Hi all,
I'm trying to conduct SEM analysis (in R) for a study with experimental design with 3 groups (1 control and 2 experimental), but I seem to get the wrong results in the regressions. Hence my question: how to correctly conduct such SEM (if it is possible)?
Best regards
Radosław
Hello Radoslaw,
if you have the set of conditions and one outcome, you will set up the model as the regression model and get the same results, namely by introducing two dummies having effects on the outcome. What made you think of the results being "wrong". How was your specification?
With best regards
--Holger
Hi Holger,
First of all thank you for your answer. Second of all, I thought that the results were wrong, as the regressions in my SEM model were greatly different from those I obtained in the regression analysis.
I specified the model according to: https://lavaan.ugent.be/tutorial/groups.html. In all I had 3 exogenous variables and 5 endogenous ones, and I treated the experimental groups as the grouping variable.
I might be misunderstanding some key concepts here or how it all works, as I am still very new to SEM, and this is the first time I am conducting SEM with multiple groups.
Thank you again,
best regards
Radosław
I still don't get it 100%. The 3 exoneous variables are not the 3 expermential conditions, right? Just three X's? And you specified a multigroup model in which you had a baseline model linking the X's to the Y's across the three conditions? If yes, than you are testing the conditions as moderatores of the X-Y-effects.
Perhaps, you could be more specific. What was the experiment, what are the conditions, what are the X's and which are the Y's.
Until then,
best regards,
--Holger
The 3 exoneous variables are not the 3 expermential conditions, right? Just three X's?
Yes. The 3 exogenous variables are not the conditions, those were sex, age and trait emphaty. Those variables were supposed to influence attitudes which in turn were supposed to influence behaviour intention. The experimental conditions were supposed to lead to greater behaviour intention compared to the control condition.
And you specified a multigroup model in which you had a baseline model linking the X's to the Y's across the three conditions?
Yes.
If yes, than you are testing the conditions as moderatores of the X-Y-effects.
I see. Does this make sense to you to do so in such a model as the one described above?
Thank you!
Best regards
Radosław
Using the three conditions as moderators makes good sense, but unless you have a measurement model, there is no need to run this analysis as SEM.
In particular, moderation is the same thing as interaction effects, so you can use two dummy variables in a regression to create a set of multiplicative interaction effects.
Here is a link to a detail discussion of how to test a similar model with just one dummy variable using SPSS
Hi Radoslaw,
so we are clear now. As your hypothesis is that the treatment(s) affect intention, the conditions should be part of the SEM in forms of dummy variables besides the 3 controls. I do not know on which theory your model is based on but when you draw on overall attitude-behavior theories (e.g., the theory of planned behaviour), then should it not be reasonable to expect that the effect of the conditions affects intention VIA attitude. That is, I personally would rather test an effect on attitude and intention but of course, this depends on the exact content of the experimental treatment. I thought whether a full mediation (treatment-->attitude--> intention) would be justified but probably there are other, unobserved mediators that lead to a direct treatment --> intention effect.
HTH
--Holger
Hi David and Holger,
In particular, moderation is the same thing as interaction effects, so you can use two dummy variables in a regression to create a set of multiplicative interaction effects.
Thank you! I shall try that approach.
Here is a link to a detail discussion of how to test a similar model with just one dummy variable using SPSS
The link did not get attached.
(...)the conditions should be part of the SEM in forms of dummy variables besides the 3 controls.
Ok! As you noticed I'm basing this research on the theory of planned behaviour, so I guess what you wrote makes sense. I wasn't sure though, whether I can somehow code the 3 conditions into one variable in SEM, so I decided on the other approach. I will try to do the dummy coding.
Thank you!
Best regards
Radosław
Hi Radoslaw,
yes, the procedure how to incorporate moderate effects or interaction effects into a statistical approach is the same (i.e. by using product terms), although conceptually they are different as when focusing on interactions you have two causally effective variables that interact and their common effect is stronger then the sum of their unique effects.
In the case of a moderator the focus is on X and the moderator Z modifies this relationship (that 's why the term moderator should be replaced by effect modifier), see
VanderWeele, T. J. (2009). On the distinction between interaction and effect modification. Epidemiology, 20(6), 863-871.
Be sure to know how to implement the dummy variables (you need two). If not ask.
All the best,
--Holger
Hi Holger,
Thank you again! I'll read up on dummy coding, and if I have any problems implementing them, I'll be sure to ask.
Best regards
Radosław
Here is the link to doing an interaction analysis in SPSS.
https://statistics.laerd.com/spss-tutorials/dichotomous-moderator-analysis-using-spss-statistics.php
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling