Message/Author |
|
|
Hello, I have a two level model, with continues and binary variables This is the model I wish to test: VARIABLE: NAMES obs lnTl Tw lnTc typ fin ini stp jolt tac pos avTW alp LavTw Nobs; USEVARIABLES lnTl fin ini tac pos lnTc LavTw obs; CATEGORICAL = fin ini tac pos; WITHIN = lnTL fin pos; BETWEEN = LavTw; CLUSTER IS obs; MODEL: %WITHIN% tac ini fin ON lnTL ; fin ON pos tac; ! no ini pos ON tac; tac ON ini; lnTc ON lnTL pos ini fin; %BETWEEN% ini tac ON LavTw; I am trying to find a way to do a proper model selection between a few similar models and also get the best estimates. What is the best way to perform it? Is it using the ML estimator (and do model selection from the AIC data?) or the WLSMV (and do the DIFFTEST?) or is it with the Bayes? Is there a way to get an estimator from the Bayes that can be used for model selection? Thanks in advance |
|
|
I would use ML and BIC in your situation. |
|
|
thank you, for some reason i get different parameter estimation using the ML, (with WLSMV i get a significant value for one of the parameters and with ML i get for the same parameter a nonsignificant value). |
|
|
is there a way to calculate the indirect effects when using ML? |
|
|
ML gives logistic regression as the default. WLSMV gives probit regression. Standard errors differ with different estimators so significance may differ. If MODEL INDIRECT is not available, you can use MODEL CONSTRAINT. |
|
|
thank you very much is there a way do get ML to use probit regression for its estimates? |
|
|
Add LINK=PROBIT; to the ANALYSIS command. |
|
|
A few more questions regarding model selection and the ML estimator 1. Is it possible to compare a two level model with a one level model using AIC/BIC? The one level model is nested within the two level model. 2. When I wish to compare two models that are not nested one within the other, (but both may be nested in a saturated model), nevertheless they have the same dependent variables and independent variables but different paths in such way that no mediator is removed do I need to fix variables to @0? or do I only need to fix variables to zero when I remove a dependent variable completely from the model? For example I have the following two models: model 1 and model 2 Model 1: A B C on x D on A B Y on D A B C x Model 2: ( I removed "D on A") A B C on x D on B Y on D A B C x Can I directly compare them using AIC/BIC? Or do I have to dix the effect of A on D to zero , so my model 2 should ectually become: Model 2: A B C on x D on B A@0 Y on D A B C x If I understand correctly when I have a model without variable D, like model 3: Model 3: A B C on x Y on A B C x And I want to use AIC/BIC to do model selection I must use the following : A B C on x D on A@0 B@0 Y on D@0 A B C x |
|
|
I think these general analysis questions are more suitable for Multilevelnet and SEMNET. |
|
Back to top |