Australasian Marketing Journal 25 (2017) 76–81

Contents lists available at ScienceDirect

Australasian Marketing Journal

j o u r n a l h o m e p a g e : w w w. e l s e v i e r. c o m / l o c a t e / a m j

Commentary

The analysis of mechanisms and their contingencies:

PROCESS versus structural equation modeling
Andrew F. Hayes *, Amanda K. Montoya, Nicholas J. Rockwood
Department of Psychology, The Ohio State University, Columbus, Ohio, USA

A R T I C L E I N F O A B S T R A C T

Article history: Marketing, consumer, and organizational behavior researchers interested in studying the mechanisms by
Available online 23 February 2017 which effects operate and the conditions that enhance or inhibit such effects often rely on statistical me-
diation and conditional process analysis (also known as the analysis of “moderated mediation”). Model
estimation is typically undertaken with ordinary least squares regression-based path analysis, such as imple-
mented in the popular PROCESS macro for SPSS and SAS (Hayes, 2013), or using a structural equation modeling
program. In this paper we answer a few frequently-asked questions about the difference between PROCESS
and structural equation modeling and show by way of example that, for observed variable models, the choice
of which to use is inconsequential, as the results are largely identical. We end by discussing consider-
ations to ponder when making the choice between PROCESS and structural equation modeling.
© 2017 Australian and New Zealand Marketing Academy. Published by Elsevier Ltd. All rights reserved.

C H I N E S E A B S T R A C T

营销、消费者和组织行为研究人员对研究这种影响经营的机制非常感兴趣，增强或抑制这种影响的条件通常依赖
于统计调节和在一定条件下的处理分析（也叫做分析“适度调节”）。 模型评估通常采用普通最小乘法基于回归
的路径分析（例如，在 SPSS 和 SAS 深受青睐的 PROCESS 宏中实现[Hayes, 2013]）或采用结构方程模型方案。 本
文回答了一些有关 PROCESS 和结构方程模型之间区别的常见问题，并举例说明，对于观察的变量模型，选择使用
哪种模型都无关紧要，因为其结果都大同小异。本文结尾讨论了在PROCESS和结构方程模型之间进行选择时应考虑哪些因素。
© 2017 Australian and New Zealand Marketing Academy. Published by Elsevier Ltd. All rights
reserved.

Marketing researchers and those who study organizational or Less common but growing in frequency are mediation models
consumer behavior strive to understand how marketing and other that allow for moderation of a mechanism, what Hayes (2013) calls
organizational effects operate, meaning the underlying cognitive, a conditional process model. Fig. 1, panels B, C, and D, represent a
social, and biological processes that intervene between a stimulus few conditional process models, also known as moderated media-
(e.g., a particular kind of packaging or promotion, or the manage- tion models. Panel A is a first stage conditional process model that
ment style of a leader) and a response (e.g., the evaluation of a allows the effect of X on M in a mediation model to depend on vari-
product, a decision or timing to purchase, or employee turnover at able W. The moderator, W, could be anything that influences or
a company). Mediation analysis is a popular statistical procedure changes the effect of X on M. For some examples, see Voola et al.
for testing hypotheses about the mechanisms by which a causal effect (2012), White et al. (2016), Shen et al. (2016), and Zenker et al.
operates. A mediation model contains at least one mediator vari- (2017). But if the moderation operates on the second stage of a me-
able M that is causally between X and Y, such that X’s effect on Y is diation process (i.e., on the effect of M on Y), as in Cassar and Briner
transmitted through the joint causal effect of X on M which in turn (2011) and Dubois et al. (2016), the result is a second stage condi-
affects Y. Fig. 1, panel A, depicts a mediation model with two me- tional process model, as in Fig. 1, panel C. If the same moderator
diators. Some examples found in the pages of Australasian Marketing influences the relationship between X and M and M and Y (Fig. 1,
Journal include Kongarchapatara and Shannon (2016), Baxter and panel D), this is a first and second stage conditional process model.
Kleinaltenkamp (2015), and Schiele and Vos (2015). Such models Examples include Shenu-Fen et al. (2012) and Etkin and Sela (2016).
are commonplace in the empirical literature. These represent only three of the many ways that mediation and
moderation can be integrated into a unified model.
Each of the models depicted in Fig. 1 looks like a path diagram,
* Corresponding author. Fax: +1 614 292 6798. with variables connected with unidirectional arrows. Such dia-
E-mail address: [email protected] (A.F. Hayes). grams, for most researchers, bring to mind structural equation

http://dx.doi.org/10.1016/j.ausmj.2017.02.001
1441-3582/© 2017 Australian and New Zealand Marketing Academy. Published by Elsevier Ltd. All rights reserved.
 A.F. Hayes et al. / Australasian Marketing Journal 25 (2017) 76–81 77

