Unfortunately, standardized regression coefficients don't determine the relative importance of individual predictors (Willett, Singer, & Martin, 1998, p. 412). Usually, r-squared change (or f-squared) is used to assess effect sizes within the multiple regression framework.
Is there a way in Mplus to determine r-squared for an outcome? If so, then presumably I can determine the r-squared without a given predictor, and then calculate the r-squared with the predictor to determine the change in r-square with the addition of the predictor. Can this be done?
The STANDARDIZED option provides an R-square for each dependent variable based on all of its predictors.
Dan Abner posted on Thursday, December 01, 2011 - 9:32 am
When I specify the STANDARDIZED option, I obtain both Std and StdYX. I, like Isaac, want a partial R-square type of measure of effect size for individual effects. For this purpose, which one (Std or StdYX) should I use?
Hi Linda, You mentioned earlier that the STANDARDIZED option provides an R-square for each dependent variable based on all of the predictors. My SEM includes interactions that are created with the xwith command. Therefore, I believe that only UNSTANDARDIZED output is generated. Is there any way to compute R-square using the UNSTANDARDIZED output, or is this not possible? Thank you, Luke
Anthony posted on Thursday, October 25, 2012 - 1:44 pm
I'm conducting a relatively simple path model and I know Mplus provides the total R-square value for each outcome variable in the model. However, my reviewers are asking for the variance accounted for by each path. Since each outcome has multiple paths, the R-square value doesn't work for this - would it simply be the sum of both the beta (i.e., STDXY estimate) value for the direct effect (none) added to the beta value for the sum of the indirect effects. For instance:
I would like to report the total variance explained (r-squared) for my latent factor outcome variable.
However, the residual variance of the latent outcome factor was a small, negative, and non-significant value. Thus, I fixed the negative residual variance of the latent outcome factor to 0 as recommended from previous posts.
Although the model ran, now I do not obtain an r-squared value for my latent outcome factor.