Moderation and Mediation in Regression

Advanced Data Analysis

[email protected]

Indian Institute of Management Kashipur

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 1 / 34
 Mediation
In regression analysis, mediation refers to a statistical process in which the relationship between
an independent variable and a dependent variable is explained, at least in part, by the inclusion
of a mediator variable.
A mediator variable, also known as an intermediate variable or a mediating variable, is a
variable that intervenes or operates between the independent and dependent variables,
mediating or explaining the relationship between them.
To understand mediation in regression, let’s consider an example.

Suppose you are interested in studying the relationship between study time (independent
variable, X ) and exam performance (dependent variable, Y ). However, you suspect that
the relationship might be influenced by a mediator variable, such as self-efficacy (M),
which represents students’ beliefs about their own abilities.
In this case, self-efficacy (M) could be a mediator between study time (X ) and exam
performance (Y ), meaning that it helps explain how study time affects exam performance.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 2 / 34
 Path Diagram

a b

c′
X Y

Figure 1: Simple mediation model with single mediator

This statistical diagram represents two equations:

M = i1 + aX + eM (1)
′
Y = i2 + c X + bM + eY (2)

where i1 and i2 are regression intercepts, eM and eY are errors in the estimation of M and Y ,
respectively, and a, b, and c ′ are the regression coefficients.
[email protected] (Indian Institute of Moderation
Managementand
Kashipur)
Mediation in Regression 3 / 34
 More Examples
a. Employee job satisfaction. Job satisfaction can mediate the relationship between
leadership style and employee performance. A supportive leadership style may lead to
higher job satisfaction, which in turn can improve performance.
b. Customer perception and attitude. Customer perceptions and attitudes toward a product
or brand can mediate the relationship between advertising spending and sales revenue.
Effective advertising can positively influence customer perceptions and ultimately lead to
higher sales.
c. Inventory turnover rate. Inventory turnover rate can mediate the relationship between
inventory management practices and supply chain performance. Efficient inventory
management can lead to a higher turnover rate, which, in turn, can improve supply chain
performance.
d. Skill development and career progression opportunities. Investments in employee
training can lead to improved skills and career opportunities, which can mediate the
relationship between training investment and employee retention.
e. Customer satisfaction. Service quality can influence customer satisfaction, which, in turn,
can mediate the relationship between service quality and customer loyalty. Satisfied
customers are more likely to remain loyal to a brand.
f. Employee creativity and motivation. Organizational culture can influence employee
creativity and motivation, which can mediate the relationship between culture and
innovation performance. A culture that encourages innovation can lead to more motivated
and creative employees, resulting in better innovation outcomes.
[email protected] (Indian Institute of Moderation
Managementand
Kashipur)
Mediation in Regression 4 / 34
 Direct Effect of X on Y
c ′ estimates the direct effect of X on Y . A generic interpretation of the direct effect is that two
cases that differ by one unit on X but are equal on M are estimated to differ by c ′ units on Y .
More formally,

c ′ = [Ŷ |(X = x , M = m)] − [Ŷ |(X = x − 1, M = m)]. (3)

A positive c ′ indicates positive direct effect and a negative c ′ means negative direct effect.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 5 / 34
 Indirect Effect of X on Y
In this model,

a = [M̂|(X = x )] − [M̂|(X = x − 1)] (4)

b = [Ŷ |(M = m, X = x )] − [Ŷ |(M = m − 1, X = x )] (5)

The indirect effect of X on Y through M is the product of a and b. The indirect effect tells us
that two cases that differ by one unit on X are estimated to differ by ab units on Y as a result
of the effect of X on M which, in turn, affects Y .

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 6 / 34
 Total Effect of X on Y
The total effect c quantifies how much two cases that differ by one unit on X are estimated to
differ on Y . That is,

c = [Ŷ |(X = x )] − [Ŷ |(X = x − 1)] (6)

In a simple mediation model, c can be derived by estimating Y from X alone:

Y = i3 + cX + eY (7)

Therefore, the total effect of X on Y is equal to the sum of the direct and indirect effects of X :

c = c ′ + ab. (8)

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 7 / 34
 Baron and Kenny (1986) Method
Baron and Kenny (1989) outlined a 4-step procedure to determine whether there is a mediation
effect.