A M1
B
W M

X Y

X Y
M2

C D
M W W M W

X Y X Y

Fig. 1. A multiple mediator model (panel A) and three conditional process models (panels B, C, and D).

modeling (SEM) as the proper analytical strategy. Yet most of the moderation, and conditional process analysis with observed
guidance offered by methodologists in the last 10 years or so on how (i.e., “manifest”) variables. It was launched with the publication of
to test the contingencies of mechanisms (i.e., whether “mediation Hayes (2013) and can be downloaded at no charge from
is moderated”) is framed in terms of ordinary regression-based path www.processmacro.org. Based on a set of conceptual and statisti-
analysis principles (e.g., Edwards and Lambert, 2007; Fairchild and cal diagrams defined by a model number, the user chooses a model
MacKinnon, 2009; Hayes, 2015; Muller et al., 2005; Preacher et al., preprogrammed into PROCESS corresponding to the model he or she
2007). Tools written for software frequently used by business and wants to estimate. Arguments are provided to the macro about what
marketing researchers (such as SPSS and SAS) that do all the nec- variables are serving which roles in the model (i.e., independent vari-
essary computations have made applying these methods rather able, dependent variable, mediator, moderator, covariate), and
painless. The PROCESS macro introduced by Hayes (2013) has PROCESS estimates all the path coefficients, standard errors, t- and
become especially popular in business and marketing (and many p-values, confidence intervals, and various other statistics.
other fields as well), as evidenced by its appearance in a variety of Except in models that contain only a moderation component,
business journals and research presented at academic conferences. every model that PROCESS estimates requires at least two regres-
We frequently get questions about how PROCESS works, what sion equations. PROCESS uses ordinary least squares regression to
it is doing, and what it can and cannot do.1 One category of these estimate the parameters of each of the equations, a common prac-
questions involves the differences between what PROCESS does and tice in observed variable path analysis. For instance, the model in
what an SEM program does and if it matters whether one tests a Fig. 1, panel A, requires three equations (one for each mediator M1
mediation or conditional process model using PROCESS or SEM. Some and M2, and one for Y), whereas the models in Fig. 1 panels B, C,
of these questions are motivated by PROCESS users who have been and D each require only two regression equations (one for M and
told by reviewers or editors that they should or must use SEM, and one for Y). PROCESS estimates each equation separately, meaning
they are not sure how to respond, or they wonder whether they have that the estimation of the regression parameters in one of the equa-
done something wrong. This short piece addresses these ques- tions has no effect on the estimation of the parameters in any other
tions. Previous publications have discussed some of the issues we equations defining the model. Regardless of how many equations
raise (Iacobucci et al., 2007; Pek and Hoyle, 2016), but without the are needed, once the PROCESS macro is activated, one line of SPSS
focus on PROCESS that is unique to our treatment. We first briefly or SAS code is all that is required to estimate the model, which makes
overview what PROCESS is and how it differs from what an SEM it a very simple and user-friendly modeling system. SPSS users can
program does. We then show by way of example (though the lessons also set up the model using a convenient point-and-click interface
learned by this example generalize beyond it) that for an ob- by installing an optional PROCESS dialog menu into SPSS.
served variable model (i.e., no latent variables), it makes little PROCESS is not needed to estimate the parameters of the re-
difference whether PROCESS or an SEM program is used. We then gression equations, as this can be done with any least squares
discuss some reasons why one might choose SEM over PROCESS. regression program (such as SPSS’s REGRESSION command or PROC
REG in SAS) and the results will be identical. But in mediation and
conditional process analysis, many important statistics useful for
1. What is PROCESS and how does it differ from SEM? testing hypotheses, such as conditional indirect effects and the index
of moderated mediation, require the combination of parameter es-
PROCESS is a computational tool — a “macro”—available for SPSS timates across two or more equations in the model. Furthermore,
and SAS that simplifies the implementation of mediation, inference about these statistics is based on bootstrapping methods,
given that many of these statistics have irregular sampling distri-
butions, making inference using ordinary methods problematic
1 The first author of this paper (Hayes) is the inventor of PROCESS. The other two (Hayes, 2013; Shrout and Bolger, 2002). PROCESS does all this behind
authors are (at the moment this paper was drafted) Ph.D. students of Hayes working the scenes and generates output that would otherwise require con-
in his lab and regularly field questions from users of PROCESS. siderable effort and programming skill to implement.
78 A.F. Hayes et al. / Australasian Marketing Journal 25 (2017) 76–81

