The Issue of Endogeneity Within Theory-Based, Quantitative Management Accounting Research
The Issue of Endogeneity Within Theory-Based, Quantitative Management Accounting Research
The Issue of Endogeneity Within Theory-Based, Quantitative Management Accounting Research
To cite this article: Robert H. Chenhall & Frank Moers (2007) The Issue of Endogeneity within
Theory-Based, Quantitative Management Accounting Research, European Accounting Review,
16:1, 173-196, DOI: 10.1080/09638180701265937
Introduction
There has developed a tradition of periodic reviews of progress and problems
related to theory-based empirical research in management accounting. These
reviews have attempted to clarify theories, reflect on findings and method result-
ing in evaluations of the extant literature (e.g. most recently Hartmann and
Moers, 1999; Hartmann, 2000; Ittner and Larcker, 2001; Chenhall, 2003; Luft
and Shields, 2003). These have been, in the main, constructive and arguably
have assisted in focusing the research agenda. While some commentators are
impatient with the lack of progress in addressing identified problems, it can be
contended that there has been heightened concern with issues related to theory
development, data collection and analysis.
Correspondence Address: Robert H. Chenhall, Department of Accounting & Finance, Monash University,
Wellington Road, Clayton, Victoria, Australia. E-mail: Robert.chenhall@BusEco.monash.edu.au
all this is that econometrics is crucial in understanding the issue, but econo-
metrics cannot solve it.
We refrain from critically reviewing past management accounting papers and
evaluating the extent of the endogeneity problem in these studies. We believe that
this is not helpful in stimulating the debate for the following reason. As noted
above, all empirical papers have endogeneity issues and a focus on past
management accounting papers would make it wrongly appear that it is a
problem unique to management accounting problem. Our purpose is to look
forward and discuss ways to tackle a general problem in a specific research
area, being management accounting. We also refrain from discussing complex
econometric tools because it is important to first have a solid understanding of
the underlying problem before these issues can be addressed in depth. Further-
more, we are skeptical about the extent to which simply using more advanced
econometrics is going to resolve endogeneity problems. It often merely gives
the researcher an unjustified feeling of assurance that a technical treatment can
fully resolve endogeneity concerns.
We offer the following areas to direct the discussion and these areas also reflect
the structure of the remainder of this paper.
. What is endogeneity?
. How important is the endogeneity assumption in theory construction?
. What is the link between endogeneity and choice variables?
. Can we deal with endogeneity?
. Is it a (management) accounting problem?
What is Endogeneity?
This paper is concerned with theory-based quantitative empirical research and,
given this focus, we make use of the predictive validity framework (Runkel
and McGrath, 1972) to discuss and explain endogeneity and to position it in
the broader context of theory-based empirical research (see Figure 1 for the pre-
dictive validity framework).1 Specifically, we consider research that develops
mathematical representations of theory that show the relationship between a
set of defined variables. These relationships are based on theories drawn from
a wide variety of disciplines including economics, behavioral and organization
sciences, and information sciences. Data are used to test these theories by way
of statistical analysis. The purpose of the theory is to develop a model, at the con-
ceptual level, that identifies one or more variables of interest, such as the useful-
ness of budgets, and other variables, such as timeliness of budgets, that are
expected to be associated with the behavior of the variable of interest. These vari-
ables are often referred to as dependent or explained variables and independent or
explanatory variables, respectively.
In theory-based quantitative research, the theory is formalized by the so-
called structural equation (link 1 in the predictive validity framework).
176 R. H. Chenhall & F. Moers
For example,
Y ¼ b 0 þ b1 X 1 þ u (1)
Y ¼ b^ 0 þ b^ 1 X1 þ 1 (2)
equal to the true value. Consistency indicates that the distribution of the estimator
b^ 1 becomes concentrated on the true value b1 as the size of the sample used to
estimate b^ 1 increases. We are concerned with unbiasedness and consistency
because this reflects the extent to which link 4 in the predictive validity frame-
work tells us something about link 1.4 The role of the error term in the regression
equation is crucial in answering this question and is central to understanding the
issue of endogeneity.
There is some variation in the meaning of endogenous and exogenous variables.
In general usage, a distinction between exogenous and endogenous variables may
be made that relates the origins of the variables to be either ‘inside’ or ‘outside’
the structural equation. A variable is endogenous if it is determined within the
context of the model, while an exogenous variable is a variable that affects
the values of endogenous variables, but whose values are determined outside the
model. To clarify this somewhat vague statement, we refer back to equation (1),
that is,
Y ¼ b0 þ b1 X1 þ u:
Now assume that X1 is also determined by a number of factors and, more specifi-
cally, assume that the following structural equation applies:
X1 ¼ g0 þ g1 Z1 þ v: (3):
Equation (3) indicates that the variable X1 is endogenous, as it is the explained vari-
able. The main question, however, is whether it is endogenous in equation (1). The
variable X1 is endogenous in equation (1) if it is correlated with the structural error
term, that is, Cov(X1, u) = 0. If X1 is correlated with the structural error term, then
X1 is determined inside the model (equation (1)), because the presence of this cor-
relation is either due to Cov(Z1, u) = 0 or due to Cov(v, u) = 0. That is, (some
of) the factors that affect X1 also affect Y and as a result equations (1) and (3) are
parts of the same model. If X1 is not correlated with the structural error term of
equation (1), then it must hold that both Cov(Z1, u) ¼ 0 and Cov(v, u) ¼ 0,
and X1 is thus determined outside the model and not endogenous. In sum, the
explained variable is, by definition, endogenous because it is always correlated
with the structural error term.5 An explanatory variable may or may not be an
endogenous variable.
