “fit term” composed of the difference
between the MCS variable (X) and the contingency vari-
able (for example, environmental uncertainty; Z) and
14
Besides the P-O fit, many different forms of fit exist, such as person-
job (P-J) fit, person-group (P-G) fit, or person-person (P-P) fit (Ostroff and
Schulte, 2007). Such forms of fit have also been labeled “need-supply” fit
or “demand-ability” fit (Edwards and Shipp, 2007).
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29 15

Panel A Panel B P an e l C

High 5 10 15 20 25 High 5 10 15 20 25 High 5 10 15 20 25

4 8 12 16 20 4 8 12 16 20 4 8 12 16 20
MCS

MCS
MCS
3 6 9 12 15 3 6 9 12 15 3 6 9 12 15

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

Low 1 2 3 4 5 Low 1 2 3 4 5 Low 1 2 3 4 5

Low High Low High Low High

Contingency Contingency Contingency

Fig. 7. Euclidean distance approach when the underlying data rather reflect a moderation form of fit. Illustration of how the three step procedure of the
deviation score technique can lead to erroneously accepting a matching form of fit when the moderation form of fit is correct: (1) Identifying the best
performing companies (Panel A). (2) Deriving empirically the fit-line based on the relationship between MCS and the contingency of the best performing
companies (Panel B). (3) Calculating the Euclidean distance of the worse performing companies from the fit-line (Panel C).

regressing performance (Y) on this term as shown in equa-
tion (1) (Gerdin and Greve, 2008; Hartmann and Moers,
1999; Venkatraman, 1989).15
 