1 Show that X is correlated with Y . Regress Y on X to estimate and test the path c. This
step establishes that there is an effect that may be mediated.
2 Show that X is correlated with M. Regress M on X to estimate and test path a. This
step essentially involves treating the mediator as if it were an outcome variable.
3 Show that M affects Y . Regress Y on both X and M to estimate and test path b. Note
that it is not sufficient just to correlate the mediator with the outcome; the mediator and
the outcome may be correlated because they are both caused by the input variable X .
Thus, the input variable X must be controlled in establishing the effect of the mediator on
the outcome.
4 To establish that M completely mediates the X − Y relationship, the effect of X on Y
controlling for M (path c ′ ) should be zero. The effects in both Steps 3 and 4 are
estimated in the same equation.

If all four of these steps are met, then the data are consistent with the hypothesis that variable
M completely mediates the X − Y relationship, and if the first three steps are met but the Step
4 is not, then partial mediation is indicated.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 8 / 34
 Baron and Kenny Method: Step 1
df <- read.csv("datasets/mediation_data.csv")
lm.xm <- df %>% lm(ExamPerformance ~ StudyTime, data = .) # Step 1
lm.xm %>% summary()

##
## Call:
## lm(formula = ExamPerformance ~ StudyTime, data = .)
##
## Residuals:
## Min 1Q Median 3Q Max
## -6.8403 -1.5503 0.0297 1.5222 5.4497
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 63.3902 0.8179 77.50 <2e-16 ***
## StudyTime 4.2900 0.1978 21.68 <2e-16 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Residual standard error: 2.538 on 98 degrees of freedom
## Multiple R-squared: 0.8275, Adjusted R-squared: 0.8258
## F-statistic: 470.2 on 1 and 98 DF, p-value: < 2.2e-16
[email protected] (Indian Institute of Moderation
Managementand
Kashipur)
Mediation in Regression 9 / 34
 Baron and Kenny Method: Step 1

In Step 1, we fit a regression model using StudyTime as the predictor and ExamPerformance as
the outcome variable. Based on this analysis, we have ĉ = 4.29(t = 21.68, p < .001) and thus
ExamPerformance is significantly related to StudyTime. Therefore, Step 1 is met: the input
variable is correlated with the outcome.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 10 / 34
 Baron and Kenny Method: Step 2
# Step 2
lm.my <- df %>% lm(SelfEfficacy ~ StudyTime, data = .)
lm.my %>% summary()

##
## Call:
## lm(formula = SelfEfficacy ~ StudyTime, data = .)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.97234 -0.85363 0.06723 0.14637 2.10680
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 1.77448 0.26898 6.597 2.15e-09 ***
## StudyTime 1.03957 0.06507 15.977 < 2e-16 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Residual standard error: 0.8345 on 98 degrees of freedom
## Multiple R-squared: 0.7226, Adjusted R-squared: 0.7198
## F-statistic: 255.3 on 1 and 98 DF, p-value: < 2.2e-16
[email protected] (Indian Institute of Moderation
Managementand
Kashipur)
Mediation in Regression 11 / 34
 Baron and Kenny Method: Step 2

In Step 2, we fit a regression model using StudyTime as the predictor and SelfEfficacy as the
outcome variable. From this analysis, we have â = 1.04(t = 15.98, p < .001) and thus
StudyTime is significantly related to SelfEfficacy. Therefore, Step 2 is met: the input variable
is correlated with the mediator.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 12 / 34
 Baron and Kenny Method: Step 3
# Step 3
lm.mxy <- df %>% lm(ExamPerformance ~ StudyTime + SelfEfficacy, data = .)
lm.mxy %>% summary()

##
## Call:
## lm(formula = ExamPerformance ~ StudyTime + SelfEfficacy, data = .)
##
## Residuals:
## Min 1Q Median 3Q Max
## -4.0628 -0.7519 0.1218 0.6812 4.0014
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 58.8971 0.5470 107.671 < 2e-16 ***
## StudyTime 1.6577 0.2091 7.929 3.79e-12 ***
## SelfEfficacy 2.5321 0.1709 14.812 < 2e-16 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Residual standard error: 1.412 on 97 degrees of freedom
## Multiple R-squared: 0.9471, Adjusted R-squared: 0.946
## F-statistic: 868.7 on 2 and 97 DF, p-value: < 2.2e-16

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 13 / 34
 Baron and Kenny Method: Step 3 & 4
In Step 3, we fit a multiple regression model with StudyTime and SelfEfficacy as predictors
and ExamPerformance as the outcome variable. For the analysis, we have
b̂ = 2.53 (t = 14.812, p < .001). Because both â and b̂ are significant, we can say
SelfEfficacy significantly mediates the relationship between StudyTime and
ExamPerformance. The mediation effect estimate is âb̂ = 1.04 × 2.53 = 2.63.

