You are on page 1of 10

Strategic Management Journal

Strat. Mgmt. J., 35: 1070–1079 (2014)


Published online EarlyView 25 June 2013 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/smj.2136
Received 5 September 2012 ; Final revision received 19 February 2013

RESEARCH NOTES AND COMMENTARIES

THE PERILS OF ENDOGENEITY AND INSTRUMENTAL


VARIABLES IN STRATEGY RESEARCH:
UNDERSTANDING THROUGH SIMULATIONS
MATTHEW SEMADENI,1 * MICHAEL C. WITHERS,2 and S. TREVIS CERTO3
1
Kelley School of Business, Indiana University, Bloomington, Indiana, U.S.A.
2
Mays Business School, Texas A&M University, College Station, Texas, U.S.A.
3
W.P. Carey School of Business, Arizona State University, Tempe, Arizona, U.S.A.

In this paper we use simulations to examine how endogeneity biases the results reported by
ordinary least squares (OLS) regression. In addition, we examine how instrumental variable
techniques help to alleviate such bias. Our results demonstrate severe bias even at low levels of
endogeneity. Our results also illustrate how instrumental variables produce unbiased coefficient
estimates, but instrumental variables are associated with extremely low levels of statistical
power. Finally, our simulations highlight how stronger instruments improve statistical power
and that endogenous instruments can report results that are inferior to those reported by OLS
regression. Based on our results, we provide a series of recommendations for scholars dealing
with endogeneity. Copyright  2013 John Wiley & Sons, Ltd.

INTRODUCTION reviewers and editors in multiple disciplines have


increasingly identified endogeneity as an alterna-
Shaver’s (1998) examination of market entry tive explanation for results presented in papers
modes introduced to strategy scholars the influ- they evaluate, and endogeneity represents a more
ence of endogeneity on statistical results. Since and more frequent reason for manuscript rejection
that time, other scholars in strategic manage- (e.g., Larcker and Rusticus, 2010; Shugan, 2004).
ment have described the consequences of endo- The objective of this paper is to provide a series
geneity as well as the techniques to circum- of simulations to illustrate the consequences of
vent it (e.g., Bascle, 2008; Hamilton and Nick- endogeneity and the robustness of the techniques
erson, 2003). Endogeneity occurs when an inde- prescribed to circumvent these consequences.
pendent variable is correlated with the error Simulations allow us to investigate the stability of
term (also known as “disturbance” or “residual”) analytical techniques while knowing what results
in an ordinary least squares (OLS) regression “should be,” and our techniques and outcome
model (for a review, see Kennedy, 2008). Endo- variables allow us to provide a number of contri-
geneity may bias the assertions that researchers butions to the literature in strategy on endogeneity.
make regarding hypothesized effects. To this end, First, the outcome measures we develop illustrate
that addressing endogeneity involves a balancing
act in which the disease (i.e., endogeneity)
Keywords: endogeneity; simulations; instrumental vari- introduces Type I error, whereas the cure (i.e.,
ables; ordinary least squares; strategy research instrumental variables) introduces Type II error.
*Correspondence to: Matthew Semadeni, Kelley School of Busi-
ness, Indiana University, 1309 East Tenth Street, Bloomington, Endogeneity—even at low levels—increases the
IN 47405, USA. E-mail: semadeni@indiana.edu likelihood of reporting statistically significant

Copyright  2013 John Wiley & Sons, Ltd.


