CHAPTER 9

Mediation Analyses in Partial Least Squares Structural Equation Modeling.

Guidelines and Empirical Examples.

by

Gabriel Cepeda, Christian Nitzl and José L. Roldán

Partial least squares structural equation modeling (PLS-SEM) is one of the options used to
analyze mediation effects. Over the past few years, the methods for testing mediation have
become more sophisticated. However, many researchers continue to use outdated methods to
test mediation effects in PLS-SEM, which can lead to erroneous results in some cases. One
reason for the use of outdated methods is that PLS-SEM tutorials do not draw on the newest
statistical findings. This chapter illustrates how to perform modern procedures in PLS-SEM by
challenging the conventional approach to mediation analysis and providing better alternatives.
These novel methods offer a wide range of testing options (e.g., multiple mediators) that go
beyond simple mediation analysis alternatives, helping researchers to discuss their studies in a
more accurate way. This chapter seeks to illustrate and help to operationalize the mediation in
Nitzl, Roldán and Cepeda’s (2016) paper about mediation in PLS, published in Industrial
Management & Data Systems, with examples of two potential mediations: a multiple mediation
with two mediators, and a multi-step multiple mediation.

1. Introduction
Partial least squares structural equation modeling (PLS-SEM) is a variance-based structural
equation modeling technique that has been used to model latent variables, specifically
composites, and the relationships between them (Henseler, 2017). Therefore, it is a useful tool
for testing hypotheses and answering research questions. One of these research questions
investigates mediation. Mediation considers the presence of an intermediate variable or
mechanism that transmits the effect of an antecedent variable to an outcome (Aguinis, Edwards,
and Bradley, 2016). For instance, mediation usually appears when the effect of reputation on
customer loyalty is transmitted by customer satisfaction, such that reputation impact on
customer satisfaction in turn influences customer loyalty (Hair, Hult, Ringle, and Sarstedt,
2014). Hence, mediation refers to underlying effects that link antecedent and consequences
variables. Despite the continuous use of mediation testing, studies in PLS-SEM often do not
consider mediation effects in their hypotheses and therefore do not analyze the relevance in
relevant structural models (Hair et al., 2017). In the worst case, researchers focus only on direct
relationships and overlook mediation effects.
While there is a large body of literature on methods for testing mediation effects (Hayes and
Scharkow, 2013), the analytical tools that researchers have used in PLS-SEM studies to test
mediation effects have generally been outdated compared to those for other statistical methods.
Nitzl et al., (2016) have recently shown the misapplication of Baron and Kenny’s procedure in
the PLS-SEM field. Whereas researchers studying covariance-based structural equation
modeling (CB-SEM) have often considered the latest findings when testing mediation (e.g.,
Iacobucci et al., 2007, Hair et al., 2010), most PLS-SEM researchers fail to do so, and in fact,
they often avoid carrying out this kind of analysis. This is somewhat surprising because state-
of-the-art applications for testing the significance of a mediator are also very suitable for PLS-
SEM.
Therefore, we can state that one of the key reasons authors do not assess mediation effects in
PLS path models is the lack of illustrative guidelines on conducting state-of-the-art mediation
analysis with PLS-SEM. Furthermore, because these publications on PLS-SEM have been
subjected to several recent changes (Henseler et al. 2016; Nitzl et al., 2016), an adequate
illustration of these new guidelines related to mediation is badly needed.
Therefore, the objective of our chapter is to provide researchers with a nice illustration to
implement mediation models in PLS-SEM. Thus, we offer complete examples and guidelines
on how to conduct mediation analysis using PLS-SEM, inspired by Nitzl et al. (2016).
Our chapter is structured as follows: we first describe an advanced procedure for mediation
analysis in PLS-SEM. We then list different types of mediation. Next, we offer illustrative
examples of how to perform and discuss a mediation analysis with PLS. We also offer detailed
guidelines for carrying out this type of analysis in PLS. Finally, we summarize our chapter and
highlight potential avenues for future research

2. Advanced Procedure for Mediation Analysis in PLS-SEM

2.1. The Mediation Effect
The core characteristic of a mediating effect (i.e., indirect effect or mediation) is that it involves
a third variable that plays an intermediate role in the relationship between the independent and
dependent variables. Technically speaking, the effect of the independent variable X on the
dependent variable Y is mediated by a third variable, M, called the mediating variable or
mediator (see Figure 1). Thus, when we formulate mediation hypotheses, we focus on “how, or
by what means, an independent variable (X) affects a dependent variable (Y) through one or

more potential intervening variables, or mediators (M)” (Preacher and Hayes, 2008). The
researcher’s objective in mediation analysis is mainly explanation (Henseler et al., 2016),
although some scholars have also recently added the purpose of prediction (Shmueli et al.,
2016).

Figure 1: Simple Cause-Effect Relationship and General Mediation

Model

Figure 1a shows the total effect c of the causal relationship between variables X and Y, and
Figure 1b shows a mediated effect in which X exerts an indirect effect a×b through M on Y.
Once we have defined the mediation effect, we briefly describe the procedure developed by
Nitzl et al. (2016) to test mediation effects on PLS-SEM and also define the different types of
mediation that researchers can find in their analysis. The procedure considers five important
statements for testing mediating effects in PLS:
1. Testing the indirect effect a×b provides researchers with all the information they need to
assess the significance of a mediation. Therefore, it is not necessary to conduct separate
tests for paths a and b by applying PLS-SEM.
2. The strength of the indirect effect a×b should determine the size of the mediation.
Therefore, it is also not necessary to test the difference between c and c’.
3. A significant indirect effect a×b is the only prerequisite for establishing a mediation effect.
4. A bootstrap test should be used to test the significance of the indirect effect a×b.
5. The significance of the direct effect (c’) has to be tested in order to determine the type of
effect and/or mediation.
These important statements are summarized in the procedure described by Nitzl et al. (2016).
The procedure has two main steps (see Nitzl et al. (2016) for a more detailed description):
Step 1: Determining the significance of indirect effects and their magnitude
The indirect effect a×b must be significant in Step 1 to establish a mediation effect. When
testing mediation effects in PLS-SEM, researchers should bootstrap the sample of the indirect
effects in order to obtain necessary information about the population distribution, in accordance
with the non-parametric PLS-SEM method where bootstrap procedures are typically used for
inference statistics, such as the calculation of the so-called pseudo t-value and confidence
intervals (Henseler et al., 2009). The bootstrapping procedure is a non-parametric inferential
technique that randomly withdraws several subsamples (e.g., 5,000) with replacement from the
original dataset. PLS-SEM uses each of the subsamples to estimate the underlying PLS path
model.

The bootstrap routines of PLS-SEM software often provide results for at least direct effects
(e.g., path a and path b). However, for a more detailed analysis of mediation, particularly in
more complex model structures (e.g., multiple mediators), it is often necessary to compute the
bootstrapping results for the indirect effects with the help of a spreadsheet application, such as
Microsoft Excel or CALC in OpenOffice. For each bootstrapping subsample, the results of path
a must be multiplied by path b to create the product term a×b of the indirect effect in a new
column. For example, the computation of k = 5,000 bootstrapping subsamples entails the
generation of k = 5,000 products a×b in a new column. The information about the
characteristics of the distribution of mediation effects is obtained by calculating a ci%
confidence interval for a×b. For that, the subsamples (k) for a×b from the bootstrapping
procedure must be arranged from smallest to largest (Hayes, 2009). In the next step, a researcher
has to select a specific alpha error; for example, for a probability of error of 5%, a 95%
confidence interval must be determined with a 2.5% probability of error at each tail when
conducting a two-sided test. The lower bound of a×b is in the k×(.5 - ci% / 2)th ordinal position
of the ordered list; for example, if one uses k = 5,000 subsamples and a 95% confidence interval,
the lower bound is the 5,000×(.5 - 0.95 / 2) = 125th ordinal position. Similarly, the
(1+k×(.5 + ci% / 2))th ordinal determines the upper bound of the bootstrap confidence, which
is the 1 + 5,000×(.5 + 0.95 / 2) = 4,876th in the previous example. If zero is not included in
the confidence interval, a researcher can assume that there is a significant indirect effect a×b.
Another problem often occurs when the mean of the bootstrapped distribution for the indirect
effect a×b is not equal to the estimated indirect effect a×b (Chernick, 2011). As a result,
researchers must correct for this bias in PLS-SEM. This can be accomplished by calculating
the difference between the estimated indirect effect aPM×bPM from the path model (PM) and the
mean value of the indirect effect aB×bB from the bootstrap sample (B). Consequently, the bias-
corrected ci% confidence interval for an indirect effect a×b can be defined as:

(𝑘×(.5 − 𝑐𝑖%/2))𝑡ℎ + (𝑎12 ×𝑏12 − 𝑎4 ×𝑏4 ); ((1 + 𝑘×(.5 + 𝑐𝑖%/2))𝑡ℎ +

(1)
(𝑎12 ×𝑏12 − 𝑎4 ×𝑏4 ) .

Hayes and Scharkow (2013) show that the bias-corrected bootstrap confidence interval is the
best approach for detecting mediation effects when a mediation effect is present (i.e., Type-II
error or power). Conversely, the percentile bootstrap confidence interval that is not bias-
corrected is a good compromise if a researcher is also concerned about Type-I errors (Hayes
and Scharkow, 2013). Thus, the bias-corrected bootstrap confidence interval is the most reliable
test if power is of the utmost concern, while the percentile bootstrap confidence interval
provides a good compromise.
Step 2: Determining the type of effect and/or mediation
Step 2 involves defining the type of effect and/or mediation. A mediating effect always exists
when the indirect effect a×b in step 1 is significant. The current mediation literature discusses
two different types of mediation, full and partial mediation. Partial mediation can be sub-
divided into complementary and competitive partial mediation. We also discuss two effects that
occur when the indirect effect is not significant, which means that only the direct effect is
significant and no effect at all is significant. The latter cases do not represent a mediating effect
in the narrow sense.
2.2 Full Mediation
A full mediation is indicated in the case where the direct effect c’ is not significant, whereas
the indirect effect a×b is significant. This means only the indirect effect via the mediator exists.



In other words, full mediation means that the effect of the variable X on Y is completely
transmitted with the help of another variable M. It also means the condition Y completely
absorbs the positive or negative effect of X. In this way, it can completely pass an effect or it
can completely hinder the effect in terms of another effect. Technically speaking, the variable
X extracts its influence only under a certain condition of M on Y. However, in the case of small
samples, a researcher is to exercise some caution when talking about full mediation. As Rucker
et al. (2011) showed, “the smaller the sample, the more likely mediation (when present) is to
be labeled full as opposed to partial, because c' is more easily rendered non-significant” (p.
364). Hence, it is advisable to ensure that the sample size is great enough so that the necessary
power of 0.8 for an alpha level of 0.05 for detecting effects in a PLS path model is obtained.
For a simple mediation model, the necessary sample size can be quite low. Notwithstanding, a
medium and small effect size would require a bigger sample. In contrast, in many cases, it can
be observed that some small direct effects, c’, remain even though the mediating effect is quite
high in relation to the mediated direct effect. However, when this relation of the direct effect to
the mediating effect becomes low but nevertheless stays significant, it can also be seen as full
mediation. A researcher could indicate this with the help of the variance accounted for (VAF)
value, which we will discuss in more detail below in our example. Conversely, when the
absolute value of the indirect path a×b is larger than the absolute value of the total effect (a×b)
+ c’, there is a suppressor effect (Cheung and Lau, 2008); this situation could also be defined
as full mediation (Hair et al., 2017).