Step 4, from the results in Step 3, we know that the direct effect
c ′ (cˆ′ = 1.66, t = 7.929, p < .0001) is also significant. Thus, there exists a partial mediation.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 14 / 34
 Sobel Test
The Sobel Test tests whether the indirect effect from X via M to Y is significant.

library(multilevel)
# run the sobel test
fit.sobel = sobel(pred = df$StudyTime,
med = df$SelfEfficacy,
out = df$ExamPerformance)

# calculate the p-value

(1 - pnorm(fit.sobel$z.value))*2

## [1] 0

The relationship between StudyTime and ExamPerformance is significantly mediated by

SelfEfficacy.

The Sobel Test is largely considered an outdated method since it assumes that the indirect effect
(ab) is normally distributed and tends to only have adequate power with large sample sizes.
Thus, again, it is highly recommended to use the mediation bootstrapping method instead.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 15 / 34
 Bootstrapping
library("mediation")

# bootstrapped mediation
fit.mediation = mediate(model.m = lm.xm,
model.y = lm.mxy,
treat = "StudyTime",
mediator = "SelfEfficacy",
boot = T)

# summarize results
fit.mediation %>% summary()

##
## Causal Mediation Analysis
##
## Nonparametric Bootstrap Confidence Intervals with the Percentile Method
##
## Estimate 95% CI Lower 95% CI Upper p-value
## ACME 10.863 9.212 12.75 <2e-16 ***
## ADE 1.658 1.203 2.10 <2e-16 ***
## Total Effect 12.520 11.085 14.06 <2e-16 ***
## Prop. Mediated 0.868 0.820 0.91 <2e-16 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Sample Size Used: 100
##
##
## Simulations: 1000

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 16 / 34
 Bootstrapping
The mediate() function provides 95% bootstrapped confidence intervals of Average Causal
Mediation Effects (ACME), Average Direct Effects (ADE), combined indirect and direct effects
(Total Effect), and the ratio of these estimates (Prop. Mediated). The ACME here is the
indirect effect of M (total effect - direct effect) and thus this value tells us that the mediation
effect is significant.

In this case, our fit.mediation bootstrap model shows a signifcant affect of SelfEfficacy on
the relationship between StudyTime and ExamPerformance, (ACME = 10.863, p < .001) with
significant direct effect of StudyTime (ADE = 1.658, p < .001) and significant total effect
(Total Effect = 12.520, p < .001).

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 17 / 34
 Bootstrapping
plot(fit.mediation)

ACME

ADE

Total
Effect

5 10 15

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 18 / 34
 Parallel Mediation Model
Parallel mediation analysis is a statistical technique used in social science research to examine
the mediating processes that explain the relationship between an independent variable and a
dependent variable.
It allows researchers to investigate multiple mediators simultaneously and assess their
independent and joint effects on the outcome variable.
In organizational behavior research, parallel mediation can be used to examine the mechanisms
through which organizational factors impact employee outcomes. For instance, a study might
investigate the mediating effects of job satisfaction, organizational commitment, and work
engagement in explaining the relationship between leadership styles and employee performance
or job satisfaction.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 19 / 34
 Parallel Mediation Model - Path Diagram

M1

a1 b1

c′
X Y

a2 b2

M2

Figure 2: Parallel mediation model with two mediators

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 20 / 34
 Parallel Mediation Model
This statistical diagram represents three equations:

M1 = iM1 + a1 X + eM1 (9)

M2 = iM2 + a2 X + eM2 (10)
Y = iY + c ′ X + b1 M1 + b2 M2 + eY (11)