Any SEM program can do path analysis with observed vari- visions) are also available and used as covariates (U1, U2, and U3)
ables as PROCESS does, although most require more code (and the in the analysis.
skill to write that code) than what is required to generate many of In this model, dysfunctional team behavior is estimated as af-
the statistics that PROCESS produces automatically. Furthermore, fecting team performance through its effect on the negativity of the
not all SEM programs can generate all of the statistics PROCESS cal- team work climate, which in turn lowers performance. But the re-
culates or implement bootstrapping in a way that facilitates inference lationship between the negativity of the work environment and
using those statistics. Although Pek and Hoyle (2016) argue that re- performance is estimated as linearly moderated by the expressivity
gression based approaches are not as easily implemented as SEM, of the team. Thus, this is a second stage conditional process model
we believe that with PROCESS, the opposite is true. Most research- as in Fig. 1, panel C. The mechanism linking X to Y through M is
ers will find PROCESS far easier to use than any SEM program. allowed to be moderated in the second stage (i.e., the M→Y path)
Other than ease of use, one of the more important differences of the process. The model also includes a direct effect from dys-
between PROCESS and SEM programs is that SEM solves the entire functional team behavior to performance. For details on the theory
system of equations simultaneously through iteration, typically using and existing research underlying this model, see Cole et al. (2008).
maximum likelihood (ML), rather than estimating the parameters This model is estimated with two equations:
of each equation independently. This involves finding an initial set
of parameter estimates for every variable in every equation defin- M̂ = iM + a1 X + a2U1 + a3U 2 + a4U 3 (1)
ing the model and then tweaking them simultaneously at each
iteration after measuring the correspondence between the covari- Ŷ = iY + c ′X + b1M + b2W + b3MW + b4U1 + b5U 2 + b6U 3 (2)
ance matrix of the variables in the model and the covariance matrix
implied by the model given the estimates derived. The estimation The index of moderated mediation, used for inference about
stops when further modification to the estimates does not improve whether the indirect effect of X on Y through M is moderated by
the correspondence more than as required by the convergence W, is a1b3, with a bootstrap confidence interval used for inference
criterion. (see Hayes, 2015). With evidence of moderation of the indirect effect,
Because SEM estimates the components of the model simulta- this moderation can be probed by estimating the conditional in-
neously, Pek and Hoyle (2016) recommend SEM and suggested that direct effect of X at various values of W using the function
the piece-wise nature of estimation with regression encourages a1(b1 + b3W), substituting values of W into this function once the
researchers to think about mediation as a procedure rather than as model coefficients are estimated, and a bootstrap confidence in-
a model. To be sure, users of the approach introduced to social sci- terval used for inference (Edwards and Lambert, 2007; Hayes, 2013;
entists by Baron and Kenny (1986) overly focus on the components Preacher et al., 2007).
of the model rather than the model as a whole. But the popularity This model was estimated using PROCESS version 2.16 and Mplus
of this procedure has been waning over the last few years. The more version 6. Mplus is a widely-used and versatile SEM program. As
modern focus on the indirect effect and other statistics in condi- noted earlier, PROCESS estimates equations 1 and 2 separately using
tional process analysis that integrate across the pieces (as PROCESS OLS regression while bootstrapping the sampling distribution of all
does) forces the analyst to contemplate the model as a whole rather parameter estimates, bringing them together across equations as
