Suyin Chang posted on Monday, January 20, 2014 - 9:38 pm
Dears Dr and Dra Muthen,
I've found many different questions in the archives about what does the Undefined R-squared means. I myself have faced a few outputs with it. But the interesting part is that sometimes I have such result without any negative residual variance being present (and very good model fit statistics).
After replicating those models in different software, I found negative R-squared where Mplus gives "undefined". However, couldn't negative R-sqd sometimes be not a problem, as it is described by Bentler and Raykov (1998) or Hayduk (1996)? Thanks!
Suyin Chang posted on Tuesday, January 21, 2014 - 4:02 pm
Thanks Dra. Muthen, but I was not thinking of a specific model in mind.
This sometimes happens and I've read many diferente questions and debates here about this topic. Then yesterday just by chance it occured to me that usually the undefined R-squared in non-recursive models estimated in Mplus are the negative R-squared in Stata and other softwares.
It made me recall those papers, where authors argue that negative R-squared are not always a problem in non-recursive models. They claim that we should try and check the square of the correlation between observed and predicted values of the endogenous variables.
Then, my main general question here is: is there a way to know in which cases the undefined R-squared in Mplus is caused by negative R-squared and in which cases it is something else? Because if there is, we could just check such mentioned correlation for more information when Mplus gives the Undefined R-square result.
Many thanks for any insights you may have on this.
I am applying a growth curve model to data at 7 time points from early childhood to adolescence. The model with a quadratic term fit the data best, but the R-squared at the last time point is "undefined" in my Mplus output (This is not the case in the linear GC model).
Preliminary investigations of the data suggested that scores on the outcome may decrease slightly over time. But at the last data point, it is evident that there is a significant decrease. Thus, I do not doubt the legitimacy of the model. But I do wish to define the "undefined" R-square. What do you suggest is the best way to do so?