In a multiple mediator model, the indirect effects are referred to as specific indirect effects.
specific indirect effect is interpreted just as in the simple mediation model, except with the
addition of “controlling for all other mediators in the model.” Thus, a model with two mediators
has two specific indirect effects, one through M1 (X → M1 → Y ) and another through
M2 (X → M2 → Y ).
In this model, the specific indirect of X on Y through M1 is a1 b1 , the specific indirect effect
through M2 is a2 b2 . When added together, the specific indirect effects yield the total indirect
effect of X on Y through all mediators in the model. In a model with two mediators:

Total indirect effect of X on Y = a1 b1 + a2 b2 , (12)

and the total effect of X ,

c = c ′ + a1 b 1 + a2 b 2 . (13)

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 21 / 34
 Moderation
Moderation in regression refers to the interaction effect between two or more independent
variables on the dependent variable.
It helps to understand whether the relationship between two variables differs across different
groups or levels of another variable.
This can provide insights into the boundary conditions of the relationship and help identify when
and for whom the relationship is stronger or weaker.
Moderation analysis is typically conducted by including an interaction term in the regression
model. The interaction term is created by multiplying the two independent variables together.
The significance and direction of the interaction term indicate whether there is a significant
moderation effect and how the relationship between the variables changes based on the values of
the moderator variable.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 22 / 34
 Examples
Cultural Differences: Moderation analysis helps researchers understand how cultural factors may
influence the strength or direction of a particular relationship. For instance, a study might
examine whether the relationship between income and happiness is moderated by cultural values
such as individualism or collectivism.
Gender Differences: Moderation analysis allows researchers to investigate potential gender
differences and understand how factors such as gender roles or socialization can moderate the
effects of certain variables.
Personality Traits: Researchers might explore whether the relationship between job demands
and job satisfaction is moderated by an individual’s level of resilience or extraversion.
Intervention Effectiveness: A study might examine whether the effectiveness of a mental health
intervention varies based on the severity of symptoms at baseline.
Organizational Context: In organizational research, researchers might explore whether the
relationship between leadership style and employee performance is moderated by the level of
team cohesion or organizational climate.
Educational Research: Moderation analysis can help explore whether the relationship between
parental involvement and academic achievement differs based on the student’s socioeconomic
status.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 23 / 34
 Moderation: Path Diagram

X a Y

Figure 3: Simple Moderation Model

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 24 / 34
 Conditional and Unconditional Effects
Consider a multiple regression model of the form Y = i1 + b1 X + b2 M + eY , which estimates Y
from two predictors X and M. More specifically, suppose i1 = 4, b1 = 1, and b2 = 2 and
therefore,
Ŷ = 4.0 + 1.0X + 2.0M
If M is held constant, as X increases by one unit Ŷ changes by 1.00 unit. The effect of a one
unit change in X on Ŷ is unconditional on M, in the sense that it does not depend on M.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 25 / 34
 Conditional and Unconditional Effects
Now, consider a model

Y = i1 + f (M)X + b2 M + eY (14)

where f (M) is any function of M. Let, f (M) = b1 + b3 M and hence, we have,

Y = i1 + (b1 + b3 M)X + b2 M + eY
= i1 + b1 X + b2 M + b3 MX + eY (15)

To see what effects adding the product of X and M as a predictor has, consider a specific
example:
Ŷ = 4.0 + 1.0X + 2.0M + 1.5XM
Now a one-unit change in X results in a change in Ŷ that depends on M. For instance, when
M = 0, changing X by one unit changes Ŷ by one unit, but when M = 1, changing X by one
unit changes Ŷ by 2.5 units.
More generally, in a model of the above form, the effect of a one-unit change in X on Ŷ is
expressed by the function

θX →Y = b1 + b3 M (16)

where θX →Y is known as conditional effect of X on Y .

[email protected] (Indian Institute of Moderation
Managementand
Kashipur)
Mediation in Regression 26 / 34
 Interpretation of the Regression Coefficients
Observe that if M is set to 0, then θX →Y = b1 . So b1 is the conditional effect of X on Y when
M = 0. That is, b1 quantifies how much two cases that differ by one unit on X but with M = 0
are estimated to differ on Y . For any value X = x ,

b1 = [Ŷ |(X = x , M = 0)] − [Ŷ |(X = x − 1, M = 0)]. (17)

Similarly, when X is set to 0, θX →Y = b2 . Thus, b2 is the conditional effect of M on Y when