Y = ˛ + ˇ1 X + ˇ2 Z + ˇ3 X − Z  + ε
This approach requires theoretically knowing the fit-
line (i.e., knowing that ‘fit’ is achieved whenever X = Z). If
ˇ3 is significant, then the matching form of fit is supported.
The second version relies on an Euclidean distance met-
ric (Drazin and Van de Ven, 1985) and is used when the
fit line is unknown (see Gerdin, 2005b, for an application).
Each observation is located in the space defined by MCS and
the contingency variables. The Euclidean distance from this
location to the fit line (where performance is optimal given
the contingency; see fit line in Fig. 2, Panel B) is expected
to be negatively associated with performance. The statis-
tical test is a bi-variate correlation between distance and
performance. As the fit line is not known theoretically, it
is estimated empirically in a subsample of the best per-
formers by regressing the MCS variable on the contingency
variables (for example, environmental uncertainty). This
procedure to determine the fit line empirically builds on the
assumption that better-performing companies are closer to
the fit line. The regression coefficients from the sample of
the high-performing companies are then used to predict
the fit line.
Though the deviation score analysis is intuitively
appealing, it has been criticized for suffering from low reli-
ability (Donaldson, 2001; Edwards, 2007). Specifically, the
reliability of the deviation score is lower than the reliabil-
ity of the MCS and the contingency variables that constitute
the deviation score, assuming these are correlated (which
is the case in most research designs). Lower reliability is
strongly associated with higher levels of type II errors and
hence is likely to result in a lack of statistically significant
support for a true matching form of fit. This limitation can
be seen as a conservative test. However, these tests can
easily lead to systematic type I errors and the incorrect
15
Relevant control variables should be included in the regression to
minimize the risk of omitted variable bias.
acceptance of a matching form of fit when, for example,
a moderation form of fit is correct instead. To illustrate this
issue, Fig. 4, Panel B shows a monotonic and ordinal mod-
eration form of fit. Let us assume that the true form of fit
(1) is a (monotonic and ordinal) moderation form of fit but
that a matching form of fit is tested using the deviation
score technique. The first step identifies the best perform-
ers (Fig. 7, Panel A). These best performers primarily come
from the upper right-hand side (performance values of, for
example, 25 and 20). In the second step, the MCS variable
is regressed on the contingency variable for this group of
best performers, and this regression is used to derive the
fit line (Fig. 7, Panel B).
A (monotonic and ordinal) moderation form of fit pre-
dicts that an increase in MCS (such as more measurement
diversity) is associated with an increase in performance. In
contrast, the matching form of fit predicts decreases in per-
formance not only when MCS is too low but also when it is
too high, given a certain level of the contingency.
If most observations are positioned below the fit-line
(Fig. 7, Panel C) (which is very likely assuming a monotonic
moderation form of fit), then the deviation score method
would erroneously find significant support for a matching
form of fit.
These issues associated with testing for a matching form
of fit lead to type I errors and illustrate how the devia-
tion score technique may systematically provide spurious
support for the wrong form of fit.
The problem with difference score methods is that
they impose untested assumptions (fit is achieved when-
ever X = Z (first version of the deviation score) or
iso-performance (second version of the deviation score
technique)). This oversimplification comes at the risk of
severely confounding main forms of fit; as Cafri et al. (2010)
explain, “(. . .) the use of difference scores (. . .) necessarily
limits the extent to which theory can develop and evolve.”
3.1.2. Residual analysis
Residual analysis, another common technique to
test the matching form of fit, implicitly assumes iso-
performance on the fit-line. Econometrically elegant, it
regresses MCS on one or several contingency factors using
16 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
Incorrectly predicted fit-line
High
Area A
Area B Correct Fit-line
MCS
Area C
Low
Low High
Contingency
Fig. 8. Illustration of potential type 2 error associated with residual anal-
ysis. Any positioning in area A and B implies that the value of the MCS
variable is too high and any positioning in area C that the value of the
MCS variable is too low.
the full sample. Next, the error term for each observation
is interpreted as the distance of each observation from the
fit line (level of mis-fit)16 and is assumed to be negatively
associated with performance.
Though this method represents a much better approach
than the deviation score method, researchers using this
method cannot know whether the predicted optimal level
of the MCS variable represents a point on, below, or above
the fit-line,17 which can lead to erroneous conclusions. For
example one might find that over-fit, but not under-fit, is
negatively associated with performance.
A hypothetical explanation for such a finding when
using the residual analysis could be as follows (Fig. 8).
If the optimal level of a MCS variable were incorrectly
predicted as a point on the bold line, companies positioned
in area B would not under-invest in MCS as predicted by
the (incorrect) bold fit line but rather would over-invest.
Therefore, the residuals in areas B and C might cancel each
other out, so that no significant result can be obtained for
the “under-investment” hypothesis.
Moreover, if using this method does not lead to signif-
icant findings, it remains unclear whether there is a true
fit-line at all. Therefore, unless this method is used very
carefully, it carries the risk of incorrectly accepting a match-
ing form of fit when another form of fit is correct instead.
We identified some studies from the management
accounting field that use this approach in a very careful
way as a supplementary analysis (Ittner et al., 2002; Ittner
16
Residual analysis has also been described by Duncan and Moore
(1989).
17
The second version of the deviation score approach and residual anal-
ysis are quite similar. The key difference is that residual analysis regresses
MCS on one or several contingency factors using the full sample, while the
second version of the deviation score approach makes this regression only
in a sub-sample of best performers. Hence, the difference between the
methods is that the fit-line is estimated differently. It is probably advan-
tageous to use the full sample, rather than a sub-sample, to derive the
fit-line.
et al., 2003; Van der Stede et al., 2006; Said et al., 2003).
These studies can serve as examples of how to use this
approach appropriately in order to avoid incorrect con-
clusions, as they investigate whether both under-fit (too
few MCS) and over-fit (too many MCS) are, independently
from each other, associated with decreased performance.
In other words, they do not compute one overall mis-fit
score but rather test the underlying theoretical assump-
tions rigorously.
Most of these studies reject a matching form of fit.
Ittner et al. (2003) and Van der Stede et al. (2006) studied
whether optimal matches between the extent of perfor-
mance measurement and company strategy lead to higher
performance compared to a mis-match. Interestingly, both
found that, as predicted, under-fit (fewer MCS than pre-
dicted) is associated with lower performance but that
over-fit (more MCS than predicted) is positively associ-
ated with better performance. Consequently, both studies
rejected a matching form of fit.
Ittner et al. (2002) studied whether an optimal match
between contingency factors (plant characteristics) and
investments in activity-based costing (ABC) is associated
with performance. Given that the MCS variable (ABC) is
measured dichotomously, their innovative application of
residual analysis is the only way to test for matching in
this specific case. They test carefully whether more under-
or over-fit is more significantly negatively associated with
performance than less under- or over-fit. The analysis
provides weak support for the matching hypothesis; specif-
ically, results showed that investing too much in ABC is
associated with decreased performance (significant at the
10% level), but not investing in ABC (hence under-fit, i.e. too
few MCS) is not significantly related to performance. There-
fore, though theory is convincingly presented, the analysis
does not clearly support the assumed matching form of fit.
Said et al. (2003) is the only study finding clear sup-
port for a matching form of fit; they found that both
under-fit and over-fit, in terms of non-financial per-
formance measurement, are associated with decreased
performance.
Meilich (2006) further demonstrates that, while resid-
ual analysis is a conservative approach when used carefully,
it nonetheless has difficulties in detecting existing match-
ing form of fit relationships. He creates 16 artificial datasets,
all perfectly reflecting a matching form of fit, and tests the
effectiveness of residual analysis in detecting this form of
fit accurately. Overall, he reports poor performance of this
approach but good performance of polynomial regression
analysis (Meilich, 2006).
3.1.3. Polynomial regression analysis as an alternative to
test matching form of fit hypotheses
Edwards and colleagues have proposed the use of poly-
nomial regression analysis (PRA) together with response
surface methodology (RSM) to test non-linear person-
environment fit relationships (Edwards, 2007; Edwards,
1995; Edwards, 1994; Edwards and Cooper, 1990; Edwards
and Parry, 1993).18 Originally used in engineering and
18
This approach dates back to the seminal paper of Box and Wilson
(1951).
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
natural sciences (Myers et al., 2009), this approach is now
also used in social and behavioral sciences (e.g., Cafri et al.,
2010; Cohen et al., 2010; Harris et al., 2008; Lee and
Antonakis, 2012).
This method can also test contingency theory predic-
tions effectively; it provides more accurate insights into
the sub-forms of fit because it tests directly what tradi-
tional methods test only indirectly. PRA does not impose
assumptions, such as the assumption of iso-performance
on the fit-line or the assumption that fit is achieved when
X equals exactly Z, which makes it more robust against type
1 and type II errors than any version of the deviation score
method.
This approach uses two steps. First, using Equation (2)
below, it tests the data that include the MCS variable (X),
the contingency variable (Z), their product term (XZ) and
the quadratic effects of both X2 and Z2 .
Y = ˛ + ˇ1 X + ˇ2 Z + ˇ3 XZ + ˇ4 X 2 + ˇ5 Z 2
 cases of polynomial regression analysis, PRA can support
+ ˇi control variables (2)
P-O-fit studies use commensurate variables—variables
that measure the same content from different perspec-
tives (for example, subordinates’ responsibilities from the
perspectives of both the subordinate and the supervisor).
Because these variables use the same numeric scale, the
results are easier to interpret (Edwards, 2002; Shanock
et al., 2010). For example, if job satisfaction is maximized
whenever X = Z, this can be interpreted to indicate that it is
optimal to have an exact balance between the subordinate’s
and the supervisor’s perspectives on the subordinate’s
responsibilities.
Variables in contingency studies measure different con-
tent and do not need to be expressed on the same scale.
Therefore, X = Z does not have any particular interpretation
in most cases.
However, it can be easier to interpret the results
when the scales have been scale-centered (instead of
mean-centered) prior to the analysis (e.g., for a 7-point
Likert scale, 4 represents the center of the scale).19
If both quadratic terms (ˇ4 and ˇ5 ) are not sig-
nificant, but the interaction term (ˇ3 ) is significant,
any matching form of fit hypotheses can be rejected,
and a moderation form of fit would be supported. If
the quadratic terms are significantly negative but the
interaction term is significantly positive, the results indi-
cate a matching form of fit, and the response-surface
methodology will precisely describe the form of the
relationship.20 This methodology comprises a graphical
visualization and further econometric analyses (Edwards,
19
Polynomial regression analysis relies on the same assumption as
regression analysis, such as error-free measurement of variables. As with
moderated regression analysis, PRA yields better results when the vari-
ables that constitute the interaction and quadratic terms have higher
reliability. Also, an adequate sample size is critical for detecting inter-
action and quadratic effects (Marsh et al., 2004).
20
