How to Control for Continuous Variables in Sem Amos
Ok firstly the two models you are using are not nested (they have different sized covariance matrices; one that includes masculinity one that does not) so directly comparing the two models fits in the way you are is not really appropriate.
You could nest them if you wanted to by having masculinity in the model but just not predicting anything and not being predicted by anything but that seems a bit silly. If you think masculinity should be in the model leave it in the model.
Yes RMSEA and TLI take into account model parsimony but again you cannot compare non-nested models using these fit indices.
MAIN MESSAGE: Do not get pushed around by fit indices OR software.
Ah I see the problem having read your link. It seems that AMOS wants you to do something different when correlated variables versus correlated residuals (I suppose making the distinction clear is kind of a nice feature). In your model you want to correlate the latent residuals of attitudes norms etc. 
Ah I see the problem having read your link. It seems that AMOS wants you to do something different when correlated variables versus correlated residuals (I suppose making the distinction clear is kind of a nice feature). In your model you want to correlate the latent residuals of attitudes norms etc.
Okay, that sort of makes sense (?). So if I correlate the residuals, that would end up looking the same statistically as having covariances between the latent variables? Also, one of the three variables I need to covary is exogenous, so I would need to also show relationships between that exogenous variable and the two endogenous ones, but it is not a causal path so I don't want a one way arrow. Would I correlate the exogenous variable with the two residuals too?
Your explanation about the nested models makes sense--so my model isn't nested because I have an entire extra variable rather than a few paths. So are there any indices or methods I can use to compare non-nested models that don't penalize for model complexity? I haven't been able to find anything so far. I wanted to be able to see if the masculinity variable contributed anything to the model, but if that contribution is obscured by the fact that there are a bunch of extra parameters, then it's hard to tell which model is better. They both exhibit adequate fit at this point.
P.S. I am totally thanking you on my dissertation acknowledgments page...I'd be so stuck on this analysis without your help.
You can use AIC, BIC, and sample size adjusted versions of the same to compare non-nested models. I am not really sure I see the point in this case though, if your theory or research questions is that masculinity should have a significant effect then just model it and report the results. I think model trimming is bad (removing non-significant paths) as it can provide a biased picture of the system under investigation...thus dropping variables is a very very naughty thing to do IMHO.
As for you first question yes you can correlate a latent variable and its residual if you like.
Okay, so I should just report the original models and not try to change them even though some of the variables aren't really contributing to it?
The AIC and BIC penalize for model complexity too, don't they? So the simpler model will always look better? The comparison is because I was testing the two different models to see if the added variable would explain more.
Also...all of the observed variables in my model ended up meeting criteria for multivariate normality except for the dummy coded experience variable, which is skewed (probably because 2/3 had experience and 1/3 didn't). Would one skewed observed variable like that significantly affect the analysis--should I have used an alternative estimation method? Or would ML still be okay?
Anyone have any advice on the last question? Is it still okay to use MLE with one skewed binary variable? I can't apply a transformation to something that is dichotomous, so I'm kind of stuck. Thanks for any help you can offer.
Thanks again! I feel like my analysis is a huge mess at this point. No one mentioned at my proposal meeting that you couldn't directly compare models with different sets of variables and they approved my proposal with the intent to compare the models, so I feel like my entire dissertation is pretty stupid now, but I can't go back and pretend that I didn't want to compare the models. My stats committee member was on sabbatical at the time, of course, but he will be back for my defense just in time to see how little I know about SEM.
Okay, more questions...modification indices and residuals this time! Per your advice, I decided to leave the models as is and not remove variables even though the standardized regression weights were super low and one of the variables isn't even significantly correlated with the dependent variable. However, I did at least want to look at any missing parameters. There are a few fairly large modification indices in my model (e.g., above 20; the rest are all quite low and wouldn't really add anything to the model). One makes theoretical sense and I plan to add it and then test the new model. However, I am not sure about the others. Some of them are between two error terms that are measured variables of the same latent variable (i.e., errors of two subscales measuring masculinity). Is it really bad form to covary these error terms? Obviously these subscales are supposed to be highly correlated, since they are all measuring different facets of masculinity.
Standardized residuals...I have read that anything over 2 or 2.5ish is bad. I have quite a few that are in the 3 and 4 range. But the model fit itself is good according to the fit indices. I have read that the modification indices are a better indicator of misspecification than the residuals, but still...I don't know what to do with all of the large residuals. I can't see anything that is horribly missing in my model.
Finally, one of the largest modification indices suggests that I should add a path between one of the measured variables for my DV and one of the latent predictors. Trying to wrap my brain around it...it makes it seem like the measured variable for the latent DV would also potentially fit as a measured variable for the latent predictor? If so, it is very problematic because it makes little sense theoretically. It is even weirder because the observed variables for the predictor and the observed variables for the latent DV don't even have a significant bivariate correlation (the correlation is actually .05 so it's very low). Does anyone have any idea why this might occur? Do the large residuals and the modification indices suggest multicollinearity?
hello all,
my question is how to enter categorical endogenous variable in amos?
hello martin,
can you please help me on bayesian estimation. In the model proposed by me i have a endogeenous variable which was orginially measured on one statement on scale. I have converted my endogeneous construct "adoption" into two categeories i.e. high=1 and low= 0.
please help me how to run bayesian extimation in the presence of three contol variables...age, gender (1=male, 0 female), Involvement( 1= high and 0=low).
thanx a lot in advance
I have a similar question to the person who initiated this thread and I saw that someone replied that if it had been the outcome variable, then it would be a different story. I have been instructed to use AMOS to analyse longitudinal data where I have one independent variable that is continuous, and an outcome variable that is binary. The independent variable have been measured at 5 time points. the dependent variable was measured 3 times. So, I am trying to predict my binary outcome variable from a continous variable. Does anybody know if/how this works in amos?
And if I later want to divide my continuous independent variable into two groups (e.g. depressed vs not depressed), what kind of analysis could I do then?
Any help would be greatly appreciated!
Source: https://www.talkstats.com/threads/sem-with-categorical-variable-parceling-how-to-enter-in-amos.55030/page-2
0 Response to "How to Control for Continuous Variables in Sem Amos"
Post a Comment