3 - PodsakoffARP2012

PS63CH21-Podsakoff ARI 31 October 2011 13:29
ANNUAL
REVIEWS Further
Sources of Method Bias
Click here for quick links to
Annual Reviews content online,
including:
in Social Science Research
Other articles in this volume
Top cited articles
and Recommendations on
Top downloaded articles
How to Control It
Annu. Rev. Psychol. 2012.63:539-569. Downloaded from www.annualreviews.org
Our comprehensive search
Philip M. Podsakoff,1 Scott B. MacKenzie,2

by Harvard University on 04/25/12. For personal use only.
and Nathan P. Podsakoff 3

1
Department of Management, Kelley School of Business, Indiana University, Bloomington,
Indiana 47405; email: podsakof@indiana.edu
2
Department of Marketing, Kelley School of Business, Indiana University, Bloomington,
Indiana 47405; email: mackenz@indiana.edu
3
Department of Management and Organizations, Eller College of Management,
University of Arizona, Tucson, Arizona 85721; email: podsakof@email.arizona.edu
Annu. Rev. Psychol. 2012. 63:53969 Keywords

First published online as a Review in Advance on common method variance, response style biases, marker variable
August 11, 2011
technique, instrumental variable technique, unmeasured latent
The Annual Review of Psychology is online at variable technique
psych.annualreviews.org
This articles doi: Abstract

10.1146/annurev-psych-120710-100452
Despite the concern that has been expressed about potential method
Copyright c 2012 by Annual Reviews. biases, and the pervasiveness of research settings with the potential to
All rights reserved
produce them, there is disagreement about whether they really are a
0066-4308/12/0110-0539$20.00 problem for researchers in the behavioral sciences. Therefore, the pur-
pose of this review is to explore the current state of knowledge about
method biases. First, we explore the meaning of the terms method
and method bias and then we examine whether method biases inu-
ence all measures equally. Next, we review the evidence of the effects
that method biases have on individual measures and on the covariation
between different constructs. Following this, we evaluate the procedural
and statistical remedies that have been used to control method biases
and provide recommendations for minimizing method bias.
539
traits (constructs). The test, or rating scale,

Contents or other device, almost inevitably elicits sys-
tematic variance due to both groups of fea-
INTRODUCTION . . . . . . . . . . . . . . . . . . 540
tures. To the extent that irrelevant method
WHAT IS METHOD BIAS
variance contributes to the scores obtained,
AND WHY IS IT A PROBLEM? . . 541
these scores are invalid. (Campbell & Fiske
What Is a Method? . . . . . . . . . . . . . . . . 541
1959, p. 84; words in parentheses added by
What Is Method Bias?. . . . . . . . . . . . . . 542
present authors)
Does Method Bias Affect All
Measures Equally? . . . . . . . . . . . . . . 543
In the years since, a number of researchers
EMPIRICAL EVIDENCE
have discussed a related problemthe biasing
OF THE EFFECTS OF
effects that measuring two or more constructs
METHOD BIASES . . . . . . . . . . . . . . . 543
with the same method may have on estimates
Effects of General Method Bias on
of the relationships between them. The ma-
Item Reliability or Validity . . . . . . 543

jor concern with measuring different constructs
Effects of General Method Bias on
with the same method is the danger that at least
the Covariation Between
some of the observed covariation between them
Constructs . . . . . . . . . . . . . . . . . . . . . . 544
may be due to the fact that they share the same
Estimates of Specic Types of
method of measurement.
Method Bias on the Covariation
This concern with method bias is potentially
Between Constructs . . . . . . . . . . . . . 545
important because the situations in which it is
WAYS TO CONTROL
likely to be a problem are quite common. For
FOR DIFFERENT SOURCES
example, Bodner (2006) reviewed the literature
OF METHOD BIAS . . . . . . . . . . . . . . 548
in six areas of psychology and found that most
Procedural Remedies . . . . . . . . . . . . . . . 548
studies (76%) involved only a single measure-
Statistical Remedies . . . . . . . . . . . . . . . . 553
ment method, and of the studies that involved
RECOMMENDATIONS . . . . . . . . . . . . 559
human subjects and adequately explained the
When Is Method Bias Likely
measurement procedures, 33% involved self-
To Be a Problem? . . . . . . . . . . . . . . 559
report questionnaires as the sole measurement
What Can Be Done To Mitigate
method. Similarly, Woszczynski & Whitman
the Problem? . . . . . . . . . . . . . . . . . . . 562
(2004) reviewed the studies reported in the
CONCLUSION . . . . . . . . . . . . . . . . . . . . . 565
top management information systems journals
from 1996 to 2000 and found that 27% of the
428 articles published in this literature during
this time period used a survey with self-reports
INTRODUCTION as the predominant method of data gathering.
Over 50 years ago, Campbell & Fiske (1959) Unfortunately, despite the concern that has
voiced their concerns about the biasing effects been expressed about method bias, and the per-
that methods of measurement may have on the vasiveness of research settings with the poten-
validity of measures: tial to produce it, there is little agreement about
whether it really is a problem for researchers.
In any given psychological measuring device, For example, although many authors believe
there are certain features or stimuli introduced that method bias is an important problem that
specically to represent the trait (construct) needs to be controlled (e.g., Campbell & Fiske
that it is intended to measure. There are other 1959; Cote & Buckley 1987, 1988; Doty &
features which are characteristic of the method Glick 1998; Podsakoff et al. 2003; Podsakoff
being employed, features which could also be & Organ 1986; Sharma et al. 2009; Williams
present in efforts to measure quite different & Anderson 1994; Williams et al. 1989, 2010),
540 Podsakoff MacKenzie Podsakoff

some claim that it is a myth or urban leg- of our knowledge about method biases in the
end (e.g., Chen & Spector 1991; Spector 1987, behavioral sciences.
2006; Spector & Brannick 2009).
Within the context of the above discussion,
the purpose of this review is to explore the cur- WHAT IS METHOD BIAS
rent state of knowledge about method biases. AND WHY IS IT A PROBLEM?
The rst issue we explore is the question of
What Is a Method?
what is method bias. It is obvious from reading
the literature that scholars have different It is obvious from reading the literature that
interpretations of what is meant by this term there are differences in how scholars dene the
(e.g., Campbell & Fiske 1959, Edwards 2008, term method. The term has traditionally been
Lance et al. 2009, Messick 1991). In addition, dened broadly to include several key aspects of
there is a difference of opinion about what con- the measurement process (Campbell & Fiske
stitutes a bias (Cote & Buckley 1987, Spector 1959, Fiske 1982). For example, according to
& Brannick 2009, Williams et al. 1989). Is it Fiske (1982, p. 82),

the effect of method factors on the validity
and reliability of individual measures, the the term method encompasses potential inu-
covariation between measures of different con- ences at several levels of abstraction. Taking
structs, or both? Finally, scholars also disagree a paper-and-pencil instrument as an example,
about whether method bias affects all measures these inuences include the content of the
equally or some measures more than others items, the response format, the general in-
(Lindell & Whitney 2001, Williams et al. structions and other features of the test-task
2010). Differences in these assumptions about as a whole, the characteristics of the exam-
the nature of method bias inuence the way in iner, other features of the total setting, and
which researchers try to control for it and the the reason why the subject is taking the test.
conclusions they reach regarding its effects. Two units that have any one of these elements
The second issue we address is what the em- in common can show convergence due to that
pirical evidence indicates about the extent to source, so the relationship obtained between
which method bias is a problem in behavioral them cannot safely be interpreted as associated
research. For the purposes of this analysis, we with the traits or constructs in those units. For
examine the evidence of the effects of method any single investigation, the only certain pro-
factors on the reliability and validity of individ- tection against this threat to validity is units
ual measures and the effects of method factors using completely independent methods.
on the covariation between measures of differ-
ent constructs. This is consistent with the views of most
The third issue we discuss is how researchers researchers (Bagozzi 1984, Baumgartner &
can control method biases. We start by review- Steenkamp 2001, Johnson et al. 2011, Messick
ing the literature on the procedural and statisti- 1991, Podsakoff et al. 2003, Siemsen et al.
cal remedies that researchers commonly use to 2010, Weijters et al. 2010c), including
control method biases, and we then discuss the Edwards (2008, p. 476), who argues that
strengths and limitations of each of these reme- method biases arise from response tendencies
dies for dealing with specic types of method that raters apply across measures, similarities in
biases. item structure or wording that induce similar
In the nal section, we address two related responses, the proximity of items in an instru-
issues: (a) when method biases are likely to be ment, and similarities in the medium, timing,
a major problem in a study and (b) what re- or location in which measures are collected.
searchers can do to mitigate their effects. We However, others have argued for a narrower
then conclude with a brief summary of the state denition (e.g., Lance et al. 2009, Sechrest
www.annualreviews.org Method Bias in Research 541

et al. 2000). For example, Lance et al. (2009, effect is that method factors can bias estimates
p. 351) argue that the term method should of construct reliability and validity (e.g.,
be restricted to those measurement facets that Bagozzi 1984, Baumgartner & Steenkamp
represent alternative approaches to assigning 2001, Cote & Buckley 1987, Williams et al.
numbers to observations to represent [an indi- 2010). A latent construct captures systematic
viduals] standing on latent constructs. Based variance among its measures. If systematic
on this denition, Lance et al. (2009, 2010) in- method variance is not controlled, this variance
clude similarities in item content, structure, or will be lumped together with systematic trait
format that induce similar responses and explic- variance in the construct. This is a problem be-
itly exclude effects due to response tendencies cause it can lead to erroneous perceptions about
that raters apply across measures, occasions of the adequacy of a scales reliability and con-
measurement, and different situations in which vergent validity (Baumgartner & Steenkamp
measurement may occur. In addition, based on 2001, Lance in Brannick et al. 2010, Williams
this denition we presume this also excludes et al. 2010), and it can lead to underestimates
item proximity and item order effects because of corrected correlations in meta-analyses be-
these are not alternative approaches to assign- cause the reliability estimates will be articially
ing numbers to observations. inated due to method variance (Le et al. 2009).
For our part, we prefer the broader In addition, Bollen (1989) demonstrated
denition of method because regardless of that in multiple regression models, uncon-
whether one considers various rater response trolled systematic or random measurement er-
styles, item characteristics, and aspects of the ror in a predictor can also bias estimates of
measurement context to be method factors, the effects of other error-free predictors on a
they are all sources of systematic measurement criterion variable even if this systematic mea-
error that threaten the validity of a studys surement error is not shared with the criterion
ndings. Indeed, if they are ignored they can variable or with any of the other predictors.
threaten construct validity, distort the dimen- The direction of the bias will depend on the
sional structure of psychological domains, and magnitude and sign of the relationships (a) be-
obscure relationships between constructs/traits tween the imperfect predictor and the criterion
(Messick 1991). Therefore, in the remainder variable and (b) between the imperfect predic-
of this review we adopt this broader concep- tor and the other predictors. Thus, although
tualization of the term method. In so doing we it is true that systematic measurement error in
acknowledge Campbell & Fiskes (1959, p. 85) a predictor that is not shared with a criterion
observation that, The distinction between trait variable will tend to attenuate estimates of the
and method is of course relative to the test con- effect of the predictor on the criterion variable
structors intent. What is an unwanted response (Spector & Brannick 2009), it can also bias esti-
set for one tester may be a trait for another who mates of the effects of other correlated predic-
wishes to measure acquiescence, willingness to tors on the criterion variable.
take an extreme stand, or tendency to attribute The second important detrimental effect of
socially desirable attributes to oneself. uncontrolled method factors is that it can bias
parameter estimates of the relationship between
two different constructs. Several researchers
What Is Method Bias? (e.g., Baumgartner & Steenkamp 2001, Cote
There is also disagreement about what consti- & Buckley 1988, Podsakoff et al. 2003,
tutes a bias. Two detrimental effects produced Siemsen et al. 2010) have demonstrated that
by method factors have been recognized in the method bias can inate, deate, or have no
literature (e.g., Cote & Buckley 1987, 1988; effect on estimates of the relationship between
Doty & Glick 1998; Podsakoff et al. 2003; two constructs. Depending upon whether the
Williams et al. 2010). The rst detrimental method bias inates or deates the relationship,

this is a serious problem because it can cases and the underestimation of them in oth-
(a) affect hypothesis tests and lead to type I or ers. Empirical tests of whether method factor
type II errors, (b) lead to incorrect perceptions loadings are equal or unequal have generally
about how much variance is accounted for found support for the assumption of unequal
in a criterion construct, and (c) enhance or effects of method bias (Rafferty & Grifn 2004,
diminish the nomological or discriminant 2006; Williams et al. 2010). Similarly, Baum-
validity of a scale. Note, however, that Siemsen gartner & Steenkamp (2001) found that the
et al. (2010) and Evans (1985) have shown proportion of variance in measures of different
that although interaction and quadratic effects types of constructs that is attributable to spe-
can be severely deated by method bias, they cic response styles ranged from 0% to 29%.
cannot be artifacts of it. Finally, Cote & Buckleys (1987) meta-analytic
It is for these reasons that, even though estimates of the proportion of method variance
Spector & Brannick (2009, p. 348) argue that in measures of different types of constructs
the effects of method factors on item validity ranged from 22% to 41%. Thus, the weight
and reliability are unimportant because they of the evidence suggests that method factors
do not speak to the issue of CMV and how it are likely to have unequal effects on different
might inate correlations, the overwhelming measureswhether they are different measures
consensus among researchers is that both of the same construct (as in Rafferty & Grifn
forms of bias are important and should be con- 2004, 2006; Williams et al. 2010) or measures
trolled whenever possible (Bagozzi & Yi 1990, of different constructs (as in Baumgartner &
Baumgartner & Steenkamp 2001, Cote & Steenkamp 2001, Cote & Buckley 1987).
Buckley 1987, Doty & Glick 1998, Podsakoff
et al. 2003, Siemsen et al. 2010, Williams et al.
2010). EMPIRICAL EVIDENCE OF THE
EFFECTS OF METHOD BIASES
Does Method Bias Affect Effects of General Method Bias