Endogeneity exists when the model includes an endogenous explanatory vari-
able. The potential for endogeneity exists in virtually all studies involving
accounting, finance and economic variables. We draw on a variety of areas
including economics (labor economics), management accounting (performance
measurement) and financial accounting (corporate governance) to illustrate
how endogeneity relates to a cross section of empirical research. Growing recog-
nition of the problems caused by endogeneity, including faulty conclusions about
178 R. H. Chenhall & F. Moers
well developed and we may not see an association between process controls and
the non-financial performance measurement system. With the processes under
control the non-financial measures adopt a role of focusing on customers, strat-
egies, benchmarking and engaging in redesign and re-engineering. Thus, while
advanced TQM may generate situations in which process controls still enhance
performance (although this may not be the case because early ‘obvious’ improve-
ments have been gained), there will be no endogeneity if process controls are not
associated with non-financial performance measures.
In the case of low TQM, as ‘process controls’ are not included explicitly as an
explanatory variable in the equation, it will be part of the structural error term u
(i.e. it will be included along with all other factors that affect performance, other
than the non-financial performance measures) and it is also correlated with ‘non-
financial performance measures’. As a result, Cov(X1, u) = 0 and the analysis is
thus subject to endogeneity.
An area of contemporary interest in financial accounting and finance is corpor-
ate governance. Corporate governance refers to the mechanisms that help align
the decisions of managers with the interests of shareholders and other stake-
holders. There is a broad array of corporate governance mechanisms.7 The com-
position and size of the board of directors has been identified as important in
providing corporate governance. Perhaps the most fundamental issue concerning
the role of the board is whether more board members provide enhanced govern-
ance and as a consequence improved performance. However, there is a view that
large boards are more ineffective at governance as they become more symbolic
and less concerned with the process of management (Lipton and Lorsch, 1992;
Jensen, 1993). These views have received some empirical support with studies
showing board size and firm value being negatively associated (Yermack,
1996; Eisenberg et al., 1998). Reflecting on why this relationship should hold
raises the potential for endogeneity problems. Successful and dominant CEOs
may be able to limit the size of their company’s board, in an attempt to decrease
the board’s influence. If there is also a significant association between the suc-
cessful track record of the CEO and performance then the role of the CEO
creates endogeneity.
Ideally, in regression analysis, the explained variable is significantly associated
with the explanatory variables (link 4), which provides support for a theoretically
proposed causal relationship between the variables (link 1). However, if endo-
geneity exists, we can no longer be confident that the results from the regression
support the causality implied in the structural equation. To see why this is so, we
need to compare the structural equation (equation (1)) to the OLS regression
equation (equation (2)). The OLS estimation method leads, by construction, to
no correlation between the explanatory variable X1 and the OLS error term 1,
that is, the explanatory variables and the OLS error term are always orthogonal.8
This implies that, if the explanatory variable X1 is correlated with the structural
error term u, then OLS will ensure that this correlation is eliminated at the oper-
ational level by adjusting the coefficient b^ 1 . Intuitively, X1 will ‘pick up’ all the
180 R. H. Chenhall & F. Moers
Omitted Variables
In theory-based quantitative research, the aim is to test a causal relationship
between explained and explanatory variables. In this case, the empirical model
will provide an estimate for prediction but the focus is on how well the explained
variable is influenced by the explanatory variables. Thus, it is the regression co-
efficients that are important and the residuals are interpreted as summing the
influence of causal variables not included in the model. In selecting which vari-
ables to include in the model, the researcher is guided by the theory that supports
the relationships within the model. Inevitably, there will be secondary, omitted
variables that may be important within the model. If these variables merely
add additional predictive power, they may be seen as complementary to the pre-
diction. For example, in examining the usefulness of balanced scorecards, find-
ings of significant relationships with size and industry would support a theory
of size and industry effects. The introduction of another variable, say strategic
orientation may be justified on theoretical grounds and when introduced into
the regression may be significantly associated with balanced scorecard usefulness
and add to the R-square. Thus, strategy is complementary and adds to the
predictive power of the model.
Of concern in testing causal relationships is the possibility that an explanatory
variable modeled as exogenous, will in fact be endogenous because of omitted
variables (link 5). That is, there will be some systematic relationship between
the explanatory variable of interest, modeled as an exogenous variable, and an
Endogeneity within Theory-Based Accounting Research 181
omitted variable not included in the model that is associated with both the expla-
natory variable and the explained variable. As noted in earlier examples, the
omitted variable of ‘process controls’ in the relationship between non-financial
performance measures and performance; and the role of a ‘successful CEO’ in
the relationship between board size and performance illustrate how omitted vari-
ables can cause endogeneity problems. In the balanced scorecard example,
decentralized structures might be associated with both size and usefulness of
balanced scorecards. Decentralized structure is an omitted variable that can
cause endogeneity problems. In these cases, the models are seen as being under-
specified as an important variable that is associated with both the explained and
the explanatory variable is excluded from the analysis. The degree to which this is
a problem, that is, leads to a substantial bias in b^ 1 , can be illustrated as follows.
Assume that the structural equation is given by
Y ¼ b0 þ b1 X1 þ b2 Z1 þ u: (4)
The method of OLS provides the Best Linear Unbiased Estimate (BLUE) if all
explanatory variables in the regression are exogenous. So, if we want to use
OLS, we need to write Y as a function of exogenous variables only, that is, we
want to know the reduced form equation for Y. The reduced form for Y is in
this case identical to the structural equation. Assume that instead of estimating
equation (4), we estimate equation (2). Using the method of OLS for estimating
equation (2) (see, e.g. Wooldridge, 2000, p. 89), the expected value of the OLS
estimate for X1(b^ 1 ) can then formally be written as
Cov(X1 ,Z1 )
E½b^ 1 ¼ b1 þ b2 (5)
Var(X1 )
Cov(X1 ,Z1 )
E½b^ 1 b1 ¼ b2 : (6)
Var(X1 )
In the above equations, b2 reflects the sign and magnitude of the impact of Z1
on Y and Cov(X1, Z1) reflects the strength of the relationship between X1 and Z1.
The implications of the above equations are relatively simple when considering
the importance of omitted variables. Ceteris paribus, the bigger the impact of Z1
on Y the larger the bias, and the stronger the relationship between X1 and Z1 the
larger the bias. If b2 and/or Cov(X1, Z1) are ‘sufficiently’ small, then endogeneity
is not a serious problem because the bias is small. More importantly, if either b2
or Cov(X1, Z1) equals zero, then there is no endogeneity. In other words, if the
purpose is to test whether a theoretically proposed relationship between Y and
X1 exists, then we are not concerned about omitted variables that are not cor-
related with X1, nor do we need to be troubled about correlated variables that
182 R. H. Chenhall & F. Moers
do not affect Y. We are only concerned about variables that are correlated with
both the explained and explanatory variable.
This observation is important given that causes of omitted variables are mainly
due to the necessity to ex ante limit the research model to a theoretically manage-
able number of variables. The key is thus to carefully identify the variables to be
included in the model following the arguments outlined above. That is, based on
theory and previous empirical evidence, one should identify those variables that
are likely to have a major impact on both the explained and explanatory variable
and are thus most likely to potentially affect the results if excluded from the
analysis. To do this, one should articulate the structural model that gives rise
to the reduced form model.10
Simultaneity
Another important cause of endogeneity related to theory construction is simul-
taneity. Simultaneity arises when one or more of the explanatory variables are
jointly determined with the explained variable. In this case, the causal relation-
ship between an explained and an explanatory variable runs both ways (link 1
is reciprocal). For example, it may be proposed that ‘usefulness of budgets’
will occur as a consequence of employing budget information that is timely.
This relationship holds if ‘timeliness of information’ is exogenously determined.
However, ‘timeliness of information’ will not be exogenous if there is a
reciprocal relationship between ‘usefulness of budgets’ and ‘timeliness of
information’. That is, as the budgets become more useful, the usefulness
causes data to be collected in a more timely way. If simultaneity exists, then
the structural error term is by definition correlated with the explanatory variable.
To illustrate this, assume we have the following two-equation structural model:
Y1 ¼ b0 þ b1 Y2 þ b2 Z1 þ u1 (7)
Y2 ¼ d0 þ d1 Y1 þ d2 Z2 þ u2 (8)
where Z1 and Z2 are assumed to be exogenous and we want to estimate the first
equation, that is, equation (7). To show that Y2 is endogenous in equation (7), we
rewrite Y2 by replacing Y1 in equation (8) by equation (7). This leads to
d 0 þ d 1 b0 d 1 b2 d2 d1 1
Y2 ¼ þ Z1 þ Z2 þ u1 þ u2 : (9)
1 d 1 b1 1 d 1 b1 1 d1 b1 1 d1 b1 1 d 1 b1
Equation (9) is the reduced form equation for Y2, that is, the endogenous vari-
able is written as a function of exogenous variables. As can be seen from this
equation, Y2 is a function of u1 and Y2 is thus endogenous in equation (7). In
general, if d1 = 0, then there exists endogeneity due to simultaneity. Intuitively,
given that the structural error term contains all ‘other’ factors that affect the
Endogeneity within Theory-Based Accounting Research 183
explained variable, and the explained variable in turn affects the explanatory vari-
able (simultaneity), the structural error is also correlated with the explanatory
variable. The OLS method does not take this simultaneity into account and
thus leads to biased estimators (simultaneity bias).11 Even if d1 ¼ 0, there can
be endogeneity if u1 and u2 are correlated, but this is identical to the omitted vari-
able bias discussed above.