Research Notes and Commentaries 1071

(yet untrue) coefficient estimates. At the same work that tests proposed theoretical perspectives.
time, correcting for endogeneity with instrumental We contribute to strategy research by providing
variables increases the likelihood of reporting recommendations to improve current practices
coefficient estimates that are near their true values, used by researchers, reviewers, and editors grap-
but these reported estimates are rarely statistically pling with the effects of endogeneity. We believe
significant. current practices must be improved to prevent both
Second, our simulations reveal and highlight Type I and Type II errors. As we describe later, our
the individual roles of instrument strength and review of articles published in SMJ between 2005
instrument endogeneity. Weaker instruments (i.e., and 2012 reveals alarming inconsistencies regard-
instruments weakly correlated with the endoge- ing how strategy researchers approach and remedy
nous variable) result in higher standard errors (i.e., endogeneity. We are hopeful that our simulations
lower efficiency), which decrease the likelihood will help to improve the rigor of the empirical
of reporting statistically significant relationships studies designed to test our theories and enhance
when they exist. In contrast, instrument endo- the overall quality of knowledge in our field that
geneity results in statistically significant coefficient is based on the results of these empirical tests.
estimates that differ from their true values (i.e.,
biased). In fact, the results of our simulations show ENDOGENEITY
that endogenous instruments often produce coeffi-
cient estimates that are inferior to those reported Endogeneity defined
by OLS regression (i.e., the “cure” is worse than While many scholars in strategic management
the “disease”). discuss endogeneity, it is important to clarify the
Third, our simulations demonstrate that stan- precise meaning of the term. Endogeneity is most
dard tests detecting the presence of endogeneity typically described in the context of ordinary least
are highly dependent on the quality of the squares (OLS) regression. Equation 1 represents a
instrumental variables used in the analyses. Our basic OLS regression equation:
results indicate that these tests will rarely detect
existing endogeneity when researchers use weak yi = α + βxi + εi (1)
instruments. These tests perform even more poorly
when the instruments are themselves endogenous. In this equation, y i represents the dependent
Our simulations highlight the ease with which variable, α represents a constant, β represents
researchers (and reviewers and editors) may the coefficient, x i represents the independent
use test results mistakenly to dismiss concerns variable, and εi represents the error term. The
regarding endogeneity. error term in an OLS regression model illustrates
Taken together, our results allow us to extend the extent to which the independent variables
previous recommendations for authors, reviewers, predict the dependent variable and should vary
and editors considering the potential impact of randomly. When the error term is correlated
endogeneity. Given our findings regarding the per- with an independent variable, however, the errors
nicious effects of endogeneity at even low lev- are not random; this leads to biased coefficient
els, our recommendations center around testing estimates (for a review see Kennedy, 2008). Bias
for endogeneity. Bascle (2008) presents a compre- occurs when the coefficient estimate based on a
hensive approach for scholars to implement when sample does not on average equal the true value
dealing with endogeneity. The first step in this of the coefficient in the population (Cohen et al.,
approach involves asking, “Is there an endogeneity 2003: 117). Therefore, a critical assumption of
problem?” (Bascle, 2008: 287). Our simulations OLS regression is that the independent variable
provide a counterintuitive result that researchers and the error term are uncorrelated.
must first identify strong and exogenous instru- According to Kennedy (2008), four different
ments before they can know whether endogeneity issues may potentially introduce endogeneity in
is problematic. Only with such instruments can OLS regression models: errors-in-variables (i.e.,
researchers effectively test for the presence of and measurement error), autoregression, omitted vari-
dismiss concerns regarding endogeneity. ables, and simultaneous causality. In each of these
The relevance of strategy research depends scenarios, OLS regression reports biased coeffi-
on both theory generation and the empirical cients. Instead of estimating the “true” relationship
Copyright  2013 John Wiley & Sons, Ltd. Strat. Mgmt. J., 35: 1070–1079 (2014)
DOI: 10.1002/smj
1072 M. Semadeni, M. C. Withers and S. Trevis Certo

between the independent variable and the depen- use estimators other than least squares, such
dent variable, OLS regression mistakenly includes as generalized method of moments (GMM) or
the correlation between the independent variable limited-information maximum likelihood (LIML).
and the error term in the estimation of the inde- Researchers employ such techniques when the
pendent variable’s coefficient. endogenous variable does not represent a dichoto-
mous decision.
Instrumental variables must fulfill two condi-
Endogeneity in strategy research tions: relevance and exogeneity (Kennedy, 2008).1
Shaver’s (1998) work on international entry modes Relevance refers to the degree to which the instru-
represents perhaps the most influential study of ment corresponds with the endogenous variable.
endogeneity in strategic management. Shaver uti- A literature on instrument strength (i.e., strong
lizes a sample of firms engaging in international vs. weak instruments) examines how relevance
expansion to study whether acquisitions or Green- influences model results, and scholars have cre-
field ventures lead to higher levels of firm per- ated recommendations based on the F-statistics
formance. According to Shaver, managers choose of first-stage regressions to determine instru-
entry modes based on their perceptions of antic- ment relevance (Stock, Wright, and Yogo, 2002).
ipated performance. It is these unobservable per- The general conclusion of this research indicates
ceptions that potentially lead to endogeneity in this that stronger (higher F-statistics)—as opposed to
empirical context. weaker (lower F-statistics)—instruments are better
Shaver’s ideas have caused scholars to con- for two-stage approaches. Complementing instru-
sider the potential endogeneity of a variety of ment relevance, exogeneity refers to the degree
independent variables. Hamilton and Nickerson to which an instrument is uncorrelated with the
(2003: 51) summarize this perspective by stat- disturbance term in the second stage. Testing for
ing: “the field of strategic management is fun- instrument exogeneity allows researchers to reduce
damentally predicated on the idea that manage- the chance that they replace one endogenous inde-
ment’s decisions are endogenous to their expected pendent variable with another (see Bascle, 2008).
performance outcomes—if not, managerial deci- Perhaps the most problematic aspect of instru-
sion making is not strategic; it is superfluous.” mental variable estimation involves identifying
Accordingly, almost all firm-level variables (e.g., suitable instruments. In practice it remains diffi-
R&D spending, acquisitions, etc.) can be con- cult to find variables that correlate strongly with
sidered decisions made by managers to influence the endogenous variable but not with the error
firm outcomes. Shaver’s introduction to endogene- term in the second stage. Instrument relevance
ity involved a dichotomous independent variable: and exogeneity often work against one another. As
the managerial decision to engage in an acqui- instrument strength increases (i.e., the instrument
sition versus a Greenfield venture. Nevertheless, becomes more like the endogenous independent
scholars have also treated continuous variables variable), it is perhaps not surprising that it may
such as employee stock ownership (Wang, He, be related to the error term in the same way as the
and Mahoney, 2009), human capital investments endogenous variable.
(Sirmon and Hitt, 2009), and CEO hubris (Li and
Tang, 2010) as endogenous variables. Instrumental variables in strategy research
To understand better how strategy researchers use
instrumental variables, we reviewed all empirical
OVERCOMING ENDOGENEITY WITH papers appearing in SMJ between 2005 and
INSTRUMENTAL VARIABLES 2012 that incorporated instrumental variables to
analyze continuous endogenous variables. We
Instrumental variables noted the degree to which researchers (1) tested for
Strategy researchers rely on instruments to model endogeneity; (2) used more than one instrument;
continuous endogenous independent variables. The (3) tested for instrument strength; and (4) tested
terms instrumental variable estimation and two-
stage least squares are often used interchange- 1 Exogeneity and endogeneity are opposites (i.e., a variable that
ably, but instrumental variable techniques may is not “ex ogenous” is “end ogenous”).