All Measures Equally? on Item Reliability or Validity
Finally, there is also some disagreement about Evidence of the impact of method biases on
how method bias affects the measures in a item validity and reliability comes from a num-
given study. Some researchers assume that if a ber of meta-analyses of the results of conrma-
method factor has any effect, it affects all mea- tory factor analyses of multi-trait multi-method
sures equally. For example, researchers who use (MTMM) matrices (e.g., Buckley et al. 1990,
the correlational marker variable technique for Cote & Buckley 1987, Doty & Glick 1998,
controlling method bias (see table 1 in Williams Lance et al. 2010, Williams et al. 1989). These
et al. 2010) implicitly assume that a method studies used previously published MTMM ma-
factor has an equal effect on all measures be- trices to estimate conrmatory factor models
cause this technique is based on the assumption with multiple trait and method factors. Typ-
that, the observed variables are contaminated ically, the correlations among the trait factors
by a single unmeasured factor that has an equal and among the method factors, but not between
effect on all of them (Lindell & Whitney 2001, the trait and method factors, were estimated.
p. 114). However, other researchers argue that A summary of these studies is provided in
method factors may have unequal effects on Table 1. Taken together, they indicate that
different measures. This is important because 18% to 32% of the total variance in the items
if equal effects are wrongly assumed when used in these studies was due to method factors.
attempting to statistically control (or test) for Scherpenzeel & Saris (1997) went one step
method bias, the result will be the overestima- further by (a) estimating conrmatory factor
tion of the effect of method factors in some models for 50 MTMM matrices involving

Table 1 Summary of studies using multi-trait multi-method matrices to partition trait and method variance in empirical
relationships
Variance Variance Variance
attributable to attributable to attributable to
Study Sample trait factors method factors error
Cote & Buckley (1987) 70 matrices examining a wide variety 42% 26% 32%
of constructs
Williams et al. (1989)a 11 matrices involving perceptions of 48% 25% 21%
jobs and work environments
Buckley et al. (1990) 61 matrices examining a variety of 42% 22% 36%
constructs
Doty & Glick (1998) 28 matrices 46% 32% 22%
Lance et al. (2010) 18 matrices 40% 18% 42%
a
Values reported for variance estimates represent medians.
601 measures, (b) calculating the validity and among the methods are high, the correlations
reliability for each item, and then (c) examining among the method factors may represent the
the effect of 15 specic method factors on these convergence of [a] general trait factor across
item validities. Among the most important methods, rather than true relationships among
predictors of the item validities and reliabilities the methods. Whether the average method
were the type of construct being measured, correlation of 0.47 found in the studies cited is
form and length of the response scale, social large enough to support this interpretation is
desirability of the item, mode of data collec- a matter of judgment. A third criticism is that
tion, position of item in a battery of questions serious problems resulting in nonconvergence
with the same instructions and response scale, and/or improper estimates can arise when
and type of information requested (judgment, attempting to t a conrmatory factor model
frequency, agree-disagree). Another study that to MTMM data (Brannick & Spector 1990).
examined the effect of specic types of method However, the results of the studies reported
factors on item validity and reliability is that of in Table 1 were based only on solutions that
Baumgartner & Steenkamp (2001). Across 60 converged and had proper estimates. Thus,
measures of 11 constructs, they found that an although these criticisms are important, we
average of 8% (ranging from 0% to 29%) of believe the MTMM-based evidence supports
the variance in an item was due to ve specic the general conclusion that method biases have
response sets/styles. an impact on individual item validities and
However, there are some potential crit- reliabilities.
icisms of this MTMM-based evidence of
method bias. One is that the estimates of the
proportion of item variance due to method Effects of General Method Bias on the
provided by these studies are not completely Covariation Between Constructs
independent because the MTMM matrices Estimates based on MTMM meta-analytic
they analyze overlap to some extent. However, studies. The results of MTMM studies can
the overlap is relatively small (about 13%). also be used to obtain estimates of the aver-
Another criticism is that trait and method vari- age effects of method biases on the correla-
ance becomes confounded as the correlations tion between different traits (or constructs).
among traits and among methods increase. Assuming that trait, method, and random error
Specically, Bagozzi (1993, p. 66) noted that interactions do not exist, Cote & Buckley (1988)
when the correlations among the traits and show that the observed correlation between two

variables x and y (Rxy ) is equal to implies a true Rmk,ml of 1.00). Indeed, if a

single method had been used to calculate these
Rx,y = (trueRti,tj tx ty ) estimates, they would have ranged from 94% to

+ (trueRmk,ml mx my ) (1) 270%. Second, the MTMM studies included in
these meta-analyses overlap to some extent, so
where true Rti,tj = average correlation between the estimates of ination are not independent.
trait i and trait j; tx = percent of trait variance Third, there is quite a bit of variance in the
in measure x; ty = percent of trait variance in estimatesranging from 38% to 92%. Never-
measure y; true Rmk,ml = average correlation theless, regardless of which estimate is used, the
between method k and method l; mx = percent bottom line is that the amount of method bias
of method variance in measure x; and my = is substantial.
percent of method variance in measure y. Estimates based on method-method pair
More importantly, they demonstrate how meta-analytic technique. Another way to use
Equation 1 and the variance estimates from meta-analytic data to estimate the impact of
MTMM meta-analytic studies can be used to method biases has been proposed by Sharma
decompose the average observed correlation et al. (2009). Their technique involves catego-
between measures of two different traits that rizing the meta-analytic correlations from pre-
share the same method into the proportion due vious studies on the basis of the susceptibility
to (a) the correlation between the traits they to method biases of the pair of methods used
represent and (b) the common method they to measure the predictor and criterion. Their
share. For example, Cote & Buckleys (1987) argument is that some method-method (M-M)
meta-analysis reports that the average true cor- pairs are more susceptible to method biases than
relation between traits across 70 MTMM sam- others and that organizing the meta-analytic
ples was 0.674, the average percentage of trait data on the basis of these pairings allows re-
variance in each measure was 0.417, the av- searchers to obtain an estimate of the effects
erage true correlation between methods was that method biases have on the relationships
0.484, and the average percentage of method of interest. To illustrate how this method can
variance in each measure was 0.263. There- be applied, Sharma et al. (2009) conducted a
fore, using Equation 1, the average observed meta-analysis of 75 samples of data reported
correlation can be decomposed into the pro- in 48 studies examining the technology accep-
portion due to (a) the correlation
between
the tance model. The results indicated that (a) the
traits they represent [0.674 0.417 0.417 = mean correlation between the focal constructs
0.281] and
(b) the common method they share was about 0.16 when the susceptibility of the
[0.484 0.263 0.263 = 0.127]. This sug- M-M pairs to method biases was low and about
gests that the correlation between the traits 0.59 when the susceptibility of the M-M pairs
was inated approximately 45% (0.127/0.281) to method biases was high, and (b) about 56%
by method bias. Similar estimates of the per- of the between-studies variance in this litera-
cent of ination due to method bias obtained ture was attributable to method biases. They
from other meta-analyses of MTMM studies concluded that method bias presents a major
are 38% in Buckley et al. (1990), 92% in Doty potential validity threat to the ndings of IS
& Glick (1998), and 60% in Lance et al. (2009). research (Sharma et al. 2009, p. 474).
Several points regarding these estimates are
worth noting. First, these estimates are conser- Estimates of Specific Types
vative because they are based on MTMM stud- of Method Bias on the Covariation
ies that used two or more less-than-perfectly Between Constructs
correlated methods, and in many cases the The effects of same versus different
biggest concern regarding method bias is in sources. In addition to using meta-analytic
studies that use only a single method (which techniques to assess the impact of method

Table 2 Summary of meta-analytic studies comparing same-source versus different-source relationships

Estimates from same Estimates from different
source source
Relationship k N r k N r % inflation
Leader behaviors outcome variables 255 2,874 0.414 0.456 255 2,354 0.156 0.191 239%
Personality variables job performance 123 1,504 0.259 0.312 139 898 0.113 0.147 212%
Job attitudes OCB 98 6,729 0.270 0.340 155 13,551 0.190 0.230 148%
Participative decision making work 91 391 0.343 0.343 140 1,453 0.165 0.165 208%
outcomes
Organizational commitment job 148 3,745 0.180 0.183 159 1,924 0.138 0.138 133%
performance
Person-organization t job 12 639 0.230 0.283 21 813 0.073 0.093 304%
performance
OCB performance evaluations
95 2,808 0.490 0.595 56 2,889 0.260 0.323 184%
Abbreviations: k, sum of number of studies used to calculate the averages across meta-analyses of a specic relationship; N, harmonic mean for the specic
relationship using reported sample sizes across meta-analyses; OCB, organizational citizenship behavior; r, unweighted average of raw correlations
reported in meta-analyses; , unweighted average of corrected correlations reported in meta-analyses;% ination was calculated as
(same source /different source ).
factors on responses to individual measures citizenship behaviors (OCBs), participative

(e.g., Cote & Buckley 1987), these techniques decision making and work outcomes, organi-
have been used to assess the effects that zational commitment and job performance,
method factors have on the strength of the person-organization t and job performance,
relationships between two or more constructs. and OCB and performance evaluations are
Table 2 reports a summary of these types 213%, 147%, 208%, 133%, 304%, and 184%
of meta-analytic studies. For this table, we larger, respectively, when these measures are
searched for meta-analytic studies that had obtained from the same source than when they
explored the moderating effect of the source of are obtained from different sources. Thus, it
the ratings on the relationships that were exam- appears that the relationships between many
ined, and we then combined the data across the widely studied constructs are strongly inu-
different meta-analyses that had examined the enced by whether their measures are obtained
same general content areas to get an estimate of from the same or different sources. However, it
the effects that same-source method biases had is important to recognize that although a large
on the strength of the relationships reported. portion of the difference in the magnitudes
The results indicate that the average corrected of these correlations is undoubtedly due to
correlation between leader behaviors and out- method bias, some portion of it may also be
come variables (employee performance, ratings due to the different perspectives of the raters
of leader effectiveness, etc.) when taken from on what constitutes job performance (Lance
the same source is 0.456, but only 0.191 when et al. 2010).
obtained from different sources. This means
that the average corrected correlation between The effects of response styles. Using
measures of leader behaviors and outcome vari- data from a large representative sample of
ables is 239% (0.456/0.191) larger when these consumers (N = 10,477) from 11 countries,
measures are obtained from the same source Baumgartner & Steenkamp (2001) examined
than when they are obtained from different the biasing effects of acquiescence, disacquies-
sources. Similarly, the corrected correlation be- cence, extreme, midpoint, and noncontingent
tween measures of personality variables and job response styles/sets on the correlations among
performance, job attitudes and organizational 14 consumer constructs. Overall, they found

that 27% of the variance in the magnitude of measuring unrelated constructs). They found
the 91 intercorrelations among the 14 con- that, on average, two items measuring unrelated
structs was due to the ve response styles (64% constructs had a correlation of only 0.04 when
of the variance was due to the traits). The effect they were positioned six items apart, but the
of the response styles on the magnitude of the correlation increased to 0.09 when the items
correlations among the constructs depended were positioned right next to each other. In
upon whether the (a) true correlation between other words, the correlation between unrelated
the constructs was positive or negative and items increases by 225% when they are posi-
(b) response style components affecting the tioned next to each other, as opposed to when
scales were positively or negatively correlated. they are positioned a few items further apart.
If the true correlation between the constructs For nonreversed item pairs, the average corre-
was positive and the correlation between the lation signicantly and substantially increases
response styles was positive, they found that with decreasing interitem distance. When po-
the magnitude of the observed correlation sitioned six or more items apart, the average
was inated by 54% (the average correlation correlation between a pair of nonreversed items
increased from 0.13 to 0.20). If the true corre- was 0.35, but this correlation increased to 0.62
lation between the constructs was negative and (an increase of 177%) when these items were
the correlation between the response styles was positioned next to each other. In contrast, when
negative, they found that the magnitude of the positioned six or more items apart, a reversed
observed correlation was inated by 67% (aver- item pair had a correlation of 0.26, which de-
age correlation increased from 0.09 to 0.15). creased in magnitude to 0.06 when these item
In contrast, if the true correlation between the pairs were placed next to each other. Thus,
constructs was positive and the correlation be- Weijters et al. (2009, p. 7) concluded that up
tween the response styles was negative so that to a point, correlations become weaker for
the substantive and response style components nonreversed items and stronger for reversed
have opposing effects, they found that the items the further items are positioned from each
observed correlation was deated by 55% (av- other . . .
erage correlation decreased from 0.11 to 0.05);
if the true correlation between the constructs The effects of item wording. Harris &
was negative and the correlation between the Bladen (1994) examined the effect of stress ver-
response styles was positive, they found that sus comfort item wording on the relationships
the average observed correlation decreased between role ambiguity, role conict, role over-
from 0.07 to 0.01 and changed signs. They load, job satisfaction, and job tension. They
also found evidence that the amount of method found that the average correlation among these
bias varied across different types of constructs. ve constructs was 0.21 when item word bias
was controlled but increased to 0.50 when it
The effects of proximity and reversed items. was not controlled (an increase of 238%). In
Weijters et al. (2009) manipulated the proxim- addition, they found that the effect of method
ity and the nature of the conceptual relationship bias also varied depending upon the constructs
between two items and examined their effects involved.
on the strength of the correlation between the
items. Next, they specied a regression model The effects of item context. Harrison et al.
that explained the correlation between all pos- (1996) manipulated the order of the questions
sible pairs of 76 items (N = 2,850) as a func- measuring four constructs (voice, options,
tion of their distance apart on the questionnaire, objectivity, and standards) to create either a
and their conceptual relationship (nonreversed positive or negative measurement context for
items measuring the same construct, reversed the questions about outcome favorability and
items measuring the same construct, or items fairness perceptions. They found that the