2.3 Partial Mediation

All other situations under the condition that both the direct effect c’ and the indirect effect a×b
are significant represent partial mediation. Two types of partial mediation can be distinguished:
a. Complementary Partial Mediation
In a complementary partial mediation, the direct effect c’ and indirect effect a×b point in the
same (positive or negative) direction (Baron and Kenny, 1986). It is an often observed result
that a×b and c’ are significant and a×b×c’ is positive, which indicates that a portion of the
effect of X on Y is mediated through M, while X still explains a portion of Y that is independent
of M. This complementary mediation hypothesis suggests that the intermediate variable
explains, possibly confounds, or falsifies the relationships between the independent and
dependent variables.
b. Competitive Partial Mediation
In a competitive partial mediation, the direct effect c’ and indirect effect a×b point in a different
direction. A negative a×b×c’ value indicates the presence of competitive mediation in Step 2.
As mentioned above, this indicates that a portion of the effect of X on Y is mediated through
M, while X still explains a portion of Y that is independent of M. In the past, researchers often
focused only on complementary mediation (Zhao et al., 2010). In the competitive partial
mediation hypothesis, it is assumed that the intermediate variable will reduce the magnitude of
the relationship between the independent and dependent variables. However, it is possible that
the intermediate variable could increase the magnitude of the relationship between the
independent and dependent variables. Competitive partial mediation has often been called a
‘negative confounding’ or an ‘inconsistent’ model. Thus, other types of mediation beyond
complementary mediation should be considered in a PLS path model.
PLS researchers might also be interested in evaluating the strength (portion) in the case of a
partial mediation. Mediation analyses regularly involve partial mediation, and therefore it can
be helpful to have further information on the mediated portion. One approach for this is
calculating the ratio of the indirect-to-total effect. This ratio is also known as the variance

accounted for (VAF) value. VAF determines the extent to which the mediation process explains
the dependent variable’s variance. For a simple mediation, the proportion of mediation is
defined as:

;×<
(2) 𝑉𝐴𝐹 = .
(;×<)=>?

2.4 No mediation
When the indirect effect is not significant, we can find another two situations. Although these
cannot be considered mediation cases in a narrow sense, two types of effects can be
distinguished:
a. Only Direct effect
If the indirect effect a×b is not significant while the direct path c’ is, the mediator variable has
no impact; this indicates that a direct, non-mediating effect is present. In this case, the study
was perhaps searching for a wrong mediation relationship. However, it is possible that an
unrecognized mediation relationship still exists and another mediation variable is present that
mediates an effect between X and Y (Shrout and Bolger, 2002). Thus, a researcher should
rethink his theoretical basis when he has not found the expected mediation relationship (cf.
Zhao et al., 2010).
b. No effect
There is no effect if neither the indirect effect a×b nor the direct effect c’ is significant. The
total effect can still be significant. First of all, in this case, a researcher should check if the
sample size has enough power to show an effect when there is an effect. Putting the last two
cases together – the indirect effect a×b is not significant and the direct path c’ is or is not –
frequently indicates a problematic or flawed theoretical framework (Zhao et al., 2010). In this
case, a researcher has to thoroughly examine his/her hypothesized model. When, for example,
the total effect c is significant, it can indicate that the mediation variable should be deleted
because it brings no further degree of explanation. In the case where the mediation variable M
has no real effect, it only dilutes the effect of the direct variable X and should be deleted.

2.5 Multiple mediation

PLS is regularly applied in complex path models. There may be multiple relationships between
one or more independent variables, one or more mediator variables, or one or more dependent
variables. For instance, a complementary mediation variable (M1) may mitigate the independent
variable (X) to a dependent variable (Y), and at the same time, a competitive mediation variable
(M2) may also exist. From a naïve perspective, someone can assume that the independent
variable is not relevant because there is no relevant total effect c. However, when one of the
mediator variables has a strong influence in a certain situation, the independent variable also
wins in terms of relevance. Such areas can become very challenging, for example, when
analyzing which process improves or hinders the influence of the external pressure to work on
the outcome in a PLS path model. However, when more than one mediating effect is present,
the abovementioned differentiation between direct and indirect effects for detecting mediation
relationships remains applicable, and the above recommendations are inalterable (Hayes,
2009).

Figure 2: Multiple Mediator Model

Figure 2 presents an example of a PLS path model with two mediators. The total effect is equal
to the direct effect of X on Y, in addition to the sum of the indirect effects of M1 and M2. A
given meditator’s indirect effect is referred to as a specific indirect effect (e.g., through M1).
The sum of the two specific indirect effects is the complete indirect effect. Thus, the total effect
is the sum of the direct effect and the complete indirect effects (i.e., the sum of the specific
indirect effects includes the relationship between M1 and M2). For the example in Figure 3, the
calculation of the total effect is:

(3) 𝑐 = 𝑐 ? + (𝑎@ ×𝑏@ ) + (𝑎A ×𝑏A ).