Copyright  2013 John Wiley & Sons, Ltd. Strat. Mgmt. J., 35: 1070–1079 (2014)
DOI: 10.1002/smj
Research Notes and Commentaries 1073

for instrument exogeneity. To code for tests METHODOLOGY


of endogeneity, we noted whether authors used
either a Hausman or Durbin-Wu-Hausman test. We To understand better the implications of endogene-
also noted whether authors examined instrument ity we provide a series of simulations. First, we
strength using thresholds such as those developed treat endogeneity as a continuous—as opposed to
by Stock et al. (2002). Finally, we noted whether dichotomous—condition and explore the implica-
a formal test was used to detect instrument tions of increasing endogeneity in OLS regres-
exogeneity. sion. Second, we examine how instrument strength
The results of our review reveal an alarm- and exogeneity influence the outcomes of instru-
ing lack of consistency in terms of how strategy mental variables. Finally, we study how instru-
researchers report instrumental variables. We iden- ment strength and exogeneity influence the tests
tified 24 articles that use two-stage least squares to detect endogeneity.
for either the primary or supplementary analyses.
Of these 24 articles, 10 test for endogeneity, 9 use
more than one instrument, 3 test for instrument Simulation design
strength, and 5 test for instrument endogeneity. Kennedy (2008) provides an excellent overview
Finally, two-thirds of these articles do not report of the intuition underlying simulation techniques.
the results corresponding to the first stage of the We used Stata for our simulations, which involved
model. two broad steps. First, we generated a dataset with
It is important to note procedures that we did 500 observations of dependent and independent
not include in our coding scheme. Many strategy variables with known properties. In this study, we
researchers focused primarily on statistical signif- generated y with the following equation:
icance to test whether instruments are relevant
and/or exogenous. Some scholars, for example, y = α + βx + e (2)
suggest that an instrument is relevant if it is sta-
tistically related to the endogenous variable and
exogenous if it is not related to the ultimate depen-
In our simulations, we assigned a value of 1
dent variable. Larcker and Rusticus (2010: 192)
to the intercept (a) and set the value of β to
declare such approaches “completely inappropri-
0.1 to represent a small effect. We then generated
ate,” so we do not count studies adopting these
normally distributed values for e, with a mean of 0.
methods as effective tests for instrument relevance
The primary issue in this simulation involves
or exogeneity.
examining various levels of endogeneity. To sim-
During our review, Hoetker and Mellewigt
ulate this, we generated independent variables (x )
(2009) stood out as providing one of the most com-
prehensive descriptions of the procedures used to that varied in terms of the correlation with the error
implement instrumental variables.2 Nevertheless, term (e). We created three categories of endogene-
the authors did not note that the F-tests regard- ity: no endogeneity (corr[x ,e] = 0), low endogene-
ing instrument strength revealed weak instruments ity (corr[x ,e] = 0.1), and medium endogeneity
(the corresponding F-test value of 3.29 was well (corr[x ,e] = 0.3).3 To do so we followed Lar-
below Stock et al.’s (2002) recommended value cker and Rusticus (2010: 193), rather than directly
of 11.59 for two instruments). The fact that one choosing the parameters, we set the population
of the most comprehensive and transparent uses correlations and then calculate the parameters nec-
of instrumental variables in strategy identified essary to obtain the desired correlations, allowing
weak instruments but still proceeded to report the for a more natural interpretation.
results should give scholars pause. In the follow-
ing sections, we use simulations to highlight the Analytical models
importance of understanding instrument strength
and exogeneity when using instrumental variables. We compared the effectiveness of two main ana-
lytic strategies: OLS regression and instrumental
2
The authors used multiple instruments, test for endogeneity
(using the Durbin-Wu-Hausman test), and reference both
3 In supplementary analyses, higher levels of endogeneity
instrument strength (using an F-test) and exogeniety (using
Hansen’s J test [1982]). resulted in even more dramatic results.

