The TECH10 residuals you are looking at are the best way to assess model fit with categorical items and maximum likelihood estimation. There are no absolute fit statistics.
tomlife posted on Wednesday, March 25, 2015 - 3:49 am
I have a similar problem.
I analyzed Data using (nine dimensional) ordinal Rasch model in ConQuest.
Due to the fact that Conquest doesn't compute standard errors or confidence intervall, which are necessary to compare the latent correlations betwenn the dimensions, I tried to reproduce the model in Mplus. So I set all factors (for each dimension) equal and the variance (for each dimension) @ 1, estimation via MLR.
For 16 of 18 relevant correlations the results by Mplus are nearly the same as by ConQuest (+/- .02), but for the other two correlations there are greater differences (> .1) which also change the result pattern. Do yor have any ideas, what causes these differences?
You want to compare the loglikelihood values that the 2 programs give to see if they have found the same solution. Although, I don't know if ConQuest uses ML (MLR) or some other estimator. With 9 dimensions using ML(R) there is also the issue of precision in the numerical integration. Are you using integration=montecarlo(5000)?