Equilibrium Conditions
Consideration of endogeneity helps focus attention on an important issue in man-
agement accounting research that relates to the assumption that firms operate in
equilibrium conditions. The most important question in this respect concerns
whether organizational performance can be used as the explained variable
when studying the effectiveness of management accounting systems. It is
argued that if equilibrium conditions are assumed, then all firms are performing
optimally, given their circumstances, and it is therefore inappropriate to examine
organizational performance. The management accounting system is a choice
variable and, in equilibrium, firms make optimal choices. That is, there is no
relationship between organizational performance and the management account-
ing system after controlling for all exogenous factors that affect firms’ accounting
choices (e.g. Demsetz and Lehn, 1985; Ittner and Larcker, 2001; Larcker, 2003).
For example, following the Demsetz and Lehn (1985) argument, there should be
no association between performance and the extent to which non-financial per-
formance measures are used. Every firm will at any time make optimal decisions
and for some firms it is optimal to emphasize non-financial measures, while for
others it is optimal to emphasize financial measures. Since both types of firms
make optimal decisions, there are no differences in economic profit between
these firms, even though there are differences in the extent to which they use
non-financial performance measures.
It is beyond the scope of this paper to fully examine whether or not there is
variation in the performance of surviving firms and whether such variation can
be associated with different types of management accounting systems. While
systems may evolve towards equilibrium, there is substantial discourse as to
how the dynamics of this takes place, if it is ever achieved given continual
shocks to the systems, and if systems have a propensity to move away from equi-
librium or sustain long periods of chaos.12 However, if equilibrium conditions are
assumed, then any empirical association between organizational performance
and attributes of management accounting systems must be due to model misspe-
cification (endogeneity), as no such relationship exists in equilibrium.
variables and equilibrium conditions. These questions are central to the criticism
that models investigating the outcomes of the choice of management accounting
systems are flawed, as the models will suffer from endogeneity.13 We identify
three questions. First, is it always problematic to put a choice variable on the
right-hand side of the regression equation, that is, does it always lead to
(severe) endogeneity? Second, is the problem solved when we put the choice
variable on the left-hand side of the regression equation? Finally, are choice vari-
ables the only potential endogenous explanatory variables? We consider each of
these questions in turn.
But even if experiments are completely free from endogeneity issues, they
have their own problems (external validity) and are unlikely to provide all the
answers. There is often a need for quasi-experimental research. In quasi-
experimental research, like survey and archival research, the researcher can
generally not determine the stimuli and achieve experimental control. In this
case, other ways of controlling for endogeneity become relevant. We discuss
these below in relation to the causes of endogeneity that we identified, that is,
omitted variables and simultaneity. For each of the two causes, we indicate (i)
the ‘textbook’ solutions and (ii) their shortcomings.
Omitted Variables
(i) If the omitted variable problem needs to be solved, then there are basically
two standard ways to do this. First, if we believe there is an omitted variable
but we are unable to include it because of data unavailability, we can include
a proxy variable in the analysis. A proxy variable is a variable that is related
to the unobserved omitted variable. For example, consider the relationship
between an explained variable, ‘usefulness of budgets’, and an explanatory
variable, ‘timeliness of information’, and we believe that ‘knowledge of
accounting’ is an important omitted variable but we are unable to obtain
data on this omitted variable. In this case, data on ‘tertiary education in
accounting’ is available and if we believe that this is related to ‘knowledge
of accounting’, it can be used as a proxy for ‘knowledge in accounting’. By
including this variable, we control for the impact of the omitted variable on
the explained variable and thus make the explanatory variable exogenous.19
This solution is called the ‘plug-in solution to the omitted variables
problem’. Second, if a suitable proxy variable is not available we can use
the method of instrumental variables. Instrumental variables are variables
that are correlated with the explanatory variable and uncorrelated with the
omitted variables (structural error term). In our example, ‘digitization of
IT systems’ can be used as an instrument for ‘timeliness of information’.
Intuitively, we can use the instrumental variable to predict the explanatory
variable and use this prediction to estimate the impact of the explanatory
variable on the explained variable. Under the instrumental variable assump-
tions stated above, the prediction is uncorrelated with the omitted variable
and the endogeneity problem is thus ‘solved’. That is, in our example, ‘digi-
tization of IT systems’ could be used as the Instrumental Variable as it is cor-
related with ‘timeliness of information’ but not with ‘knowledge of
accounting’. The resulting estimator is called the Instrumental Variable
(IV) estimator and, in the case of multiple instruments, sometimes the
Two Stage Least Squares (2SLS) estimator. This estimator is, under the
above-mentioned assumptions, consistent, though generally never unbiased.
It is important to note that the use of instrumental variables as an approach to
188 R. H. Chenhall & F. Moers
omitted variables has some special conditions (see, Wooldridge, 2002, pp.
105– 107).