Copyright  2013 John Wiley & Sons, Ltd. Strat. Mgmt. J., 35: 1070–1079 (2014)
DOI: 10.1002/smj
1074 M. Semadeni, M. C. Withers and S. Trevis Certo

variables.4 We used Stata’s “regress” command to generate the data for each simulation) value. An
invoke OLS regression, which we refer to as OLS unbiased estimator will report confidence intervals
when discussing the results. We then used Stata’s that include the true value in approximately
“ivreg” command to invoke a two-stage least 95 percent of the cases. PercSig refers to the
squares approach.5 With this approach, researchers percentage of simulations that report statistically
must specify two stages. The first stage involves significant coefficients. This measure assesses an
using an instrumental variable, z , to determine the estimator’s power (e.g., Cohen, 1992). Estimators
endogenous independent variable, x . The second reporting PercSig levels that exceed their “true”
stage then uses the predicted value from the first values (based on power calculations) indicate
stage as an independent variable in the second biased coefficients and/or standard errors.7
stage. The intuition behind this approach is that Examining 95int and PercSig together allows
the first stage “partials out” any common variance us to uncover interesting relationships. Although
between x and z , so the predicted value does not it is impossible to know a “true” value, fundamen-
share any variance that is related to the error term tally researchers should be interested in approxi-
in the second stage. We vary the strength of the mating the coefficient’s true value (95int). Practi-
instrument by modifying the correlation between cally speaking, however, researchers are also inter-
x and z . We report results for two alternative ested in identifying statistically significant rela-
instrumental variable approaches: (1) IVweak refers tionships (Bettis, 2012). PercSig reveals the condi-
to a two-stage approach using weak instruments tions under which endogeneity increases the prob-
(i.e., the correlation between x and z is 0.1), and ability of finding significance while decreasing the
(2) IVmod refers to a two-stage approach using probability of finding the “true” relationship.
moderate instruments (i.e., the correlation between
x and z is 0.33).6
RESULTS
Outcome measures Endogeneity and OLS
We examined how endogeneity influences the bias Table 1 illustrates the results of our simulations
and efficiency of OLS regression by creating four comparing OLS, IVWeak , and IVMod . This table
different outcome measures. We simulated 1,000 includes three different endogeneity conditions: no
iterations of each condition. For each condition, we endogeneity, low endogeneity, and medium endo-
saved the estimated β and the estimated standard geneity. For each endogeneity condition, Table
error. Beta Med represents the median estimate of 1 displays the four outcome measures associated
β for the 1,000 iterations in our simulations. with each of the estimators. We use two panels
SEMed represents the median standard error for to contrast between simulations with a true effect
the 1,000 iterations in our simulation. Values of (true β = 0.1 in the top panel) and without a
these measures that exceed (are less than) their true effect (true β = 0 in the bottom panel). Col-
true values suggest positive (negative) bias. umn 1 in Table 1’s top panel illustrates the effects
We also included two outcome measures that of endogeneity on OLS when the true value of
incorporated both the reported betas and standard B should equal 0.1. When there exists no endo-
errors to examine the significance of the results geneity, OLS reported an unbiased Beta Med of
reported by each estimator. The expression 95int 0.10. This unbiased beta, coupled with the reported
denotes the extent to which an estimator reports SE Med of 0.045 resulted in a 95int value of 95
a 95 percent confidence interval that includes the percent and PercSig of 60 percent.8
coefficient’s “true” (i.e., the coefficient used to When endogeneity was low, the reported beta
for OLS was twice its true value and increased
4
In supplementary analyses, we generated panel data and to 0.43 for moderate levels of endogeneity.
examined the effects of endogeneity on fixed- and random-
effects models. Our results were substantively similar to those
reported by OLS regression. 7 The regression simulations with no endogeneity provide the
5 In supplementary analyses, results for Stata’s GMM and LIML “true” values that are used for comparison purposes.
options for ivreg were virtually identical. 8
The reported power of 60 percent is consistent with statistical
6 The effect sizes we examined are consistent with standards power calculators using a sample size of 500 and an effect size
used in other disciplines (e.g., Stock et al. 2002). of 0.10.