An interesting situation occurs when a1×b1 and a2×b2 in Equation 2 have an opposite sign; this
indicates that one effect functions as a complementary effect, and the other functions as a
competitive mediator effect. Such a model is called an inconsistent mediation model
(MacKinnon et al., 2007). Consequently, even though significant specific indirect effects exist,
the complete indirect effect (e.g., (a1×b1) + (a2×b2)) may not be significant.
Preacher and Hayes (2008) argue that the incorporation of multiple mediators and the
comparison of their specific mediating effects is also useful for comparing different competing
theories. Given this background, researchers are interested in comparing the strengths of
specific mediating effects (e.g., (a1×b1) and (a2×b2)) in complex models (Williams and
MacKinnon, 2008). For example, a researcher could test for two complementary mediator
variables if mediator (M1) has a stronger mediator effect than mediator (M2). The previous
explanation of how to compute bootstrap confidence intervals in PLS can be extended to test
the significance of the difference between two specific mediating effects (Lau and Cheung,
2012). For that purpose, a researcher must calculate the following equation:

(4) 𝐷2 = 𝑀@ − 𝑀A ,

where M1 and M2 are the specific indirect effects and DM is the difference between these two
specific indirect effects. In this way, we test whether two specific indirect effects are equal or
if they amount to zero. In the case examined in this study, the equation for Figure 2 would be
DM = (a1×b1) – (a2×b2). Again, researchers can calculate the equation using a spreadsheet
application to build a confidence interval with the help of the bootstrapping results of the PLS
program (cf. Chin et al., 2013; Rodríguez-Entrena et al, 2016).
A frequently encountered case is that in which two mediators are connected to each other. This
indicates an additional relationship between M1 and M2 in Figure 2. Next, we provide examples
of how to test such multiple mediation relationships in a PLS path model. In such a case, the

total effect c can be calculated as follows: 𝑐 = 𝑐 ? + (𝑎@ ×𝑏@ ) + (𝑎A ×𝑏A ) + (𝑎@ ×𝑎D ×𝑏A ) ,
where a3 stands for the relation between M1 and M2. An interesting case in this situation is when
a2, b2, and c’ are not significantly different from zero, but the indirect effect (a1×a3×b2) is (e.g.,
when M1 is the causal predecessor of M2); this would mean that M1 fully mediates the direct
effect between X and M2 and that M2 fully mediates the direct effect between M1 and Y, thus
establishing a direct causal chain X→M1→M2→Y (Mathieu et al., 2008). Next, we illustrate
this in our second example.

3. Illustrative examples
3.1. A Case of a Multiple Mediator Model
In this first example, we take data from Roldan, Cegarra and Cepeda (2014). This research
examines the relationship between a key component of the absorptive capacity, the realized
absorptive capacity (RACAP) and the organizational outcomes, this link being mediated by the
unlearning context and the organizational agility (Figure 3). These connections are examined
through an empirical investigation of 112 large companies.

Unlearning
context
a1 b1

Realized c’ Organizational
absorptive outcomes
capacity

a2 b2
H1 = c’
H2 = a1b1 Organizational
agility
H3 = a2b2

Figure 3: An example of a multiple mediator model. Source: Roldán, Cegarra and Cepeda
(2014)

3.1.1. Data collection and Measures

The population of this study consists of Spanish organizations that use Editran TM, and which
have more than 100 employees. EditranTM is a software used to enhance communications over
different platforms, and is a de facto standard in the Spanish banking system. This population
is suitable for our study, because these businesses are more familiar with knowledge and
technology management. 464 companies were identified from the SABI (Sistema de Análisis
de Balances Ibéricos) database and invited to participate in the study. 121 companies agreed. A
total of 112 valid and completed questionnaires were collected
We modeled RACAP and unlearning context as multidimensional constructs (composites). We
measured RACAP by two first-order dimensions (composites): transformation and exploitation.
The unlearning context variable was assessed using three first-order dimensions (composites):
The examination of lens fitting (ELF), the framework for changing individual habits (CIH), and

the framework for consolidation of emergent understandings (CEU). The example’s constructs
were estimated in Mode A, and the characteristics of the scales are the following:
a) RACAP. Items were measured using a seven-point Likert scale from the study by Jansen et
al. (2005). RACAP includes the transformation and exploitation of new external knowledge.
The final cleansed scale consists of four items for the transformation dimension and three items
for the exploitation dimension.
b) Unlearning context. At the organizational level it is viewed as memory elimination in
general, and as changing beliefs, norms, values, procedures and routines in particular. As
described above, the unlearning context has three dimensions: consolidation of emergent
understandings, the examination of lens fitting, and the framework for changing individual
habits. The measures relating to consolidation of emergent understandings consisted of six
items taken from a scale designed by Cegarra and Sanchez (2008). Five items were used to
measure the examination of lens fitting. Finally, we measured the framework for changing
individual habits using seven items.
c) Organizational agility. A business-wide capability to deal with changes that often arise
unexpectedly in business environments via rapid and innovative responses that exploit changes
as opportunities to grow and prosper. The indicators of organizational agility are based on the
measures of organizational agility used by Lu and Ramamurthy (2011). The scale was
composed of six items.
d) Organizational outcomes. It is understood as an assessment of the global performance of the
business. The scale for organizational outcomes consisted of ten reflective items adapted from
Quinn and Rohrbaugh (1983).

