# Mediation

## Explanation

In a conceptual model, the concepts are normally placed in a rectangular. We have three concepts, X, the independent variable, Me, the mediator, and Y, the dependent variable.
We assume a causal relationship running from X to Y (See Figure 1). We assume that in part this relationship can be explained by a Mediator, Me (See Figure 2). If you want to graphically depict a mediation mechanism, it may be useful to show the model without the mediator as well but this is a bit a matter of taste.
It is not always necessary to label the paths but for this tutorial it will turn out to be handy. Normally, when there is no sign (or label) it is assumed that the path has a positive valence. It is, however, good practice to include the valence of the paths in your conceptual models.

### path names

• path a is called the total effect of X on Y.
• path a’ is called the direct effect of X on Y.
• the path running through the mediator, bc, is the indirect effect of X on Y.

## Abstract hypothesis/hypotheses

### Mediation

• Hypo1: X leads to Y ($$a>0$$).
• Hypo2: X leads to Me ($$b>0$$).
• Hypo3: Me leads to Y ($$c>0$$).
• Hypo4: Me mediates the relationship of X to Y ($$a' < a \equiv b*c>0$$)

Please note, that:

• hypothesis 4 may hold even if hypothesis 2 and 3 are false.
• hypotheses 2, 3 and 4 may be true but that hypothesis 1 is false.

It is thus good practice to accept your mediation mechanism only if all four hypotheses hold true. Or more precisely, if you cannot refute any of the four hypotheses.

### Suppression

Supression is closely related to mediation. It means that after taking into account Me the initial relations between X and Y becomes stronger.

This may occur when:

• $$a>0$$ and ($$a' > a \equiv b*c<0$$)

or:

• $$a<0$$ and ($$a' < a \equiv b*c>0$$)

## Real life example

X is educational success Me is occupational success.
Y is health

• Hypo1: Educational success leads to a better health.
• Hypo2: Educational success leads to occupational success.
• Hypo3: Occupational success leads to a better health.
• Hypo4: The relationship between educational success and health becomes weaker once we take into account that educational success causes occupational success and occupational success leads to a better health.

You will often encounter Hypo4s formulated as: The causal relation between educational success and health is (in part) explained by occupational success.

I hope you notice that a mediation model is very similar to the model with two direct effects. The only difference is that we now assume a causal path between X and Me and not a correlation between $X_1$ and $X_2$.

## Structural equations

• Y=X
• Y=Me
• Me=X

or, following the syntax of the R package Lavaan

• Y~X + Me
• Me~X

## Formal test of hypotheses

Load the NELLS data.

rm(list = ls())  #empty environment
require(haven)
nells <- read_dta("../static/NELLS panel nl v1_2.dta")  #change directory name to your working directory

Operationalize concepts.

# We will use the data of wave 2.
nellsw2 <- nells[nells$w2cpanel == 1, ] # As an indicator of occupational success we will use income in wave 2. table(nellsw2$w2fa61, useNA = "always")
attributes(nellsw2$w2fa61) # recode (I will start newly created variables with cm from conceptual models) nellsw2$cm_income <- nellsw2$w2fa61 nellsw2$cm_income[nellsw2$cm_income == 1] <- 100 nellsw2$cm_income[nellsw2$cm_income == 2] <- 225 nellsw2$cm_income[nellsw2$cm_income == 3] <- 400 nellsw2$cm_income[nellsw2$cm_income == 4] <- 750 nellsw2$cm_income[nellsw2$cm_income == 5] <- 1250 nellsw2$cm_income[nellsw2$cm_income == 6] <- 1750 nellsw2$cm_income[nellsw2$cm_income == 7] <- 2250 nellsw2$cm_income[nellsw2$cm_income == 8] <- 2750 nellsw2$cm_income[nellsw2$cm_income == 9] <- 3250 nellsw2$cm_income[nellsw2$cm_income == 10] <- 3750 nellsw2$cm_income[nellsw2$cm_income == 11] <- 4250 nellsw2$cm_income[nellsw2$cm_income == 12] <- 4750 nellsw2$cm_income[nellsw2$cm_income == 13] <- 5250 nellsw2$cm_income[nellsw2$cm_income == 14] <- 5750 nellsw2$cm_income[nellsw2$cm_income == 15] <- 6500 nellsw2$cm_income[nellsw2$cm_income == 16] <- 7500 nellsw2$cm_income[nellsw2$cm_income == 17] <- NA # let us scale the variable a bit and translate into income per 1000euro nellsw2$cm_income <- nellsw2$cm_income/1000 # from household income to personal income attributes(nellsw2$w2fa62)
table(nellsw2$w2fa62, useNA = "always") nellsw2$cm_income_per <- nellsw2$w2fa62 nellsw2$cm_income_per[nellsw2$cm_income_per == 1] <- 0 nellsw2$cm_income_per[nellsw2$cm_income_per == 2] <- 10 nellsw2$cm_income_per[nellsw2$cm_income_per == 3] <- 20 nellsw2$cm_income_per[nellsw2$cm_income_per == 4] <- 30 nellsw2$cm_income_per[nellsw2$cm_income_per == 5] <- 40 nellsw2$cm_income_per[nellsw2$cm_income_per == 6] <- 50 nellsw2$cm_income_per[nellsw2$cm_income_per == 7] <- 60 nellsw2$cm_income_per[nellsw2$cm_income_per == 8] <- 70 nellsw2$cm_income_per[nellsw2$cm_income_per == 9] <- 80 nellsw2$cm_income_per[nellsw2$cm_income_per == 10] <- 90 nellsw2$cm_income_per[nellsw2$cm_income_per == 11] <- 100 nellsw2$cm_income_per[nellsw2$cm_income_per == 12] <- NA nellsw2$cm_income_ind <- nellsw2$cm_income * nellsw2$cm_income_per/100