Copyright  2013 John Wiley & Sons, Ltd. Strat. Mgmt. J., 35: 1070–1079 (2014)
DOI: 10.1002/smj
Research Notes and Commentaries 1075
Table 1. Main findings OLS vs. single instrument

Weak and Moderate and


OLS Weak inst. Moderate inst. endo instr. endo instr.

PANEL A: Sample size: 500, True B : 0.1


No endogeneity
Beta 0.100 0.088 0.100 1.095 0.399
SE 0.045 0.469 0.138 0.633 0.141
95% interval 95% 100% 96% 84% 43%
% significant 60% 1% 9% 26% 84%
Low endogeneity (0.1)
Beta 0.200 0.076 0.107 1.073 0.406
SE 0.045 0.478 0.135 0.598 0.139
95% interval 39% 100% 96% 78% 40%
% significant 69% 1% 12% 34% 86%
Medium endogeneity (0.3)
Beta 0.430 0.097 0.092 1.142 0.404
SE 0.042 0.475 0.136 0.556 0.129
95% interval 0% 100% 95% 66% 36%
% significant 100% 4% 13% 45% 86%
PANEL B: Sample size: 500, True B : 0
No endogeneity
Beta 0.001 −0.009 −0.004 0.940 0.335
SE 0.045 0.474 0.137 0.610 0.157
95% interval 95% 100% 95% 84% 44%
% significant 5% 1% 5% 16% 56%
Low endogeneity (0.1)
Beta 0.100 0.051 −0.009 1.016 0.336
SE 0.044 0.489 0.136 0.608 0.154
95% interval 39% 100% 95% 77% 40%
% significant 61% 0% 5% 23% 60%
Medium endogeneity (0.3)
Beta 0.330 −0.001 −0.007 0.987 0.338
SE 0.042 0.466 0.135 0.536 0.143
95% interval 0% 98% 95% 63% 35%
% significant 100% 2% 5% 37% 65%

Complementing the biased betas, Table 1’s top that endogeneity dramatically increases the extent
panel shows that endogeneity also biases the to which the reported coefficient is statistically
standard errors reported by OLS regression, significant, but this value is much less likely to
but this bias is opposite of that associated with be the true value.
coefficients. Although the standard errors remain Table 1’s bottom panel displays the same
constant for low levels of endogeneity, the information for simulations in which β should
reported standard errors are actually lower for equal 0 (i.e., there is no true effect). The results
moderate levels of endogeneity. As shown in illustrate that when endogeneity is low, OLS
Table 1, 95int decreases as endogeneity increases; will report a positive and statistically significant
relationship in 61 percent of the cases. When
the positively biased coefficient coupled with the
endogeneity increases to 0.33, OLS reports a
increasingly narrower confidence interval suggests
positive and statistically significant relationship in
the reported confidence interval for OLS will be 100 percent of the cases. These results illustrate
less likely to include the true coefficient value as how endogeneity leads researchers to report results
endogeneity increases. At the same time, Table 1 that support relationships that do not exist.9
shows that OLS is more likely to report a
statistically significant coefficient as endogeneity 9
We also ran the simulations modeling a negative relationship
increases. Taken together, these results indicate (i.e., B = −0.1). In these simulations, OLS was able to detect
Copyright  2013 John Wiley & Sons, Ltd. Strat. Mgmt. J., 35: 1070–1079 (2014)
DOI: 10.1002/smj
1076 M. Semadeni, M. C. Withers and S. Trevis Certo