3.1.2. Hypotheses development

Once the conceptual framework is shown, the next step is hypotheses development. The
research model depicted in Figure 3 includes one direct and two mediating hypotheses:
H1: RACAP is positively associated with organizational outcomes
H2: The relationship between RACAP and organizational outcomes is positively mediated by
the unlearning context
H3: The relationship between RACAP and organizational outcomes is positively mediated by
organizational agility

3.1.3. PLS-SEM practical considerations

This chapter’s goal is not to illustrate the complete data analysis with PLS, but to focus on the
structural model, specifically on how to test this mediation model with PLS. Guidelines for a
complete analysis with PLS can be found in Henseler et al. (2016) and Hair et al. (2017).

(i) Significance of direct and indirect effects

We assess the significance of one direct (c’) and two indirect effects (a1×b1, and a2×b2). The
critical issue is that if the significance of each indirect effect cannot be established, there is no
mediating effect. Consequently, having a significant indirect effect is the key to determining
the type of mediation effect and its magnitude. Considering that our hypotheses have been
formulated with direction (+), we will use a one-sided test. Accordingly, we will estimate 90%
confidence intervals (CI).

Nitzl et al. (2016) suggested a procedure using a spreadsheet and multiplying the bootstrapping
outputs (i.e., a1×b1, and a2×b2) to calculate the percentile and the bias-corrected confidence
intervals. Therefore, once we run the model, we next perform the bootstrapping procedure with
5,000 subsamples and no sign changes. In Figure 4 we can see the estimates for direct effects.

Unlearning
context
R2 = 0.684
a1= 0.827sig b1 = 0.259sig

Realized c’ = 0.006nsig Organizational

absorptive outcomes
capacity R2 = 0.513

a2 = 0.637sig b2 = 0.509sig
H1 = c’ Organizational
H2 = a1b1 agility
H3 = a2b2 R2 = 0.405

sig: significant based on one-sided test

nsig: not significant based on one-sided test

Figure 4: An example of a multiple mediator model. Results. Source: Roldán, Cegarra and
Cepeda (2014)

In order to calculate the specific indirect effects and the different confidence intervals, we are
going to use a spreadsheet application (i.e. Excel or CALC), to obtain the significance of
mediator effects (a1×b1, and a2×b2) in the relationship between RACAP and organizational
outcomes. We suggest carrying out the following steps:
1. Take the 5,000 path coefficients from all direct effects created by the bootstrap procedure
and copy and paste into a spreadsheet’s columns (Figure 5)

Figure 5: Example 1. Step 1.

2. Create a new column for each indirect effect under assessment. In this case, we generate
two new columns (a1×b1, and a2×b2), and explicitly calculate the product of the direct paths

that form such indirect paths. In addition, we include another column for estimating the
total indirect effect (a1×b1) + (a2×b2) (Figure 6).
2) Insert a column for estimating the total
indirect effect (a1 b1)+(a2 b2)

1) Create two new columns a1 b1 and

a2 b2, and explicitly calculate the product
of the direct paths that form the indirect
paths under assessment

Figure 6: Example 1. Step 2.

3. Copy the original values (O) provided by PLS for the direct effects. Then calculate the
product of the direct paths that form each indirect path. In the line below, calculate the mean
(M) for each column of the paths obtained with the bootstrapping process (Figure 7).
1) Copy the original values (O) 2) Calculate the product of 3) estimating the total
provided by PLS for the direct the direct paths that form the indirect effect =G2+H2
effects. indirect paths

4) Calculate the mean (M) for each column, =AVERAGE(range).

E.g. for c’, =AVERAGE(B10:B5009)

Figure 7: Example 1. Step 3.

4. Insert a new line where you estimate the bias as Original (O) – Mean (M) for each column
(Figure 8).

Calculate the Bias: Original (O) – Mean (M)

Figure 8: Example 1. Step 4.

5. Estimate the percentile bootstrap CI for each column using the function
PERCENTILE(range,k), k being the percentile value between 0 and 1. In our case, given
our hypotheses are postulated with direction (+), we will use one-sided test and we will
estimate 90% CI (Figure 9).

Percentile LOWER (5%) for c’ =PERCENTILE(B10:B5009,0,05)

Percentile LOWER (95%) for c’ =PERCENTILE(B10:B5009,0,95)

Figure 9: Example 1. Step 5.

6. Estimate the bias-corrected CI adding the bias to the previously calculated percentile CI
(Figure 10).

Figure 10: Example 1. Step 6.

7. If the confidence interval (CI) for a mediation effect (products) does not include 0 value, it
means the mediating effect is significantly different from 0. In our example, both indirect
effects are significant. In addition, the total indirect effect is also significant (Table 1).

The key point to determine a mediation effect is the evaluation of the significance of the indirect
effect (Table 1). In our example, both indirect effects are significant, therefore H2 and H3 are
supported. However, the direct effect is not significant, consequently H1 is not supported.

Table 1. Example 1. Summary of mediating effects tests

Coefficient Bootstrap 90% CI
Percentile BC
Direct effects
H1: c' 0.006nsig -0.189 0.194 -0.191 0.192
sig
a1 0.827 0.757 0.884 0.758 0.885
a2 0.637sig 0.509 0.748 0.506 0.745
sig
b1 0.259 0.022 0.474 0.030 0.482
sig
b2 0.509 0.365 0.670 0.356 0.661
Indirect effects Point estimate Percentile BC VAF
sig
H2: a1×b1 0.214 0.018 0.396 0.025 0.403 39.3%
sig
H3: a2×b2 0.324 0.217 0.459 0.208 0.451 59.6%
sig
Total indirect effect 0.538 0.373 0.715 0.372 0.714 98.9%
Notes: sig: Significant. nsig: Not significant. BC: bias corrected. VAF: variance accounted for.