# as an indicator of educational success we will use highest completed level of education in years.
# the rationale behind this coding this I will take the maximum for university as 16.5 (taking into
# account that some masters are 2 years and some 1 year) and subsequently subtract the years needed
# to obtain a university degree given the degree under consideration.

attributes(nellsw2$w2fa102) table(nellsw2$w2fa102, useNA = "always")
nellsw2$cm_education <- nellsw2$w2fa102
nellsw2$cm_education[nellsw2$w2fa102 == 1] <- 6
nellsw2$cm_education[nellsw2$w2fa102 == 2] <- 9
nellsw2$cm_education[nellsw2$w2fa102 == 3] <- 10
nellsw2$cm_education[nellsw2$w2fa102 == 4] <- 11
nellsw2$cm_education[nellsw2$w2fa102 == 5] <- 12
nellsw2$cm_education[nellsw2$w2fa102 == 6] <- 10
nellsw2$cm_education[nellsw2$w2fa102 == 7] <- 11
nellsw2$cm_education[nellsw2$w2fa102 == 8] <- 14
nellsw2$cm_education[nellsw2$w2fa102 == 9] <- 15
nellsw2$cm_education[nellsw2$w2fa102 == 10] <- 16.5
nellsw2$cm_education[nellsw2$w2fa102 == 11] <- 16.5
nellsw2$cm_education[nellsw2$w2fa102 == 12] <- 7
nellsw2$cm_education[nellsw2$w2fa102 == 13] <- 11
nellsw2$cm_education[nellsw2$w2fa102 == 14] <- 14.5
nellsw2$cm_education[nellsw2$w2fa102 == 15] <- 4