For contingency style studies, finding ˇ4 to be significantly negative is
particularly important, as this indicates that too much MCS can be harmful
from a certain point on.
17
http://public.kenan-flagler.unc.edu/faculty/edwardsj, and
Shanock et al., 2010, provide a comprehensive descrip-
tion of the method as well as the necessary Excel tools
and SPSS commands to conduct the econometric analy-
ses). The graphical analysis can be performed by plotting
the coefficients from Equation 2 for a number of specific
values for X and Z in a three-dimensional graph. The graph
indicates (a) whether there is a fit line at all so that both
under-fit and over-fit are harmful, (b) whether there is iso-
or hetero-performance along the fit-line and (c) where the
fit-line is located. Depending on the coefficients, it may
well be that the ‘fit-line’ is not linear but rather has some
curvature.21
PRA simultaneously tests for all bi-variate effects of two
variables on an outcome variable, including all sub-forms
of bi-variate fit relationships. Thus, finding support for one
sub-form of fit conclusively eliminates alternative expla-
nations (i.e., other forms of fit). Because traditional ways
of testing for matching fit as well as MRA represent special
a certain form of fit while simultaneously rejecting alter-
native forms (Edwards, 1994). Therefore, using PRA rather
than these traditional methods offers the opportunity to
more rigorously test (or explore) which model best fits the
data.
The application of this method does not depend on
knowing exactly where the fit-line lies ex ante but instead
allows it to be detected ex post.22 Statistical analysis can
complement graphical inspection to better characterize the
surface (Edwards and Parry, 1993; Edwards, 2002, 2007).
A useful extension provides additional technical details for
studying the surface (Cohen et al., 2010).
In order to assess whether results are robust, robustness
checks can probe for additional higher-order terms to test
other, more sophisticated interaction models (Edwards,
1994). For example, a second product term interacting with
the quadratic term of the MCS variable by the contingency
ZX2 could be added to the product term (XZ). If ˇ3 , ˇ4 and
ˇ5 in Equation (3) (below) are significant, the form of fit
represented curvilinear by linear interaction (Cohen et al.,
2003, pp. 292–293):
Y = ˛ + ˇ1 X + ˇ2 Z + ˇ3 XZ + ˇ4 ZX 2 + ˇ5 X 2 + ε (3)
Depending on the sign and magnitude of the obtained
coefficients (Equation (3)), the relationship between MCS
and performance may be curvilinear at low levels of the
contingency but linear at high levels of the contingency
(Fig. 9).
We are not aware of an example from the field of man-
agement accounting research of a relationship described by
such a more model. However, future studies might include
a test of this equation to provide more evidence for their
suggested form of fit.
21
Edwards and Parry (1993) and Edwards (2001) show many different
surfaces that can result from testing the coefficients of Equation 2.
22
It is advisable to plot and interpret the surface only within the relevant
region of data (e.g., where 90% or 95% of actual data points lie) to avoid
extrapolating into regions with no/few data points (Edwards, 2002, pp.
388–389).
18 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
Chigh
High
Performance
Cmiddle
Clow
Low
Low High
MCS
Fig. 9. Curvilinear by linear interaction (adapted from Cohen et al. (2003),
p. 292).
In this section, we give only an overview highlight-
ing the strength of PRA for precisely testing the different
sub-forms of fit and link this method to contingency the-
ory. While we provide numerical examples and additional
technical details in Appendix A, we do not repeat all the
technical details that are well-described in the literature
cited above.
Researchers in organizational behavior who have used
polynomial regression analysis have not found support for
the traditional matching form of fit with iso-performance
on the fit line (Kristof-Brown and Jansen, 2007). This is not
surprising, given the many restrictions on the coefficients
required to obtain the shape depicted in Fig. 1, Panel A.
The polynomial regression approach is often used in
the field of organizational behavior (see Edwards, 2001,
for an overview) and is preferred to the deviation score.
It simultaneously tests competing forms of fit and elim-
inates alternative explanations. Analyzing exactly where
the fit line is located and studying performance for optimal
matches on the fit-line also offers a way to test whether
hetero-performance exists on the fit-line.23 This finding
supports the use of this more powerful approach in man-
agement accounting to provide more accurate insights and
thereby to develop and refine contingency-based theory.
However, when using PRA in an exploratory rather than
a confirmatory manner, it is important to follow guide-
lines for exploratory research in order to avoid “post hoc
interpretations of complex surfaces that are not both gen-
eralizable and conceptually meaningful” (Edwards, 1994,
p. 74).24
23
However, it is not easy to study whether mis-fit has asymmetric effects
on performance.
24
Edwards (2001) presents a detailed discussion about the myths sur-
rounding the deviation score and the polynomial regression approach.
These myths suggest that the deviation score approach is a conserva-
tive test and that there are procedures to mitigate the fundamental
problems associated with it. Other myths, which suggest that the polyno-
mial regression approach represents a data-driven procedure and suffers
from multicollinearity, are likewise invalidated.
3.2. Contingency theory’s moderation form of fit
3.2.1. Main methods to test for moderation form of fit
A moderation form of fit implies that a linear relation-
ship between two variables is moderated by a contingency
variable (Fig. 4). Hence, the contingency variable alters the
form of the relationship. The literature identifies two main
ways to test moderation forms of fit: (1) moderated regres-
sion analysis (MRA) and (2) sub-group regression analysis.
MRA tests moderation, including the product-term of
the structure variable (X) and the contingency variable (Z)
in the regression, together with its constituting variables.
Formally, MRA tests the following equation:
Y = ˛ + ˇ1 X + ˇ2 Z + ˇ3 XZ + ε (4)
If ˇ3 is significant, moderation will be supported.
Sub-group regression analysis splits the sample into
several sub-samples based on the contingency variable and
separately regresses performance on the MCS variable for
each sub-group. It has been argued that both methods
are appropriate to test for moderation. However, MRA is
preferred if the moderator variable is continuous because
forming sub-groups leads to a loss of information. Such
information losses are associated with increased type I and
type II errors (Irwin and McClelland, 2001).
When using MRA, it is useful to expand Equation (4)
and use Equation (2), hence including the quadratic terms
(Klein et al., 2009). The reason for this is that increas-
ing correlation between the structure variable (X) and the
contingency variable (Z) and decreasing reliability of meas-
ures increase the risk of finding a spurious moderation
effect when a curvilinear relationship between structure
variable and performance is correct instead (Lubinski and
Humphreys, 1990; MacCallum and Mar, 1995).
A moderation form of fit requires ˇ3 to be significant
while ˇ4 and ˇ5 are non-significant; otherwise, a matching
form of fit could be correct (see previous section), particu-
larly if ˇ4 is significant and negative. If ˇ4 is significant and
ˇ3 is not, then there is solely a curvilinear effect of the MCS
variable on performance without any interaction with the
contingency variable.
3.2.2. Criticism of MRA: The “no relation” argument
Most prior criticisms of the use of MRA have been
well-stated (Hartmann and Moers, 1999, 2003; Shields
and Shields, 1998); however, Chenhall (2003, pp. 155–156)
suggests that contingency-based management accounting
studies might have ignored moderation models too often
in favor of path models: “Given recent criticism directed
towards moderating variable models there is a danger that
researchers will try and force arguments about interaction
effects into intervening variable models.” Following this
reasoning, we will address some arguments put forward
in the literature that might have caused confusion.
Hartmann and Moers (2003, pp. 807-808) correctly
point out that a moderation effect can be spurious,
confounding a simple curvilinear relationship of the MCS
variable in the presence of strong correlations between
the moderator and the potential contingency variable.
Hartmann and Moers (2003) suggest that, in such cases,
“(. . .) a positive and significant ˇ3 in (. . .) Eq. [4] may cause
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
MRA to mistakenly present a path model as a moderated
relationship.”
However, this potential source of confusion can easily
be addressed by controlling for the quadratic term using
equation (2) as explained above; if the interaction term
remains significant, there is no reason to discard the inter-
action effect. There is a certain tradition to propose that the
moderator (contingency) variable should not be conceptu-
ally related and hence correlated with the MCS variable
(Duh et al., 2006; Gerdin, 2005a; Gerdin and Greve, 2004;
Hartmann and Moers, 1999, 2003; Shields and Shields,
1998, pp. 52–53).25 If they are correlated, Gerdin and Greve
(2004, p. 310) explicitly propose excluding moderation
form of fit in favor of mediation form of fit.26 From a the-
oretical perspective, however, the moderator variable can
be related and hence statistically correlated with the MCS
variable. Outside the management accounting commu-
nity, a correlation between the contingency (moderator)
variable and the structure (MCS, independent) variable
is not believed to either validate or reject a moderation
model. In fact, if contingency theory holds, causal rela-
tions and hence significant correlations are likely to be
present in cross-sectional study designs. This is referred to
as selection forces, which represent long-term adaptation
processes that lead to significant correlations (Donaldson,
2001; Meilich, 2006). In the presence of selection forces,
both the matching or moderation form of fit can be cor-
rect as long as a sufficient number of companies are still
in a state of ‘mis-fit’ (Luft and Shields, 2003). Also, more
generally, it is common to assume and handle levels of sig-
nificant correlations between moderator and independent
variables (Ganzach, 1997; Klein et al., 2009; Marsh et al.,
2004; Moosbrugger et al., 2009).
Failing to test for a correct moderation form of fit model
can likely lead to the acceptance of a theoretically false
mediation form of fit model. For example, if an ordinal
monotonic moderation is present, as illustrated in Fig. 4,
Panel A, MCS will have a positive direct effect on per-
formance, and the magnitude (but not the sign) of the
relationship will differ at different levels of the moderator
25
In this context, the definition of Shields and Shields (1998, p. 51): “A
moderator variable is defined as having nonsignificant, bivariate relation-
ships with both the independent and dependent variables.” might have
caused some confusion. We agree that a variable labeled as a modera-
tor should not have a direct effect on the dependent variable. If it does,
then the literature makes a distinction between a “moderating” and a
“joint” effect. Though the statistical test remains the same, the concep-
tual difference is that the former implies a non-significant relationship
between the moderator variable and the dependent variable, while the
latter assumes that both independent variables forming the interaction
term have individual effects on the dependent variable. Luft and Shields
(2003) label this independent variable “interaction,” and the marketing
literature has referred to such cases as “quasi-moderations” (Sharma et al.,
1981). However, claiming that a moderator variable should not be related
to the independent variable with which it forms the interaction term does
not represent a commonly accepted assumption.
26
Gerdin and Greve (2008) do not repeat this “no relation” argument,
but they do mention that mediation form of fit models with performance
as the dependent variable “do not test for the existence of interaction
effects between context and MAS on performance.” (p. 1004). While this
sentence is obviously true, it is not clear whether it also implies that the
“no relation” argument is invalid.
19
variable. Omitting the moderation effect from the equation
will result in finding a positive direct effect between MCS
and performance, thereby supporting a false path model,
because the direct effect is a weighted average of the direct
effects of MCS on performance at the different levels of the
moderator variable.
To illustrate this issue, we performed a number of Monte
Carlo simulations (reported in detail in Appendix B). We do