(ii) Type of mediation and magnitude

Once we have determined the significance of the two mediation effects. We can go for the
second step to determine the type of mediation and its magnitude. Table 1 shows the point
estimate for the direct effect (c’), the indirect effects (a1×b1, a2×b2) and the total indirect effect

((a1×b1) + (a2×b2)). Given that c’ is not significant, and both the indirect and the total indirect
effects are significant, a full mediation can be defended. In addition, we can calculate VAF to
assess the magnitude for each mediation. It can be said that almost 99% of the total effect is
due to two mediation effects jointly. Because the VAF exceed 80%, this implies an additional
argument for a full mediation.

(iii) Comparison of mediating effects

When we evaluate a multiple mediator model, we can go further comparing the different
mediating effects. In our example, we want to test whether the unlearning context (M1) has a
stronger mediator effect than the organizational agility (M2) variable. With this aim in mind,
we will assess the potential statistical difference between a1×b1 and a2×b2 following the
guidelines provided by Chin et al. (2013) and Rodríguez-Entrena et al. (2016). Thus, we will
include a new column where we estimate the difference between a1×b1 and a2×b2 and calculate
percentile and bias-corrected CI. Because we have not postulated any hypothesis about the
differential impact of both indirect effects, we will carry out a two-sided test (95% CI) (Figure
11).

Two-sided test = 95% CI

Insert a column for

estimating the differential
impact M1 – M2 =
(a1 b1) – (a2 b2)

Figure 11: Example 1. Comparison of mediating effects. Spreadsheet illustration.

The test (Table 2) shows there is not a differential impact between M1 and M2 since both CIs
contain the zero value. Accordingly, we cannot state that the unlearning context (M1) has a
stronger mediator effect than the organizational agility (M2) variable, and vice versa.

Table 2. Example 1. Comparison of mediating effects.

Coefficient Bootstrap 95% CI
Percentile BC
Differential effect
M1-M2 = (a1×b1) –(a2×b2) -0.110 -0.455 0.174 -0.440 0.189

3.2. An example of a multi-step multiple mediator model

Our second example has been extracted from Roldán, Sánchez-Franco and Real (2017). This
study examines post-adoption behaviors (i.e., frequency of use, routinization and infusion) and
their effects on the sense of community in the domain of social network sites. Specifically, this
contribution formulates a multi-step mediator model where frequency of use affects social
integration via routinization and infusion (Figure 12). The data was collected from 278 users of
Tuenti, a popular social network site among the Spanish college student population during the
period 2006-2012.

a3
Routinization Infusion

a1 b2

a2 b1

Frequency of c’ Social
use integration

H1 = Frequency of use à Social Integration = c’

H2 = Frequency of use à Routinization à Social Integration = a1b1
H3 = Frequency of use à Infusion à Social Integration = a2b2
H4 = Frequency of use à Routinization à Infusion à Social Integration = a1a3b2

Figure 12: An example of a model with a three-path mediated effect. Source: Roldán, Sánchez-
Franco and Real (2017)

3.2.1. Data collection and Measures

Undergraduate students, users of the Tuenti social network, were recruited from social studies
at a public University in Southern Spain. A total of 278 questionnaires were collected from
members who responded to an offline survey.
Frequency of use is defined as the number of times that an individual uses a social network site
(SNS). It was operationalized by two self-reported measures. Routinization describes the state
in which SNS use is no longer perceived as out-of-the-ordinary but becomes institutionalized,
being associated with habitual and standardized usage; that is, the integrating of the SNS into
daily routines. We measure it by adapting a scale developed by Sundaram et al. (2007). Infusion
is conceptualized as the extent to which a person uses an SNS to its highest level to maximize
its potential, implying the notion of a deeper use. We use an adaptation of the measure
developed by Jones, Sundaram and Chin (2002). Finally, social integration measures both the
sense of belongingness to, and the identification with, the SNS and the social community’s
interactivity level. Consequently, social integration is modeled as a multidimensional construct
composed of two dimensions: community participation and community integration. All
variables have been estimated in Mode A.

3.2.2. Hypotheses development

Considering the research model described in Figure 12, we have postulated one direct and three
mediating hypotheses, one of them proposing a three-path mediated effect:
H1: Frequency of use is positively related to social integration
H2: The relationship between frequency of use and social integration is positively mediated by
routinization.
H3: The relationship between frequency of use and social integration is positively mediated by
infusion.
H4: The relationship between frequency of use and social integration is sequentially and
positively mediated by routinization and infusion.

3.2.3. PLS-SEM practical considerations

We follow the guidelines described in the previous example. Therefore, we will show the final
results of our analyses in order to avoid excessive redundancy in the explanation (Figure 13).

a3 = 0.404sig
Routinization Infusion
R2= 0.376 R2= 0.334

a1 = 0.613sig b2 = 0.498sig

a2 = 0.235sig b1 = 0.277sig

Frequency of c’ = .014nsig Social

integration
use R2= 0.489

H1 = Frequency of use à Social Integration = c’

H2 = Frequency of use à Routinization à Social Integration = a1b1
H3 = Frequency of use à Infusion à Social Integration = a2b2
H4 = Frequency of use à Routinization à Infusion à Social Integration = a1a3b2
sig: significant based on one-sided test
nsig: not significant based on one-sided test

Figure 13: An example of a model with a three-path mediated effect. Results. Source: Roldán,
Sánchez-Franco and Real (2017).