(ii) The shortcomings of the ‘plug-in solution’ relate to typical measurement
issues such as validity and we therefore refrain from discussing these in
detail. Of more importance to this paper is to discuss the shortcomings of
the IV estimator. In the previous example, we stated that the use of ‘digitiz-
ation of IT systems’ as an instrument solves the endogeneity problem.
Unfortunately, such a solution rarely happens in empirical research, since
it is difficult, if not impossible, to find variables that satisfy the requirements
of an instrumental variable (Ittner and Larcker, 2001; Larcker, 2003). The
least problematic requirement is the relevance of the instrument, that is,
the extent to which the instrument and the explanatory variable are corre-
lated. To illustrate, we return to equations (1) and (3) without the intercepts,
that is,
Y ¼ b1 X1 þ u ¼ b1 g1 Z1 þ u (10 )
X1 ¼ g1 Z1 þ v: (30 )
Assume that Cov(X1, u) = 0 and we want to use the IV estimator to correct for
this endogeneity. If Cov(Z1, u) ¼ 0, that is, Z1 is exogenous in equation (1), and
Cov(X1, Z1) = 0, that is, X1 and Z1 are correlated (Z1 is relevant), then the con-
sistency of the IV estimator can be seen from the probability limit ( p lim) of the
estimator, that is,
Cov(Z1 , u)
p lim bIV
1 ¼ b1 þ (10)
Cov(Z1 , X1 )
where the second term is asymptotically zero. Thus, under these assumptions, the
IV estimator is consistent, irrespective of the relevance of Z1. However, the IV
estimator is not unbiased. Intuitively, in finite samples, the IV estimator is biased
because we need to estimate the impact of the instrumental variable on the expla-
natory variable, since we do not observe the ‘true’ relationship (e.g. Bound et al.,
1995; Hahn and Hausman, 2002).20 Hahn and Hausman (2002) provide an
approximation of the bias in the IV estimator, which can be roughly character-
ized by21
Cov(u, v)
E½bIV
1 b1 : (11)
R2f X12
This equation shows that bias increases with decreases in the first-stage
R-square (R2f ), that is, the lower the relevance of the instrument the greater the
bias in the IV estimator. Although Hahn and Hausman (2002) indicate that the
Endogeneity within Theory-Based Accounting Research 189
bias in the IV estimator is lower than the bias in the OLS estimator, the situation
is more complicated when the instrument is not exogenous in the structural
equation, that is, Cov(Z1, u) = 0. If Cov(Z1, u) = 0, then both the OLS estima-
tor and the IV estimator are biased and inconsistent. Larcker and Rusticus (2005)
show that in these circumstances IV estimators are unlikely to be preferred over
OLS estimators. In sum, the major shortcoming of IV estimation is that this ‘text-
book’ solution to the endogeneity problem is worse than the problem itself if the
instrument is weak (low relevance) and endogenous.
Simultaneity
(i) Simultaneity can be taken into account by using simultaneous equation
models. The technique of Two Stage Least Squares can be used to test simul-
taneity if a number of conditions are satisfied (see, e.g. Wooldridge, 2000).
The necessary condition for the system to be identified is the order condition,
while the sufficient condition is the rank condition. In a two-equation setting,
the order condition states that for identifying the first (second) equation, at
least one exogenous variable should be excluded from the first (second)
equation. The rank condition additionally requires that at least one of the
exogenous variables excluded from the first (second) equation has a
nonzero (population) coefficient in the second (first) equation.
(ii) The shortcomings of the 2SLS estimator have been discussed above. These
shortcomings are even more severe in the simultaneous equation setting,
since we now have to have valid instruments for two equations instead of
one (rank condition).
Cov(X1 , X2 )
E½b^ 1 b1 ¼ b2 :
Var(X1 )
190 R. H. Chenhall & F. Moers
d1
E½b^ 1 b1 ¼ Var (u1 ): (12)
1 d1 b 1
Given that Var(u1) . 0, the sign of the bias is given by d1 =(1 d1 b1 ). If opposite
effects are expected (see, e.g. Nagar, 2002), that is, b1 . 0 and d1 , 0, then there
exists an attenuation bias (potential Type II error). If both b1 . 0 (b1 , 0) and
d1 . 0 (d1 , 0), then the sign of the bias depends on whether b1d1 is lower or
greater than 1.
In general, the issue is whether the bias works in favor or against the hypoth-
esized relationship. However, establishing the direction of the bias becomes
(almost) impossible when the model has multiple endogenous explanatory vari-
ables. Thus, in a ‘real-life’ context, that is, actual empirical research, it is difficult
to predict the bias in OLS estimators. In addition, the paradox in this analysis is
that the sign and the magnitude of the bias is unknown ex ante, since it requires
knowledge of the true parameters that the researcher is trying to estimate in the
first place (Moers, 2006). Finally, the 2SLS solution is questionable in small
samples, which characterizes most management accounting studies. So, the
issue remains as to what can be done to address the potential for this bias?