not find much evidence of a significant moderation effect
when a path model is correct instead. These type I errors
remain at commonly accepted levels close to 5%. However,
testing a path model if there is a significant moderation
effect leads to type I error rates in up to 100% of the cases.
The literature includes several examples of the use of
the “no-relation” argument to discard moderation in favor
of the mediation form of fit. Gerdin (2005a) and Duh et al.
(2006) explicitly use this criterion to discard the moder-
ation form of fit.27 In their path models, the contingency
variable influences the MCS variables, which in turn have
an effect on performance. Duh et al. (2006) report that they
also find 4 statistically significant moderation effects (out
of 5 tested) in their sample, with the interaction term being
significant. However, referring to the arguments made in
the literature cited above, they discard these moderation
effects, arguing that they were ‘spurious,’ and conclude
that a path model (without having tested for mediation)
is correct instead. Similarly, Gerdin (2005a) argues that the
moderator variable and the independent variable should
not be correlated and therefore proposes a mediation form
of fit model; however, this is a path model only, as a test
for mediation is missing.
Hartmann (2005) outlines in a reply to Gerdin’s (2005a)
article that any path (or mediation) model does not capture
states of misfit between the contingency and MCS variable
on performance. Replying to this comment, Gerdin (2005b)
uses the deviation score method to test for a matching form
of fit, which seems to produce evidence for this form of fit.
However, it is not reported in the paper whether not only
‘under-fit’ but also ‘over-fit’ (i.e., too much MCS) is neg-
atively associated with performance; only an aggregated
misfit score is negatively associated with performance. As
explained in Section 3.1.1, such a procedure can easily pro-
duce spurious support for the wrong form of fit (matching
with iso-performance). Moreover, Gerdin (2005a, p. 318)
reports the result of subgroup regression analysis, with the
results rather indicating a moderation form of fit.
3.2.3. Usefulness of MRA and the Johnson-Neyman
technique for probing the nature of the interaction
In order to assess whether there is a significant
interaction effect, it is sufficient to look at the output
MRA provides. However, prior literature has correctly
pointed out that the coefficient of the interaction term
alone is not sufficient to explain whether the relation-
ship between MCS and performance is monotonic or
non-monotonic, nor whether it is ordinal or disordinal
(Hartmann and Moers, 1999; Gerdin and Greve, 2008).
27
The models tested in these papers do not use any procedure to test
for mediation.
20 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
Panel A
C7
High
C6
Performance
C5
C4
C3
C2
C1
Low
Low High
MCS
Slope significantly different from zero
Slope not significantly different from zero
Fig. 10. Graphical illustration how to probe for interaction effects using the Johnson-Neyman technique.
However, recommendations on how to best probe for these
additional characteristics are not specific. Dichotomizing
continuous variables to probe for non-monotonicity of an
interaction (see Hartmann and Moers, 1999, section 4.6) is
no longer a state-of-the art method (Irwin and McClelland,
2001; Preacher et al., 2005), nor is complementing MRA
with any “unprecise” method to detect a possible disordi-
nal interaction (Gerdin and Greve, 2008, pp. 1006–1007).
However, Gerdin and Greve (2008, p. 1007) tentatively sug-
gest: “An alternative is to perform additional analyses of the
MRA regression output to examine whether the interaction
effect is non-monotonic (Schoonhoven, 1981, pp. 376–377)
and disordinal (Cohen et al., 2003, p. 288).”
The uncertainty regarding how to probe for these addi-
tional characteristics may have resulted partly from a
critical assessment of MRA and the Johnson-Neyman tech-
nique (Hartmann and Moers, 1999, section 4.5).
While Hartmann and Moers (1999) make several good
arguments, their overall conclusion unnecessarily down-
plays this technique (p. 306 and 307): “The use of the
Johnson-Neyman technique in MRA is therefore ques-
tionable, since it plays no role in the test of hypotheses
of the interaction format.” This misperception may be
attributable to the lack of comprehensive explanations
of how to best probe for those additional character-
istics of an interaction that are now available (Bauer
and Curran, 2005; Hayes and Matthes, 2009; Preacher
et al., 2006). This rigorous probing includes using outputs
from MRA, the Johnson-Neyman technique, and graphical
techniques.
Fig. 10 illustrates the Johnson-Neyman technique (see
the above-cited literature for a more technical explana-
tion). Fig. 10, Panel A plots a conjectured relationship
between MCS and performance for seven different val-
ues of a contingency variable C (e.g., size, strategy, or
Panel B
C7
High
C6
Point III Point V
Performance
Point I Point VII
Point II Point VIII
Point IV Point VI
C2
C1
Low
Low High
MCS
1 2 3 4 5 6 7
Region A Region B
Region C Region D
environmental uncertainty). As described in Hayes and
Matthes (2009), researchers can test separately for every
single value of the contingency variable whether the slope
is significantly different from zero (referred to as a pick-
a-point approach). Alternatively, the Johnson-Neyman [JN]
technique can be used to identify the regions of significance
of the values of the contingency variable for which the slope
is significantly different from zero. In this example, the JN
technique, together with MRA and graphing of the rela-
tionships, would inform the researcher that the slopes are
significantly positive in the upper range of values of the
contingency (values C6, and C7), non-significant between
C5 and C3, and significantly negative for C1 and C2 (Bauer
and Curran, 2005; Hayes and Matthes, 2009; Preacher et al.,
2006). There is an openly accessible R-based online tool
that facilitates the application of this advanced method of
probing for interactions, which includes a graphical visu-
alization (http://www.quantpsy.org/interact/index.html).
Practicing researchers simply use the output from MRA,
and the tool calculates the regions of significance, finally
offering the opportunity to plot the relationship and allow-
ing the researcher to precisely conclude whether the
interaction is (non-)monotonic and/or (dis-)ordinal.
The JN technique can also be used to test at which val-
ues of MCS the difference in performance is significantly
different from zero at the 5% level, given specific values of
the contingency. Fig. 10, Panel A indicates that performance
significantly differs when MCS is at C1 and C2 as well as C6
and C7. Fig. 10, Panel B depicts four pairs of points where
performance differences are significant. For instance, if
Points III and IV (where the value of MCS is 3) lead to sig-
nificant differences in performance, then this significance
holds for MCS in the (1, 3) range. Similarly, if Points V and VI
are significantly different at exactly 5%, then the difference
in performance is significant in the (5, 7) range. When the
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
contingency variable takes values C6 or C2, the region of
significance in performance is smaller. If the significance is
exactly 5% (Points I and II on one side and Points VII and VIII
on the other side), then the ranges of significance are (1, 2)
and (6, 7). Therefore, if MCS is equal to 3, the difference in
performance is significant when the contingency variable
is C7 or C1 but not when the contingency variable is C6 or
C2. Around the point where the lines intersect, there is no
significant difference in performance. These extra analyses
are relevant from a contingency perspective as they allow
the researcher to conclude, for example, whether, given a
certain value of the MCS variable, performance is signifi-
cantly higher at high levels of the contingency variable than
at low levels of the contingency variable.
4. Structural equation modeling (SEM) to test
contingency hypotheses
4.1. Testing moderation form of fit in covariance-based
SEM
Covariance-based structural equation modeling
(SEM)28 techniques belong to the advanced statistical
approaches in social sciences. In the past few years, an
increasing number of publications in the field of manage-
ment accounting have used these techniques (Henri, 2007;
Smith and Langfield-Smith, 2004). Covariance-based SEM
can test more complex models with several dependent
and/or mediating variables at the same time (Smith and
Langfield-Smith, 2004). It can also analyze moderating
effects within a more complex model structure (Jaccard
et al., 1990, p. 64). A second central advantage is that
measurement error can be accounted for, leading to
more precise parameter estimations (Baumgartner and
Homburg, 1996).
SEM’s better handling of measurement error is partic-
ularly advantageous for testing moderation effects using a
product term approach, because the reliability of the prod-
uct term is generally much lower than the reliability of the
constituting variables in MRA (Busemeyer and Jones, 1983;
Jaccard and Wan, 1995). The lower the reliabilities, the
more difficult it is to detect moderating effects, resulting
in higher rates of type II errors (Jaccard and Turrisi, 2003).
Like traditional regression-based techniques,
covariance-based SEM offers two possibilities for analyz-
ing moderating effects: multi-sample analysis and the use
of a product-term.
4.1.1. The multi-sample analysis in SEM for testing
moderating effects
Management accounting researchers often use multi-
group analysis in covariance-based SEM to analyze
moderating effects (Burkert et al., 2011; Burney and
Widener, 2007; Henri, 2006; Marginson and Bui, 2009;
28
The most popular software packages that allow applied researchers
to use covariance-based SEM for their analyses are AMOS (Arbuckle,
1994), LISREL (Jöreskog and Sörbom, 2001), EQS (Bentler, 1995) and MPlus
(Muthén and Muthén, 2007). Besides covariance-based SEM, the variance-
based partial least squares (PLS) approach is often used, particularly when
sample size is rather small (Burkert and Lueg, 2013).
21
Viator, 2001). If the moderator variable is not a dummy,
groups are formed based on a potential moderator vari-
able, similar to regression sub-group analysis. This variable
can be nominal (for example, gender, nationality), interval
(Likert-scaled) or cardinal. In the latter two cases, a median
split is typically used to form two groups.
For covariance-based SEM, Jaccard and Wan (1996)
explain how a nested goodness-of-fit strategy is applied.
To substantiate a significant moderating effect, the rela-
tionships theorized to be moderated by a third variable are
estimated in both sub-groups separately. If the model fit
significantly improves, a significant moderation effect is
supported.
To test whether model fit has significantly improved,
the Chi-square test should be used and reported. Further-
more, in an attempt not to lose statistical power, Marsh
et al. (2006) recommend using multi-group analysis only
if the moderator variable is not Likert-scaled or cardinal so
that a small number of distinct groups can be formed (e.g.,
male versus female) (see also Rigdon et al., 1998). More
powerful alternatives are now available to test for modera-
tion in covariance-based SEM when the moderator variable
is Likert-scaled or cardinal. The next subsection provides an
overview of these alternatives.
4.1.2. SEM approaches for testing latent moderation
effects
If the moderator variable is a cardinal or Likert-scaled
variable, covariance-based SEM offers various alternatives
for analyzing moderating effects. Covariance-based SEM
has been developed to analyze linear relationships
between reflective constructs that are multivariate and
normally distributed. If these assumptions are violated,
there are now effective ways to deal with non-normally
distributed and non-continuous scaled data in covariance-
based SEM (Finney and DiStefano, 2006). However,
analyzing moderation effects using a product term affects
the assumption of multi-normally distributed indicators
in a unique way. Even if the indicators of the indepen-
dent and the moderator variable are normally distributed
(which is seldom the case), their product term can lead to a
high degree of non-normality (Algina and Moulder, 2001;
Kenny and Judd, 1984), making it hard to estimate latent
interaction effects in covariance-based SEM.
Because the moderation term is estimated by multiply-
ing the indicators for the moderator and the independent
variables, each of these indicators represents an indicator
for the moderating term. One option is to use all possible
combinations of indicators. If the moderator variable and
the independent variable are measured with four indica-
tors each, this results in 16 indicators for the moderation
