Improving Model Fit in Mplus
Вставка
- Опубліковано 23 лип 2024
- The lecture contains a practical guide on how to increase model fit in Mplus. First we show how removing items with poor loadings affect model fit, and then we look at how correlating error terms improves fit
The resources for this series of lectures (Slides, syntaxes, data) can be downloaded here: bit.ly/31iGzfc
Good job! 👍
I am running a NOCOVARIANCES model (CFAs using a MLR estimator) and my CFIs are suspiciously low (some are 0.4, some are 0) considering the wide use of the scale I am using. Do you have any ideas as to why that might be? I triple checked and the data in the data file is correct, everything adds up. With the other CFAs we ran from the same database but for other variables, there is no problem. Only this one variable has extremely poor fit. Any ideas? Thanks in advance!
Thank you! I try correlate the items suggested bu MI, and I only correlate those belong to the same consturct. The model then fit well, however, mplus gives a warning that:" WARNING: THE RESIDUAL COVARIANCE MATRIX (THETA) IS NOT POSITIVE DEFINITE.
THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE FOR AN OBSERVED
VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN TWO OBSERVED
VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO OBSERVED VARIABLES.
CHECK THE RESULTS SECTION FOR MORE INFORMATION.
PROBLEM INVOLVING VARIABLE T3_ENGA2."
What does it mean and does it matter? Thank you so much!
It means one of your variables has a negative residual variance. You cant have a negative "error term". Look in your Standardized Results (second last part of the default output), item T3+ENGA2 will have a negative number. You need to figure out why. IF all else fails constrain the value manually to be just over zero like T3_ENGA2@0.04;
Thank you so much for your clear explanation. If you co-vary items in the scale, does it impact how you ultimately calculate the scale score? eg. degree of happiness
Hi Tabitha. Yes it does affect how you calculate the scale scores. Thats why you should never create mean scores. What I suggest you do is based on the best fitting model, you save the FACTOR scores and use the factor scores for any (traditional) analysis going forward. These factor scores takes everything into account.
@@MplusforDummies Thank you so much for replying to my query. Could you possibly provide me with more info about what Factor scores are and how I would use them to calculate individual participant scores? perhaps you could point me towards a paper on another UA-cam video - I don't want to use up too much of your time. Thank you again for your help. I have been stuck on this for so long!
Thank you very much for this eye-opening video! I have a question. To improve the model fit, can we remove a factor loading that is less than 0.4 but statistically significant? Should it be both less than 0.4 and non-significant?
Hi Ipek, Remember this is your study and you set the criteria. Psychometric theory indicates that there should be a well defined item, that loads significantly and highly onto a predefined latent factor. What significance, well defined and highly means is debatable. If I were to develop a new instrument for example, I would want all the items to load really well (so Id even go as far as to say 0.60-0.70). If it was a well established instrument, Id look if there was parsimony (atleast three items on each latent factor) and then start removing items till we hit that point,. The only thing thats important is that you are consistent with what every choice you make. So if you are going to remove items below 0.50 for example, but then one factor only has 3 items of which one is 0.40, and you keep it in for just that latent factor, then it would be an issue. TLDR: Make a choice, describe the reason in your methods section, be consistent. 0.50 and below is fine! Hope that helps
Thank you very much for the video. Would it also be okay to correlate error terms that belong to different factors in a latent change score model (e.g., correlate the error terms of TP_1 at different measurement time points)?
You would artificially inflate the model fit if you'd do that. Correlation the same item.at two time points would lead to a massive decrease in Chi2 because it's the same thing being measured. But theoretically things have changed over time..so in the latent change model, I wouldnt do that
Rather look at other means, like constraining or freeing up paths for example
resources cant be download