In general, the ‘only’ thing we can do is articulate the structural model of inter-
est, deduct the reduced form equation and develop an empirical model based on
this. To the extent that the empirical model differs from the reduced form
equation, because of, for example, empirical restrictions, we need to analyze
this difference using the above simple procedures. If these simple procedures
indicate that bias is unlikely (likely) to be a major problem, then we are better
off using OLS (OLS and 2SLS). Obviously this recommendation is too general
to help researchers tackle the endogeneity issue. Below we provide some more
specific recommendations for different types of studies.
more concerned with the possibility that the effect might be driven by other
choice variables. Consider the example where we are interested in the
impact of non-financial performance measures on managerial short-term orien-
tation. Given that the underlying idea relies on motivational (incentive) effects,
the question is: what other motivational (incentive) mechanisms are chosen by
the ‘firm’ in conjunction with the use of non-financial performance measures
and what effects do these other mechanisms have on managerial short-term
orientation? By spending more effort on articulating the structural model,
these questions can be answered and used to improve the empirical model.
. Choice variable on the left-hand side (LHS) of the equation: this usually
involves a typical analysis of correlated omitted variables. The basic question
here is whether the proposed determinant on the RHS really is the determinant,
or is there some ‘indirect’ effect (path)? Thus, in this case the focus is on the
relationship between the determinants!
. Choice variables on RHS and LHS: in this case, we suggest that researchers
follow both of the above two points, in addition to establishing the simultaneity
bias using the ‘simple’ rule above.
In all of the above cases, it is theory and logic that needs to be used as a first step in
addressing the endogeneity issue. That is, researchers first need to build better
arguments as a prelude to building econometric models. Empirical researchers
should be more careful to tell the reader something about their thinking about devel-
oping the structural model and (more importantly) identify which variables are
assumed to be endogenous and exogenous in whatever economic model they have
in mind. This would enable the reader to judge the reasonableness of the assumptions
about the variables and perhaps the degree to which endogeneity is a problem.
In addition to the above recommendations, we provide the following final
recommendations:
. If a decision is made to use 2SLS, then we recommend that the procedures dis-
cussed by Larcker and Rusticus (2005) are followed. Most, if not all, studies in
accounting that apply 2SLS provide (almost) no assurance that the use of this
method is actually valid. The Durbin –Wu –Hausman test that is typically
reported to substantiate the use of 2SLS is useless if 2SLS is itself not valid.
In essence, for this method to be valid, the instruments used must be (1) rel-
evant and (2) exogenous. Larcker and Rusticus (2005) state that, in addition
to providing some theoretical argument for why the instruments are expected
to be relevant and exogenous, specification tests need to be performed to cor-
roborate this argument. The specification tests for relevance are the first-stage
exclusion restriction and the partial R-square, which examine whether the
instruments are significantly related to the endogenous explanatory variable
and how much variance is explained by the instruments, respectively. The spe-
cification test for exogeneity is the overidentifying restriction, which can only
be used if the number of instruments exceeds the number of endogenous
192 R. H. Chenhall & F. Moers
Conclusion
In this paper, we have provided an overview and discussion of endogeneity. The
issue of endogeneity warrants discussion because we believe that some ‘readers’
of papers have overestimated the potential problems of endogeneity, while some
‘writers’ of papers have underestimated the potential problems. Restricting the
problem of endogeneity to studies that examine the effects of choice variables,
as appears to be the case in accounting, is an unsubstantiated restriction and an
overgeneralization. Much can be gained by examining the econometric definition
and treatments for endogeneity.
We have indicated that endogeneity has as much to do with putting a choice vari-
able, such as a management accounting system, on the right-hand side as with
putting a choice variable on the left-hand side. Importantly, we have indicated
that endogeneity has nothing to do with putting a choice variable on the right-
hand side. In other words, any attempt to define endogeneity in terms of what
type of variable is put on what side of the equation is likely to be unhelpful. It is
suggested that clarification is achieved if we focus on a single all-inclusive definition
of endogeneity derived from the econometric definition, that is, endogeneity is
present when an explanatory variable is correlated with the structural error term.
We have discussed the different causes of endogeneity and possible solutions
to the problem. At a practical level, it is unlikely that any single study is complet-
ely free of endogeneity issues and we therefore argue that the initial consideration
should be sought in careful theory construction. By providing an explanation of
what endogeneity actually is and what its consequences are, we hope that
researchers will be encouraged to appreciate more explicitly the extent to
Endogeneity within Theory-Based Accounting Research 193
which there may or may not be problems associated with endogeneity. As indi-
cated at the outset, the main aim of this paper is to assist in promoting debate
about endogeneity among management accounting researchers undertaking
theory-based quantitative research.
Notes
1
See Libby et al. (2002) for an application of the predictive validity framework to experimental
research in financial accounting.
2
In the literature, u is labeled the ‘structural residual’, ‘structural error term’ or ‘structural
disturbance’.
3
Link 5 in Figure 1 refers to controls. This is where ‘other potentially influential’ variables besides
the explanatory variables are controlled (Libby et al., 2002, p. 800). In, for example, experimental
research, this is possible by holding these variables constant or through randomization.
4
Unbiasedness is a feature of an estimator for a given sample size (finite sample property), while
consistency is a feature of the estimator when the sample size gets large (asymptotic property).
The assumptions needed to assure unbiasedness of the OLS estimator also assure its consist-
ency, although the opposite does not apply. That is, consistency does not necessarily imply
unbiasedness. However, when we state that the estimator is unbiased (biased) we also mean
it is consistent (inconsistent), unless we explicitly state otherwise. Consistency, sometimes
called asymptotic unbiasedness, is a minimal requirement for an estimator.
5
This follows as the structural error term reflects all factors that affect the explained variable
other than the included explanatory variables.
6
One explanation for this finding is that all firms are operating in equilibrium. We address this
issue later in the paper.
7
Other mechanisms of corporate governance include the provision of financial accounting infor-
mation to provide corporate transparency, executive compensation to align manager’s actions
with those of shareholders and direct monitoring by outside parties with a dominant shareholding.
8
This is why one needs to be careful of characterizing endogeneity as a correlation between the
explanatory variable and ‘the error term’ (see Greene, 2000, pp. 656– 657). It is the correlation
with the ‘structural error term’ that is of importance, since the model can always be rewritten to
have no correlation with the error term, as is the case with OLS.
9
In a strict sense, ‘functional misspecification’ (link 1 vs. link 4) and ‘measurement error’ (links
2 and 3) can cause the explanatory variable(s) to be correlated with the structural error term and
can thus lead to endogeneity. Given that these two types of problems deviate from the core issue
that we address, we refrain from discussing these problems.
10
Another common example of endogeneity that is caused by an omitted variable is self-selection
(Heckman, 1979). For example, the explained variable of ‘usefulness of sophisticated capital bud-
geting’ may be theoretically associated with ‘tertiary education in accounting’. However, there
may be an unobserved variable of ‘ability’ that relates to how useful managers find ‘sophisticated
capital budgeting’ which is also associated with ‘tertiary education in accounting’. This is caused
by individuals with high ‘ability’ self-selecting into higher ‘tertiary education in accounting’. As a
result, there is a correlation between ‘tertiary education in accounting’ and the structural error
term thereby generating endogeneity. Although the econometric ‘solution’ to the self-selection
problem is specific (see Heckman, 1979), the essence of the problem is identical to the omitted
variable problem discussed and we therefore refrain from addressing self-selection in more detail.
11
Note that the OLS method causes no concerns when estimating equation (9), which is the
reduced form equation.
12
Milgrom and Roberts (1992) argue that it is unlikely that everybody optimizes at the same time
and there should therefore be cross-sectional variation around the optimal choice (see also Ittner
and Larcker, 2001; Larcker, 2003).
194 R. H. Chenhall & F. Moers
13
It is important when considering choice variables to clarify if the managers, who are being con-
sidered in the theory, can in fact take decisions influencing the choice variables of interest.
There may be a theoretical relationship between strategy, incentive use of financial measures
and short-term decision orientation. However, the divisional manager may only have autonomy
over the division’s strategy with the incentive use of financial measures being determined and
fixed by senior managers. In this case, the incentive use of financial measures is not a choice
variable at the divisional level.
14
These types of direct and indirect effects underlie most path analyses and structural equation models,
where antecedents lead to choices and then to consequences (see, e.g. Van der Stede, 2000).
15
This logic is not only based on economics, it also forms the basis for Kerr’s (1975) oft-cited
behavioral paper ‘On the Folly of Rewarding A while Hoping for B’.
16
Only if we assume that there is simultaneity between effects and choice variables will putting a
choice variable on the right-hand side always lead to endogeneity. However, this is not only an
unrealistic assumption, it is also not the assumption underlying the arguments made in the account-
ing literature that drives this theory about incentives, strategy and short-term oriented behavior.
17
Note that it would also not be a very relevant analysis if we were ultimately interested in
knowing the impact of ‘timeliness of information’ on ‘usefulness of budgets’.
18
A similar example can be found in the area of international accounting research, which often
uses the variable ‘legal origin’ (common law vs. code law). In these types of studies, the
legal origin of a country is predetermined and can, to some extent, be considered theoretically
exogenous. However, if legal origin influences other institutions, and these other institutions are
not controlled for, then we still face endogeneity issues due to correlated omitted variables.
19
Note that, if we are able to observe and measure the omitted variable, we would obviously use
this measure in the analysis instead of the proxy variable.
20
This is not a problem when the sample size gets ‘very large’ (law of large numbers), which
means that there is ‘asymptotic unbiasedness’, that is, the estimator is consistent.
21
Note that this is a rough characterization of the bias, not consistency, which is represented by the
probability limit (see equation (10)). Conventional asymptotics (n ! 1) ignores the influence
of Cov(u, v) and R2f .
22
It is important to note that an implicit assumption underlying this method is that a residual of
‘zero’ represents the optimum. This requires that firms are equally likely to overshoot the
optimum as they are to undershoot the optimum.
References
Ashenfelter, O. and Card, D. (Eds) (2001) Handbook in Labor Economics (Amsterdam: North-
Holland).