term. Because each single indicator is used four times,
there is redundant information to estimate the moderation
term. Several techniques have been proposed to deal with
these issues (Marsh et al., 2006 and Moosbrugger et al.,
2009); however, as they are software-intensive and their
efficiency has been questioned, they have been used only
rarely in empirical research. Recently, Marsh et al. (2004)
analyzed the performance of a simpler approach, referred
to as the unconstrained approach to test for latent mod-
eration effects. They propose using each indicator once
22 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
to form an indicator product. Given four indicators each
for the independent and moderator variables, this would
lead to four indicators for the moderation term, which are
then used to estimate the moderating effect. The overall
good performance of this approach in Monte Carlo simu-
lations compared to more complex options (Marsh et al.,
2004) makes this approach a suitable choice for empirical
research.
A new approach referred to as latent moderated struc-
tural (LMS) equations (Klein and Moosbrugger, 2000;
Moosbrugger et al., 1997) does not rely on product indica-
tors or on the normal distribution of the moderating term
but still requires that the component variables (moder-
ator and independent variable) be normally distributed.
Available Monte Carlo simulation supports the perfor-
mance of this approach (Klein and Moosbrugger, 2000;
Schermelleh-Engel et al., 1998). When the moderator
and independent variables are correlated (which is often
the case in contingency-based management accounting
research), LMS has been shown to be superior to the uncon-
strained approach (Marsh et al., 2004) in terms of type I and
type II error rates (Moosbrugger et al., 2009).29 LMS is now
part of new versions of the Mplus software package.30
4.2. Testing matching form of fit hypotheses and SEM
Whereas SEM now offers effective ways to test moder-
ation form of fit models, testing for matching form of fit
is still a challenge. Recently Cheung (2009a,b) proposed
testing matching form of fit models using a latent con-
gruence model (LCM), which uses the mean of the latent
variables and their difference. This model is conceptually
similar to Equation (1); the difference is that measurement
errors of the variables used to build the difference score
can be more adequately addressed. The LCM approach thus
differs from the polynomial regression analysis in that it
does not rely on interaction and quadratic terms. Like the
deviation score method discussed in Section 3, LCM is not
able to test either for rival hypotheses (matching versus
moderation form of fit) or for iso- or hetero-performance
(Edwards, 2009). Therefore, Edwards’s (2009) critique of
LCM is relevant (p. 35): “From a measurement perspective,
the LCM is a step forward in congruence research, given
that measurement error can bias coefficient estimates in
polynomial regression analysis and lead to erroneous con-
clusions. However, from a substantive perspective, the LCM
is a step backward.”
Ideally, Equation (2) would be tested using SEM. How-
ever, an important assumption underlying the covariance-
based SEM approach is that relationships are linear
(Edwards, 2009). Therefore, matching form of fit cannot
easily be tested in SEM because its main feature is the
29
Klein et al. (2009) find that including the quadratic term in covariance-
based SEM when testing for moderation might be an effective way to
minimize spurious moderating effects as well.
30
Another promising approach which might be suited to test for sev-
eral moderation effects simultaneously is the Quasi-Maximum Likelihood
(QML) method (Klein and Muthén, 2007). This method is not currently part
of any standard software package, but it may become available as a new
option or stand-alone software in the future (Marsh et al., 2006).
curvilinear relationship between MCS and performance
that depends on the contingency (Klein et al., 2009).
Edwards (2009) indicates that there are ongoing efforts
to develop quadratic structural equations, which represent
a potentially appropriate approach to testing Equation (2)
in covariance-based SEM (Edwards and Kim, 2002). These
efforts might result in such an approach in the future.
5. Summary and implications for future research
This paper addressed important theoretical differences
of contingency fit and the methods to test them in man-
agement accounting.
We first discussed different forms of fit, arguing that
overview papers in management accounting have men-
tioned sub-types of fit that do not belong to contingency
theory (mediation form of fit), while specific sub-forms of
matching fit that appear in the broader management liter-
ature (matching forms of fit with iso-performance versus
hetero-performance on the fit-line and/or with asymmetric
effects of mis-fit on performance) have not been explicitly
addressed. We also discussed the methods used in man-
agement accounting research to test ‘fit’ and assessed their
ability to effectively test the specific sub-form of fit they
postulate as well as to discriminate between rival forms
of fit. We focused on the methods used to test matching
and moderation forms of fit that entail effects of (mis-) fit
on performance. Our review and analyses lead to several
recommendations for future empirical contingency work.
First, polynomial regression analysis represents a pow-
erful approach to testing matching versus moderation form
of fit models that represent competing theoretical explana-
tions. Polynomial regression analysis simultaneously tests
for any bivariate moderation and matching form of fit and
thereby avoids the weaknesses related to the traditional
methods to test fit outlined in Section 3. Second, we rec-
ommend accepting moderation effects after ruling out the
possibility of a matching form of fit and a curvilinear effect
only of the MCS variable. As explained earlier, any path
model with performance as a dependent variable (referred
to as mediation form of fit) is not a suitable alternative.
This recommendation is based on theoretical and empir-
ical arguments that question the “no relation” condition
between the moderator and the independent variables. The
central insight of our Monte Carlo simulations is that simple
path models easily conceal existing moderation effects.
Moreover, we presented an overview of the computa-
tional procedures now available for probing interactions,
including the application of the Johnson-Neyman tech-
nique (the suitability of which had previously been
questioned in the management accounting literature).
These procedures allow researchers to conclusively test
sub-forms of moderation fit.
Third, we provided an overview and recommenda-
tions on how to use covariance-based SEM when testing
contingency theory. In particular, we recommend not
using covariance-based SEM to test matching form of
fit hypotheses, but instead using polynomial regression
analysis. Due to the linearity constraints of covariance-
based SEM, this technique is not yet advanced enough to
provide reliable estimates similar to those obtainable from
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
polynomial regression analysis. Future methodological
work is required before SEM can handle the non-linearity
implied by the polynomial equation.
In contrast to matching form of fit models, covariance-
based SEM is now sufficiently well-developed to test
moderation form of fit models. Finally, the fact that
covariance-based SEM has been developed to analyze lin-
ear and not non-linear effects is reflected in the results of
the Monte Carlo simulation, as the fit indices in covariance-
based SEM do not warn against an un-modeled moderator
(or quadratic) effect. To prevent incorrect theoretical con-
clusions, studies that investigate complex path models
should verify as part of their robustness checks that mod-
erator and/or quadratic effects have not been overlooked.
The conclusions from this work imply two main
avenues for fruitful future research grounded on con-
tingency theory. First, future contingency-style studies
could attempt to develop models based on the concept
of hetero-performance of different “matches” between the
contingency and MCS variables. It would also be interest-
ing to investigate potential asymmetric effects of mis-fit on
performance. Such work may not necessarily be quantita-
tive in nature, but could also be qualitative, as theory may
not yet be developed well enough for more sophisticated
models.31
Second, carefully discussing when moderation forms
of fit might represent truncated versions of a match-
ing form of fit might contribute to a more fine-graded
view of contingency relationships. However, any empirical
contingency-related work should pay particular attention
to the limits of studying ‘fit,’ as any research design can
only consider a limited number of variables at a time.
Also, future contingency studies may critically reflect on
whether the underlying assumption of contingency the-
ory (that companies are always seeking to regain ‘fit’)
always holds true. If, for example, companies are found
to adopt practices just to appear legitimate (DiMaggio and
Powell, 1983), researchers may find quite a few companies
remaining in misfit without attempting to regain fit.
Acknowledgments
The paper has benefited from comments made by Raul
Barroso, Franz Michael Fischer, Rainer Lueg, Martin Klar-
mann, Karl Schuhmacher and participants of the research
seminars of IESE Barcelona and the University of Innsbruck.
We moreover thank the editors Michael Bromwich and
Robert Scapens as well as the five reviewers for their excel-
lent comments and suggestions.
Appendix A. Technical example of Polynomial
Regression Analysis and Response Surface
Methodology
In the first example, we assume that both the MCS and
the contingency variable were measured on a 7-point Likert
scale and that fit was theorized to be achieved whenever
31
See Rossing (2013) for an excellent example of qualitative
contingency-based research.
23
X = Z. In the second example, the exact location of the fit
line is not clear ex ante; instead, we derive the fit-line from
regression output. The second example might be more rel-
evant for contingency researchers, as in most cases the
meaning of optimal fit is not clear and variables may be
measured on quite different scales.
Example 1:32 If fit is achieved whenever X = Z33 , iso-
performance on the fit-line can be directly assessed. For
matching form of fit, ˇ3 needs to be significantly positive,
while minimally ˇ4 must be significantly negative. In order
to assess iso-performance on the fit-line, we replace Z with
X in Equation (2) to get Equation (A1):
Y = ˛ + (ˇ1 + ˇ2 )X + (ˇ3 + ˇ4 + ˇ5 )X 2 + ε (A1)
For iso-performance on the fit line, Y must be a constant
for all values of X, thus (ˇ1 + ˇ2 ) and (ˇ3 + ˇ4 + ˇ5 ) must be
equal to 0 (Cohen et al., 2010; Edwards and Parry, 1993, p.
1593).
If (ˇ1 + ˇ2 ) is positive and (ˇ3 + ˇ4 + ˇ5 ) is 0, the fit-line
is linearly increasing. Thus, higher values of Z = X are associ-
ated with higher performance, reflecting a matching form
of fit with hetero-performance.
Example 2. Optimal matches between MCS and the con-
tingency variable may not occur when X = Z. Alternatively,
they may occur in lines parallel to the one where X = Z or
the fit-line is rotated. PRA and RSM determine whether and
where a fit-line exists. For instance, assume a conjectured
fit relationship between strategy (Z) and measurement
diversity (X) and optimal matches lead to a higher sales
volume:34
Sales volume = 5.628 + 0.314X − 0.118Z + 0.299XZ
−0.219X 2 − 0.102Z 2 + ε (A2)
The positive interaction term of Equation (A2) and the
negative quadratic terms indicate a matching form of fit.
Inspecting the lines where X = Z and where X = −Z leads to
the two two-dimensional graphs in Fig. A.1). The line for the
values when X = Z (Panel A) indicates an upward orientation
of the surface along this line, indicating a matching form
of fit with hetero-performance. The inverted U shape for
X = −Z suggests a matching form of fit where performance
drops the more X and Z differ (Panel B).
However, the three dimensional surface for this exam-
ple (Fig. A.2) indicates that the fit-line shifts to the right of
the X = Z line and is slightly rotated counter-clockwise.
In this case, as any matching form of fit implies that
there is only one performance-maximizing value of the
MCS variable given a certain level of the contingency, the
fit-line can be derived from the derivative of the Y variable
to the X (MCS) variable (Meilich, 2006). Formally expressed,
32
This example is taken from Edwards and Parry (1993).
33
This corresponds to the first version of the deviation score methods
where the X–Z term is included (see Equation (1)).
34
We use the slightly adapted numerical results of the 4th polynomial
regression from Table 11.3 (p. 273) used by Edwards (2002) to create
Figure 11.8d (p. 375).
24 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29