Endogeneity and instrumental variables Testing for endogeneity


Columns two and three of Table 1 display Addressing the deleterious effects of endogeneity
the results of weak and moderate instrumental begins with testing for its presence, since endo-
variables, respectively. When endogeneity is zero, geneity remediation in its absence yields less effi-
IVWeak and IVMod both produce Beta Med values cient estimates. The Hausman and Durbin-Wu-
that are close to the true value. In contrast, the Hausman (DWH) tests both examine whether the
SE Med for both instrumental variable techniques independent variable of interest is in fact endoge-
are dramatically higher than the corresponding nous. The quality of these tests, however, depends
SE Med values for OLS, but Table 1 also shows that on the appropriateness of the instruments. As Lar-
SEMed decreases as instrument strength increases. cker and Rusticus (2010: 191) suggest, such tests
This combination results in a PerSig value of are valid “[u]nder the assumption of the appropri-
less than 10 percent, underscoring the efficiency ateness of the instruments.”
problems associated with instrumental variables. Panel A of Table 2 illustrates the effectiveness of
When endogeneity is at low or medium levels, the DWH test as instrument strength and number
similar patterns remain. of observations vary.11 With weak instruments, the
Like OLS, this combination of Beta Med and DWH test identifies endogeneity in less than 20
SE Med influence outcome measures associated percent of cases. With moderate instruments the
with confidence intervals, but this combination DWH is more effective, but it still remains difficult
presents the exact opposite effects. Columns 2 and to detect low levels of endogeneity. In contrast,
3 in Table 1 show that 95int is essentially greater Panel B of Table 2 illustrates that instrument
than or equal to 95 in all cases; these results are endogeneity has the opposite effect. Endogenous
rarely statistically significant. This pattern directly instruments cause the test to report endogeneity
contradicts OLS. For instrumental variables, the even when it is not present. Taken together,
unbiased betas are associated with such large stan- weak instruments provide results suggesting that
dard errors and confidence intervals that the true endogeneity is not present (even when it is), and
value is almost always included, but these large endogenous instruments provide results suggesting
confidence intervals almost always include zero. endogeneity is present (even when it is not).
Complementing instrument strength, we also
create simulations to investigate how instrument
endogeneity influences the results of instrumental DISCUSSION
variable analysis. We create simulations to exam-
ine the outcomes at low levels of instrumental vari- Contributions
able endogeneity (i.e., the correlation between z
and e was 0.10).10 As illustrated in Columns 4 and Strategy scholars are concerned about the effects
5 of Table 1, instrument endogeneity substantively of endogeneity, and our results suggest that such
biases Beta Med for both weak and medium instru- concerns are warranted. In full disclosure, as
ments. As compared to exogenous instruments, researchers we were hopeful to present research
Beta Med is nearly 1,000 percent larger for weak suggesting that concerns about endogeneity were
instruments and 300 percent larger for moderate perhaps overstated. Our simulations suggest, how-
instruments for low levels of instrument endo- ever, that even low levels of endogeneity can bias
geneity. Interestingly, the effects of these changes reported coefficient estimates by as much as 100
are worse for moderate, as compared to weak, percent. And when the hypothesized true relation-
instruments in terms of 95int. Nevertheless, both ship is negative, low levels of endogeneity can cre-
types of endogenous instruments produce coeffi- ate positive, negative, or no relationships.12 Con-
cient estimates that are far more biased than those sequently, it is difficult for us to conclude that
reported by OLS. concerns about endogeneity are excessive.

a statistically significant relationship in only 5 (0) percent of 11 We also ran the same simulations for the Hausman test, and
the cases with low (moderate) levels of endogeneity. When the
direction of endogeneity is opposite of the coefficient, the erratic the results were substantively similar.
12
results of OLS are much like those related to suppression. In other words, when empirical results are significant and in
10 This approach was conservative, as higher levels of endogene- the opposite direction from that suggested by well-established
ity produced dramatically more biased results. theory, negative endogeneity may be the culprit.