than just its pieces. We believe most users of PROCESS and users needed for computation of conditional indirect effects, the index
of SEM are thinking similarly and interact with the output of their of moderated mediation, and bootstrap confidence intervals for these
chosen method in much the same way. statistics. Mplus estimates the coefficients simultaneously and it-
eratively using maximum likelihood (the default, which we used).
The MODEL CONSTRAINT command in Mplus allows for computa-
2. Does the choice of PROCESS rather than SEM matter? tion of the index of moderated mediation and conditional indirect
effects using estimates from each equation, and the BOOTSTRAP
Does it matter whether one uses PROCESS as opposed to an SEM option generates bootstrap confidence intervals. Mplus code for this
program? Specifically, will one’s results be influenced by whether illustration can be found in an online supplement available through
PROCESS or an SEM program is used? Given that SEM and PROCESS afhayes.com or processmacro.org.
are based on different estimation methods and theory, some dif- The results are displayed in Table 1. The bolded top entry in each
ferences can be expected. However, for models of observed variables cell is the point estimate, and the middle entry is the estimated stan-
(i.e., nothing latent), differences in results tend to be trivial, and rarely dard error. As can be seen, the point estimates of model parameters,
will the substantive conclusions a researcher arrives at be influ- conditional indirect effects, and the index of moderated media-
enced by the decision to use PROCESS rather than SEM. Here we tion are largely identical, with a few discrepancies at the third
illustrate by example, recognizing that one demonstration of a phe- decimal place. Clearly, whether one uses separate OLS regression
nomenon does not equate to a general proof. equations or an SEM program to estimate the coefficients makes
An example of the similarity in results between PROCESS and no difference even in a small sample such as this. There are differ-
an SEM program is provided in Hayes (2013) for a multiple medi- ences in the standard errors, but these are expected, as the
ator model (such as in Fig. 1, panel A). In this demonstration, we underlying statistical theory of OLS versus ML sampling variance
show this similarity with a more complex conditional process model. estimation is different and relies on different assumptions. There
The data come from Cole et al. (2008), who examined the role of are also some discrepancies in the bootstrap confidence intervals
affect in the link between dysfunctional team behavior and team for the conditional indirect effects and index of moderated medi-
performance. The variables in the model are measured at the team ation. But this has nothing to do with the smaller standard errors
level from 60 teams working at an automobile parts manufactur- from SEM, as standard errors are not used in the construction of
ing facility. The variables include dysfunctional team behavior (X), these bootstrap confidence intervals. These interval estimates are
team performance (Y), negative emotional tone of the team work based on 10,000 bootstrap samples randomly constructed, and the
environment (M), and how emotionally expressive the team is from 10,000 generated by PROCESS are, of course, not the same random
the perspective of the supervisor (W), all of which are continuous samples generated by Mplus. Regardless, substantively, the results
dimensions and based on aggregates (average or sum scores) of are the same, with evidence that the indirect effect varies linearly
various questions the employees or their supervisor were asked. with the moderator (the index of moderated mediation), and the
Three dummy variables coding manufacturing parts division (4 di- conditional indirect effect definitively different from zero among
 A.F. Hayes et al. / Australasian Marketing Journal 25 (2017) 76–81 79

