Can anyone answer me, is it indispensable to have control variables in SEM AMOS significant? How we interpret control variables?
Dear Colleague:
It is not "indispensable" to include control/mediator variables in a SEM model to be fitted with AMOS or any other software.
To postúlate direct or indirect effects on your outcome variables depends on the theory you endorse (adopt) to construct the model to be fitted and on the properties of your measures.
As the link provided by Shaban correctly points out, mediator variables are often needed to explain and predict outcomes, but then again the decision to include them is based on more on theory than on the method employed.
One thing to keep in mind in SEM is the principle of parsimony, you should postúlate only those entities (variables) that are strictly necessary for the data to fit your model. That is simpler (less complex) models will result in a better fit.
Although model fit per se is not an end in itself, if you achieve better fit with or without mediator variables, you accumulate evidence in favor of the correctness of the model postulated in support of the theory you are testing.
Hi,
variables have to be incorporated in models (regression or SEM), if they are (a) related to you independent variable of interest and (b) by themselves have an effect on the outcome.
Both issues lead to the respective variable creating a non-causal connection between the IV and the outcome. If you not control for these variables, this non-causal connection is added to the estimated effect and, hence, biases this effect. Think - for extample of a confounder bias - where some variable(s) Z affect both the IV and the outcome. If you omit Z, then the product of both effects (Z-->IV * Z--> outcome) is added to the IV-->outcome effect.
Important issues (learning goals) in this regard is to understand "path tracing" and "d-seperation".
What is important: You MUST NOT control for common dependent variables (for instance, controling for a variable that is affected by the outcome and the IV). This is called collider bias.
You may also introduce yourself to endogeneity and instrumental variable methods that can help in situtions in which you fail to include all variables that adher to criteria a + b above (which is almost always the case ;)
HTH
Holger
- Badges
- Science topic
- Similar topics
- Thesis Research