Copyright  2013 John Wiley & Sons, Ltd. Strat. Mgmt. J., 35: 1070–1079 (2014)
DOI: 10.1002/smj
Research Notes and Commentaries 1077
Table 2. Durbin-Wu-Hausman endogeneity test resultsa be driven by endogeneity as opposed to the theo-
rized relationships. Our results also highlight how
Observations even low levels of instrument endogeneity increase
100 (%) 500 (%) 1000 (%) reported betas by nearly 1,000 percent—a bias that
is worse than even OLS regression. While instru-
Panel A ment endogeneity increases the likelihood of statis-
No instrument endogeneity tically significant coefficients, it also decreases the
Instrument strength zero likelihood that the confidence intervals surround-
No endogeneity 4.9 5.0 5.2
Low endogeneity 5.0 4.8 5.0
ing these estimates include the true value.
Moderate endogeneity 5.4 5.4 5.5
Instrument strength weak Instrumental variables and type II errors
No endogeneity 5.3 4.7 4.6
Low endogeneity 5.6 5.6 5.9 Our simulations also provide a number of results
Moderate endogeneity 6.7 12.4 19.1 regarding instrumental variables that require dis-
Instrument strength moderate cussion. First, our simulations demonstrate that
No endogeneity 4.6 5.5 4.5 both weak and moderate instruments provide coef-
Low endogeneity 7.5 13.1 20.0
Moderate endogeneity 21.5 78.5 97.4 ficient estimates that closely approximate their true
values, but the associated standard errors greatly
Panel B exceed those of OLS. As a result, instrumental
Low instrument endogeneity variable techniques provide unbiased coefficient
Instrument strength zero estimates but are associated with extremely low
No endogeneity 17.6 60.9 87.6
Low endogeneity 17.5 60.7 88.5 levels of statistical power. Directly contrasting our
Moderate endogeneity 18.2 66.7 92.8 results regarding OLS, instrumental variables are
Instrument strength weak likely to produce confidence intervals that contain
No endogeneity 16.1 61.6 89.7 the true value of beta, but these estimates are rarely
Low endogeneity 14.1 53.3 81.4 statistically significant.
Moderate endogeneity 10.3 35.2 60.9
Instrument strength moderate
No endogeneity 19.6 66.3 92.1 Testing for endogeneity
Low endogeneity 10.2 36.9 61.7
Moderate endogeneity 6.1 6.3 5.9 Finally, our simulations suggest that the effec-
tiveness of endogeneity tests depends on instru-
a
Values in the cells denote the percentage of the simulated DWH ment quality. In other words, weak and/or endoge-
tests that were significant (i.e., finding endogeneity present). nous instruments yield suspect results whereas
stronger, exogenous instruments reveal endogene-
OLS and type I errors ity. Accordingly, tests that rely on weak and/or
endogenous instruments may mislead authors,
The simulations reveal two important results asso- reviewers, editors, and general readers.
ciated with endogeneity in the context of OLS.
First, OLS coefficients become more biased (i.e.,
Recommendations
they become increasing larger than their true val-
ues) as endogeneity increases. Second, the stan- The results of our simulations allow us to provide
dard errors reported by OLS become smaller (i.e., a number of recommendations to researchers
they become increasingly smaller than their true confronting endogeneity. Our results indicate that
values) as endogeneity increases. It is not sur- stronger instrumental variables result in more
prising, then, that as endogeneity increases the 95 accurate betas as compared to OLS, and the benefit
percent confidence intervals reported by OLS are of stronger instruments involves reduced standard
less likely to include the true value of the coeffi- errors. Because of the importance of instrument
cient. At the same time, as endogeneity increases, relevance, strategy researchers should start by
OLS becomes more likely to report statistically identifying at least moderately strong instruments.
significant results. Endogeneity, then, may make Authors should always report instrument strength
it easier for researchers to find statistically sig- by noting the F-statistic in the first stage associated
nificant relationships, but such significance may with the addition of the instrumental variable(s)
Copyright  2013 John Wiley & Sons, Ltd. Strat. Mgmt. J., 35: 1070–1079 (2014)
DOI: 10.1002/smj
1078 M. Semadeni, M. C. Withers and S. Trevis Certo

(e.g., Larcker and Rusticus, 2010; Stock et al., instruments, these tests will almost surely fail to
2002). Without this detail, it is difficult for readers report endogeneity—even when it exists. In our
to understand the influence of the instruments in view, approaches such as “We tested for endogene-
the models. Consistent with the recommendations ity and found that it was not an issue” are not
of econometricians (Angrist and Pischke, 2009; acceptable. Instead, reviewers, editors, and ulti-
Kennedy, 2008), researchers should report the full mately readers need to understand clearly the prop-
results for the first-stage models, and these models erties of the instruments used in endogeneity tests.
should include the controls used in the second-
stage (i.e., structural model).
Finding instrumental variables
After identifying moderately strong instruments,
researchers should also examine the potential Researchers interested in endogeneity understand
endogeneity of the instruments. Compared to that the most difficult aspect of instrumental vari-
exogenous instruments, endogenous instruments able analysis involves identifying suitable instru-
dramatically raise the likelihood of reporting a sta- ments. Nevertheless, economists may have some
tistically significant result. Given such enormous suggestions in this regard. Angrist and Pischke
bias, we recommend testing for instrument endo- (2009: 117) suggest that, “Good instruments come
geneity using the Sargan (1958) test or Hansen’s from a combination of institutional knowledge and
(1982) J-statistic (for other tests of instrument ideas about the processes determining the vari-
endogeneity, see Bascle, 2008). able of interest.” Because researchers in strategy
It should be noted that multiple instruments address a wide array of topics and contexts, it
are required to test for instrument endogeneity is difficult to provide advice that applies equally
(e.g., Kennedy, 2008). In supplementary analyses to all researchers. We propose that researchers in
not reported, our simulations also show that strategy can learn from the insights and examples
multiple instruments help to decrease standard offered by economists, who have been dealing with
errors. For these reasons, we cannot overempha- endogeneity for far longer. Kennedy (2008) sum-
size the importance of identifying multiple—as marizes a number of creative ideas in this regard.
opposed to single—instruments.13 Nevertheless, Distance from a college, for instance, can serve as
our results should give strategy researchers an instrument for years of education as a predictor
pause regarding the number of existing studies of wages. More generally, he also discusses how
in strategic management that have advanced researchers may employ lagged exogenous vari-
theory while using only a single—and potentially ables as instruments for endogenous predictors.
endogenous—instrument. If authors can only find In our view, though, the key to remedying endo-
one instrument of sufficient strength, they must geneity is to consider its effects before submit-
present compelling theoretical evidence that the ting manuscripts for review. If reviewers highlight
instrument is not itself endogenous. In this regard, endogeneity after the original submission, it will
authors should keep in mind that the instrumental likely prove difficult for authors to identify suitable
variable should be uncorrelated with the residuals instruments that fit neatly within existing models.
associated with the dependent variable—and not For researchers using archival data, for instance,
the dependent variable itself. it may prove difficult to identify a natural exper-
Finally, but perhaps most importantly, authors iment after the fact that helps to reframe a study
should compare their results to OLS (or a related examining the effects of a strategic decision (e.g.,
procedure such as logistic, fixed-effects regression, market entry) on firm performance. Correcting for
etc.) and test for endogeneity. We emphasize this is endogeneity in a post-hoc fashion is even more
the final step in the process, because quality (i.e., problematic for researchers relying on survey data.
relevant and exogenous) instruments are required
for the Hausman or DWH tests. Without quality
CONCLUSION
13 Furthermore, it is important to note that multiple endogenous
In sum, given noted publication biases that focus
instruments will not be able to detect instrument endogeneity. on statistical significance (Bettis, 2012), we can
Supplemental analyses found that detection requires at least one
moderately strong exogenous instrument among the multiple speculate that endogeneity has led to a myriad
instruments. of type I errors among published papers in
Copyright  2013 John Wiley & Sons, Ltd. Strat. Mgmt. J., 35: 1070–1079 (2014)
DOI: 10.1002/smj
Research Notes and Commentaries 1079