Table 1
Path coefficients, standard errors, conditional indirect effects, and index of moderated mediation (with percentile bootstrap confidence intervals) for the second stage con-
ditional process model in Fig. 1C, estimated using PROCESS and Mplus with the data from Cole et al. (2008).

Negative tone of the Team performance (Y)

work climate (M)

PROCESS Mplus PROCESS Mplus

Constant Coeff −0.206 −0.206 −0.175 −0.177

s.e. (n = 60) 0.131 0.125 0.131 0.121
s.e. (n = 180) 0.073 0.072 0.072 0.070
Dysfunctional behavior (X) Coeff 0.610 0.611 0.373 0.372
s.e. (n = 60) 0.167 0.160 0.181 0.169
s.e. (n = 180) 0.093 0.092 0.099 0.097
Negative tone of the work climate (M) Coeff – – −0.489 −0.489
s.e. (n = 60) 0.138 0.128
s.e. (n = 180) 0.076 0.074
Team expressiveness (W) Coeff – – -0.022 -0.022
s.e. (n = 60) 0.118 0.110
s.e. (n = 180) 0.065 0.063
M×W Coeff – – −0.450 −0.450
s.e. (n = 60) 0.245 0.228
s.e. (n = 180) 0.135 0.132
Division1 (U1) Coeff 0.349 0.348 0.182 0.182
s.e. (n = 60) 0.172 0.164 0.172 0.160
s.e. (n = 180) 0.096 0.095 0.095 0.092
Division 2 (U2) Coeff 0.295 0.296 0.084 0.085
s.e. (n = 60) 0.212 0.203 0.210 0.195
s.e. (n = 180) 0.119 0.117 0.115 0.113
Division 3 (U3) Coeff 0.251 0.251 0.282 0.283
s.e. (n = 60) 0.166 0.159 0.165 0.153
s.e. (n = 180) 0.093 0.092 0.091 0.089

PROCESS Mplus

Index of moderated mediation Value −0.274 −0.275

95% CI (n = 60) −0.705, −0.025 −0.704, −0.022
95% CI (n = 180) −0.469, −0.127 −0.480, −0.129
Conditional indirect effect at 1 SD below the sample mean of team expressiveness Value −0.146 −0.147
95% CI (n = 60) −0.439, 0.191 −0.425, 0.201
95% CI (n = 180) −0.288, 0.007 −0.292, 0.010
Conditional indirect effect at the sample mean of team expressiveness Value −0.296 −0.297
95% CI (n = 60) −0.564, −0.072 −0.560, −0.068
95% CI (n = 180) −0.431, −0.172 −0.435, −0.171
Conditional indirect effect at 1 SD above the sample mean of team expressiveness Value −0.445 −0.446
95% CI (n = 60) −0.783, −0.147 −0.783, −0.143
95% CI (n = 180) −0.619, −0.274 −0.630, −0.279

teams moderate to high in expressivity but not among teams low trivial. This example illustrates what we believe to be a general phe-
in expressivity. nomenon. In small samples, there will be differences in the standard
If you think that the smaller standard errors when using SEM errors, but in bigger samples, it generally makes no difference
would favor simultaneous estimation using ML, think again. SEM whether you use ML-based SEM or PROCESS. The choice is yours,
is generally regarded as a large sample technique. The default es- at least for observed variable models such as this. So we disagree
timation methods used by most SEM programs rely on large sample with Iacobucci et al. (2007), who advocated the use of SEM rather
asymptotic theory. Although the performance of ML estimation in than separate regressions in mediation analysis. We believe the in-
small samples is highly context dependent and depends on a variety terpretation of their findings fail to acknowledge the downward bias
of factors, in general, ML standard errors tend to be biased down- in ML standard errors when using SEM in smaller samples.
ward in small samples (Hoogland and Boomsma, 1998). So the
apparent advantage of SEM evidenced by the smaller standard errors
is likely illusory in this case and similar ones. Smaller standard errors
3. Considerations when making the choice
are not better when they are wrong.2
To illustrate the effect of sample size on discrepancies in results,
We do not intend to suggest above that there no advantages to
we repeated the analysis, pretending the sample size was 180 team
using an SEM program rather than PROCESS when conducting a me-
members. This was accomplished by copying the data set twice,
diation or conditional process analysis. Hayes (2013, pp. 161–162)
stacking each copy below the original in the data set, prior to es-
outlines some of these advantages, and here we elaborate on some
timation. The path coefficients don’t change, but as can be seen in
of them while adding a few new ones. Our discussion of this topic
Table 1, the standard errors, of course, do. Most important, the dis-
is not exhaustive.
crepancies in standard errors between PROCESS and Mplus are now
PROCESS is an observed-variable modeling tool that relies on OLS
regression. One of the weaknesses of regression analysis is its sus-
2 But the differences in standard errors in this example are a bit larger than would
ceptibility to bias in the estimation of effects due to random
be expected if small sample bias in ML standard errors was the only culprit. Many
measurement error (see Darlington & Hayes, 2017, pp. 525–532).
things can affect OLS and ML standard errors, and affect them differently, such as This weakness generalizes to regression-based mediation analy-
non-normality and heteroscedasticity. sis. The extent of the bias, which can be positive or negative, depends
80 A.F. Hayes et al. / Australasian Marketing Journal 25 (2017) 76–81