The evaluation of our research model involves estimating the significance of one direct (c’) and
the three indirect effects (a1×b1, a2×b2, and a1×a3×b2). Considering that our hypotheses have
been formulated with direction (+), we will use a one-sided test, calculating 90% confidence
intervals (Figure 14)

Figure 14: Example 2. Final spreadsheet with the estimation of indirect effects and confidence
intervals.

(i) Significance of direct and indirect effects

Frequency of use has no significant direct effect on social integration (H1: c’) (Table 3).
Therefore, H1 is not supported. On the other hand, all the indirect effects of frequency of use
on social integration are significant. This means that H2 to H4 have been supported. Thus,
routinization positively mediates the relationship between frequency of use and social
integration (H2: a1×b1). Likewise, infusion mediates the path between frequency of use and
social integration (H3: a2×b2). Finally, we find that frequency of use is positively associated
with higher routinization and infusion, which relates to higher levels of social integration (H4:
a1×a3×b2).
Table 3. Example 2. Summary of mediating effects tests
Coefficient Bootstrap 90% CI
Percentile BC
Direct effects
H1: c' 0.014 nsig -0.073 0.107 -0.074 0.106
sig
a1 0.613 0.548 0.672 0.547 0.671
sig
a2 0.235 0.132 0.339 0.131 0.339
sig
a3 0.404 0.286 0.521 0.286 0.521
sig
b1 0.277 0.165 0.386 0.166 0.386
sig
b2 0.498 0.408 0.584 0.408 0.584
Indirect effects Point estimate Percentile BC VAF
H2: a1×b1 0.170 0.101 0.240 0.101 0.240 40.0%
H3: a2×b2 0.117 0.063 0.176 0.063 0.176 27.6%
H4: a1×a3×b2 0.123 0.080 0.172 0.079 0.171 29.0%
Total indirect effect 0.410 0.343 0.479 0.343 0.478 96.6%
Notes: sig: Significant. nsig: Not significant. BC: bias corrected. VAF: variance accounted for.

(ii) Type of mediation and magnitude

Table 3 indicates that c’ is not significant, and all postulated indirect effects are significant.
Consequently, this means that routinization and infusion fully and jointly mediate the influence

of frequency of use on social integration. This is also supported by applying the variance
accounted for (VAF) index. When the VAF has an outcome above 80%, a full mediation can
be assumed. This occurs when we assess the total indirect effect of frequency of use on social
integration (VAF = 96.6%).

(iii) Comparison of mediating effects

Finally, we will test whether routinization (M1) has a stronger mediator effect than infusion
(M2). As we did in the previous example, we evaluate the statistical difference between a1×b1
and a2×b2 (Table 4). In this case, we do not observe a significant difference between both
indirect effects.
Table 4. Example 2. Comparison of mediating effects.
Coefficient Bootstrap 95% CI
Percentile BC
Differential effect
M1-M2 = (a1×b1) –(a2×b2) 0.053 -0.067 0.175 -0.067 0.175

4. Conclusion
PLS-SEM is a statistical procedure for structural equation modeling that social science
researchers can consider when conducting research. This chapter helps readers to understand
how PLS-SEM can be applied in mediation analysis through two illustrative examples. PLS-
SEM seems not to be “a panacea for flaws in research design or execution” (Rigdon, 2016:
604), but research must not ignore the proper model assessment prior to drawing a conclusion.
This kind of advanced modeling (i.e. mediation analysis) can be performed by PLS-SEM as
illustrated by this chapter. The adoption of these guidelines is advised for researchers who use
PLS-SEM, particularly when they tackle multiple mediation models.

5. References
Aguinis, H., Edwards, J.R., and Bradley, K.J. (2016): Improving Our Understanding of
Moderation and Mediation in Strategic Management Research. Organizational Research
Methods, forthcoming.

Baron, R. M. and Kenny, D. A. (1986), "The Moderator-Mediator Variable Distinction in Social

Psychological Research: Conceptual, Strategic and Statistical Considerations", Journal of
Personality and Social Psychology, Vol. 51 No. 6, pp. 1173-1182.

Cegarra, J.G. and Sánchez, M.T. (2008) “Linking the individual forgetting context with
customer capital from a seller’s perspective”, Journal of the Operational Research Society,
Vol 59, pp 1614-1623.

Chernick, M. R. (2011), Bootstrap Methods: A Guide for Practitioners and Researchers, Wiley.

Cheung, G. W. and Lau, R. S. (2008), "Testing Mediation and Suppression Effects of Latent
Variables: Bootstrapping With Structural Equation Models", Organizational Research
Methods, Vol. 11 No. 2, pp. 296-325.

Chin, W. W., Kim, Y. J. and Lee, G. (2013). “Testing the Differential Impact of Structural Paths
in PLS Analysis: A Bootstrapping Approach”. In H. Abdi et al. (eds.), Springer Proceedings in
Mathematics and Statistics (Vol. 56, pp. 221–229). New York: Springer Science + Business
Media.

Hair, J. F., Black, W. C., Babin, B. J. and Anderson, R. E. (2010), Multivariate Data Analysis,
Prentice Hall, Englewood Cliffs.

Hair, J. F., Hult, G. T. M., Ringle, C. M., and Sarstedt, M. (2017). A Primer on Partial Least
Squares Structural Equation Modeling (PLS-SEM), 2nd edition. Thousand Oaks: Sage.