strategy. While it is not our purpose to highlight Hamilton BH, Nickerson JA. 2003. Correcting for endo-
specific cases, we can safely presume that even geneity in strategic management research. Strategic
Organization 1(1): 51–78.
small amounts of endogeneity have resulted in Hansen LP. 1982. Large sample properties of generalized
a number of published papers that have resulted method of moments estimators. Econometrica 50(4):
in statistically significant results driven not by 1029–1054.
the purported independent variables but instead Hoetker G, Mellewigt T. 2009. Choice and performance
by endogeneity. We are hopeful that the results of governance mechanisms: matching alliance gover-
nance to asset type. Strategic Management Journal
and recommendations we provide are helpful to 30(10): 1025–1044.
authors, reviewers and editors. Kennedy P. 2008. A Guide to Econometrics (2nd edn).
Blackwell: Oxford, UK.
Larcker DF, Rusticus TO. 2010. On the use of instru-
ACKNOWLEDGEMENTS mental variables in accounting research. Journal of
Accounting and Economics 49(3): 186–205.
Li J, Tang Y. 2010. CEO hubris and firm risk taking in
We thank Brian Connelly, Ryan Krause, Don China: the moderating role of managerial discretion.
Lange and Phil Podsakoff for their constructive Academy of Management Journal 53(1): 45–68.
comments on earlier versions of this manuscript. Sargan JD. 1958. The estimation of economic rela-
We also thank Margarethe Wiersema and two tionships using instrumental variables. Econometrica
anonymous reviewers for their guidance in shaping 26(3): 393–415.
Shaver JM. 1998. Accounting for endogeneity when
the final manuscript. assessing strategy performance: does entry mode
choice affect fdi survival? Management Science 44(4):
571–585.
REFERENCES Shugan SM. 2004. Editorial: endogeneity in marketing
decision models. Marketing Science 23(1): 1–3.
Angrist JD, Pischke JS. 2009. Mostly Harmless Econo- Sirmon DG, Hitt MA. 2009. Contingencies within
metrics: an Empiricist’s Companion. Princeton Uni- dynamic managerial capabilities: interdependent
versity Press: Princeton, NJ. effects of resource investment and deployment on
Bascle G. 2008. Controlling for endogeneity with instru- firm performance. Strategic Management Journal
mental variables in strategic management research. 30(13): 1375–1394.
Strategic Organization 6(3): 285–327. Stock JH, Wright JH, Yogo M. 2002. A survey of weak
Bettis RA. 2012. The search for asterisks: compromised instruments and weak identification in generalized
statistical tests and flawed theories. Strategic Manage- method of moments. Journal of Business & Economic
ment Journal 33(1): 108–113. Statistics 20(4): 518–529.
Cohen J. 1992. A power primer. Psychological Bulletin Wang HC, He J, Mahoney JT. 2009. Firm-specific knowl-
112(1): 155–159. edge resources and competitive advantage: the roles
Cohen J, Cohen P, West SG, Aiken LS. 2003. Applied of economic- and relationship-based employee gov-
Multiple Regression/Correlation Analysis for the ernance mechanisms. Strategic Management Journal
Behavioral Sciences. Erlbaum: Mahwah, NJ. 30(12): 1265–1285.

Copyright  2013 John Wiley & Sons, Ltd. Strat. Mgmt. J., 35: 1070–1079 (2014)
DOI: 10.1002/smj

You might also like