Fig. A.1. Panel A Panel B, Line of congruence Line of incongruence. Panel A plots the line of perfect congruence where X equals Z. In this example performance
is increasing along this line indicating an upward orientation of the surface. Panel B plots the line of incongruence which is inverted U-shaped.

Fig. A.2. Illustration of the response surface methodology.

for a given Z we are looking for the value of X that satisfies Further transformation produces:
the following equation:
ˇ1 ˇ3
X=− − ·Z
2 · ˇ4 2 · ˇ4
∂Y !
= ˇ1 + ˇ3 · Z + 2 · ˇ4 · X = 0
∂X Inserting the values −3, −2, −1, 0, 1, 2 and 3 in Z (con-
tingency variable) in the above equation, we obtain the
following optimally “matching values” for the MCS variable
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29 25

(X): −1.22, −0.65, 0.03, 0.72, 1.4, 2.08, 2.76. Inserting the Table B.1
Design of first Monte Carlo simulation: “True” mediation model, no
obtained values for the pairwise matches (X/Z) produces
moderation.
the following performance levels on the fit-line for these
seven matches: Model characteristics Specifications

Composite reliability of 0.65 (medium) 0.8 (high)

• First match (−1.33/−3): 5.45 variables
• Second match (−0.65/−2): 5.55 constituting the
• interaction term
Third match (0.03/−1): 5.64
• Fourth match (0.72/0): 5.74 Composite reliability of 0.9
• Fifth match (1.4/1): 5.84 dependent variable
• Sixth match (2.08/2): 5.93 Number of indicators 4
• Seventh match (2.76/3): 6.03 per construct

Sample size 100/250/500

These results for performance levels of the matches on
the fit-line further corroborate the conclusion obtained
from inspecting Fig. A.1 and the X = Z line of having a match-
ing form of fit with hetero-performance.
Appendix B. Monte-Carlo Simulation on potential
confusion of moderation form of fit and mediation
form of fit models
We performed a Monte Carlo simulation to assess the
risk that, with increasing correlations between the mod-
erator variable and the independent variable, an existing
moderation form of fit will be confused with an incorrect
path model, and vice versa. Monte Carlo simulations offer a
powerful way to test the quality of proposed methods and
decision rules (Kennedy, 2003).
B.1. First simulation: Correct path model and wrong
moderation form of fit
First, we define the form and the strength of the relation-
ships between the variables in the underlying model. This
underlying model contains a direct effect of the indepen-
dent (potential contingency) variable on the MCS variable
and a direct effect of the MCS variable on the performance
variable. Based on this specified underlying model, random
samples are generated. Our first Monte Carlo simulation
examines how often a moderation model is spuriously
accepted when the underlying model is a path model (type
I error).
We systematically vary the strength of the effect of
the contingency variable on the MCS variable. We also
vary the sample size, the composite reliability of the con-
structs used, and the statistical technique to test the extent
of spurious interaction terms. We examine two meth-
ods commonly used in contingency-based management
accounting research: moderated regression analysis (MRA)
and covariance-based structural equation modeling (SEM).
Table B.1 describes the parameters of the first simulation.
This leads to a 4 × 3 × 2 × 2 research design with 48 cells.
We create in total 12,000 random samples that we analyzed
with MRA and SEM to test each specification (cell). These
datasets are analyzed first using MRA and second using SEM
in Mplus 4.0. We use both approaches to ensure that our
findings are not attributable to the choice of method. This
is relevant in the context of moderating effects, as there
is an ongoing discussion regarding how to test modera-
tion effects effectively, particularly in SEM (Jaccard and
Effect of independent 0.2/0.4/0.6/0.8
variable on
mediating variable
Effect of the mediating 0.4
variable on the
dependent variable
Effect of the interaction 0.0
term on dependent
variable
Methods used to test Moderated regression analysis
for interaction effect & SEM in Mplus 4.0
Wan, 1996). For the SEM analysis, we use the unconstrained
approach, which requires using the indicators of the mod-
erating variable and the independent variable to create
indicator product-terms that are then used as items for the
interaction term (Section 5 presents alternative approaches
to testing moderating effects in SEM). This procedure has
been shown to represent a powerful way to test moder-
ation effects in SEM when the variables are Likert-scaled
(Marsh et al., 2006).
Table B.2 provides information on the convergence rates
(i.e. how often the method used converges to a solution).
The convergence rate is always 100% for the MRA analysis.
If composite reliability is low and sample size is small,
covariance-based SEM does not always converge. This
result underscores simulation-based evidence that shows
covariance-based SEM requires samples with more than
100 observations as well as adequate levels of construct
reliability (Reinartz et al., 2009).
Table B.2, Panel B provides information on type I error
rate (i.e. the percentage of cases in which a significant mod-
eration effect was found when a path model is actually
correct instead).
Increasing sample size and construct reliability slightly
increases type I error rates, because moderation effects are
more likely to be significant if sample size is large and
construct reliability is high. We also find that SEM overall
leads to lower error rates than MRA.
Overall, type I error rates remain at commonly accepted
levels close to 5%; thus, the risk of mistakenly accept-
ing a moderation model instead of a correct path model
(the mediation form of fit) is small. In particular, increas-
ing correlations between the independent (the “spurious”
moderator in this example) and structure (MCS) vari-
ables does not lead to high levels of spurious moderation
effects.
26 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29

Table B.2
True path model and wrong moderation model.