on many factors, such as the complexity of the model, the degree researchers would find the task of estimating latent variable inter-
of unreliability in measurement, and the correlation between the actions so daunting that the unknown effects that can result from
variables in the model (see Cole and Preacher, 2014, for a discus- ignoring measurement error would seem an acceptable price to pay
sion). Given perfect reliability of measurement can rarely be assumed, in exchange for the ease of the analysis and interpretation when
this means that estimates of indirect and direct effects (condition- using an observed-variable modeling tool like PROCESS. But if you
al or unconditional) based on OLS regression probably are biased are up to the challenge, go for it. Perhaps in time, the methodolo-
to some extent, in one direction or another. gy and technology for estimating latent variable interactions will
Structural equation modeling can help manage the effects of mea- become more intuitive and user-friendly.
surement error, but it won’t do so automatically. It only does so when Another reason a researcher might choose SEM over PROCESS
the analyst combines a structural model (i.e., a model of the rela- is the greater flexibility an SEM program offers for model specifi-
tionship between latent variables) with an explicit measurement cation. The current release of PROCESS (v2.16) has 76 pre-
model, with multiple indicators for each latent variable or using a programmed models that differ with respect to which and how many
single indicator with constraints in the model determined by an es- paths are moderated and by how many moderators, as well as the
timate of or guess about the reliability of the indicator (see e.g., Kline, presence of various constraints such as effects fixed to zero. But if
2016). Both these approaches can reduce bias in the estimation of your model is different than these pre-programmed models, the only
effects in a mediation or conditional process analysis (see Cheung option available is to move to an SEM environment, which would
and Lau, 2015; Ledgerwood and Shrout, 2011). This cannot be done allow you to estimate the model you desire at the expense of the
in PROCESS. No latent variables are allowed by PROCESS that haven’t additional effort required to write the code corresponding to that
first been reduced to observed variable proxies (e.g., sum scores or model. However, this benefit of SEM relative to PROCESS is dimin-
averages of indicators), which make them, by definition, observed ishing. The next version of PROCESS (to be released in late 2017)
and not latent. includes options for modifying preprogrammed models or con-
Researchers, reviewers, and editors who are aware of the prob- structing a model from scratch.
lems produced by measurement error may question the legitimacy SEM programs also offer measures of fit of the model to the data
of results generated by PROCESS and insist that an SEM program (e.g., RMSEA, CFI, etc.), whereas PROCESS offers no omnibus measure
with a combined measurement and structural model be used instead. of model fit, so one cannot go away from an analysis using PROCESS
Although this is a defensible position, three things are important with information about how well the complete model describes the
to recognize. First, if that is the position one takes, it should be data. However, several of the more popular models built into
applied consistently, meaning such a critic should doubt the legit- PROCESS are saturated, so fit by some measures would be perfect
imacy of any analysis that can be expressed in the form of a linear when these models are estimated using SEM. Furthermore, there
regression model (even if not actually described as such in a manu- are more important things in analytical life than good fit, especial-
script). This would include regression analysis itself, analysis of ly given that any good fitting model typically has several minor
variance and analysis of covariance, the independent group t-test, variations with different interpretations that fit equally well (the