Bound, J. et al. (1995) Problems with instrumental variables estimation when the correlation between
the instruments and the endogenous explanatory variable is weak, Journal of the American Stat-
istical Association, 90, pp. 443 –450.
Chenhall, R. H. (2003) Management control systems design within its organizational context: findings
from contingency-based research and directions for the future, Accounting, Organizations and
Society, 28, pp. 127 –168.
Core, J. E. (2001) A review of the empirical disclosure literature: discussion, Journal of Accounting
and Economics, 31, pp. 441–456.
Demsetz, H. and Lehn, K. (1985) The structure of corporate ownership: causes and consequences,
Journal of Political Economy, 93, pp. 1155–1177.
Eisenberg, T. et al. (1998) Larger board size and decreasing firm value in small firms, Journal of
Financial Economics, 48, pp. 35–54.
Engle, R. F. et al. (1983) Exogeneity, Econometrica, 51, pp. 277–304.
Fields, T. D. et al. (2001) Empirical research on accounting choice, Journal of Accounting and Econ-
omics, 31, pp. 255– 307.
Endogeneity within Theory-Based Accounting Research 195
Greene, W. H. (2000) Econometric Analysis (Upper Saddle River, NJ: Prentice Hall).
Hahn, J. and Hausman, J. (2002) Notes on bias in estimators for simultaneous equation models,
Economics Letters, 75, pp. 237 –241.
Hamilton, B. H. and Nickerson, J. A. (2003) Correcting for endogeneity in strategic management
research, Strategic Organization (London: Sage).
Hammer, M. (1996) Beyond Reengineering (New York: Harper Business).
Hartmann, F. G. H. (2000) The appropriateness of RAPM: toward the further development of theory,
Accounting, Organizations and Society, 25, pp. 451– 482.
Hartmann, F. G. H. and Moers, F. (1999) Testing contingency hypotheses in budgetary research: an
evaluation of the use of moderated regression analysis, Accounting, Organizations and Society,
24, pp. 291–315.
Hechman, J. (1979) Sample selection bias as a specification error, Econometrica, 47(1), pp. 153–161.
Ittner, C. D. and Larcker, D. F. (1995) Total quality management and the choice of information and
reward systems, Journal of Accounting Research, 33(supplement), pp. 1–34.
Ittner, C. D. and Larcker, D. F. (1998) Innovations in performance measurement, trends and research
implications, Journal of Management Accounting Research, 10, pp. 205–238.
Ittner, C. D. and Larcker, D. F. (2001) Assessing empirical research in managerial accounting: a value-
based management perspective, Journal of Accounting and Economics, 32, pp. 349–410.
Ittner, C. D. et al. (2002) The association between activity-based costing and manufacturing perform-
ance, Journal of Accounting Research, 40, pp. 711 –726.
Jensen, M. (1993) The modern industrial revolution, exit, and the failure of internal control systems,
Journal of Finance, 48, pp. 831–880.
Kerr, S. (1975) On the folly of rewarding A, while hoping for B, Academy of Management Journal, 18,
pp. 769–783.
Larcker, D. F. (2003) Discussion of ‘Are executive stock options associated with future earnings?’,
Journal of Accounting and Economics, 36, pp. 91 –103.
Larcker, D. F. and Rusticus, T. O. (2005) On the use of instrumental variables in accounting research,
Working Paper, University of Pennsylvania.
Libby, R. et al. (2002) Experimental research in financial accounting, Accounting, Organizations and
Society, 27, pp. 775–810.
Lipton, K. and Lorsch, J. (1992) A modest proposal for improved corporate governance, Business
Lawyer, 48, pp. 59–77.
Luft, J. and Shields, M. D. (2003) Mapping management accounting: graphics and guidelines for
theory-consistent empirical research, Accounting, Organizations and Society, 28, pp. 169–249.
Milgrom, P. and Roberts, J. (1992) Economics, Organization and Management (Englewood Cliffs,
NJ: Prentice Hall).
Moers, F. (2006) Performance measure properties and delegation, The Accounting Review, 81,
pp. 897–924.
Nagar, V. (2002) Delegation and incentive compensation, The Accounting Review, 77, pp. 379–395.
Runkel, P. and McGrath, J. (1972) Research on Human Behavior: A Systematic Guide to Method
(New York: Holt, Rinehart and Winston).
Shugan, S. M. (2004) Endogeneity in marketing decision models, Marketing Science, 23, pp. 1– 3.
Van der Stede, W. A. (2000) The relationship between two consequences of budgetary controls:
budgetary slack creation and managerial short-term orientation, Accounting, Organizations
and Society, 25, pp. 609–622.
Wooldridge, J. M. (2000) Introductory Econometrics: A Modern Approach (Cincinnati, OH:
South-Western College Publishing).
Wooldridge, J. M. (2002) Econometric Analysis of Cross Section and Panel Data (Cambridge, MA:
MIT Press).
Yermack, D. (1996) Higher market valuation of companies with a small board of directors, Journal of
Financial Economics, 40, pp. 185– 211.