Panel A: Convergence rate

N = 100 N = 250 N = 500

CR: 0.65 CR: 0.80 CR: 0.65 CR: 0.80 CR: 0.65 CR: 0.80

Relationship between contingency (independent) and structure (mediation) variable: 0.2

SEM 75.6% 98.2% 92.0% 100% 99.2% 100%
MRA 100% 100% 100% 100% 100% 100%

Relationship between contingency (independent) and structure (mediation) variable: 0.4

SEM 81.4% 99.6% 96.4% 100% 99.8% 100%
MRA 100% 100% 100% 100% 100% 100%

Relationship between contingency (independent) and structure (mediation) variable: 0.6

SEM 93.2% 99.8% 99.8% 100% 100% 100%
MRA 100% 100% 100% 100% 100% 100%

Relationship between contingency (independent) and structure (mediation) variable: 0.8

SEM 97.6% 99.8% 100% 100% 100% 100%
MRA 100% 100% 100% 100% 100% 100%

Panel B: Percentage of spurious moderation effects found

N = 100 N = 250 N = 500

CR: 0.65 CR: 0.80 CR: 0.65 CR: 0.80 CR: 0.65 CR: 0.80

Relationship between contingency (independent) and structure (mediation) variable: 0.2

SEM 0.0% 2.2% 0.9% 4.4% 1.6% 5.2%
MRA 5.0% 4.6% 3.4% 4.8% 4.8% 5.2%

Relationship between contingency (independent) and structure (mediation) variable: 0.4

SEM 0.2% 3.2% 1.2% 5.0% 3.4% 3.8%
MRA 6.8% 6.6% 3.0% 7.4% 4.6% 5.4%

Relationship between contingency (independent) and structure (mediation) variable: 0.6

SEM 0.0% 4.0% 2.4% 3.6% 3.4% 5.2%
MRA 6.8% 5.8% 6.8% 4.4% 7.4% 5.4%

Relationship between contingency (independent) and structure (mediation) variable: 0.8

SEM 0.2% 3.6% 2.0% 4.2% 3.0% 6.0%
MRA 6.0% 6.0% 5.2% 4.4% 5.6% 6.2%

CR: Composite Reliability, SEM: Structural Equation Model; MRA: Moderated Regression Analysis.

B.2. Second simulation: Correct moderation form of fit
and wrong mediation form of fit
The second Monte Carlo simulation addresses the oppo-
site case: how often a path model is erroneously supported
when a moderation form of fit relationship is correct
instead. An incorrect path model is accepted if both
the relationship between the contingency factor and the
structure variable and the relationship between the struc-
probability that a path model (mediation form of fit model)
is erroneously accepted. Another important finding is that
the fit indices typically used to evaluate covariance-based
SEM models–2 /df, RMSEA (root mean square error of
Table B.3
Design of second Monte Carlo simulation: “True” (statistically significant)
moderation, no mediation.
Model characteristics Specifications

ture variable and performance are significant. Composite reliability (CR) of 0.65 (medium) 0.8 (high)
Again we systematically vary sample size, composite independent and intervening (true
moderator) variable
reliabilities, and the strength of the selection forces (cor-
relation coefficients between the constructs forming the Composite reliability (CR) of 0.9
product term). Table B.3 provides a summary of the param- dependent variable

eters of the second simulation study. Number of indicators per construct 4

This design leads to a 4 × 3 × 2 research design with 24 Sample size 100/250/500
cells. We create 500 random samples for each set of spec-
Effect of independent variable on 0.3
ifications (cell), for a total of 12,000 datasets. We analyze
dependent variable
these datasets with covariance-based SEM in Mplus 4.0.
The models always converge (Table B.4, Panel A). With Effect of the interaction term on 0.3
dependent variable (moderating
increasing levels of sample size, composite reliability and effect)
strength of selection forces, the rate of type 1 errors dras-
Correlation between independent 0.0/0.2/0.4/0.6/0.8
tically increases (Table B.4, Panel B). Given medium-size
variable and moderator variable
samples of 250 and medium levels of correlations of 0.4,
the type I error rates are above 80%. Hence, there is a high Methods used to test the path model SEM in Mplus
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29 27

Table B.4
True moderation model and wrong path model.

Panel A: Convergence rate

N = 100 N = 250 N = 500

CR: 0.65 CR: 0.80 CR: 0.65 CR: 0.80 CR: 0.65 CR: 0.80

Correlation between contingency (moderator) and structure (independent) variable: 0.2

SEM 100% 100% 100% 100% 100% 100%

Correlation between contingency (moderator) and structure (independent) variable: 0.4

SEM 100% 100% 100% 100% 100% 100%

Correlation between contingency (moderator) and structure (independent) variable: 0.6

SEM 100% 100% 100% 100% 100% 100%

Correlation between contingency (moderator) and structure (independent) variable: 0.8

SEM 100% 100% 100% 100% 100% 100%

Panel B: Percentage of spurious mediation form of fit models

N = 100 N = 250 N = 500

CR: 0.65 CR: 0.80 CR: 0.65 CR: 0.80 CR: 0.65 CR: 0.80

Correlation between contingency (moderator) and structure (independent) variable: 0.2

SEM 4.8% 21.7% 45.5% 70.1% 77.4% 93.6%

Correlation between contingency (moderator) and structure (independent) variable: 0.4

SEM 22.4% 55.3% 89.6% 95.2% 99.6% 100%

Correlation between contingency (moderator) and structure (independent) variable: 0.6

SEM 37.3% 59.9% 88.2% 92.6% 99.0% 100%

Correlation between contingency (moderator) and structure (independent) variable: 0.8