correlation between outcome favorability dispositional mood states, and tendencies on

and fairness was only 0.10 in a positive mea- the part of the rater to acquiesce or respond in a
surement context but increased to 0.50 in a lenient, moderate, or extreme manner because
negative measurement context. This difference they make it impossible for the mindset of a
was signicant and found to be due to the effect common rater to bias the predictor-criterion
of the measurement context manipulation on relationship.
the variance in fairness perceptions. They Evidence of the effectiveness of obtaining
concluded that researchers would have come the predictor measure(s) from one person and
to different substantive conclusions about the criterion measure(s) from another person
the existence and strength of inuences on is summarized in Table 2. The data reported
fairness, solely because of the position in in this table indicate that although the average
which proposed antecedents were measured corrected correlation between predictor and
(Harrison et al. 1996, p. 257). criterion variables was 0.359 when they were
Taken together, the evidence presented in obtained from the same source, it decreased to
this section is not consistent with Spector & 0.184 when they were obtained from different
Brannicks (2009) assertion that the effects of sources (a 49% decrease). Using a variation
method have generally been small and rarely of this procedure, Ostroff et al. (2002) found
pose a threat and instead supports Johnson that obtaining the predictor and criterion
et al.s (2011) conclusion that CMV is not variables from different sources (rather than
an urban legend, but rather a specter that has the same source) decreased the average split-
the potential to haunt interpretations of ob- level correlations between several dimensions
served relationships. Thus, it is no surprise of work climate and satisfaction by 71%.
that editors of major journals in several dis- More specically, their results indicated that
ciplines (Chang et al. 2010, Kozlowski 2009, separating the sources decreased the average
Straub 2009, Zinkhan 2006) consider method split-level correlation from 0.07 to 0.02 when
biases an important problem that needs to be the individual was the unit of analysis, and
addressed. from 0.24 to 0.07 when the department was the
unit of analysis. Evidence of the effectiveness
of controlling for method bias by obtaining
WAYS TO CONTROL FOR either the predictor or criterion measure(s)
DIFFERENT SOURCES OF from one person and the other measure from
METHOD BIAS secondary data sources comes from two meta-
analyses. First, a meta-analysis of research on
Procedural Remedies
the relationship between leadership style and
Obtain measures of predictor and criterion effectiveness by Lowe et al. (1996) found that
variables from different sources. One obtaining both the predictor and criterion
obvious way to help control for method bias is variables from different sources decreased the
to obtain the measures from different sources. correlation between leadership style and effec-
There are two main ways this can be done: tiveness by 67% (from 0.57 to 0.19) compared
(a) obtain the predictor measure(s) from one to when both measures were obtained from
person and the criterion measure(s) from the same source. Second, a meta-analysis by
another; or (b) obtain either the predictor or Hulsheger et al. (2009) on the relationship
criterion measure(s) from one person and the between four team-process variables and team
other measure from secondary data sources innovation found that obtaining both the
(e.g., company records, annual reports). These predictor and criterion variables from different
procedures can diminish or eliminate the effects sources decreased the relationship by about
of consistency motifs, idiosyncratic implicit 49% (from 0.45 to 0.23) compared to when
theories, social desirability tendencies, both measures came from the same source.

Despite the fact that this approach seems to temporal separation does this by allowing previ-
control for several important sources of method ously recalled information to leave short-term
bias, it may not be appropriate to use in all cases. memory, whereas a proximal separation does
For example, this procedure is not appropriate this by eliminating common retrieval cues, and
when both the predictor and criterion variables a psychological separation does this by reduc-
are capturing an individuals perceptions, ing the perceived relevance of the previously
beliefs, judgments, or feelings. Beyond this, recalled information in short-term memory.
Chan (in Brannick et al. 2010) noted that this Evidence of the effectiveness of introducing
procedure is problematic for self-referential at- a temporal separation between the measure-
titude and perception constructs because (a) the ment of the predictor and criterion variables
individuals perceptions may not translate into comes from several studies. First, Ostroff et al.
observable behaviors, (b) others may not have (2002) compared predictor-criterion variable
the opportunity to observe these behaviors, and correlations for concurrent ratings of both
(c) valid measurement by others requires them variables to ratings obtained after a one-hour or
to accurately infer the individuals attitudes one-month delay. Although they found no sig-
or perceptions based on the observation of nicant differences in the average correlations
the individuals behavior. Furthermore, this between the concurrent and one-hour delay
technique may not be feasible to use in all conditions, they reported that the average
cases. For example, in some situations it may correlations were 32% lower after a one-
not be possible to obtain archival data that month delay than they were in the concurrent
adequately represent one of the constructs of condition. Second, Johnson et al. (2011, study
interest. In other situations, this technique 2) examined the effects of a three-week delay
may require more time, effort, and/or cost than on the correlation between a latent predictor
the researcher can afford. In addition, when construct and a latent criterion construct.
the sample size is small and the individual Their results indicated that the correlation
is the unit of analysis, the split-group procedure between the constructs was 43% smaller after
used by Ostroff et al. (2002) may not be feasible a three-week delay than it was when both were
because it requires cutting the sample size in measured at the same time (although the design
half and can result in too little power to detect makes it unclear how much of this difference
the effects hypothesized. was due to the time delay rather than sample
differences). Finally, in a rm-level analysis,
Temporal, proximal, or psychological sepa- Rindeisch et al. (2008, table 5) found few sig-
ration between predictor and criterion. An- nicant differences in the correlations between
other way to control for method bias is to in- several predictor and criterion variables after
troduce a separation between the measures of either no delay or a 30- to 36-month delay.
the predictor and criterion variables (Feldman Although the weight of the evidence sug-
& Lynch 1988, Podsakoff et al. 2003). This sep- gests that introducing a temporal separation
aration may be (a) temporal (i.e., a time delay is an effective means of controlling for some
between measures is introduced), (b) proximal method biases, there are several disadvantages
(i.e., the physical distance between measures of this approach. First, introducing a tempo-
is increased), or (c) psychological (i.e., a cover ral separation will obviously increase the com-
story is used to reduce the salience of the linkage plexity of the research design and potentially
between the predictor and criterion variables). its cost. Second, when a temporal separation
Podsakoff et al. (2003) noted that these types of is introduced, it may allow other nonmethod-
separation should reduce the respondents abil- ological factors to inuence the level of the out-
ity and/or motivation to use previous answers come variable. Third, the longer the temporal
to ll in gaps in what is recalled, infer miss- delay, the greater the chance of respondent at-
ing details, or answer subsequent questions. A trition. Fourth, it is difcult to determine what

the appropriate delay should be for any given method bias is not readily available. However,
relationship, and it is likely that the appropri- there is no shortage of studies that recommend
ate delay varies across types of relationships. If this procedure. For example, Aronson et al.
the delay is too short, the temporal separation (1998) note that one way to psychologically sep-
may be ineffective; and if the delay is too long, arate the predictor and criterion variables is to
intervening factors are likely to affect the crite- use a multiple study cover story, in which
rion variable. Fifth, and most importantly, the participants are told that for reasons of con-
temporal separation procedure is based on the venience or efciency several unrelated stud-
assumption that the true relationship between ies are being conducted at the same time. This
the constructs is relatively stable over the time ruse is frequently employed in priming exper-
period of the delay and that method bias will iments (e.g., Higgins et al. 1977) and attitudi-
dissipate over time. If it is suspected that either nal research (e.g., Rosenberg 1965). Other ways
of these assumptions is inaccurate, this method to psychologically separate the predictor and
of control should not be used. Indeed, re- criterion measures might be to (a) camouage
cent empirical research (Alessandri et al. 2010, interest in the criterion or predictor variable
Weijters et al. 2010a) suggests that the assump- by embedding it in the context of other ques-
tion that the method bias dissipates over time tions so that it is less psychologically prominent
may be questionable. (i.e., diminishing the salience of the measure) or

Indirect evidence of the effectiveness of (b) disguise the reasons for obtaining the pre-
introducing a proximal separation between dictor or criterion measure by leading respon-
the measures of the predictor and criterion dents to believe that it is tangential to the main
variables comes from studies demonstrating purpose of the study (i.e., making respondents
that separation attenuates method biases due think it is unimportant).
to context effects (Tourangeau et al. 2000) The principal disadvantage of this technique
and question order effects (Tourangeau et al. is that its effectiveness is dependent upon the
2003). However, the most direct evidence credibility of the cover story, but a considerable
comes from the study by Weijters et al. (2009) amount of creativity and ingenuity is required
that found that proximal separation prevents to develop a convincing cover story. Conse-
the correlation between nonreversed (reversed) quently, it is essential to thoroughly pretest the
items from being articially inated (deated). cover story in order to ensure its effectiveness.
Based on their analysis, they concluded that
researchers should try to position measures Eliminate common scale properties. Sev-
of the same-construct at least six items apart, eral authors (e.g., Campbell & Fiske 1959,
separated by measures of other constructs Cronbach 1946, Feldman & Lynch 1988,
using the same or different formats, or by Podsakoff et al. 2003, Tourangeau et al. 2000)
means of dedicated buffer items. have observed that method bias can result from
Although there is some evidence that proxi- common scale properties (i.e., scale type, num-
mal separation is an effective means of control- ber of scale points, anchor labels, polarity, etc.)
ling for some method biases, there are some shared by the items used to measure different
disadvantages of this approach. First, it can in- constructs. For example, Feldman & Lynch
crease the length of the questionnaire and that (1988, p. 427) note that method bias will occur
may cause fatigue, decrease response rates, or to the extent that the question formats are per-
increase costs. Second, if the ller items are ceived to be similar by respondents, because
conceptually related to the measures of interest, the similarity of the response format enhances
they could create context effects that increase the probability that cognitions generated in
method bias. answering one question will be retrieved to
Empirical evidence of the effectiveness of answer subsequent questions. The obvious
psychological separation as a means of reducing remedy to this problem is to try to minimize

the scale properties shared by the measures of & Bernstein 1994). In these instances, it is
the predictor and criterion variables. important to give priority to maintaining the
Evidence of the effectiveness of this rem- content validity of the items because a lack of
edy comes from several studies. Kothandapani content validity poses an even bigger threat to
(1971) measured three constructs using four construct validity than does common method
different scale formats (Likert, Thurstone, bias (MacKenzie et al. 2011). Thus, although
Guttman, and Guilford) and found that the av- minimizing common scale properties is always
erage correlation was 0.45 when the criterion a good idea, there are practical limits to the
and predictor shared the same scale format and extent to which this can be done.
dropped to 0.18 when they did not share this
scale property, a decrease of 60%. Arora (1982)
measured three constructs using three differ- Improving scale items to eliminate ambi-
ent scale formats (Likert, semantic differential, guity. Ambiguous items are ones that are dif-
and Stapel) and found that the average corre- cult to interpret and require people to construct
lation was 0.34 when the criterion and predic- their own idiosyncratic meanings for them.
tor shared the same scale format and dropped Johnson (2004) identies several causes of item
to 0.23 when they did not share this scale ambiguity, including the presence of indetermi-
property, a decrease of 32%. Flamer (1983) nate words such as many and sometimes,
measured three constructs using three different words with multiple meanings, multiple ideas
scale formats (Likert, semantic differential, and linked together with conjunctions or disjunc-
Thurstone) in two different samples and found tions, or complex constructions such as double
that the average correlation was 0.06 in sam- negatives.
ple A and 0.09 in sample B when the criterion According to several authors (Cronbach
and predictor shared the same scale format and 1950, Feldman & Lynch 1988, Podsakoff et al.
dropped to 0.04 in sample A and 0.08 in sample 2003), the problem with ambiguous items is
B when they did not share this scale property, a that they cause respondents to be uncertain
decrease of 33% in sample A and 11% in sample about how to respond on the basis of the items
B. Finally, Weijters et al. (2010c) examined the content, which increases the likelihood that
effect of common scale labeling on the correla- their responses will be inuenced by their
tion between attitudes and intentions and found systematic response tendencies (e.g., acquies-
that the average correlation was 0.69 when both cent, extreme, or midpoint response styles).
the predictor and criterion variables had only The best solution to this problem is to make
the extreme end points of the scale labeled, and every effort to: keep questions simple, specic,
dropped to 0.60 when the criterion had the ex- and concise; dene ambiguous or unfamiliar
treme end points labeled and the predictor had terms; decompose questions relating to more
all points on the scale labeled, a decrease of than one possibility into simpler, more focused
15%. questions; avoid vague concepts and provide
An advantage of this procedure is that it is of- examples when such concepts must be used;
ten easy to translate some types of scale formats avoid double-barreled questions; and avoid
(e.g., Likert) into other formats (e.g., semantic complicated syntax (see Tourangeau et al.
differential) without changing the content of 2000). In addition, Krosnick (1991) notes that
the item or other properties of the item (e.g., labeling every point on the response scale
number of scale points). However, that is not (rather than only the end points) is also an
always the case. For example, although it can effective means of reducing item ambiguity.
be done, translating Likert or semantic differ- Unfortunately, we were unable to nd any
ential items into Thurstone or Guttman scales empirical evidence that specically examined
is often difcult to do without altering the the effect of item ambiguity on estimates of the
conceptual meaning of the measures (Nunally relationships between two different constructs.

Reducing social desirability bias in item the relations across all items, traits, and trait
wording. There is a great deal of evidence levels. This would suggest that the linear
that items differ in perceived social desirability correlation between responses to an item and
and that this affects responses to the item. For responses to a social desirability scale may not
example, Edwards (1970) measured the impact always be a valid indication of the tendency of
of item social desirability on responses to the the item to evoke socially desirable responses.
Minnesota Multiphasic Personality Inventory.
He found that ratings of the social desirability Balancing positive and negative items. A
of self-descriptive items made by a sample of number of authors (Baumgartner & Steenkamp
judges were correlated 0.87 with endorsements 2001, Billiet & McClendon 2000, Mirowsky &
of the items by another sample of subjects. Ross 1991, Weijters et al. 2010b) have noted
This suggests that item wording can poten- that scale formats that ask respondents how
tially undermine the accuracy of responses strongly they agree or disagree with statements
by causing subjects to edit their responses for may be susceptible to acquiescence or disac-
social acceptability. If this editing affects both quiescence response style biases. Respondents
predictor and criterion measures in a similar who exhibit acquiescence response styles tend
manner, it could possibly bias the relationship to disproportionately use the positive side of
between them. There are two commonly used the scale, whereas those that exhibit disacqui-
procedures for controlling item social desir- escence response styles tend to disproportion-
ability (e.g., Kuncel & Tellegen 2009). The ately use the negative side of the scale. As noted
rst is to obtain an assessment of the perceived by Mirowsky & Ross (1991), these response
social desirability of the items from judges and style tendencies are problematic because they
to revise the wording of the highly rated items inate the estimates of the reliability of mea-
to minimize or reduce the perceived level of sures, may produce misleading factor analytic
social desirability. The other is to calculate solutions, and may inate or deate correlation
the correlation between subjects responses and regression coefcients, depending on the
to each item and responses to a recognized type of questions that are asked. One procedural
social desirability scale (e.g., Paulhus 1984) and remedy that has been used to try to reduce this
to revise the wording of items that correlate type of bias is balancing the positively worded
highly with this scale to minimize or reduce (i.e., agreement with the item indicates a higher
the perceived level of social desirability. score on the underlying construct) and neg-
Although these procedures have been atively worded (i.e., agreement with the item
widely used and seem to have few disad- indicates a lower score on the underlying con-
vantages, we were unable to nd any direct struct) measures of each construct. According
empirical evidence of their ability to prevent to Baumgartner & Steenkamp (2001, p. 147),
item social desirability from biasing the correla- although balanced scales do not eliminate the
tions between measures of different constructs. occurrence of acquiescence per se, they contain
Moreover, implementing these procedures may a built-in control for contamination of observed
be more difcult than it appears for two rea- scores by yea-saying, because the bias is upward
sons. First, revising the items to eliminate their for half of the items and downward for the other
social desirability without compromising their half.
content validity may be easier said than done. The advantage of this technique is that
Second, Kuncel & Tellegen (2009, p. 201) have it is a proactive way to control for acquies-
shown that the relation between degree of en- cence and disacquiescence biases. However,
dorsement of an item and its judged desirability there are several limitations of this technique
level is often nonlinear and varies across items (Baumgartner & Steenkamp 2001, Mirowsky
such that no general model of item desirability & Ross 1991, Weijters et al. 2010b). First,
can be adopted that will accurately represent many existing scales do not contain an equal

number of positively and negatively worded control for all systematic sources of bias when
items, and reversing the wording of some items testing hypotheses about the relations between
may alter their content. Second, reversed items constructs. Indeed, Phillips & Lord (1986)
may be confusing for some respondents. Fi- noted a similar confounding of method and
nally, empirical research suggests that because substantive variance when trying to control
this technique does not always completely for halo effects and concluded that there are
control for these biases, it should be used in advantages to controlling for both. Second,
conjunction with the statistical methods of if the ratio of the number of indicators to
control described in the next section. the number of substantive constructs is low,
the addition of a method factor can cause
identication problems. Finally, this procedure
Statistical Remedies is based on the assumption that the method
Although it is possible that the use of the pro- factor does not interact with the trait factors; an
cedural remedies discussed above will minimize assumption that has been questioned by several
the detrimental effects of method biases, re- researchers (see Bagozzi & Yi 1990, Campbell
searchers may not always be able to implement & OConnell 1967, Wothke & Browne 1990).
them beforehand. In these circumstances, they
may nd it useful to use one of the statistical Correlation-based marker variable tech-
remedies. nique. This approach (Lindell & Whitney
2001) ideally requires researchers to (a) identify
Unmeasured latent method factor tech- a marker variable that is expected for the-
nique. This is perhaps the oldest latent vari- oretical reasons to be completely unrelated
able control technique (Bagozzi 1984, Bagozzi to the substantive variables of interest, (b) use
& Phillips 1982, Widaman 1985), and it has the smallest correlation between the marker
been used in approximately 50 studies (see variable and the substantive variables as an esti-
Richardson et al. 2009). This technique in- mate of the effects of method bias, (c) adjust the
volves adding a rst-order method factor whose zero-order correlation between every pair of
only measures are the indicators of the theoret- substantive variables of interest by subtracting
ical constructs of interest that share a common this estimate from the zero-order correlation
method. This technique has several advantages: between any pair of substantive variables and di-
(a) it does not require the researcher to measure viding by the quantity of 1 minus this estimate,
the specic factor responsible for the method and (d ) examine whether the resulting partial
effect; (b) it models the effect of the method correlation is signicantly different from zero.
factor at the measurement level, rather than at They argue that if this partial correlation re-
the latent construct level (Schaubroeck et al. mains signicant, the substantive relationships
1992, Williams et al. 1996); and (c) it does not still hold even after controlling for method
require the effects of the method factor on each bias. According to Williams et al. (2010),
measure to be equal. this technique has been widely used in recent
However, this approach has been criticized years.
for several reasons. First, as noted by Podsakoff The primary advantage of this technique
et al. (2003, p. 894), the unmeasured latent is that it is easy to implement. The disad-
method factor may reect not only different vantages are many. First, Lindell & Whitney
types of common method variance but also (2001) do not require the marker variable
variance due to relationships between the to share any method characteristics with the
constructs other than the one hypothesized. substantive variables. Indeed, they suggest that
This is considered to be a serious aw (e.g., the variable with the smallest correlation with
Richardson et al. 2009), but it could also be the substantive variables can be arbitrarily se-
considered a virtue, since it is desirable to lected in an ad hoc manner as a marker variable.

As noted by Williams et al. (2010, p. 507) this is is evidence that different types of variables
problematic because if the marker variable does contain differing amounts of method variance.
not share method characteristics with the sub-
stantive variables, it cannot provide the vehicle Regression-based marker variable tech-
for partialling out these biases from estimates nique. Siemsen et al. (2010) recently proposed
of relations among substantive variables so as to that common method bias can be eliminated
obtain a truer estimate of the relation, which when estimating a regression equation subject
is the goal behind the use of marker variables. to method bias by adding a marker variable that
Second, it assumes that method bias can only (a) is uncorrelated with the substantive variables
inate and never deate relationships among of interest and (b) suffers from some type of
the substantive variables. Several researchers method bias. In the event that the marker vari-
have demonstrated that this assumption is able is modestly correlated with the substantive
incorrect (Baumgartner & Steenkamp 2001, variables, their numerical analysis suggests that
Cote & Buckley 1988, Podsakoff et al. 2003). the addition of 3 to 5 variables is necessary.
Third, this technique ignores measurement Perhaps the greatest advantage of this
error that could attenuate the correlations be- technique is that it is easy to implement. How-
tween the marker variable and the substantive ever, there are several disadvantages. Like the
variables that are used to obtain an estimate correlational marker variable technique, this
of method bias (Lance et al. 2010, Podsakoff technique (a) ignores measurement error that
et al. 2003, Williams et al. 2010). Fourth, could attenuate the correlation between the
this approach controls for method bias at the marker variable(s) and the substantive variables
scale level rather than the item level (Williams of interest, (b) controls for method bias at
et al. 2010). Fifth, this method is based on the the scale level rather than the item level, and
assumption that the method factor represented (c) is based on the assumption that the method
by the marker variable does not interact with factor represented by the marker variable(s)
the substantive variables of interest, which has does not interact with the substantive variables
been disputed by several researchers (Bagozzi of interest. In addition, this technique controls
& Yi 1990, Campbell & OConnell 1967, for only the net effect of the sources of method
Wothke & Browne 1990). Sixth, it is based on bias common to the marker variable(s) and the
the assumption that the smallest correlation substantive variables and applies only to single-
between the marker variable and the substan- equation models. Furthermore, it is unclear
tive variables is a reasonable estimate of the what is being controlled by the addition of the
effects of all types of method bias, which is marker variables. It is assumed to be method
not justied because the marker variable is not bias based on the subjective judgment that the
required to share any measurement charac- marker variables are theoretically unrelated
teristics (e.g., common scale format, anchors) to the substantive variables. However, that
with the substantive variables (cf. Podsakoff may not be true. Finally, Siemsen et al. (2010,
et al. 2003, Williams et al. 2010). Finally, p. 472) limited their analysis to a single method
this technique assumes that the method factor factor even though they note that, In practice,
represented by the marker variable has an observed variables may suffer from multiple
identical effect on every substantive variable of different methods factors . . . Although we
interest in the study. However, this assumption expect our insights to hold if these methods
has been widely criticized (e.g., Podsakoff et al. factors are uncorrelated with each other,
2003, Richardson et al. 2009, Sharma et al. examining multiple correlated methods factors
2009). Indeed, Williams et al. (2010) note that may lead to different results.
an analytical technique that can incorporate
unequal method effect is needed in most Instrumental variable technique. This
organizational research settings because there technique is based on the fact that the presence

of a method factor that inuences both the pre- direction as OLS in small samples. Even in
dictor and the criterion variable in a model will large samples, Kennedy (2008, p. 145) notes
cause the structural error term for the equation that the magnitude of this bias (a) can be quite
to be correlated with the predictor. In this large if the IV is not strongly correlated with
instance, the supposedly exogenous predictor the endogenous predictor it represents and
variable is really an endogenous predictor. This (b) becomes even worse if several weak IVs are
is an important problem because it violates used. Indeed, even a slight correlation between
an assumption of many estimation techniques a weak IV and the structural error term for
[e.g., ordinary least squares (OLS), maximum the equation (perhaps caused by method bias)
likelihood (ML)] and causes the estimate of the can cause the IV estimate to exhibit more bias
effect of the predictor on the criterion variable (even asymptotically) than an OLS estimate.
to be biased (i.e., inconsistent). Antonakis Another reason why strong IVs are required,
et al. (2010) point out that method bias can be according to Kennedy (2008, p. 145), is that A
controlled, and an unbiased (i.e., consistent) weak instrument also causes the IV variance to
estimate of the effect of the predictor on the be underestimated in small samples; this causes
criterion can be obtained, by adding appropri- the true type I error rate to be higher than its
ate instrumental variables (IVs) to the model chosen level. Thus, weak IVs lead to biased
and estimating the effect of the predictor on the estimates and unreliable inference.
criterion variable using two-stage least squares Second, the IV must be completely uncor-
(2SLS). Briey, in the rst stage of the 2SLS related with the structural error term for the
estimation process the endogenous predictors equation. Antonakis et al. (2010) emphasize
are regressed on the IVs (and any other truly that the instruments must rst pass a theoret-
exogenous variables included in the model) to ical overidentication test before an empirical
obtain predicted values for the endogenous one because if all the modeled instruments
predictors. In the second stage, the criterion are not truly exogenous the overidentication
variable is regressed on the predicted values test will not necessarily catch the misspeci-
of the endogenous predictors obtained in the cation. According to Kennedy (2008), this
rst stage (and any other truly exogenous theoretical overidentication test for each IV
variables in the model). An instrumental should involve the use of existing literature and
variable is a truly exogenous variable (i.e., it theory to (a) defend the implicit assumption
does not depend on other variables) that is that the IV is not an explanatory variable in
(a) correlated with the endogenous predictor the equation being estimated (i.e., that the IV
for which it is to serve as an instrument and does not have any direct effect on the criterion
(b) uncorrelated with the structural error term variable) and (b) explain why the IV could not
for the equation. Thus, an IV is indirectly be inuenced by any of the method factors
related to the criterion variable through the that inuence the criterion variable or by any
endogenous predictor but not directly related other omitted variables that affect the criterion
to the criterion variable. Antonakis et al. variable (because if the IV is affected by these
(2010) recommend adding at least one more factors, it would be correlated with the struc-
IV than there are endogenous predictors in the tural error for the equation). Assuming that the
model. IVs pass these theoretical tests, Antonakis et al.
In order to be useful, each IV must satisfy (2010) recommend using a Sargan chi-square
two essential requirements (Antonakis et al. test of overidentication to test empirically
2010, Kennedy 2008). First, the IV must the assumption that the IVs are uncorrelated
be signicantly and strongly related to the with the structural error term. If the IVs are
predictor it represents. This is required for two unrelated to the structural error term, the over-
reasons. One is that IV estimators (although identication tests will all be nonsignicant. If
asymptotically unbiased) are biased in the same they are not, one must nd better instruments.

