I have a model (attached) which is looking at stability of humour across time, and also looking at whether a best-friend's humour influences your later humour use. The model is based on an APIM approach, hence the parameter constraints (I have indistinguishable dyads).
- When I run this with my full sample (N=457) I get a good fit.
- When I run it with a large subsample, I get good fit (n=356)
- When I run it with a small subsample (the group that I want to do the analysis on!) I get the following warning "The model is probably unidentified. In order to achieve identifiability, it will probably be necessary to impose 2 additional constraints."
Normally I can identify where the problem lies when I get that kind of error, but I have never seen it appear as a function of subsample. It doesn't make sense to me that the model can be missing constraints only when a smaller sample is being used. I've tried creating an SPSS file with only the subsample I want instead of selecting them but this gives exactly the same response.
Please also note that I do not get the warning when I delete the two gender variables. However, I need to include these in response to reviewers' comments. The issue seems to be linked to those two variables though because I still get the warning if I delete the age variables instead of the gender variables.
Thanks in advance :)
Simon
Let me say back what you have told us:
1) With the full sample, you do not get the error.
2) With the subsample, you do not get the error when you remove the two gender variables.
So, I would wonder if the subsample has some distribution problem with gender. Have you run descriptives on your predictors for your subsample to see how (and if) they are distributed as the full sample?
Simon, I am not very familiar with APIM, but here are my two cents:
First, in a small sample one effect may not be significantly different from zero anymore. Then, identification problems may occur. I would just fix all effects to an arbitrary value, e.g., 1.0, and then relax the constraints one by one in order to detect the problem.
Second, did you try different small samples build from your large sample, and can you replicate the problem?
Third, I wonder why you did not control for the covariates at measurement time point 1. Is there a reason that they should have effects only on time point two? Wouldn't it be of interest to investigate a moderator effect of gender?
Regards,
Karin
Let me say back what you have told us:
1) With the full sample, you do not get the error.
2) With the subsample, you do not get the error when you remove the two gender variables.
So, I would wonder if the subsample has some distribution problem with gender. Have you run descriptives on your predictors for your subsample to see how (and if) they are distributed as the full sample?
The information is not sufficient. What is your effect a.l.a. likelihood estimate and other fit indices. if your LE is unusually high and fit indices are still under control (with in limits) the problem is with the data. See if there are negative values in your covariance matrix. Reliability measure might be handy in this situation. Usually SEM always will be in favor of small samples. In your case it is opposite. For small samples you may confirm with either Bartlett's, Levene's or Brown-Forsythe test. However, simple F-test might do well.
Thank you for these useful suggestions. I'll try to get back on to report how I get on with these.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS