Hi Witold, Chris Stride's web page is a great resource for Mplus code for testing mediation, moderation, and moderated-mediation: http://offbeat.group.shef.ac.uk/FIO/mplusmedmod.htm Just scroll down, Mplus code is available for more than 90 models! Kind regards, Karin

Hi Witold, Chris Stride's web page is a great resource for Mplus code for testing mediation, moderation, and moderated-mediation: http://offbeat.group.shef.ac.uk/FIO/mplusmedmod.htm Just scroll down, Mplus code is available for more than 90 models! Kind regards, Karin

Thank you again dear Karin

Witold

Hi Witold. I had a problem with an interaction term in one of my previous models. I got an extremely high change in the X2 values and asked Linda Muthen about it (se exchange below). I didn't quite understand her explanation, but be aware of this possible problem with interaction terms.

Best

John-Kåre

From: John-Kåre Vederhus Sent: Saturday, July 26, 2014 11:28 AM To: [email protected] Subject: Licence SCB....: Strange output

Dear Dr Muthen

I've done a path analysis With Mplus.

The theoretic framework was the Theory of Planned behavior.

In the last step I wanted to include the interaction between PBC (perceived behavior Control) and INT (intention) = PBC_INT.

Without this interaction, my model fit was fine and evertything worked well. But With the interaction, my model fit went over the cliffs, with an extreme change in X2 and bad RMSEA and CFI fit indices.

I've read in the Byrne book (2011) that fit indices can be unstable when variables are not normally distributed, but the change was enormous (>750 increase in X2, see enclosed output).

Some of the previous ON results (INT on PBC AANAever) were also change to negative values even if the PBC_INT statement was not introduced until later in the analysis).

Strangely, the last

O_BEHAV ON INT PBC_INT

AANAEVER GROUP

statement came out as aspected.

Is there some explanation for this large change in X2? Is there something I could do that would allow me to include the PBC_INT term?

Best regards

John-Kåre Vederhus (Ph.D.)

Addiction Unit

Sorlandet hospital

Reply (Linda Muthen):

The chi-square test of overall fit cannot be relied on when a mediator variable such as Int or PBC has an interaction. Consider the simple case of

y on m x mz z; ! y = b1*m+b2*x+b3*mz+b4*z+e1

m on x z; ! m = a1*x+a2*z+e2

where mz represents the interaction m*z. In this case, the conditional variance

V[y | x, z, mz] = V[b1*(a1*x+a2*z+e2)+b2*x+ b3*(a1*x+a2*z+e2)*z+b4*z+e1]

involves the term V[b2*e2*z] = b3^2*z^2 V[e2], that is, the variance varies with the z values. In contrast, the chi-square test considers a covariance structure where this

variance is assumed to be constant, that is, a single value.

______________________________________________________________

Linda K. Muthen

Muthen & Muthen website: www.statmodel.com