and even hypothesis tests involving the simple correlation between equivalent models problem; MacCallum et al., 1993). When con-
two variables. Though mediation analysis without a latent vari- ducting a mediation or conditional process analysis, just as important
able measurement model receives much criticism, all of these (if not more) are the estimates of and inference about the effects
methods suffer from problems caused by random measurement error, involving the variables in the model and related statistics that are
to varying degrees and with various effects on inference. Our guess sensitive to the hypotheses of interest (such as conditional indi-
is that most who take this position have published their own re- rect effects and the index of moderated mediation).
search (or the work of others) using methods that are just as Finally, some SEM programs offer more options than PROCESS
susceptible to measurement error-induced problems such as bias. for dealing with missing data, a fact of life in most empirical re-
Second, structural equation modeling with a combined mea- search. PROCESS requires complete data. If it doesn’t get it, it will
surement and structural model is not a panacea to problems make it complete through listwise deletion. Missing data solu-
produced by imperfect reliability of measurement. Accounting prop- tions that rely on some kind of literal imputation must be
erly for measurement error requires a proper model of that error, implemented before the use of PROCESS, and procedures that involve
one that typically carries assumptions that may not be met. Fur- creating data where there are none all have flaws. Many SEM pro-
thermore, as Ledgerwood and Shrout (2011) discuss, biases in the grams can implement more sophisticated missing data procedures
estimation of one parameter can be offset by biases in another in that don’t require actually creating data, such as full information
the opposite direction, which is important given that indirect effects maximum likelihood (FIML). Between these two extremes (listwise
are the products of parameter estimates. In addition, standard errors deletion and FIML) is multiple imputation, not available in PROCESS
in latent variable mediation analysis can actually be larger in some but an option in some SEM programs.
circumstances, even though bias in the estimation of the structur-
al parameters is reduced compared to when using observed variable
mediation analysis. So latent variable mediation analysis may be, 4. Conclusion
at least in some circumstances, more accurate in the estimation of
effects than observed variable analysis, but less powerful in detect- Our intention was to provide guidance to researchers inter-
ing them. ested in understanding the mechanisms of effects and their
Third, conditional process models have at least one interaction contingencies about how the PROCESS macro and SEM differ both
between variables. The proper estimation of interactions between in operation and results, and to offer considerations to ponder when
latent variables remains highly controversial, and there are many making the choice between them. The greater flexibility of SEM, both
methods available that require various (and sometimes different) in terms of model specification and handling missing data, as well
assumptions (see Marsh et al., 2013). In our experience, it can be as its ability to account for random measurement error when es-
difficult to trust a model which involves estimating latent variable timating relevant effects involving latent variables all make it an
interactions because it is difficult to determine whether the result- attractive choice. But that comes at the price of greater effort and
ing estimates of interactions are reasonable. Making things worse, programming skill required to calculate relevant statistics and
different methods can produce different results and are vulnera- methods of inference that PROCESS does automatically and pain-
ble to assumption violations (e.g., Cham et al., 2012). We think most lessly. For models that are based entirely on observed variables,
 A.F. Hayes et al. / Australasian Marketing Journal 25 (2017) 76–81 81