Hayes, A. F. (2009), "Beyond Baron and Kenny: Statistical Mediation Analysis in the New
Millennium", Communication Monographs, Vol. 76 No. 4, pp. 408-420.

Hayes, A. F. and Scharkow, M. (2013), "The Relative Trustworthiness of Inferential Test of the
Indirect Effect in Statistical Mediation Analysis: Does Method Rally Matter? ", Psychological
Science, Vol. 24 No. 10, pp. 1918-1927.

Henseler, J. (2017): “Bridging Design and Behavioral Research With Variance-Based

Structural Equation Modeling”. Journal of Advertising, (forthcoming).

Henseler, J., Hubona, G. and Ray, P. A. (2016), "Using PLS Path Modeling in New Technology
Research: Updated Guidelines", Industrial Management & Data Systems, Vol. 116 (1), 2-20.

Henseler, J., Ringle, C. M. and Sinkovics, R. R. (2009), "The Use of Partial Least Squares Path
Modeling in International Marketing", in Sinkovics, R. R. and Ghauri, P. N. (Eds.) Advances
in International Marketing, Emerald Bingley, pp. 277-320.

Iacobucci, D., Saldanha, N. and Deng, X. (2007), "A Mediation on Mediation: Evidence That
Structural Equation Models Perform Better Than Regression", Journal of Consumer
Psychology, Vol. 17 No. 2, pp. 140-154.

Jansen, J.J.P., George, G., van den Bosch, F.A.J. and Volberda, H.W. (2008) “Senior team
attributes and organizational ambidexterity: The moderating role of transformational
leadership”. Journal of Management Studies, Vol 45, No. 5, pp. 982–1007.

Jones, E., Sundaram, S. and Chin, W.W. (2002) “Factors leading to sales force automation
use: A longitudinal analysis”, Journal of Personal Selling & Sales Management, Vol. 22, pp.
145-156.

Lau, R. S. and Cheung, G. W. (2012), "Estimating and Comparing Specific Mediation Effects
in Complex Latent Variable Models", Organizational Research Methods Vol. 15 No. 1, pp. 3-
16.

Lu, Y. and Ramamurthy, K. (2011) "Understanding the link between information technology
capability and organizational agility: An empirical examination", MIS Quarterly, Vol 35, No.
4, pp 931-954.

MacKinnon, D. P., Fairchild, A. J. and Fritz, M. S. (2007), "Mediation Analysis", Annual

Review of Psychology Vol. 58, pp. 593-614.

Mathieu, J. E., DeShon, R. P. and Bergh, D. D. (2008), "Mediational Inferences in

Organizational Research: Then, Now, and Beyond", Organizational Research Methods, Vol.
11 No. 2, pp. 203-223.

Nitzl, C., Roldan, J.L. and Cepeda-Carrion, G. (2016): Mediation Analysis in Partial Least
Squares Path Modeling: Helping Researchers Discuss More Sophisticated Models. Industrial
Management & Data Systems, forthcoming.

Preacher, K. J. and Hayes, A. F. (2008), "Asymptotic and Resampling Strategies for Assessing
and Comparing Indirect Effects in Multiple Mediator Models", Behavior Research Methods,
Vol. 40 No. 3, pp. 879-891.

Quinn R.E. and Rohrbaugh J. (1983) “A spatial model of effectiveness criteria: Towards a
competing values approach to organizational analysis”, Management Science, Vol 29, No. 3,
pp 363-377.

Rigdon, E. (2016): Choosing PLS path modeling as analytical method in European

Management Research: A realist perspective. European Management Journal, forthcoming

Rodríguez-Entrena, M., Schuberth, F. and Gelhard, C. (2016). “Assessing statistical

differences between parameters estimates in Partial Least Squares path modeling”, Quality &
Quantity, forthcoming.

Roldán, J.L., Cegarra, J. G., and Cepeda, G. (2014). “Building organisational agility through
an unlearning Context”. Paper presented at the European Conference on Knowedge
Management, 2, 834-842.

Roldán, J.L., Sanchez-Franco and Real (2017). “From frequency of use to social integration:
The mediation of routinization and infusion in Tuenti community”. European Research on
Management and Business Economics, forthcoming.

Rucker, D. D., Preacher, K. J., Tormala, Z. L. and Petty, R. E. (2011), "Mediation Analysis in
Social Psychology: Current Practices and New Recommendations", Social and Personality
Psychology Compass, Vol. 5 No. 6, pp. 359-371.

Shmueli, G., Ray, S., Velasquez Estrada, J. M. and Chatla, S. B. (2016), "The Elephant in the
Room: Evaluating the Predictive Performance of PLS Models", Journal of Business Research,
Vol. forthcoming.

Shrout, P. E. and Bolger, N. (2002), "Mediation in Experimental and Nonexperimental Studies:

New Procedures and Recommendations", Psychological Methods, Vol. 7 No. 4, pp. 422-445.

Sundaram, S., Schwarz, A., Jones, E. and Chin, W.W. (2007), “Technology use on the front
line: How technology enhances individual performance”, Journal of the Academy of Marketing
Sciences, Vol. 35, pp. 101-112.

Williams, J. and MacKinnon, D. P. (2008), "Resampling and Distribution of the Product

Methods for Testing Indirect Effects in Complex Models", Structural Equation Modeling Vol.
15 No. 1, pp. 23-52.

Zhao, X., Lynch, J. G. and Chen, Q. (2010), "Reconsidering Baron and Kenny: Myths and
Truths about Mediation Analysis", Journal of Consumer Research, Vol. 37 No. 3, pp. 197-206.