Dear My Le,

I recommend that (1) you fix both A12 and A13 parameters as 1, not only A12. (2) Also, you could remove the covariations between your independent variables (or make a second-order latent variable). There are a large number of parameters. (3) You could try to use your dependent variable (DV) as a "direct/observable measure" if you can assume that it is not a latent variable.

Best regards, Lucas.

If it would be advisable to use the dependent variable as a direct observable variable

Hi all,

I would agree with Lucas that the problem is in the two-indicator latent variable. I guess that the latent variable "wifi" does not correlate significantly with any other of the variables?

The latent dependent variable requires fixing both the loading and the error variance. Whether to use the manifest indicator of the DV as variable or the latent formulation does not matter for the unstandardized estimate or its standard error but the standardized coefficient.

Finally, eliminating the correlations among the predictors will signify a gross misspecification if in fact these variables correlate and changing the structure to a second-order structure simply implies a different model with different assumptions. Again this will not work if the structure among the independent variable is not caused by a latent second-order factor. In the end, the relevant question is what you think the structure is.

That having said, please note that your model is essentially a regression model--that is the structure is saturated and has zero df (is not tesable). The df of you overall model completely stems from the measurement model. That's something you have to keep in mind when you use the fit as a support for the model.

All the best with your work,

--Holger

Thanks you for all your advices.

I try to run the dependent variable as a direct observable variable.

But I am not confident with the model because I'm lack of experiences and this is the first time I run multivariate analysis. Can you have a look at my model and give me some assessments? I appreciate your work.

My study includes EFA, CFA, SEM. After I ran EFA, it had 6 factors with 23 variables. Next, I ran CFA, I eliminated variables having standardized regression weights below 0,7 because when I multiplied themselves, variances were below 0,5. The result of CFA included 5 factors with 15 variables. Final, I ran SEM with dependent variable (H45 showed in pic).

Did I do wrong at any steps? Please show me solutions. Thanks all.

Hi My Le,

opinions may vary on this, but even after purification, your model does not fit. Forget the ratio of chisquare and df and optimally the fit indexes, too.

The problem is that you use AMOS and this is very limited with regard to diagnostics. Of course you can ignore everything I said as the fit indexes will most satisfy most reviewers.

In case you want to develop, I recommend using Mplus or R (esp. the lavaan package, see lavaan.org for tutorials).

A few weeks, I had a similar case to yours and the diagnostics in R (the "standardized residuals" in addition to some detective work led me to a hypothesis of a better model.

The annotated code can be seen here:

https://htmlpreview.github.io/?https://github.com/IcarusAE/Mixed-open/blob/master/COSOMO%20validation%20study%20analysis.html

The background of the study was validating a 5-factor model with worries about the corona crisis. As it turned out, we had some false assumptions about the meaning and factor structure of one of the latent factors.

Sometimes it pays out to look to the details :)

All the best,

--Holger

Thank you for your advice. Now, I dont have enough time to use another program except AMOS.

Can you have me answer one more question?

As you know, I have several datasets with latent variables which each have 3-4 items. However, I also have several single observed variables and I am wondering whether I can use these as dependent variables in SEM or not. I am under the impression that in SEM structural models dependent variables have to be latent variables. Is that correct? These are the types of variables that I am wondering to use as dependent variables:

1) A single Likert Scale variable.

2) A normally distributed count variables.

3) A non-normally distributed count variable (with a large number of 0 counts and then a normal distribution centering around a count of 3)

Many thanks in advance,

My Le.

Hi,

you can have any kind of variable in a SEM. There are a few statistical issues, though

a) For the likert scale, it depends on the number of response scales and whether you can assume that the responses are semi-equally spaced intervals. With 5 or (better) more this should work

b) With the count variables, you again would have to switch to another program. These cannot currently be handled by AMOS or Lavaan. The issue is the error distribution which is usually a poison or negative-binomially variable. For these cases, I would use sum scores of the independent variables and go for a manifest count regression model. This should also work in SPSS. With the large number of zero's, you would possibly use a hurdle model or ZIP/ZINB model (zero-inflated poison or neg-bin).

Don't simply use a count variable in AMOS. Your standard errors will blow up and you get non-significant effects despite a population effect. True count models are much more efficient.

Lachenbruch, P. A. (2002). Analysis of data with excess zeros. Statistical Methods in Medical Research, 11(4), 297-302. doi:10.1191/0962280202sm289ra

Delucchi, K. L., & Bostrom, A. (2004). Methods for analysis of skewed data distributions in psychiatric clinical studies: Working with many zero values. American Journal of Psychiatry, 161(7), 1159-1168. doi:10.1176/appi.ajp.161.7.1159

Atkins, D. c., & Gallop, R. J. (2007). Rethinking how family researchers model infrequent outcomes: A tutorial on count regression and zero-inflated models. Journal of Family Psychology, 21(4), 726-735. doi:10.1037/0893-3200.21.4.726

Blevins, D. P., Tsang, E. W. K., & Spain, S. M. (2015). Count-based research in management: Suggestions for improvement. Organizational Research Methods, 18(1), 47-69. doi:10.1177/1094428114549601

All the best,

--Holger

My dependent variable (student satisfaction) is an observed variable. And independent variables are dimensions of service quality, which are latent.

My hypothesis: There is a significant positive relationship between each dimension of service quality and student satisfaction. Is it acceptable to the relationship between DV (observed variable) and IV (latent variable)? I can't find any articles supporting this relationship.

Hair (2010) stated: "With SEM, we examine relationships between latent constructs much as we examined the relationships between independent and dependent variables in multiple regression analysis". It means that my dependent variable should have been latent?

That's what I'm worried. So sad because I couldn't collect more data about student satisfaction while it only had one 1 question.

What should I do? Many thanks in advance.