SEM 48.6% 54.9% 88.2% 91.6% 99.4% 100%

approximation), TLI (Tucker Lewis index), CFI (compar-
ative fit index), SRMR (standardized root mean square
residual–do not indicate that the model is mis-specified.
If an important linear effect between two variables is
missing, covariance-based SEM indicates that these fit
indices would improve if the path were specified. As
covariance-based SEM has been developed for the analysis
of linear effects, moderation (and quadratic) effects (see
Section 4 for more details) remain undetected. It is there-
fore important as part of a robustness check to exclude
potential moderation effects by simply testing for them.
References
Algina, J., Moulder, B.C., 2001. A note on estimating the Jöreskog-Yang
model for latent interactions using LISREL 8.3. Structural Equation
Modeling 8, 40–52.
Arbuckle, J., 1994. AMOS: Analysis of moment structures. Psychometrika
59, 135–137.
Bauer, D.J., Curran, P.J., 2005. Probing interactions in fixed and multilevel
regression: Inferential and graphical techniques. Multivariate Behav-
ioral Research 40, 373–400.
Baron, R.M., Kenny, D.A., 1986. The moderator-mediator variable dis-
tinction in social psychological research: Conceptual, strategic and
statistical considerations. Journal of Personality and Social Psychology
51, 1173–1182.
Baumgartner, H., Homburg, C., 1996. Applications of structural equation
modeling in marketing and consumer research: A review. Interna-
tional Journal of Research in Marketing 13, 139–161.
Bentler, P.M., 1995. EQS Structural Equation Program Manual. Multivariate
Software, Encino, CA.
Bisbe, O., Malagueno, R., 2012. Using strategic performance measurement
systems for strategy formulation: Does it work in dynamic environ-
ments? Management Accounting Research 23, 296–311.
Box, G.E.P., Wilson, K.B., 1951. On the experimental attainment of opti-
mum conditions. Journal of the Royal Statistical Society, Series B 13,
1–45.
Burkert, M., Fischer, F., Schäffer, U., 2011. Application of the controllability
principle and managerial performance: The role of role perceptions.
Management Accounting Research 22, 143–159.
Burkert, M., Lueg, R., 2013. Differences in the sophistication of value-based
management–the role of top executives. Management Accounting
Research 24, 3–22.
Burney, L., Widener, S.K., 2007. Strategic performance measurement sys-
tems, job-relevant information, and managerial behavioral responses
– role stress and performance. Behavioral Research in Accounting 19,
43–69.
Burney, L.L., Henle, C.A., Widener, S.K., 2009. A path model examining
the relationships among strategic performance measurement system
characteristics, organizational justice, and extra- and in-role perfor-
mance. Accounting, Organizations and Society 34, 305–321.
Busemeyer, J.R., Jones, L.E., 1983. Analysis of multiplicative combination
rules when the causal variables are measured with error. Psychologi-
cal Bulletin 93, 549–562.
Cafri, G., van den Berg, P., Brannick, M.T., 2010. What have the difference
scores not been telling us? A critique of the use of self-ideal discrep-
ancy in the assessment of body image and evaluation of an alternative
data-analytic framework. Assessment 17, 361–376.
Cardinal, L.B., Sitkin, S.B., Long, C.P., 2004. Balancing and rebalancing in the
creation and evolution of organizational control. Organization Science
15, 411–431.
Chenhall, R.H., 2003. Management control systems design within its orga-
nizational context: Findings from contingency-based research and
directions for the future. Accounting, Organizations and Society 28,
127–168.
Chenhall, R.H., Chapman, C.S., 2006. Theorizing and testing fit in contin-
gency research on management control systems. In: Hoque, Z. (Ed.),
Methodological Issues in Accounting Research. Spiramus, London, pp.
35–54.
Chenhall, R.H., Langfield-Smith, K., 1998. The relationship between strate-
gic priorities, management techniques and management accounting:
an empirical investigation using a systems approach. Accounting,
Organizations and Society 23, 243–264.
Chenhall, R.H., Morris, D., 1986. The impact of structure, environment,
and interdependence on the perceived usefulness of management
accounting systems. The Accounting Review 61, 16–35.
Cheung, G.W., 2009a. Introducing the latent congruence model for
improving the assessment of similarity, agreement, and fit in
organizational research. Organizational Research Methods 12,
6–33.
28 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
Cheung, G.W., 2009b. A multiple-perspective approach to data analysis in
congruence research. Organizational Research Methods 12, 63–68.
Chong, V.K., Chong, K.M., 1997. Strategic choices, environmental uncer-
tainty and SBU performance: A note on the intervening role of
management accounting systems. Accounting and Business Research
27, 268–276.
Cohen, J., Cohen, P., West, S.G., Aiken, L.S., 2003. Applied Multiple Regres-
sion/Correlation Analysis for the Behavioral Sciences, 3rd ed. Erlbaum,
Hillsdale.
Cohen, A., Nahum-Shani, I., Doveh, E., 2010. Further insight and addi-
tional inference methods for polynomial regression applied to the
analysis of congruence. Multivariate Behavioral Research 45, 828–
852.
Davila, A., Foster, G., 2005. Management accounting systems adoption
decisions: Evidence and performance implications from early-
stage/startup companies. The Accounting Review 80, 1039–1068.
Davila, A., Foster, G., 2007. Management control systems in early-stage
startup companies. The Accounting Review 82, 907–937.
DiMaggio, P.J., Powell, W., 1983. The iron cage revisited institutional iso-
morphism and collective rationality in organizational fields. American
Sociological Review 48, 147–160.
Donaldson, L., 2001. The Contingency Theory of Organizations. Sage,
Thousands Oaks, CA.
Donaldson, L., 2006. The contingency theory of organizational design:
Challenges and opportunities. In: Burton, R.M., Eriksen, B., Hakonsson,
D.D., Snow, C.C. (Eds.), Organization Design: The Evolving State-of-
the-Art. Springer, pp. 19–40.
Drazin, R., Van de Ven, A.H., 1985. Alternative forms of fit in contingency
theory. Administrative Science Quarterly 30, 514–539.
Duh, R.-R., Chen, H., Chow, C.W., 2006. Is environmental uncertainty
an antecedent or moderating variable in the design of budgeting
systems? An exploratory study. International Journal of Accounting,
Auditing and Performance Evaluation 3, 341–361.
Duncan, K., Moore, K., 1989. Residual analysis: A better methodology for
contingency studies in management accounting. Journal of Manage-
ment Accounting Research 1, 89–104.
Edwards, J.R., 1994. The study of congruence in organizational behavior
research: Critique and a proposed alternative. Organizational Behav-
ior and Human Decision Processes 58, 51–100.
Edwards, J.R., 1995. Alternative to difference scores as dependant variables
in the study of congruence in organizational research. Organizational
Behavior and Human Decision Processes 64, 307–324.
Edwards, J.R., 2001. Ten difference score myths. Organizational Research
Methods 4, 265–287.
Edwards, J.R., 2002. Alternatives to difference scores: Polynomial regres-
sion analysis and response surface methodology. In: Drasgow, F.,
Schmitt, N.W. (Eds.), Advances in Measurement and Data Analysis.
Jossey-Bass, San Francisco, pp. 350–400.
Edwards, J.R., 2007. Polynomial regression and response surface method-
ology. In: Ostroff, C., Judge, T.A. (Eds.), Perspectives on Organizational
Fit. Taylor & Francis Group, LLC, New York, NY, pp. 361–372.
Edwards, J.R., 2009. Latent variable modeling in congruence research.
Organizational Research Methods 12, 34–62.
Edwards, J.R., Cooper, C.L., 1990. The person-environment fit approach to
stress: Recurring problems and some suggested solutions. Journal of
Organizational Behavior 10, 293–307.
Edwards, J.R., Kim, T.Y., 2002. Moderation in structural equation mod-
eling: Specification, estimation, and interpretation using quadratic
structural equations. In: Paper Presented at the Annual Meeting of
the Society for Industrial and Organizational Psychology, Toronto,
Ontario, Canada.
Edwards, J.R., Parry, M.E., 1993. On the use of polynomial regression equa-
tions as an alternative to difference scores in organizational research.
Academy of Management Journal 36, 1577–1613.
Edwards, J.R., Shipp, A.J., 2007. The relationship between person-
environment fit and outcomes: An integrative theoretical frame-
work. In: Ostroff, C., Judge, T.A. (Eds.), Perspectives on Orga-
nizational Fit. Taylor & Francis Group, LLC, New York, NY,
pp. 209–258.
Ennen, E., Richter, A., 2010. The whole is more than the sum of its part,
or is it? A review of the empirical literature on complementarities in
organizations. Journal of Management 36, 207–233.
Finney, S.J., DiStefano, C., 2006. Non-normal and categorical data in
structural equation modeling. In: Hancock, G.R., Mueller, R.O. (Eds.),
Structural Equation Modeling: A Second Course. Information Age,
Greenwich, CT, pp. 269–314.
Fisher, J., 1995. Contingency-based research on management control sys-
tems: Categorization by level of complexity. Journal of Accounting
Literature 14, 24–53.
Franco-Santos, M., Lucianettib, L., Bourne, M., 2012. Contemporary per-
formance measurement systems: A review of their consequences
and a framework for research. Management Accounting Research 23,
79–119.
Galbraith, J., 1973. Designing Complex Organizations. Addison-Wesley,
Reading, MA.
Ganzach, Y., 1997. Misleading interaction and curvilinear terms. Psycho-
logical Methods 2, 235–247.
Gerdin, J., 2005a. The Impact of departmental interdependencies and man-
agement accounting system use on subunit performance. European
Accounting Review 14, 297–327.
Gerdin, J., 2005b. The impact of departmental interdependencies and
management accounting system use on subunit performance: A sec-
ond look. European Accounting Review 14, 335–340.
Gerdin, J., Greve, J., 2004. Forms of contingency fit in management
accounting research – a critical review. Accounting, Organizations and
Society 29, 303–326.
Gerdin, J., Greve, J., 2008. The appropriateness of statistical methods for
testing contingency hypotheses in management accounting research.
Accounting, Organizations and Society 33, 995–1009.
Greiner, L.E., 1998. Commentary and revision of HBR classic, “Evolution
and revolution as organizations grow”. Harvard Business Review 76,
55–68.
Hall, M., 2008. The effect of comprehensive performance measurement
systems on role clarity, psychological empowerment and managerial
performance. Accounting, Organizations and Society 33, 141–163.
Hall, M., 2011. Do comprehensive performance measurement systems
help or hinder managers’ mental model development? Management
Accounting Research, 1–16.
Harris, M.M., Anseel, F., Lievens, F., 2008. Keeping up with the Joneses: a
field study of the relationships among upward, lateral, and downward
comparisons and pay level satisfaction. Journal of Applied Psychology
93, 665–673.
Hartmann, F.G.H., 2005. The impact of departmental interdependencies
and management accounting system use on subunit performance: A
comment. European Accounting Review 14, 329–334.
Hartmann, F.G.H., Moers, F., 1999. Testing contingency hypotheses in bud-
getary research: An evaluation of the use of moderated regression
analysis. Accounting, Organizations and Society 24, 291–315.
Hartmann, F.G.H., Moers, F., 2003. Testing contingency hypotheses in bud-
getary research using moderated regression analysis: a second look.
Accounting, Organizations and Society 28, 803–809.
Hayes, A.F., Matthes, J., 2009. Computational procedures for probing inter-
actions in OLS and logistic regression: SPSS and SAS implementations.
Behaviour Research Methods 41, 924–936.
Henri, J.-F., 2006. Management control systems and strategy: A
resource-based perspective. Accounting, Organizations and Society
31, 529–558.
Henri, J.-F., 2007. A quantitative assessment of the reporting of structural
equation modeling information: The case of management accounting
research. Journal of Accounting Literature 26, 76–115.
Irwin, J.R., McClelland, G.H., 2001. Misleading heuristics and moder-
ated multiple regression models. Journal of Marketing Research 38,
100–109.
Ittner, C.D., Lanen, W.N., Larcker, D.F., 2002. The association between
activity-based costing and manufacturing performance. Journal of
Accounting Research 40, 711–726.
Ittner, C.D., Larcker, D.F., Randall, T., 2003. Performance implications
of strategic performance measurement in financial services firms.
Accounting, Organizations and Society 28, 715–741.
Jaccard, J., Turrisi, R., 2003. Interaction Effects in Multiple Regression. Sara
Miller McCune. Sage Publications, Inc, Thousands Oaks.
Jaccard, J., Wan, C.K., 1995. Measurement error in the analysis of interac-
tion effects between continous predictors using multiple regression:
Multiple indicator and structural equation approaches. Psychological
Bulletin 117, 348–357.
Jaccard, J., Wan, C.K., 1996. Lisrel Approaches to Interaction Effects in
Multiple Regression. Sage Publications, Thousand Oaks.
Jaccard, J.J., Turrisi, R., Wan, C.K., 1990. Interaction Effects in Multiple
Regression. SAGE Publications, Newbury Park, CA.
Jacobucci, D., 2008. Mediation Analysis. SAGE Publications, Inc, Thousand
Oaks, CA.
Jöreskog, K.G., Sörbom, D., 2001. LISREL 8: User’s Reference Guide, 2nd ed.
Scientific Software International, Inc., Lincolnwood, IL.
Kennedy, P., 2003. A Guide to Econometrics, 5th ed. Cambridge.
Kenny, D.A., Judd, C.M., 1984. Estimating the nonlinear and interactive
effects of latent variables. Psychological Bulletin 96, 201–210.
Klaas, P., Donaldson, L., 2009. Underfits versus overfits in the contin-
gency theory of organizational design: Asymmetric effects of misfits
 M. Burkert et al. / Management Accounting Research 25 (2014) 6–29
on performance. In: Bollingtoft, A., Hakonsson, D.D., Nielsen, J.F., Snow,
C.C., Ulhoi, J. (Eds.), New Approaches to Organization Design. Springer,
Dordrecht, pp. 147–168.
Klein, A., Moosbrugger, H., 2000. Maximum likelihood estimation of latent
interaction effects with the LMS method. Psychometrika 65, 457–474.
Klein, A., Muthén, B.O., 2007. Quasi-maximum likelihood estimation of
structural equation models with multiple interaction and quadratic
effects. Multivariate Behavioral Research 42, 647–673.
Klein, A.G., Schermelleh-Engel, K., Moosbrugger, H., Kelava, A., 2009.
Assessing spurious interaction effects. In: Teo, T., Khine, M.S. (Eds.),
Structural Equation Modeling in Educational Research. Sense Publish-
ers, Rotterdam, Boston, Taipei.
Kristof-Brown, A.L., Jansen, K.J., 2007. Issues of person-organization fit. In:
Ostroff, C., Judge, T.A. (Eds.), Perspectives on Organizational Fit. Taylor
& Francis Group, LLC, New York, NY, pp. 123–153.
Kristof-Brown, A.L., Zimmerman, R.D., Johnson, E.C., 2005. Conse-
quences of individuals’ fit at work: A meta-analysis of person–job,
person–organization, person–group, and person–supervisor fit. Per-
sonnel Psychology 58, 281–342.
Lee, C.-L., Yang, H.-J., 2011. Organization structure, competition and
performance measurement systems and their joint effects on perfor-
mance. Management Accounting Research 22, 84–104.
Lee, Y.-T., Antonakis, J., 2012. When preference is not satisfied but the
individual is: how power distance moderates person–job fit. Journal
of Management, http://dx.doi.org/10.1177/0149206311436080.
Lubinski, D., Humphreys, L.G., 1990. Assessing spurious “moderator
effects”: Illustrated substantively with the hypothesized (“synergis-
tic”) relation between spatial and mathematical ability. Psychological
Bulletin 107, 385–393.
Luft, J., Shields, M.D., 2003. Mapping management accounting: Graphics
and guidelines for theory-consistent empirical research. Accounting,
Organizations and Society 28, 169–249.
MacCallum, R.C., Mar, C.M., 1995. Distinguishing between moderator and
quadratic effects in multiple regression. Psychological Bulletin 118,
405–421.
Malmi, T., Brown, D.A., 2008. Management control system as package
– Opportunities, challenges and research directions. Management
Accounting Research 19, 287–300.
Marginson, D., Bui, B., 2009. Examining the human cost of multiple role
expectations. Behavioral Research in Accounting 21, 59–81.
Marsh, H.W., Wen, Z., Hau, K. -T., 2004. Structural equation models of
latent interactions: Evaluation of alternative estimation strategies and
indicator construction. Psychological Methods 9, 275–300.
Marsh, H.W., Wen, Z., Hau, K.-T., 2006. Structural equation models of
latent interaction and quadratic effects. In: Hancock, G., Mueller, R.O.
(Eds.), Structural Equation Modeling: A Second Course. Information
Age Publishing, Greenwich, CT, pp. 225–265.
Meilich, O., 2006. Bivariate models of fit in contingency theory: Critique
and a polynomial regression alternative. Organizational Research
Methods 9, 161–193.
Milgrom, P., Roberts, J., 1990. The economics of modern manufacturing:
Technology, strategy and organization. American Economic Review
80, 511–528.
Milgrom, P., Roberts, J., 1995. Complementarities and fit: Strategy,
structure, and organizational change in manufacturing. Journal of
Accounting and Economics 19, 179–208.
Moosbrugger, H., Schermelleh-Engel, K., Kelava, A., Klein, A.G., 2009. Test-
ing multiple nonlinear effects in structural equation modeling: A
comparison of alternative estimation approaches. In: Teo, T., Khine,
S. (Eds.), Structural Equation Modelling in Educational Research: Con-
cepts and Applications. Sense Publishers, Rotterdam, Boston, Taipei,
pp. 103–136.
Moosbrugger, H., Schermelleh-Engel, K., Klein, A., 1997. Methodological
problems of estimating latent interaction effects. Methods of Psycho-
logical Research Online 2, 95–111.
Muthén, L.K., Muthén, B.O., 2007. Mplus User’s Guide, 5th ed. Muthén &
Muthén, Los Angeles.
Myers, R.H., Montgomery, D.C., Anderson-Cook, C.M., 2009. Response Sur-
face Methodology, 3rd ed. John Wiley & Sons, Inc, Hoboken NJ.
Naman, J.L., Slevin, D.P., 1993. Entrepreneurship and the concept of fit:
A model and empirical results. Strategic Management Journal 14,
137–153.
29
Ostroff, C., Judge, T.A., 2007. Perspectives on Organizational Fit. Taylor and
Francis Group, LLC, New York, NY.
Ostroff, C., Schulte, M., 2007. Multiple perspectives of fit in organizations
across levels of analysis. In: Ostroff, C., Judge, T.A. (Eds.), Perspectives
on Organizational Fit. Taylor & Francis Group, LLC, New York, NY, pp.
3–69.
Preacher, K.J., Curran, P.J., Bauer, D.J., 2006. Computational tools for pro-
bing interactions in multiple linear regression, multilevel modeling,
and latent curve analysis. Journal of Educational and Behavioral Statis-
tics 31, 437–448.
Preacher, K.J., Rucker, D.D., Nicewander, W.A., 2005. Use of the extreme
groups approach: A critical reexamination and new recommenda-
tions. Psychological Methods 10 (2), 178–192.
Preacher, K.J., Hayes, A.F., 2004. SPSS and SAS procedures for estimat-
ing indirect effects in simple mediation models. Behavior Research
Methods, Instruments, and Computers 36, 717–731.
Reinartz, W., Haenlein, M., Henseler, J., 2009. An empirical comparison
of the efficacy of covariance-based and variance-based SEM. Interna-
tional Journal of Research in Marketing 26, 332–344.
Rigdon, E.E., Schumacker, R.E., Wothke, W., 1998. A comparative review
of interaction and nonlinear modeling. In: Schumacker, R.E., Mar-
coulides, G.A. (Eds.), Interaction and Nonlinear Effects in Structural
Equation Modeling. Erlbaum, Mahwah, NJ, pp. 1–16.
Rossing, C.P., 2013. Tax strategy control: The case of transfer pricing tax
risk management. Management Accounting Research 24, 175–194.
Said, A.A., Hassabelnaby, H.R., Wier, B., 2003. An empirical investigation
of the performance consequences of nonfinancial measures. Journal
of Management Accounting Research 15, 193–223.
Schermelleh-Engel, K., Klein, A., Moosbrugger, H., 1998. Estimating non-
linear effects using a latent moderated structural equations approach.
In: Schumacker, R.E., Marcoulides, G.A. (Eds.), Interaction and Non-
linear Effects in Structural Equation Modeling. Lawrence Erlbaum
Associates, Mahwah, NJ, pp. 203–238.
Schoonhoven, C.B., 1981. Problems with contingency theory: Testing
assumptions hidden within the language of contingency “theory”.
Administrative Science Quarterly 26, 349–377.
Shanock, L.R., Baran, B.E., Gentry, W.A., Pattison, S.C., Heggestad, E.D.,
2010. Polynomial regression with response surface analysis: A
powerful approach for examining moderation and overcoming lim-
itations of difference scores. Journal of Business Psychology 25, 543–
554.
Sharma, S., Durand, R.M., Gur-Arie, O., 1981. Identification and analysis of
moderator variables. Journal of Marketing Research 18, 291–300.
Shields, J.F., Shields, M.D., 1998. Antecedents of participative budgeting.
Accounting, Organizations and Society 23, 49–76.
Smith, D., Langfield-Smith, K., 2004. Structural equation modeling in man-
agement accounting research: Critical analysis and opportunities.
Journal of Accounting Literature 23, 49–86.
Sobel, M.E., 1982. Asymptotic confidence intervals for indirect effects
in structural equation models. In: Leinhardt, S. (Ed.), Sociological
Methodology. American Sociological Association, Washington DC, pp.
290–312.
Southwood, K.E., 1978. Substantive theory and statistical interaction: Five
models. American Journal of Sociology 83, 1154–1203.
Tushman, M.L., Nadler, D.A., 1978. Information processing as an inte-
grating concept in organizational design. Academy of Management
Review 3, 613–624.
Van der Stede, W.A., Chow, C.W., Lin, T.W., 2006. Strategy, Choice of
Performance Measures, and Performance. Behavioral Research in
Accounting 18, 185–205.
Venkatraman, N., 1989. The concept of fit in strategy research: Toward ver-
bal and statistical correspondence. Academy of Management Review
14, 423–444.
Viator, R.E., 2001. The association of formal and informal public account-
ing mentoring with role stress and related job outcomes. Accounting,
Organizations and Society 26, 73–93.
Zatzick, C.D., Moliterno, T.P., Fang, T., 2012. Strategic (mis)fit: The
implementation of TQM in manufacturing organizations. Strategic
Management Journal 33, 1321–1330.
Zhao, X., Lynch Jr., J.G., Chen, Q., 2010. Reconsidering Baron and Kenny:
Myths and truths about mediation analysis. Journal of Consumer
Research 37, 197–206.