Hello,
I have survey data that I am attempting to use in IBM AMOS to create a SEM UTAUT model. However, during output, for "Result" I get:
Minimum was achieved
The model is probably unidentified. In order to achieve identifiability, it will probably be necessary to impose 7 additional constraints.
Chi-square = 2417.406
Degrees of freedom (corrected for nonidentifiability) = 82
Probability level = .000
I know close to nothing about statistics, and am a total newbie when it comes to AMOS, but from what I gather, that chi-square is bad; also for P CMIN/DF I get 29.481, which I also think is not great since it is greater than 3.
The SEM Model is supposed to be UTAUT model. Worst case scenario, can I still use the data as is? If so, how to correctly interpret the data? I have provided the model.
Any help is appreciated. Thank you.
In general, you should not interpret the parameter estimates or fit statistics for an underidentified model. They may be completely incorrect/misleading.
SEM is a complex statistical methodology. You mention that you "know close to nothing about statistics." I would say, given that, it is not a good idea for you to use SEM unless you are guided by a statistician who is an expert in SEM and AMOS. There is a lot that can go wrong with SEM.
In your case, you have two latent variables (PerformanceExpectancy and EffortExpectancy) that each have only one indicator (observed/measured variable). These latent variables are not identified unless you add more indicators or fix the error variances of the indicators to a meaningful value. Also, your endogenous (dependent) latent variable does not seem to have an indicator (measured variable) at all.
Christian Geiser, thank you for your reply and advice. I had been studying how to use the software and the idea for quite some time now and have read papers that use software for analyzing UAUT models. Applying the UTAUT Model to Understand Factors Affecting the Use of E-books in Fujian, China by Chen Lin (2019), as suggested by Kai Heinrich was the best resource for me. Link to resource: https://www.researchgate.net/deref/https%3A%2F%2Fwww.diva-portal.org%2Fsmash%2Fget%2Fdiva2%3A1447041%2FFULLTEXT01.pdf
As per your advice, and reading this resource: https://www.ibm.com/support/pages/should-single-indicator-latent-variables-be-used-amos, I have set my single indicator latent variables to 0. I have also moved one indicator from one latent variable to the dependent latent variable. The question could technically be applied to either, so this was fortunate. Further, I added covariances to error variables belonging to the same construct by examining modification indices and removed some indicators with low estimates for standardized regression weights as suggested by this video: https://www.youtube.com/watch?v=3Y7dOXINV2E and others, while testing the changes along the way.
The biggest problem, it turned out to be, was my data, which was very porous. Cleaning the data up by removing responses with missing answers and those with insincere answers, along with doing the above resulted in much better-looking results (see attached images).
I looked at the following results (same as Lin 2019) and the numbers seem good: CMIN/DF, GFI, AGFI, TLI, CFI, RMSEA. I also looked at Standardized RMR, which was .0439; this was suggested by the maker of the YouTube video above.
I am using UTAUT and IBM SPSS for my dissertation as it has become more widely used in my field, Human/Machine Interaction.
I hope these results are acceptable, please let me know if not and any suggestions or thoughts you might have.
Thank you again!
I would be skeptical about including so many data-driven modifications (error covariances) based on modification indices alone. This is typically not something that one would recommend, and also something that cannot be tested with the same data (i.e., the tests of model fit for the modified model are misleading/incorrect when conducted on the same data). In essence, you cannot explore and confirm model structures on the same set of data. You also risk capitalizing on chance when adding error covariances purely based on modification indices. It is not something that one should do in my opinion.
I would also think about whether it is a good idea to delete missing data. Typically, it is better to include cases with missing values/partial data and to use modern missing data techniques such as full information maximum likelihood (FIML; which I believe is available in AMOS) or multiple imputation.
- Badges
- Science topic
- Similar topics
- Thesis Research