investigators can rest assured that it generally makes no differ- Hoogland, J.J., Boomsma, A., 1998. Robustness studies in covariance structure models:
an overview and meta-analysis. Sociol. Methods Res. 26, 329–367.
ence which is used, as the results will be substantively identical.
Iacobucci, D., Saldanha, N., Deng, J.X., 2007. A meditation on mediation: evidence
The choice, in that case, is inconsequential. that structural equation models perform better than regressions. J. Consum.
Psychol. 17, 140–154.
Kline, R.B., 2016. Principles and Practice of Structural Equation Modeling, fourth ed.
References The Guilford Press, New York.
Kongarchapatara, B., Shannon, R., 2016. The effect of time stress on store
loyalty:
Baron, R.B., Kenny, D.A., 1986. The moderator-mediator variable distinction in social a case for food and grocery shopping in Thailand. Aust. Mark. J. 24, 267–274.
psychological research: conceptual, strategic, and statistical considerations. J. Pers. Ledgerwood, A., Shrout, P.E., 2011. The trade-off between accuracy and precision in
Soc. Psychol. 51, 1173–1182. latent variable models of mediation processes. J. Pers. Soc. Psychol. 101,
Baxter, R., Kleinaltenkamp, M., 2015. How relationship conditions affect suppliers’ 1174–1188.
resource inputs. Aust. Mark. J. 23, 117–123. MacCallum, R.C., Wegener, D.T., Uchino, B.N., Fabrigar, L.R., 1993. The problem of
Cassar, V., Briner, R.B., 2011. The relationship between psychological contract breach equivalent models in applications of covariance structure analysis. Psychol. Bull.
and organizational commitment: exchange imbalance as a moderator of the 114, 185–199.
mediating role of violation. J. Vocat. Behav. 78, 283–289. Marsh, H.W., Wen, Z., Hau, K., Nagengast, B., 2013. Structural equation models of latent
Cham, H., West, S.G., Ma, Y., Aiken, L.S., 2012. Estimating latent variable interactions interactions and quadratic effects. In: Hancock, G.R., Mueller, R.O. (Eds.), A Second
with nonnormal observed data: a comparison of four approaches. Multivariate Course in Structural Equation Modeling, second ed. Information Age, Greenwich,
Behav. Res. 47, 840–876. CT.
Cheung, G.W., Lau, R.S., 2015. Accuracy of parameter estimates and confidence Muller, D., Judd, C.M., Yzerbyt, V.Y., 2005. When moderation is mediated and
intervals in moderated mediation models: a comparison of regression and latent mediation is moderated. J. Pers. Soc. Psychol. 89, 852–863.
moderated structural equations. Organ. Res. Methods doi:10.1177/ Pek, J., Hoyle, R.H., 2016. On the (in)validity of tests of simple mediation: threats
1094428115595869. and solutions. Soc. Personal. Psychol. Compass 10, 150–163.
Cole, D.A., Preacher, K.J., 2014. Manifest variable path analysis: potential serious and Preacher, K.J., Rucker, D.D., Hayes, A.F., 2007. Assessing moderated mediation
misleading consequences due to uncorrected measurement error. Psychol. hypotheses: theory, methods, and prescriptions. Multivariate Behav. Res. 42,
Methods 19, 300–315. 185–227.
Cole, M.S., Walter, F., Bruch, H., 2008. Affective mechanisms linking dysfunctional Schiele, H., Vos, F.G.S., 2015. Dependency on suppliers as a peril in the acquisition
behavior to performance in work teams: a moderated mediation study. J. Appl. of innovations? The role of buyer attractiveness in mitigating potential negative
Psychol. 93, 945–958. dependency in buyer-supplier relations. Aust. Mark. J. 23, 139–147.
Darlington, R.B., Hayes, A.F., 2017. Regression Analysis and Linear Models: Concepts, Shen, H., Zhang, M., Krishna, A., 2016. Computer interfaces and the “direct-touch”
Applications, and Implementation. The Guilford Press, New York. effect: can iPads increase the choice of hedonic food? J. Mark. Res. 53,
Dubois, D., Rucker, D.D., Galinsky, A.D., 2016. Dynamics of communicator and 745–758.
audience power: the persuasiveness of competence and warmth. J. Consum. Res. Shenu-Fen, Y., Sun-May, L., Yu-Ghee, W., 2012. Store brand proneness: effects of
43, 68–85. perceived risks, quality, and familiarity. Aust. Mark. J. 20, 48–58.
Edwards, J.R., Lambert, L.S., 2007. Methods for integrating moderation and mediation: Shrout, E., Bolger, N., 2002. Mediation in experimental and nonexperimental studies:
a general analytical framework using moderated path analysis. Psychol. Methods new procedures and recommendations. Psychol. Methods 7, 422–445.
12, 1–22. Voola, R., Casimir, G., Carlson, J., Agnihotri, M.A., 2012. The effects of marketing
Etkin, J., Sela, A., 2016. How experience variety shapes postpurchase product orientation, technological opportunism, and e-business adoption on performance:
evaluation. J. Mark. Res. 53, 77–90. a moderated mediation analysis. Aust. Mark. J. 20, 136–146.
Fairchild, A.J., MacKinnon, D.P., 2009. A general model for testing mediation and White, K., Lin, L., Dahl, D.W., Ritchie, R.J.B., 2016. When do consumers avoid
moderation effects. Prev. Sci. 10, 87–99. imperfections? Superficial packaging damage as a contamination cue. J. Mark.
Hayes, A.F., 2013. Introduction to Mediation, Moderation, and Conditional Process Res. 53, 110–123.
Analysis: A Regression-Based Approach. The Guilford Press, New York. Zenker, S., Braun, E., Petersen, S., 2017. Branding and destination versus the place:
Hayes, A.F., 2015. An index and test of linear moderated mediation. Multivariate the effects of brand complexity and identification for residents and visitors. Tour.
Behav. Res. 50, 1–22. Manage. 58, 15–27.