

Comparing models with different varia... 

Message/Author 

Sally Czaja posted on Tuesday, June 20, 2006  12:03 pm



I would like to compare models (using AIC and AIC weights) in which a 2nd and 3rd predictor is added to a base model that has just 1 predictor. I'm confused on how to specify the simpler models. If the base model is "Y on P;" and the 2nd model is "Y on P C;" (C is a control variable), then the base model has a better (lower) AIC and an AIC weight of 1 (i.e., they're not even close), even though the Rsquared is better (higher) in the 2nd model. Should C be in the base model? (if so, how exactly?). The 3rd model introduces a mediator: "Y on P C M;" "M on P C;". If the comparison model is "Y on P C;", then the simpler model is preferred again. I want to make sure I'm doing this right before I interpret these results. Thanks. 


You need to include the full set of variables in each model for the loglikelihood values to be on the same scale. y ON p c m; y ON p c m@0; o ON p c@0 m@0; 


I want to run a multigroup single mediation SEM analysis with two groups (all continuous data). The problem is that when I initially fit the model separately for each group, the same model does not fit. For one group, I had to remove an indicator for one of the latent variables, because even with the suggested modification indices, the model was still poor fitting. So, the groups have the same number of latent variables, but different numbers of indicators for one of the latent variables. Are there any current solutions for this issue? Would it be inappropriate to do the analysis using a multigroup approach to establish invariance since both groups need the same number of latent variables and indicators? In this case, it is more appropriate to just perform a single group analysis for each group? 


This general analysis question is suitable for SEMNET. 

Back to top 