The primary advantage of the instrumental a latent method factor and estimates all possi-
variable technique is that it provides a straight- ble correlations among the method factor and
forward solution to the problem of common the substantive constructs of interest, but it sets
method bias in situations where its causes the loadings from the method factor to the in-
cannot be identied or measured directly dicators of the substantive constructs to zero.
(Antonakis et al. 2010). However, perhaps the The second model requires that (a) the cor-
biggest disadvantage of this technique is the relations between the method and substantive
difculty of selecting IVs that are related to latent factors be set to zero, (b) the indicator
the endogenous predictors and completely loadings of the latent method factor be xed at
uncorrelated with the structural error term for the estimates obtained from the CFA model,
the equation. Indeed, if (as is often the case) and (c) the loadings from the latent method
all of the possible sources of method bias that factor to the indicators of the substantive con-
might affect the endogenous predictors and the structs be set to zero. This model is called the
criterion variable cannot be identied, then it baseline model because it serves as the base-
is unclear how the IVs could pass Antonakis line against which the method effects are as-
et al.s theoretical overidentication test and sessed. The third model is called the method-C
how one could be condent that they were not model (i.e., a constrained model). This model
affected by these unidentied method biases as estimates the loadings from the latent method
well. In addition, it may prove to be difcult factor to the indicators of the substantive con-
to identify IVs that are strongly related to the structs but constrains these loadings to be equal
endogenous predictors, and as noted above, to each other. A comparison of the t of this
weak IVs lead to biased estimates and unre- model to the t of the Baseline model pro-
liable inference. Unfortunately, if these two vides a test of the assumption that the latent
requirements are not met, this technique will method factor has equal (tau equivalent) effects
produce biased estimates and inate the type I on the indicators of the substantive constructs
error rates, and researchers would be better off of interest. In contrast to this assumption, the
using another technique to control for method fourth model (called the method-U, or uncon-
biases. A nal disadvantage is that because the strained, model) allows the loadings from the
results are dependent upon the IVs selected, method factor to the indicators of the substan-
a test of the robustness of the second-stage tive variables to be freely estimated (i.e., uncon-
estimates should probably be conducted. strained). A comparison of the t of this model
to the t of the method-C model provides a test
CFA marker technique. To address some of of the assumption that the method factor has
the problems with the correlation-based marker unequal effects on the indicators of the substan-
variable technique, Williams et al. (2010) rec- tive constructs. Finally, the fth model, which
ommend using a series of marker variables is referred to as the method-R model (to repre-
that share measurement characteristics with sent restrictions on the parameters) is specied.
the substantive variables of interest as indica- This model is identical to the method-C and
tors of a latent method factor. They propose a method-U models, with the exception that the
three-phase conrmatory factor analysis (CFA) correlations among the substantive constructs
marker technique to identify and control for are constrained to the values estimated in the
method biases. Phase I of this analysis tests for baseline model. A comparison of the t of this
the presence and quality of method effects as- model to the t of either the method-C model
sociated with the latent marker variable. This or the method-U model (depending on which
phase requires specifying ve different latent of these models ts the best) provides a test of
variable models and comparing their relative t the bias in the correlations among the substan-
to each other. The rst model (the CFA model) tive constructs that is due to the latent method
estimates loadings for each marker variable on factor.

Phase II of this analysis is devoted to method biases affect all measures equally or
quantifying how method variance affects differentially.
the reliability of the substantive constructs. Despite these advantages, there are a few po-
This is important because if method variance tential problems regarding this approach that
is not controlled, it will bias the reliability should be noted. First, it doesnt identify the
estimates of the substantive constructs. First, nature of the method bias being controlled. In-
the completely standardized estimates of the deed, Williams et al. (2010, p. 507) note that,
factor loadings and error variances for each without conceptual analysis of the nature of
substantive construct from the baseline model the marker variable, the meaning of its covaria-
are used to obtain reliability estimates for tion with substantive variables cannot be under-
each construct (Werts et al. 1974). Next, the stood. Related to this, a second problem is that
completely standardized substantive construct the conceptual meaning of the latent method
factor loadings, method factor loadings, and factor is ambiguous. Empirically, this construct
the error variances (from either the method-C is dened as the common variance among the
or method-U model, depending upon which marker variables. Although this technique re-
was supported) are used to decompose the total quires the marker variables to be theoretically
reliability calculated in the rst step into the unrelated to the substantive constructs, it places
proportion due to the substantive construct no constraints on their theoretical relationships

and the proportion due to the method factor. to each other. This means that potentially the
Finally, phase III is used to conduct a marker variables could all come from a scale
sensitivity analysis to increase condence in for a recognized construct (albeit one that is
the ndings. Briey, Williams et al. (2010) theoretically unrelated to the substantive con-
argue that since the amount of method variance structs of interest). Consequently, it is unknown
associated with each indicator of the substan- whether the common variance that empirically
tive constructs is represented by the magnitude denes the marker variable construct is due to
of their loadings on the method factor, the method artifacts or to some theoretically mean-
sensitivity of the estimates of the correlations ingful construct that is confounded with it. This
between the substantive constructs to method would affect the loadings of the marker variables
bias can be examined by substituting larger on this latent construct in Williams et al.s CFA
alternative values for these method factor load- model as well as their baseline model (since it
ings. The specic alternative values selected uses these estimates from the CFA model as
should be based on the condence intervals xed parameters).
of the unstandardized method factor loadings Another problem is that the results are sen-
from either the method-C or method-U sitive to the specic variables used as indicators
models (depending on which one was sup- of the latent method factor. This technique will
ported). The examination of this model allows only control for the net effect of the method
researchers to determine the sensitivity of characteristics that are shared by all of the
their results to increasing amounts of method marker variables and the indicators of the sub-
variance associated with sampling error in the stantive constructs. If there are many relatively
indicators. important method characteristics shared, this
Williams et al. (2010) have identied several procedure will provide a strong test of method
advantages of this approach over the partial bias, but if there are only a few relatively unim-
correlation approach proposed by Lindell & portant method characteristics shared, this pro-
Whitney (2001). First, it models the effects cedure will provide only a weak test of method
of method biases at the indicator level (rather bias.
than construct level). Second, it provides a A nal problem is that in phases I and III,
statistical test of method bias based on model this procedure requires xing parameter esti-
comparisons. Third, it permits a test of whether mates in one model at specic values obtained

from the estimation of an alternatively spec- Measured response style technique. An-
ied model. This two-step estimation process other promising technique is to systematically
may not provide correct standard errors and measure common response styles and partial
goodness-of-t statistics to test the t of the re- out their effects on responses. This procedure
sulting model (Kennedy 2008, Joresk og
1998). requires several steps. First, the relevant item
population must be dened and a random
sample taken of it to produce a representative
Directly measured latent method fac- heterogeneous set of items. As noted by
tor technique. To apply this technique, Weijters et al. (2010b, p. 118), The items
researchers must be able to anticipate the should relate to constructs that do not form a
potential source of method bias and obtain meaningful nomological network. In order to
measures of it. If direct measures of this par- develop reliable measures of the response styles,
ticular source of method bias are available, bias they recommend that a minimum of three sets
can be controlled by adding to the theoretical of ve items each should be used. Second, this
model a method factor that has both the direct random sample of heterogeneous items should
measures and the measures of the substantive be inserted as buffer items between the scales of
constructs of interest as reective indicators. substantive interest using the same scale format
This technique has been used in several studies as for the other items on the questionnaire.
(e.g., Bagozzi 1984, Schaubroeck et al. 1992, Third, as many researchers have noted (e.g.,
Williams & Anderson 1994, Williams et al. Baumgartner & Steenkamp 2001; Weijters
1996). For example, Williams et al. (1996) et al. 2008, 2010a,b), the most common
used this technique to control for the effects of response styles can be measured for each set of
negative affectivity on the relationship between items as follows: (a) acquiescence response style
job attitudes and role perceptions. In general, (ARS)calculate the extent of agreement with
it can be used to control for any contaminating both positively and negatively worded items
factor for which direct measures are available in each set (before negatively worded items
(e.g., social desirability, positive affectivity). have been reverse-scored), (b) disacquiescence
The advantages of this approach are that response style (DRS)calculate the extent of
(a) it unambiguously identies the source of disagreement with both positively and nega-
the method bias, (b) it controls for measure- tively worded items in each set (before nega-
ment error, (c) it models the effects of the bi- tively worded items have been reverse-scored),
asing factor at the item level rather than at the (c) extreme response style (ERS)calculate
construct level, and (d ) it does not constrain the proportion of items in each set on which
the effects of the methods factor on the mea- the respondent endorses the most extreme
sures of the substantive construct to be equal. (positive or negative) scale categories, and
Perhaps the biggest disadvantage of this tech- (d ) midpoint response style (MRS)calculate
nique is that it requires researchers to anticipate the proportion of items in each set on which
the most important sources of method biases in the respondent endorses the middle scale cat-
their studies and to include measures of these egory. Weijters et al. (2008, p. 414) provide an
sources. This is a serious problem because it excellent illustration of how these operational-
is often difcult to identify the key sources of izations can be applied to a 7-point Likert scale
method bias in a given situation, and valid mea- (1 = strongly disagree, 7 = strongly agree).
sures for these sources may not exist. In addi- For each set of k items, they compute the mea-
tion, this technique assumes that the method sures of each response style as follows: ARS =
factor does not interact with the substantive [ f(5) 1 + f(6) 2 + f(7) 3]/k; DRS = [ f(1)
constructs, which has been questioned by sev- 3 + f(2) 2 + f(3) 1]/k; ERS = [ f(1) +
eral researchers (Bagozzi & Yi 1990, Campbell f(7)]/k; MRS = f(4)/k; where f(o) refers to the
& OConnell 1967, Wothke & Browne 1990). frequency of response option o. This results in

a measure of each of the four response styles evaluation of the method construct(s) in
for each of the sets of k items. Fourth, for each terms of convergent and discriminant validity.
of the response styles, the measures obtained Second, it allows for unique variances in the
from each set of items are used as indicators of response-style item sets. Therefore, these
a latent construct. This means that, if there are unique variances are not confounded with
three sets of items, there would be three indi- the method construct(s) itself. Third, using
cators of each response-style latent construct.1 multiple indicators of the method construct(s)
Fifth, the latent constructs representing each enhances the stability of the model. Finally,
response style are added to a latent variable unlike the unmeasured latent variable, marker
model and their effects on the measures of the variable, or CFA-marker variable approaches,
substantive constructs of interest are added.2 it species the nature of the method construct
A few words of caution are in order. First, (e.g., ARS, ERS) whose effects are being
in order to ensure that the response-style mea- controlled. However, despite these advantages,
sures only capture method variance, it is essen- there are also some limitations of this ap-
tial for the content of the set of items used to proach. First, it only controls for the response
measure the response styles to be independent styles explicitly measured. Second, it requires
of the content of the measures of the substantive the researchers to collect additional data to
constructs. De Beuckelaer et al. (2010) found measure the response styles.

that using ad hoc sets of items is suboptimal
at detecting ARS and ERS compared to using RECOMMENDATIONS
random heterogeneous items. Second, it is im-
In this section, we suggest strategies for
portant to use a complete prole of response
(a) identifying when method bias is likely to be a
styles because it is difcult to decide a priori
problem and (b) mitigating its effects. However,
which response style may cause bias (Weijters
as Podsakoff et al. (2003, p. 899) emphasize,
et al. 2010a).
The key point to remember is that the pro-
With these caveats in mind, if the effects of
cedural and statistical remedies selected should
the response-style constructs on the measures
be tailored to t the specic research question
of the substantive constructs are signicant, it is
at hand. There is no single best method for
evidence of method bias. However, if the esti-
handling the problem of common method vari-
mate of the relationship between the constructs
ance because it depends on what the sources of
of interest is signicant after controlling for
method variance are in the study and the fea-
these response styles, then one can be con-
sibility of the remedies that are available. The
dent that the relationship is not solely due to
goal is to reduce the plausibility of method bi-
these forms of method bias.
ases as a rival explanation for the relationships
Using multiple indicators to measure
observed in a study.
response-style constructs has several advan-
tages (Weijters et al. 2010b). First, it facilitates
When Is Method Bias Likely
To Be a Problem?
1
Weijters et al. (2010b) found that the loadings of the
response-style indicators on the method factors are essential There is widespread agreement that generating
tau equivalent (complemented with a time-invariant autore- an optimal answer to even a single question
gressive effect): This means that ARS and ERS are largely can require a great deal of cognitive work, and
but not completely consistent over the course of a question-
naire (p. 105). Note that they did not examine whether the the effort required to answer a long series of
method factor had tau equivalent effects on the measures of questions on a wide range of topics is sub-
any other substantive constructs. stantial (Krosnick 1999, Sudman et al. 1996,
2
When specifying this model, the covariances among the re- Tourangeau et al. 2000). Although we may wish
sponse styles should be estimated, and the indicators that are
based on the same sets of items should have correlated error otherwise, not all respondents will be willing
terms across response styles. and able to exert the cognitive effort required

to generate accurate answers to the questions result from a respondents low cognitive ability,
on a typical research instrument. What then? poorly differentiated cognitive structure, or
Krosnick (1999) argues that when the difculty uncertainty about how to respond to the
of the task of generating an optimal answer is question. They also provide an excellent sum-
high but a respondents ability or motivation to mary of several personality characteristics that
expend the required amount of cognitive effort are associated with biased responding. They
are low, respondents may satisce rather than cite research indicating that (a) stimulation-
generate the most accurate answers by simply seeking extroverts may have a tendency to
being less thorough in question comprehen- accept statements impulsively and agree with
sion, memory retrieval, judgment, and response them regardless of content (i.e., ARS or positiv-
selection. In our view, when respondents are ity bias), (b) controlled and reective introverts
satiscing rather than optimizing, they will be who try to avoid external stimulation may
more likely to respond stylistically and their have a tendency to disagree with items regard-
responses will be more susceptible to method less of content (i.e., DRS or negativity bias),
bias. In other words, we expect that responses and (c) respondents who are rigid, dogmatic,
will be more strongly inuenced by method anxious, or intolerant of ambiguity may have a
bias when the respondents cant provide tendency to endorse the most extreme response
accurate responses (which is a function of their categories regardless of content (i.e., ERS).
ability and the difculty of the task) or when Thus, method biases and stylistic respond-
they are unwilling to try to provide accurate ing may be more likely to the extent that the
responses (which is a function of motivation). respondents in a study have these ability limi-
tations or possess these personality characteris-
Ability factors that may cause biased tics. Consequently, under these circumstances
responding. The rst question to consider researchers would be wise to implement the
is whether respondents are able to provide appropriate procedural and statistical remedies
accurate answers, because if they are not, discussed below.
they may respond stylistically or be more
susceptible to method bias. For example, Motivational factors that may cause biased
Krosnick (1999) summarizes research that responding. A second question to consider is
shows that respondents who are low in verbal whether respondents are motivated to provide
ability or education are more likely to respond accurate answers. Method biases and stylistic
in a nondifferentiated manner when asked to responding should be less likely to the extent
rate objects on a single response scale (i.e., that respondents are motivated to provide opti-
by giving all objects the same rating) and mal responses to the questions and more likely
that nondifferentiated responding is more to the extent that respondents are motivated
prevalent toward the end of a questionnaire to expend less effort by satiscing. Krosnick
due to fatigue. Similarly, there is evidence (e.g., (1999) notes several factors that increase a
Schwarz et al. 1992) that the amount of expe- respondents motivation to exert the cognitive
rience a respondent has had thinking about the effort required to provide optimal answers
topic of a question decreases his/her tendency including the need for cognition; the desire
to select the most recent of several response for self-expression, intellectual challenge,
alternatives mentioned (regardless of content), self-understanding, or emotional catharsis; and
presumably because it makes the respondents the desire to help employers improve working
knowledge of the topic more accessible. conditions, manufacturers produce better
Beyond this, Baumgartner & Steenkamp quality products, or governments make better-
(2001) note that the tendency to agree with informed policy decisions. To the extent that
items regardless of content (i.e., ARS) can respondents possess these needs/desires, they

may be more likely to expend the effort required to rebel against a testing procedure, or are
to generate an optimal answer, and the threat not motivated to invest the cognitive energy
of method bias should be lower. In contrast, re- required to read and interpret questionnaire
spondents may be motivated to minimize effort items ( Jackson 1967). Finally, researchers
when they feel the questions are unimportant; should also consider whether respondents are
believe their responses will not have useful likely to believe that two constructs are related
consequences; feel compelled to participate in by an implicit theory, because if they are, then
a survey to fulll a course requirement; become the respondents may be motivated to provide
fatigued by a seemingly unending stream of answers that are consistent with that theory.
questions; or dislike the interviewer, experi-
menter, or source of the survey. To the extent Task factors that may cause or facilitate
that these things are true, respondents may be biased responding. A third question that re-
more likely to minimize their effort and rely on searchers should consider is the impact of
stylistic tendencies or other decision heuristics the task on respondents. More specically, re-
to arrive at a merely satisfactory answer. searchers should evaluate the extent to which
In addition to considering the general respondents will have difculty generating ac-
factors that might motivate respondents to curate answers to the questions and the extent to
attempt to minimize effort by satiscing, which the measurement conditions may make
researchers should also consider aspects of the it easy for them to minimize their effort by re-
measurement conditions that might increase sponding in a stylistic manner. For example,
the threat of specic types of bias. For example, Doty & Glick (1998) argue that one reason why
researchers should consider the magnitude of Cote & Buckley (1987) found that some types
the social consequences of a respondents an- of measures contain more method variance than
swers and the extent to which the measurement others is that responding to complex, abstract
conditions make those consequences salient questions is a more difcult task for respondents
(see Paulhus 1984, Steenkamp et al. 2010). than answering simple, concrete questions. In
The more serious the social consequences of a addition, they note that complex, abstract ques-
particular response, the stronger a respondents tions are more likely to trigger social psycho-
desire to provide a socially acceptable response logical processes that increase the covariation
is likely to be. Similarly, the more that the mea- among the systematic error variance compo-
surement conditions threaten a respondents nents, thereby increasing the bias in the ob-
self-esteem, heighten his/her defensiveness, or served relationships between constructs (Doty
increase the benets (costs) of presenting a good & Glick 1998, p. 381).
(bad) impression, the more the respondent is Another task characteristic that makes
likely to be motivated to respond in a socially it more difcult for respondents to provide
desirable manner. Baumgartner & Steenkamp accurate responses is item ambiguity. Be-
(2001) suggest that researchers should also cause ambiguity makes respondents less certain
consider whether aspects of the measurement about how to accurately answer a question (e.g.,
context motivate respondents to conceal their Podsakoff et al. 2003), it increases the likeli-
true opinion by using the middle scale category hood that they will rely on their own stylistic
regardless of their true feelings (MRS) or by re- response tendencies to generate a merely
sponding to items carelessly, randomly, or non- satisfactory answer and increases the sensi-
purposefully (NCR). The former may happen tivity of their answers to context effects (see
because respondents become suspicious about Tourangeau et al. 2000). Consequently, when
how their data will be used (Schmitt 1994), and evaluating the potential threat of method bias,
the latter may happen because respondents are researchers should consider the extent to which
motivated to leave the testing situation, wish their questions fail to dene ambiguous or

unfamiliar terms, refer to vague concepts with- that decrease their motivation to respond
out providing clear examples, have complicated stylistically by increasing the effort required to
syntax, or are double-barreled. In addition, do so.
Krosnick (1991) notes that item ambiguity is The key thing that must be done to make
greater if only the end points of a response sure that respondents have the ability to an-
scale are labeled (rather than every point). swer questions accurately is to match the dif-
In contrast, rather than making the task of culty of the task of answering the questions
providing an accurate response more difcult, with the capabilities of the respondents. One
other aspects of the measurement context may obvious way to do this is to make sure that you
enhance the threat of method bias by making dont ask respondents to tell more than they
it easier to provide an alternative, merely satis- can know (Ericsson & Simon 1980, Nisbett &
factory, response. For example, it is easier for Wilson 1977). This can be avoided by exercis-
respondents to provide answers that are consis- ing caution when asking respondents about the
tent with each other or with an implicit theory motives for their behavior, the effects of situa-
if the answers to previous questions are readily tional factors on their behavior, or other things
available (physically or in memory) at the time pertaining to cognitive processes that they are
of answering a later question. This is likely to be unlikely to have attended to or stored in short-
the case in a self-administered paper and pencil term memory. Beyond this, researchers can de-
questionnaire and is often (but need not be) the crease the difculty of responding accurately
case for online questionnaires. This may also be by using clear and concise language, avoid-
the case when questions are grouped together ing complicated syntax, dening ambiguous or
in close proximity by construct on the ques- unfamiliar terms, not referring to vague con-
tionnaire. Alternatively, it seems plausible that cepts without providing clear examples, avoid-
ERS or MRS response styles would be easier ing double-barreled items, and labeling all scale
to implement if the measures were grouped to- points rather than just the end points.
gether by scale type, with the same number of Perhaps the easiest way to increase the
scale points, with common anchor labels, and probability that respondents will try to pro-
without any reversed item wording. vide accurate answers to the questions is by
developing a good cover story and instructions
(Aronson et al. 1998). For example, the desire
What Can Be Done To Mitigate for self-expression or emotional catharsis may
the Problem? be enhanced by explaining in the cover story or
Procedural remedies. Generally, studies instructions that we value your opinion, we
should be designed to maximize respondent need your feedback, or that we want respon-
motivation and ability and minimize task dents to tell us what you think. The tendency
difculty so that respondents are more likely to to respond in a socially desirable manner,
respond accurately. To increase the probability threats to self-esteem, and defensiveness may be
that respondents can provide accurate answers diminished through anonymity, telling respon-
to the questions, it is necessary to implement dents in the cover story or instructions there are
procedures that ensure that respondents have no right or wrong answers, and assuring them
the ability to answer the questions asked, that people have different opinions about the
decrease the difculty of responding accu- issues addressed in the questionnaire. The mo-
rately, and increase the difculty of responding tivation of respondents to provide accurate an-
stylistically. To increase the probability that swers may also be increased by explaining how
respondents will try to provide accurate the information will be used or how it will ben-
answers, it is necessary to implement not only et them or their organization (e.g., by men-
procedures that increase their motivation to tioning that the data will help their employer to
provide accurate answers, but also procedures improve working conditions or make their job

easier). Promising feedback to respondents may responding. However, if these conditions are
motivate them to respond more accurately so not met, this procedure may also introduce in-
that they can gain greater self-understanding. formation biases or attribution biases.
Motivation can also be increased through en- If separating the sources is not feasible or
dorsement of the study by senior management. desirable, another procedure that should help
Finally, motivation to respond accurately can be to diminish method bias is to separate the
maintained by keeping the questionnaire short measurement of the predictor and criterion
and minimizing redundancies to the extent constructs temporally, methodologically, or
possible. However, because multiple measures psychologically. Temporal separation involves
of the same construct are usually essential, the introducing a time lag between measurement
best approach may be to vary the wording of of the predictor and criterion variables. This
the items rather than just using synonyms. procedure is appropriate if (a) the phenomenon
In addition to increasing the motivation to is not ephemeral, short lived, or rapidly chang-
respond accurately, it is also important to de- ing; (b) the phenomenon is based on long-term
crease the motivation to respond stylistically by (rather than short-term) memory effects;
increasing the effort required to do so. This (c) a signicant amount of respondent attrition
can be done in several ways. The rst is by re- is not likely to occur; and (d ) it is nancially
versing the wording of some of the items to and logistically feasible. To be effective, it is
balance the positively and negatively worded important for the temporal delay to be long
items. Of course, this is only a good idea if it enough to produce forgetting, clear short-term
can be done without altering the content va- memory, or to disassociate cues in the two
lidity or conceptual meaning of the scale and measurement occasions.
if the reverse-worded items are not confusing Methodological separation involves having
to respondents. A second way is by separating respondents complete the measurement of the
items on the questionnaire to eliminate prox- predictor variable under different methodolog-
imity effects. However, this may not be feasible ical conditions than the criterion variables.
if the questionnaire is too short. A third way is For example, researchers can use different
by varying the scale types and anchor labels to scale properties, response modes, and data
the extent that it is conceptually appropriate. collection locations for the predictor and crite-
Of course, the procedures outlined above rion measures, or they can physically separate
are not likely to fully control for every type the predictor and criterion measures on the
of method bias. For example, it is unlikely questionnaire. Methodological separation is
that self-deception biases, memory biases (e.g., appropriate provided that varying the scale
things that were recently activated are more properties or response mode does not alter
accessible), or perceptual biases (e.g., Gestalt the conceptual meaning of the measures and
principles of perception) would be controlled that the questionnaire is of sufcient length
by these efforts. To the extent that these things to separate the measures. This can diminish
are a concern, try to obtain the measures of the method bias by increasing the difculty of re-
predictor and criterion constructs from differ- sponding stylistically, eliminating the saliency
ent sources. This is most easily done if there of any contextually provided retrieval cues,
are multiple observers of the phenomenon of and/or reducing the respondents ability to
interest who have access to the same infor- use previous answers to ll in gaps in what is
mation and if the phenomenon is not self- recalled or to use prior responses to answer
referential (Chan in Brannick et al. 2010). subsequent questions (Podsakoff et al. 2003).
Under these circumstances, separating sources The measures of the predictor and criterion
should help to diminish the effects of involun- variables can be separated psychologically by
tary memory-based and perceptual biases and using a multiple study cover story, camou-
may help to reduce the biasing effects of stylistic aging interest in the criterion or predictor

variable (i.e., by embedding it in the context of CFA marker technique or the common method
other questions so that it is less psychologically factor technique because these approaches con-
prominent), or disguising the reasons for ob- trol for measurement error, even though they
taining the predictor or criterion measure (i.e., do not clearly specify the nature of the method
by leading respondents to believe that it is tan- bias. The CFA marker technique requires the
gential to the main purpose of the study). These researcher to include appropriate marker vari-
procedures diminish method biases by reduc- ables that are theoretically unrelated to any of
ing the perceived diagnosticity of responses the measures of the focal constructs of interest
to the measures of the predictor variable as in the questionnaire. The common method fac-
cues for how to respond to the measures of the tor technique does not require the inclusion of
criterion variable (cf. Feldman & Lynch 1988). any additional measures, but it is problematic
However, psychological separation is unlikely because it may capture irrelevant trait variance
to diminish biases due to the accessibility of in addition to systematic method variance.
responses to the measures of the predictor A nal technique that could be used to
variable in memory. For this reason, it may be control statistically for method biases is the
wise to use this procedure in conjunction with instrumental variable technique. Although
a temporal separation long enough to clear it provides no insight into the nature of the
short-term memory. Of course, implementing method bias, if it could be properly imple-

this procedure (i.e., psychological separation) mented it would be effective. However, as we
is contingent upon the researchers ability to noted earlier, it is extremely difcult to identify
create a credible cover story, and it is only use- instrumental variables that are strongly related
ful if the means of producing the psychological to the endogenous predictor variables but
separation does not cause a temporal delay that completely uncorrelated with the structural
is longer than the phenomenon of interest. error term for the equation. This is a serious
barrier to implementing this technique because
Statistical remedies. In situations where if these requirements are not met, this tech-
method bias is still an important concern, even nique can produce biased estimates and inate
after implementing procedural methods of con- the type I error rates; even a slight correlation
trol, we recommend that researchers follow this between a weak IV and the structural error
up with appropriate statistical remedies. More term for the equation can cause the IV estimate
specically, we recommend that researchers to exhibit more bias (even asymptotically) than
rst try to use the directly measured latent fac- an OLS estimate (Kennedy 2008). Therefore,
tor technique or the measured response style although the use of this technique to control
technique because both of these techniques for method bias is possible in principle, it may
control for measurement error and specify the be difcult to put into practice.
nature of the method bias. The former would be
used if a researcher is concerned about a partic- Additional approaches. Two nal ap-
ular source of bias for which a valid measure of proaches that might help rule out method bias
the biasing factor is available or could be devel- as a rival explanation for a studys ndings have
oped. The latter would be used if a researcher is been identied recently. The rst alternative
concerned with the biasing effects of response approach is based on the simulation ndings
styles (e.g., ARS, ERS). In this case, we rec- of Evans (1985) and a proof by Siemsen et al.
ommend following the guidelines outlined in (2010), which demonstrate that although
Weijters et al. (2008). method bias can inate (or deate) bivariate
If the specic source of the method bias is linear relationships, it cannot inate (but
unknown or valid measures of the source of bias does deate) quadratic and interaction effects.
are not available, then we recommend using the Consequently, if a study is designed to test

hypotheses about quadratic or interaction CONCLUSION

effects, rather than main effects, then method The purpose of this article has been to review
bias would not be able to account for any sta- the current state of knowledge about method
tistically signicant effects observed. Although biases. Our review indicates that although there
this may not be possible or desirable in many is some disagreement about the way method
instances, in those cases where it is conceptually and method biases are dened, the evidence
appropriate and possible, it may be a reasonable shows that method biases can signicantly in-
alternative to the procedural and statistical uence item validities and reliabilities as well as
remedies described above. The second alter- the covariation between latent constructs. This
native approach, recently suggested by Chan suggests that researchers must be knowledge-
(in Brannick et al. 2010), is to (a) identify one able about the ways to control method biases
or more potential sources of method bias, that might be present in their studies. Conse-
(b) manipulate them in the design of the study, quently, we recommend procedural and statis-
and (c) test whether the hypothesized estimates
tical remedies that can be used to achieve this

of the relationships among the constructs control. Although space constraints prevent us
generalize across conditions. Importantly, from addressing all of the issues regarding this
Chan notes that when used in combination important topic, we hope that we have pro-
with the statistical techniques described above, vided some recommendations that researchers
this method provides a powerful means of can use to deal with the detrimental effects of
detecting and controlling method bias. method biases in their research.
DISCLOSURE STATEMENT
The authors are unaware of any afliation, funding, or nancial holdings that might be perceived
as affecting the objectivity of this review.
LITERATURE CITED
Alessandri G, Vecchione M, Fagnani C, Bentler PM, Barbaranelli C, et al. 2010. Much more than model
tting? Evidence for the heritability of method effect associated with positively worded items of the life
orientation test revised. Struct. Equ. Model. 17:64253
Antonakis J, Bendahan S, Jacquart P, Lalive R. 2010. On making causal claims: a review and recommendations.
Leadersh. Q. 6:108620
Aronson E, Wilson TD, Brewer MB. 1998. Experimentation in social psychology. In The Handbook of Social
Psychology, ed. DT Gilbert, ST Fiske, G Lindzey, Vol. 1, pp. 99142. Boston, MA: McGraw-Hill. 4th ed.
Arora R. 1982. Validation of an S-O-R model for situation, enduring, and response components of involvement.
J. Mark. Res. 19:50516
Bagozzi RP. 1984. A prospectus for theory construction in marketing. J. Mark. 48:1129
Bagozzi RP. 1993. Assessing construct-validity in personality research: applications to measures of self-esteem.
J. Res. Personal. 27:4987
Bagozzi RP, Phillips LW. 1982. Representing and testing organizational theoriesa holistic construal. Admin.
Sci. Q. 27:45989
Bagozzi RP, Yi Y. 1990. Assessing method variance in multitrait-multimethod matrices: the case of self-
reported affect and perceptions at work. J. Appl. Psychol. 75:54760
Baumgartner H, Steenkamp JBEM. 2001. Response styles in marketing research: a cross-national investigation.
J. Mark. Res. 38:14356
Billiet JB, McClendon MJ. 2000. Modeling acquiescence in measurement models for two balanced sets of
items. Struct. Equ. Model. 7:60828

Bodner TE. 2006. Designs, participants, and measurement methods in psychological research. Can. Psychol.
47:26372
Bollen KA. 1989. Structural Equations with Latent Variables. New York: Wiley
Brannick MT, Chan D, Conway JM, Lance CE, Spector PE. 2010. What is method variance and how can we
cope with it? A panel discussion. Organ. Res. Methods 13:40720
Brannick MT, Spector PE. 1990. Estimation problems in the block-diagonal model of the multitrait-
multimethod matrix. Appl. Psychol. Meas. 14:32539
Buckley MR, Cote JA, Comstock SM. 1990. Measurement errors in the behavioral sciences: the case of
personality attitude research. Educ. Psychol. Meas. 50:44774
Campbell DT, Fiske D. 1959. Convergent and discriminant validation by the multitrait-multimethod matrix.
Psychol. Bull. 56:81105
Campbell DT, OConnell EJ. 1967. Methods factors in multitrait-multimethod matrices: multiplicative rather
than additive? Multivar. Behav. Res. 2:40926
Chang S-J, van Wittleloostuijn A, Eden L. 2010. From the editors: common method variance in international
business research. J. Int. Bus. Stud. 41:17884
Chen PY, Spector PE. 1991. Negative affectivity as the underlying cause of correlations between stressors and
strains. J. Appl. Psychol. 76:398407
Cote JA, Buckley R. 1987. Estimating trait, method, and error variance: generalizing across 70 construct
validation studies. J. Mark. Res. 24:31518

Cote JA, Buckley R. 1988. Measurement error and theory testing in consumer research: an illustration of the
importance of construct validation. J. Consum. Res. 14:57982
Cronbach LJ. 1946. Response sets and test validity. Educ. Psychol. Meas. 6:47594
Cronbach LJ. 1950. Further evidence on response sets and test validity. Educ. Psychol. Meas. 10:331
De Beuckelaer A, Weijters B, Rutten A. 2010. Using ad hoc measures for response styles: a cautionary note.
Qual. Quant. 44:76175
Doty DH, Glick WH. 1998. Common methods bias: Does common methods variance really bias results?
Organ. Res. Methods 1:374406
Edwards AL. 1970. The Measurement of Personality Traits by Scales and Inventories. New York: Holt, Rinehart
& Winston
Edwards JR. 2008. To prosper, organizational psychology should . . . overcome methodological barriers to
progress. J. Organ. Behav. 29:46991
Ericsson KA, Simon HA. 1980. Verbal reports as data. Psychol. Rev. 87:21557
Evans MG. 1985. A Monte Carlo study of the effects of correlated method variance in moderated multiple
regression analysis. Organ. Behav. Hum. Decis. Process. 36:30523
Feldman JM, Lynch JG. 1988. Self-generated validity and other effects of measurement on belief, attitude,
intention, and behavior. J. Appl. Psychol. 73:42135
Fiske DW. 1982. Convergent-discriminant validation in measurements and research strategies. In Forms of
Validity in Research, ed. D Brinbirg, LH Kidder, pp. 7792. San Francisco, CA: Jossey-Bass
Flamer S. 1983. Assessment of the multitrait-multimethod matrix validity of Likert scales via conrmatory
factor analysis. Multivar. Behav. Res. 18:275308
Harris MM, Bladen A. 1994. Wording effects in the measurement of role conict and role ambiguity: a
multitrait-multimethod analysis. J. Manage. 20:887901
Harrison DA, McLaughlin ME, Coalter TM. 1996. Context, cognition, and common method variance: psy-
chometric and verbal protocol evidence. Organ. Behav. Hum. Decis. Process. 68:24661
Higgins ET, Rholes WS, Jones CR. 1977. Category accessibility and impression formation. J. Exp. Soc. Psychol.
13:14154
Hulsheger UR, Anderson N, Salgado JF. 2009. Team-level predictors of innovation at work: a comprehensive
meta-analysis spanning three decades of research. J. Appl. Psychol. 94:112845
Jackson DN. 1967. Acquiescence response styles: problems of identication and control. In Response Set in
Personality Assessment, ed. IA Berg, pp. 71114. Chicago: Aldine
Johnson JA. 2004. The impact of item characteristics on item and scale validity. Multivar. Behav. Res. 39:273
302

Johnson RE, Rosen CC, Djurdevic E. 2011. Assessing the impact of common method variance on higher-order
multidimensional constructs. J. Appl. Psychol. 96:74461
Joresk
og
KG. 1998. Interaction and nonlinear modeling: issues and approaches. In Interaction and Nonlinear
Effects in Structural Equation Modeling, ed. RE Schumacker, GA Marcoulides, pp. 23950. Mahwah, NJ:
Erlbaum
Kennedy P. 2008. A Guide to Econometrics. Malden, MA: Blackwell. 6th ed.
Kozlowski S. 2009. Editorial. J. Appl. Psychol. 94:14
Kothandapani V. 1971. Validation of feeling, belief, and intention to act as three components of attitude and
their contribution to prediction of contraceptive behavior. J. Personal. Soc. Psychol. 19:32133
Krosnick JA. 1991. Response strategies for coping with the cognitive demands of attitude measures in surveys.
Appl. Cogn. Psychol. 5:21336
Krosnick JA. 1999. Survey research. Annu. Rev. Psychol. 50:53767
Kuncel NR, Tellegen A. 2009. A conceptual and empirical reexamination of the measurement of the social
desirability of items: implications for detecting desirable response style and scale development. Pers.
Psychol. 62:201228
Lance CE, Baranik LE, Lau AR, Scharlau EA. 2009. If it aint trait it must be method: (mis)application of
the multitrait-multimethod design in organizational research. In Statistical and Methodological Myths and
Urban Legends: Doctrine, Verity, and Fable in the Organizational and Social Sciences, ed. CE Lance, RL
Vandenberg, pp. 33760. New York: Routledge

Lance CE, Dawson B, Birkelbach D, Hoffman BJ. 2010. Method effects, measurement error, and substantive
conclusions. Organ. Res. Methods 13:40720
Le H, Schmidt FL, Putka DJ. 2009. The multifaceted nature of measurement artifacts and its implications for
estimating construct-level relationships. Organ. Res. Methods 12:165200
Lindell MK, Whitney DJ. 2001. Accounting for common method variance in cross-sectional designs. J. Appl.
Psychol. 86:11421
Lowe KB, Kroeck KG, Sivasubramaniam N. 1996. Effectiveness correlates of transformational and transac-
tional leadership: a meta-analytic review of the MLQ literature. Leadersh. Q. 7:385425
MacKenzie SB, Podsakoff PM, Podsakoff NP. 2011. Construct measurement and validity assessment in be-
havioral research: integrating new and existing techniques. MIS Q. 35: 293334
Messick S. 1991. Psychology and methodology of response styles. In Improving the Inquiry in Social Science: A
Volume in Honor of Lee J. Cronbach, ed. RE Snow, DE Wiley, pp. 161200. Hillsdale, NJ: Erlbaum
Mirowsky J, Ross CE. 1991. Eliminating defense and agreement bias from measures of the sense of control:
a 2 2 index. Soc. Psychol. Q. 54:12745
Nisbett RE, Wilson TD. 1977. Telling more than we can know: verbal reports on mental processes. Psychol.
Rev. 84:23159
Nunally JC, Bernstein IH. 1994. Psychometric Theory. New York: McGraw-Hill. 3rd ed.
Ostroff C, Kinicki AJ, Clark MA. 2002. Substantive and operational issues of response bias across levels of
analysis: an example of climate-satisfaction relationships. J. Appl. Psychol. 87:35568
Paulhus DL. 1984. Two-component models of socially desirable responding. J. Personal. Soc. Psychol. 46:598
609
Phillips JS, Lord RG. 1986. Notes on the theoretical and practical consequences of implicit leadership theories
for the future of leadership measurement. J. Manage. 12:3141
Podsakoff PM, MacKenzie SB, Lee J-Y, Podsakoff NP. 2003. Common method biases in behavioral research:
a critical review of the literature and recommended remedies. J. Appl. Psychol. 88:879903
Podsakoff PM, Organ DW. 1986. Self-reports in organizational researchproblems and prospects. J. Manage.
12:53144
Rafferty AE, Grifn MA. 2004. Dimensions of transformational leadership: conceptual and empirical exten-
sions. Leadersh. Q. 15:32954
Rafferty AE, Grifn MA. 2006. Rening individualized consideration: distinguishing developmental leadership
and supportive leadership. J. Occup. Organ. Psychol. 79:3761
Richardson HA, Simmering MJ, Sturman MC. 2009. A tale of three perspectives: examining post hoc statistical
techniques for detection and correction of common method variance. Organ. Res. Methods 12:762800

