I am planning on running a structural equation model using AMOS. Currently I am in the process of checking whether my measurement models fulfill the necessary prerequisites in order to qualify for the calculation of the SEM model. I encountered three issues which give me trouble.
1) Sample size
The majority of my indicator reliabilities are just fine, but I have two borderline cases. I have read that when the sample size is greater than 1000, the minimum value for the indicator reliability is 0,2. My sample size is indeed greater than 1000, but I have missing values and use the Maximum Likelihood method including the FIML-estimation. When I check the sample size in my AMOS output, it gives me an N of over 2000- how is this possible? I really just want to check whether my N is still bigger than 1000 after accounting for the missing values- if it is, my borderline indicator reliabilities are just fine, if my sample size is smaller than 1000 they cause trouble.
1) Chi square difference test
I am also performing a chi square difference test and set the covariance between my 7 factors to 1 subsequently. That worked well except with the covariance between two factors. When running the model with the covariance set to 1 (as I proceeded with all other variances), I get the following error: “Iteration limit reached. The results that follow are therefore incorrect. The model is probably unidentified. In order to achieve identifiability, it will probably be necessary to impose 2 additional constraints.” I have checked for several issues, but I have set all error variance to 0, I have fixed the path from the indicator to the variable to 1 and have fixed the path from the error to the indicator to 1- I am not sure how to handle this, as increasing the maximum number of iterations also does not help.
2) No output for Test for normality and outliers
I would like to check my data for its multivariate normality. I tick the required box in the analysis properties, but AMOS just won’t give me the provided output. Has anybody else encountered this?
I would very much appreciate any kind of help!
Sajid Khan pointed out that your model in AMOS will not run with missing values in the dataset. I usually put the averages instead of the missing values .
Regarding your missing data; this should be a serious assessment before you being your analysis technique. With a sample size of 1000, you are on defensible high ground to consider different techniques.
First, estimate the % of missing data points. You have 2 options here: A. If your missing data is under 5%, you are on defensible high ground to use a listwise deletion technique (there are may citations available). In which case, you'd remove the participants that have missing data. B. If you are over 5%, I recommend using multiple imputations, not averages of missing values, as this is a more rigorous strategy to handle missing data. This can be done using SPSS, and there are online tutorials for it, however, I recommend MPLUS version 6 or higher, because MPLUS does this automatically.
After you have dealt with missing data, You should really look at the distributions of your data. If your data are reasonably normally distributed (this isn't the most rigorous cut off, but some researchers use the rule of less than 25% variables non-normally distributed) then ML is a reasonable estimator. You can check the univariate and multivariate normality of your data in SPSS using the DeCarlo macro. However, at a sample size of 1000, if your data is not normally distributed, you may wish to consider using the asymptotically distribution-free estimator, which is appropriate in cases of non-normal data.
This may clear up some of the issues you are having, of course, you should use the model specification resources posted on this thread as well.
Cheers,
Pat
AMOS is very sensitive to missing values so if you code missing values as 99 AMOS still reads those as missing values. Use the mean score instead of missing values and AMOS will run. I generally make 2 files a regular file with missing values coded as 99 and an AMOS file with the mean score for that variable entered into the data box instead of 99. I do that for every missing value and then use that file to run AMOS.
Thank you very much for your useful comments!
I have dealt with my missing values (under 5%, no pattern to be identified, so I went for listwise deletion) and now managed to get my tests on the multivariate normal distribution. Unfortunately, my data is far from being normally distributed so I wanted to use distribution-free estimator once I start the equation modelling. In Hair et al (2007) they recommend to have ratio of 15 respondents: 1 Parameter if data is nonnormally distributed. Even if I would use multiple imputations for my missing data, I would not achieve that ratio. Is this ratio a very strict cut-off or would a 10:1 ratio be tolerable?
I agreed with Michael that it is about understanding the foundations to the rules.
Here is an interesting article attempting to lay out the various positions on the issues of total sample size and ratios of respondents to "survey" parameters.
http://pareonline.net/getvn.asp?v=9&n=11.
It might help your journey to the question your asked. Rather than answer it directly.
regards Murray
Hi, Can I try your data and get the AMOS output. I think the data may be copied from Excel and paste it in SPSS. If you are interested , kindly send your SPSS data file, your proposed model and your questionnaire. This is my ID [email protected]
Hi,
I have kind of the same problem as Stephanie. I had already used multiple imputations to account for the missing data and my variables have a normal distribution (at least the skewness/ curtosis is at max 1.09). However, I also get the message "Iteration limit reached. The results that follow are therefore incorrect."
I´ve watched one of the Gaskin tutorials on this issue (https://www.youtube.com/watch?v=B7YOv7hSohY). In the tutorial the problem is with one of the manifest indicators, which is than just deleted. But in my case the problem seems to be with three regression weights between latent variables, since they have too high standartized estimates (larger than 1).
Might my model just be too bad? I would be glad for any advice how to interpret or handle the problem. Many thanks in advance.
As for point 2, AMOS will not give output with any missing data. I usually code missing data as 99. However, AMOS still reads this as missing data and will not run analyses. I usually create an AMOS file and put the column mean score for missing values rather than 99. When I have done that AMOS will run fine. This is just a quirk of the AMOS program.
I also concur with Patrick that online tutorials are often very helpful esp on utube which gives screen shots and explanations of how to interpret outcomes. Often they don't go far enough but some have been quite helpful esp to get started.
Stephanie, Did you ever find out why your normality/outlier output was missing? I'm having the same problem. Thanks. Angela
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling