6
$\begingroup$

I'm trying to build a SEM that looks like the picture shown below:

enter image description here

Where:

  • $X$ and $X^2$ are the independent variables
  • $Y$ is the dependent variable
  • $M_1$ and $M_2$ are the mediators
  • $W$ is a moderator

None of the default mediation packages that I've tried so far support such a structure so I'm using lavaan SEM instead. I've been reading up on Hayes and have figured out the model without the moderator:

  # direct effect
  Y ~ cprime1*X + cprime2*X2 + controls

  # mediator 1 paths
  M1 ~ a1*X+a2*X2  + controls          
  Y ~ b1*M1

  # mediator 2 paths
  M2 ~ a1prime*X + a2prime*X2 + d*M1 + controls
  Y ~ b2*M2
       
  # indirect effect (a*b) using the Hayes2010 formula (Table 1) 
  ab :=  (a1 + 2*a2*X)*b1 + (a1prime + 2*a2prime + (a1+2*a2*X)*d ) * b2
  
  # total effect
  cprime := (cprime1+2*cprime2)
  total := cprime + (ab)

I'm aware of how to moderate with 1 input variable, for instance for M1 you'd have something like:

M1 ~ a1*X1 + a2*W + a3*X1:W + controls

But it's unclear to me how one would add the moderators for X2 in these and the other parts of the model.

$\endgroup$
2
  • $\begingroup$ I am quite unsure about your indirect effect. You are looking to 2 indirect effects and the equation only compute a single one. $\endgroup$
    – POC
    Apr 2, 2021 at 16:51
  • $\begingroup$ I'm following Hayes to compute the total indirect effect as the sum of the individual indirect effects. $\endgroup$
    – ciri
    Apr 2, 2021 at 20:04

2 Answers 2

1
+50
$\begingroup$

I simulated some irrelevant data:

library(lavaan)

dat <- as.data.frame(mvrnorm(5e2, rep(0, 6), matrix(.25, 6, 6) + .75 * diag(6)))
colnames(dat) <- c("X", "X2", "M1", "M2", "W", "Y")
dat$X2 <- dat$X ^ 2
head(dat)
#            X         X2         M1         M2          W          Y
# 1 -1.2556613 1.57668526  0.9558917 -0.6155703 -0.4540778 -0.3747208
# 2 -0.8463471 0.71630340  0.2066086 -0.1680590 -0.9181445 -0.2819832
# 3  0.1081230 0.01169059 -0.1934604  1.2662057  0.4817797 -0.5342619
# 4  0.6180336 0.38196555  0.3301925 -0.1277026  1.3222274  0.3626635

Given these data, this is the model code you want:

summary(sem(
  "M1 ~ a11 * X + a12 * X2 + w10 * W + w11 * X:W + w12 * X2:W
  M2 ~ a21 * X + a22 * X2 + w20 * W + w21 * X:W + w22 * X2:W + d * M1
  Y ~ c1 * X + c2 * X2 + w30 * W + w31 * X:W + w32 * X2:W + b1 * M1 + b2 * M2
  a11db2 := a11 * d * b2
  a12db2 := a12 * d * b2
  a11b1 := a11 * b1
  a12b1 := a12 * b1
  a21b2 := a21 * b2
  a22b2 := a22 * b2",
  dat), rsquare = TRUE)

With this, you can work out the exact combinations that get you the total indirect, direct and total effects. The nature of W makes things a bit difficult. Right now, the computed effects at the bottom are for when W = 0. You might want to play around with the computed effects depending on the type of variable W is.

$\endgroup$
4
  • $\begingroup$ Thank you, this is very helpful! How would one get the effect size for one standard deviation above/below the value of a continuous W? $\endgroup$
    – ciri
    Apr 9, 2021 at 1:08
  • $\begingroup$ @ciri standardize W before adding it to the model. Then add its interaction coefficient wherever it might be of interest. $\endgroup$ Apr 9, 2021 at 2:32
  • $\begingroup$ I think library(MASS) is required for running mvrnorm. $\endgroup$
    – Galen
    Oct 18, 2022 at 21:39
  • $\begingroup$ I have requested the developers of semopy to reproduce this example. $\endgroup$
    – Galen
    Oct 18, 2022 at 23:47
2
$\begingroup$

Moderation is basically the product of the IVs ($X_i$) with the moderator ($W$). Since you have two IVs, you have two products $X_1W$ and $W_2W$, you just have to do it once for each predictor $X_i$ on each outcome ($M_1,M_2,Y$), since you want the moderator on each of them.

I used the notation in the graph for the code,

Y ~ cprime1*X + cprime2*X2 + w3_1*X1:W + w3_2*X2:W + controls    
M2 ~ aprime1*X1 + aprime2*W + w2_1*X1:W + w2_2*X2:W + + d*M1 + controls
M1 ~ a1*X1 + a2*W + w2_1*X1:W + w2_2*X2:W + controls

plus your indirect effects.

You can also consider a triple interaction X1:X2:W. You would have to add the double interaction of $X_1X_2$ and the triple interaction $X_1X_2W$. At this point you would have an IV and two moderators (conceptually speaking; statistically, it is arbitrary because there is no difference between the variables). I don't think it is what you want though.

$\endgroup$
2
  • $\begingroup$ Could you further clarify the full model? What would the implication be of adding a triple interaction? $\endgroup$
    – ciri
    Apr 3, 2021 at 15:52
  • $\begingroup$ I edited my answer. $\endgroup$
    – POC
    Apr 7, 2021 at 17:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.