X = 0. For any value M = m,

b2 = [Ŷ |(M = m, X = 0)] − [Ŷ |(M = m − 1, X = 0)]. (18)

For any value X = x and M = m,

b3 = ([Ŷ |(X = x , M = m)] − [Ŷ |(X = x − 1, M = m)])−

([Ŷ |(X = x , M = m − 1)] − [Ŷ |(X = x − 1, M = m − 1)]). (19)

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 27 / 34
 Example
df <- read.csv("datasets/mathmod.csv")
lm.mod <- lm(math~training*gender, data=df)
lm.mod %>% summary()

##
## Call:
## lm(formula = math ~ training * gender, data = df)
##
## Residuals:
## Min 1Q Median 3Q Max
## -2.6837 -0.5892 -0.1057 0.7811 2.2350
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 4.98999 0.27499 18.146 < 2e-16 ***
## training -0.33943 0.05387 -6.301 8.70e-09 ***
## gender -2.75688 0.37912 -7.272 9.14e-11 ***
## training:gender 0.50427 0.06845 7.367 5.80e-11 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Residual standard error: 0.9532 on 97 degrees of freedom
## Multiple R-squared: 0.3799, Adjusted R-squared: 0.3607
## F-statistic: 19.81 on 3 and 97 DF, p-value: 4.256e-10

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 28 / 34
 Interpretation
Since the regression coefficient (0.504) for the interaction term training:gender is significant
(p < 0.001), there exists a significant moderation effect. In other words, the effect of training
intensity on math performance significantly depends on gender.

When gender=0 (male students), the estimated effect of training intensity on math performance
is βˆ1 = −.34. When gender=1 (female students), the estimated effect of training intensity on
math performance is βˆ1 + βˆ3 = −.34 + 0.50 = 0.16. The moderation analysis tells us that the
effects of training intensity on math performance for males (-.34) and females (.16) are
significantly different for this example.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 29 / 34
 Interaction Plot

plot(df$training, df$math, type='n') ## create an empty frame

abline(5, -.34) ## for male
abline(2.23, .16, lty=2, col='red') ## for female
legend('topright', c('Male', 'Female'), lty=c(1,2), col=c('black', 'red'))

## add scatter plot

points(df$training[df$gender==0], df$math[df$gender==0])
points(df$training[df$gender==1], df$math[df$gender==1], col='red')

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 30 / 34
 Interaction Plot

Male
Female
6
5
4
df$math

3
2
1

0 2 4 6 8 10

df$training

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 31 / 34
 Moderated Mediation
Moderated mediation, also known as conditional indirect effects or mediated moderation, is a
statistical analysis technique that combines elements of mediation analysis and moderation
analysis.
It allows researchers to examine how the indirect effect of an independent variable on a
dependent variable through a mediator variable is influenced by a moderator variable.
Moderated mediation is commonly used in social science research to gain insights into complex
relationships and understand under what conditions certain indirect effects occur.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 32 / 34
 Moderated Mediation: Path Diagram

M
W
w
b
a

c′
X Y

Figure 4: A moderated Mediation Model

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 33 / 34
 Examples
Health Psychology: Researchers may examine how the relationship between a specific
intervention or treatment (independent variable) and health outcomes (dependent variable) is
mediated by factors such as self-efficacy or coping mechanisms (mediator variable), and how
these mediation effects differ based on demographic variables like age or gender (moderator
variable).
Organizational Behavior: In studies of workplace behavior, moderated mediation can be used to
explore how the indirect effect of job satisfaction (mediator) on job performance (dependent
variable) through motivation is influenced by factors like leadership style (moderator). This can
shed light on the conditions under which job satisfaction has a stronger or weaker impact on job
performance.
Education Research: Researchers may investigate how the relationship between a teaching
intervention (independent variable) and academic achievement (dependent variable) is mediated
by student engagement (mediator variable), and how this mediation effect varies depending on
factors such as socioeconomic status (moderator variable).
Social Psychology: In studies examining the effects of prejudice reduction interventions,
moderated mediation can be used to understand how the indirect effect of the intervention on
reducing prejudice (mediator variable) is influenced by individual differences such as empathy
(moderator variable), providing insights into the conditions under which these interventions are
most effective.

[email protected] (Indian Institute of Moderation

Managementand
Kashipur)
Mediation in Regression 34 / 34