# as an indicator of health we will use subjective well being from 5 (excellent) to 1 (bad) thus we
# have to reverse code original variable
attributes(nellsw2$w2scf1) table(nellsw2$w2scf1, useNA = "always")
nellsw2$cm_health <- 6 - nellsw2$w2scf1
##
##    1    2    3    4    5    6    7    8    9   10   11   12   13   14   15   16   17 <NA>
##   55   78  103  204  338  326  282  272  276  205  133   62   48   22   22   29  374    0
## $label ## [1] " wat is het netto inkomen per maand van u en uw partner samen?/van u?/ " ## ##$format.stata
## [1] "%8.0g"
##
## $labels ## Minder dan ¤150 per maand ¤150 - ¤299 per maand ¤300 - ¤499 per maand ## 1 2 3 ## ¤500 - ¤999 per maand ¤1.000 - ¤1.499 per maand ¤1.500 - ¤1.999 per maand ## 4 5 6 ## ¤2.000 - ¤2.499 per maand ¤2.500 - ¤2.999 per maand ¤3.000 - ¤3.499 per maand ## 7 8 9 ## ¤3.500 - ¤3.999 per maand ¤4.000 - ¤4.499 per maand ¤4.500 - ¤4.999 per maand ## 10 11 12 ## ¤5.000 - ¤5.499 per maand ¤5.500 - ¤5.999 per maand ¤6.000 - ¤6.999 per maand ## 13 14 15 ## ¤7.000 of meer per maand weet niet, wil niet zeggen ## 16 17 ## ##$class
## [1] "haven_labelled" "vctrs_vctr"     "double"
##
## $label ## [1] " hoe groot is uw bijdrage in dit inkomen ongeveer? kunt u een percentage noemen " ## ##$format.stata
## [1] "%8.0g"
##
## $labels ## vrijwel geen bijdrage ongeveer 10% ongeveer 20% ongeveer 30% ## 1 2 3 4 ## ongeveer 40% ongeveer 50% ongeveer 60% ongeveer 70% ## 5 6 7 8 ## ongeveer 80% ongeveer 90% ongeveer 100% weet niet ## 9 10 11 12 ## ##$class
## [1] "haven_labelled" "vctrs_vctr"     "double"
##
##
##    1    2    3    4    5    6    7    8    9   10   11   12 <NA>
##  253   48   89  259  233  242  183  229  114   63  887  229    0
## $label ## [1] " wat is uw hoogst voltooide opleiding, dat wil zeggen waarvan u een diploma heef" ## ##$format.stata
## [1] "%8.0g"
##
## $labels ## lagere school ## 1 ## lbo, vmbo-kb\\bbl ## 2 ## mavo, vmbo-tl ## 3 ## havo ## 4 ## vwo\\gymnasium ## 5 ## mbo-kort (kmbo), primair leerlingwezen, bol\\bbl niveau 1 of ## 6 ## mbo-tussen\\lang (mbo), secundair\\tertiar leerlingwezen, bol\\ ## 7 ## hbo ## 8 ## universiteit (bachelor) ## 9 ## universiteit (master, doctoraal) ## 10 ## promotietraject ## 11 ## buitenlandse opleiding, niet goed in te delen, lager onderwi ## 12 ## buitenlandse opleiding, niet goed in te delen, middelbaar on ## 13 ## buitenlandse opleiding, niet goed in te delen, hoger onderwi ## 14 ## geen opleiding ## 15 ## ##$class
## [1] "haven_labelled" "vctrs_vctr"     "double"
##
##
##    1    2    3    4    5    6    7    8    9   10   11   12   13   14   15 <NA>
##  118  223  202  205  117  223  737  586   89  208   12    8   20   17   34   30
## $label ## [1] " wat vindt u, over het algemeen genomen, van uw gezondheid? " ## ##$format.stata
## [1] "%8.0g"
##
## $labels ## uitstekend zeer goed goed matig slecht ## 1 2 3 4 5 ## ##$class
## [1] "haven_labelled" "vctrs_vctr"     "double"
##
##
##    1    2    3    4    5 <NA>
##  438  853 1211  247   48   32

And test the model with Lavaan.

require(lavaan)
var(nellsw2\$cm_income_ind, na.rm = TRUE)

model <- "
# direct effect
cm_health ~ a*cm_education
# mediator
cm_income_ind ~ b*cm_education
cm_health ~ c*cm_income_ind
# indirect effect
bc := b*c
# total effect
total := a + (b*c)
"

fit <- sem(model, data = nellsw2)
summary(fit, standardized = TRUE)
inspect(fit, "r2")  #to obtain r-squared
## [1] 1.021848
## lavaan 0.6-7 ended normally after 12 iterations
##
##   Estimator                                         ML
##   Optimization method                           NLMINB
##   Number of free parameters                          5
##
##                                                   Used       Total
##   Number of observations                          2326        2829
##
## Model Test User Model:
##
##   Test statistic                                 0.000
##   Degrees of freedom                                 0
##
## Parameter Estimates:
##
##   Standard errors                             Standard
##   Information                                 Expected
##   Information saturated (h1) model          Structured
##
## Regressions:
##                    Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##   cm_health ~
##     cm_educatn (a)    0.055    0.008    7.044    0.000    0.055    0.154
##   cm_income_ind ~
##     cm_educatn (b)    0.137    0.008   17.972    0.000    0.137    0.349
##   cm_health ~
##     cm_incm_nd (c)    0.019    0.020    0.959    0.338    0.019    0.021
##
## Variances:
##                    Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##    .cm_health         0.815    0.024   34.103    0.000    0.815    0.974
##    .cm_income_ind     0.887    0.026   34.103    0.000    0.887    0.878
##
## Defined Parameters:
##                    Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##     bc                0.003    0.003    0.957    0.338    0.003    0.007
##     total             0.058    0.007    7.873    0.000    0.058    0.161
##
##     cm_health cm_income_ind
##         0.026         0.122

We need to refute our mediation mechanism. Do you see why?
Well, because income is not related to SWB, at least not when we take into account educational success. Our c-path is not significant. Thus, the reason why we observed a positive relation between income and SWB previously was because of omitted variable bias.