Rindeisch A, Malter AJ, Ganesan S, Moorman C. 2008. Cross-sectional versus longitudinal research: con-
cepts, ndings, and guidelines. J. Mark. Res. 45:26179
Rosenberg MJ. 1965. When dissonance fails: on eliminating evaluation apprehension from attitude measure-
ment. J. Personal. Soc. Psychol. 1:2842
Scherpenzeel A, Saris W. 1997. The validity and reliability of survey questions: a meta-analysis of MTMM
studies. Soc. Methods Res. 25:34183
Schaubroeck J, Ganster DC, Fox ML. 1992. Dispositional affect and work-related stress. J. Appl. Psychol.
77:32235
Schmitt N. 1994. Method bias: the importance of theory and measurement. J. Organ. Behav. 15:39398
Schwarz N, Hippler HJ, Noelle-Neumann E. 1992. A cognitive model of response-order effects in survey
measurement. In Context Effects in Social and Psychological Research, ed. N Schwarz, S Sudman, pp. 187
201. New York: Springer-Verlag
Sechrest L, Davis MF, Stickle TR, McKnight PE. 2000. Understanding method variance. In Research Design:
David Campbells Legacy, ed. L Bickman, Vol. 2, pp. 6388. Thousand Oaks, CA: Sage
Sharma R, Yetton P, Crawford J. 2009. Estimating the effect of common method variance: the method-method
pair technique with an illustration from TAM research. MIS Q. 33:47390

Siemsen E, Roth A, Oliveira P. 2010. Common method bias in regression models with linear, quadratic, and
interaction effects. Organ. Res. Methods 13:45676
Spector PE. 1987. Method variance as an artifact in self-reported affect and perceptions at work: myth or
signicant problem. J. Appl. Psychol. 72:43843

Spector PE. 2006. Method variance in organizational research: truth or urban legend? Organ. Res. Methods
9:22132
Spector PE, Brannick MT. 2009. Common method variance or measurement bias? The problem and possible
solutions. In The Sage Handbook of Organizational Research Methods, ed. DA Buchanan, A Bryman, pp. 346
62. Los Angeles, CA: Sage
Steenkamp JBEM, De Jong MG, Baumgartner H. 2010. Socially desirable response tendencies in survey
research. J. Mark. Res. 47:199214
Straub DW. 2009. Creating blue oceans of thought via highly citable articles. MIS Q. 33:iiivii
Sudman S, Bradburn NM, Schwarz N. 1996. Thinking About Answers: The Application of Cognitive Processes to
Survey Methodology. San Francisco, CA: Jossey-Bass
Tourangeau R, Rips LJ, Rasinski KA. 2000. The Psychology of Survey Response. London: Cambridge Univ. Press
Tourangeau R, Singer E, Presser S. 2003. Context effects in attitude surveyseffects on remote items and
impact on predictive validity. Soc. Methods Res. 31:486513
Weijters B, Cabooter E, Schillewaert N. 2010c. The effect of rating scale format on response styles: the
number of response categories and response category labels. Int. J. Mark. 27:23647
Weijters B, Geuens M, Schillewaert N. 2009. The proximity effect: the role of inter-item distance on reverse-
item bias. Int. J. Res. Market. 26:212
Weijters B, Geuens M, Schillewaert N. 2010a. The stability of individual response styles. Psychol. Methods
15:96110
Weijters B, Geuens M, Schillewaert N. 2010b. The individual consistency of acquiescence and extreme
response style in self-report questionnaires. Appl. Psychol. Meas. 34:10521
Weijters B, Schillewaert N, Geuens M. 2008. Assessing response styles across modes of data collection.
J. Acad. Mark. Sci. 36:40922
Werts CE, Linn RL, Joreskog KG. 1974. Intraclass reliability estimatestesting structural assumptions. Educ.
Psychol. Meas. 34:2533
Widaman KF. 1985. Hierarchically nested covariance structural models for multitrait-multimethod data. Appl.
Psychol. Meas. 9:126
Williams LJ, Anderson SE. 1994. An alternative approach to method effects by using latent-variable models:
applications in organizational behavior research. J. Appl. Psychol. 79:32331
Williams LJ, Cote JA, Buckley MR. 1989. Lack of method variance in self-reported affect and perceptions at
work: reality or artifact? J. Appl. Psychol. 74:46268
Williams LJ, Gavin MB, Williams ML. 1996. Measurement and nonmeasurement processes with negative
affectivity and employee attitudes. J. Appl. Psychol. 81:88101

Williams LJ, Hartman N, Cavazotte F. 2010. Method variance and marker variables: a review and compre-
hensive CFA marker technique. Organ. Res. Methods 13:477514
Woszczynski AB, Whitman ME. 2004. The problem of common method variance in IS research. In Handbook
of Information Systems Research, ed. ME Whitman, AB Woszczynski, pp. 6677. Hershey, PA: Idea Group
Wothke W, Browne MW. 1990. The direct product model for the MTMM matrix parameterized as a second
order factor analysis model. Psychometrika 55:25562
Zinkhan GM. 2006. Research traditions and patterns in marketing scholarship. J. Acad. Mark. Sci. 34:28183

PS63-FrontMatter ARI 10 November 2011 9:52
Annual Review of
Psychology
Volume 63, 2012 Contents
Prefatory
Working Memory: Theories, Models, and Controversies
Alan Baddeley p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 1
Developmental Psychobiology
Learning to See Words
Brian A. Wandell, Andreas M. Rauschecker, and Jason D. Yeatman p p p p p p p p p p p p p p p p p p p p p31

Memory
Remembering in Conversations: The Social Sharing
and Reshaping of Memories
William Hirst and Gerald Echterhoff p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p55
Judgment and Decision Making
Experimental Philosophy
Joshua Knobe, Wesley Buckwalter, Shaun Nichols, Philip Robbins,
Hagop Sarkissian, and Tamler Sommers p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p81
Brain Imaging/Cognitive Neuroscience
Distributed Representations in Memory: Insights from Functional
Brain Imaging
Jesse Rissman and Anthony D. Wagner p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 101
Neuroscience of Learning
Fear Extinction as a Model for Translational Neuroscience:
Ten Years of Progress
Mohammed R. Milad and Gregory J. Quirk p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 129
Comparative Psychology
The Evolutionary Origins of Friendship
Robert M. Seyfarth and Dorothy L. Cheney p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 153
Emotional, Social, and Personality Development
Religion, Morality, Evolution
Paul Bloom p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 179
vi
Adulthood and Aging

Consequences of Age-Related Cognitive Declines
Timothy Salthouse p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 201
Development in Societal Context
Child Development in the Context of Disaster, War, and Terrorism:
Pathways of Risk and Resilience
Ann S. Masten and Angela J. Narayan p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 227
Social Development, Social Personality, Social Motivation, Social Emotion
Social Functionality of Human Emotion
Paula M. Niedenthal and Markus Brauer p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 259
Social Neuroscience
Mechanisms of Social Cognition
Chris D. Frith and Uta Frith p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 287
Personality Processes
Personality Processes: Mechanisms by Which Personality Traits
Get Outside the Skin
Sarah E. Hampson p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 315
Work Attitudes
Job Attitudes
Timothy A. Judge and John D. Kammeyer-Mueller p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 341
The Individual Experience of Unemployment
Connie R. Wanberg p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 369
Job/Work Analysis
The Rise and Fall of Job Analysis and the Future of Work Analysis
Juan I. Sanchez and Edward L. Levine p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 397
Education of Special Populations
Rapid Automatized Naming (RAN) and Reading Fluency:
Implications for Understanding and Treatment of Reading Disabilities
Elizabeth S. Norton and Maryanne Wolf p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 427
Human Abilities
Intelligence
Ian J. Deary p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 453
Research Methodology
Decoding Patterns of Human Brain Activity
Frank Tong and Michael S. Pratte p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 483
Contents vii
Human Intracranial Recordings and Cognitive Neuroscience

Roy Mukamel and Itzhak Fried p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 511
Sources of Method Bias in Social Science Research
and Recommendations on How to Control It
Philip M. Podsakoff, Scott B. MacKenzie, and Nathan P. Podsakoff p p p p p p p p p p p p p p p p p p p p 539
Neuroscience Methods
Neuroethics: The Ethical, Legal, and Societal Impact of Neuroscience
Martha J. Farah p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 571
Indexes
Cumulative Index of Contributing Authors, Volumes 5363 p p p p p p p p p p p p p p p p p p p p p p p p p p p 593

Cumulative Index of Chapter Titles, Volumes 5363 p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 598
Errata
An online log of corrections to Annual Review of Psychology articles may be found at

http://psych.AnnualReviews.org/errata.shtml
viii Contents

3 - PodsakoffARP2012

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

3 - PodsakoffARP2012

Transféré par

Droits d'auteur :

Formats disponibles

PS63CH21-Podsakoff ARI 31 October 2011 13:29

Our comprehensive search

Philip M. Podsakoff,1 Scott B. MacKenzie,2

and Nathan P. Podsakoff 3

Annu. Rev. Psychol. 2012. 63:53969 Keywords

This articles doi: Abstract

traits (constructs). The test, or rating scale,

Item Reliability or Validity . . . . . . 543

540 Podsakoff MacKenzie Podsakoff

& Brannick 2009, Williams et al. 1989). Is it Fiske (1982, p. 82),

www.annualreviews.org Method Bias in Research 541

542 Podsakoff MacKenzie Podsakoff

Does Method Bias Affect Effects of General Method Bias

www.annualreviews.org Method Bias in Research 543

544 Podsakoff MacKenzie Podsakoff

variables x and y (Rxy ) is equal to implies a true Rmk,ml of 1.00). Indeed, if a

www.annualreviews.org Method Bias in Research 545

Table 2 Summary of meta-analytic studies comparing same-source versus different-source relationships

95 2,808 0.490 0.595 56 2,889 0.260 0.323 184%

factors on responses to individual measures citizenship behaviors (OCBs), participative

546 Podsakoff MacKenzie Podsakoff

www.annualreviews.org Method Bias in Research 547

correlation between outcome favorability dispositional mood states, and tendencies on

548 Podsakoff MacKenzie Podsakoff

www.annualreviews.org Method Bias in Research 549

may be questionable. (i.e., diminishing the salience of the measure) or

550 Podsakoff MacKenzie Podsakoff

www.annualreviews.org Method Bias in Research 551

552 Podsakoff MacKenzie Podsakoff

www.annualreviews.org Method Bias in Research 553

554 Podsakoff MacKenzie Podsakoff

www.annualreviews.org Method Bias in Research 555

556 Podsakoff MacKenzie Podsakoff

proportion due to the substantive construct no constraints on their theoretical relationships

www.annualreviews.org Method Bias in Research 557

558 Podsakoff MacKenzie Podsakoff

constructs. De Beuckelaer et al. (2010) found measure the response styles.

www.annualreviews.org Method Bias in Research 559

560 Podsakoff MacKenzie Podsakoff

www.annualreviews.org Method Bias in Research 561

562 Podsakoff MacKenzie Podsakoff

www.annualreviews.org Method Bias in Research 563

short-term memory. Of course, implementing method bias, if it could be properly imple-

564 Podsakoff MacKenzie Podsakoff

hypotheses about quadratic or interaction CONCLUSION

tical remedies that can be used to achieve this

www.annualreviews.org Method Bias in Research 565

validation studies. J. Mark. Res. 24:31518

566 Podsakoff MacKenzie Podsakoff

Vandenberg, pp. 33760. New York: Routledge

www.annualreviews.org Method Bias in Research 567

pair technique with an illustration from TAM research. MIS Q. 33:47390

signicant problem. J. Appl. Psychol. 72:43843

568 Podsakoff MacKenzie Podsakoff

www.annualreviews.org Method Bias in Research 569

Volume 63, 2012 Contents

Brian A. Wandell, Andreas M. Rauschecker, and Jason D. Yeatman p p p p p p p p p p p p p p p p p p p p p31

Adulthood and Aging

Human Intracranial Recordings and Cognitive Neuroscience

Cumulative Index of Contributing Authors, Volumes 5363 p p p p p p p p p p p p p p p p p p p p p p p p p p p 593

An online log of corrections to Annual Review of Psychology articles may be found at

Vous aimerez peut-être aussi