Message/Author |
|
Rick Fulton posted on Tuesday, April 14, 2009 - 4:02 am
|
|
|
Is this a proper way to specify a nonlinear, quadratic effect of a latent predictor in Mplus, using the LMS approach? DV BY dv1 dv2 dv3; IV BY y1 y2 y3; IV_sq | IV XWITH IV; DV ON IV IV_sq; Thanks! |
|
|
Yes. |
|
|
I've added a quadratic effect of a latent predictor to a survival model and am getting a strange result. When the linear effect of the latent variable is entered without the quadratic effect, it is significant with a parameter estimate of .871 and a SE of .184. When both the linear effect and the quadratic effect are in the model, neither is significant and they both have parameter estimates of .000. This suggests to me that the linear and quadratic effects are highly collinear but as latent variables are centered in Mplus, I don't understand how that can be possible. Any thoughts you might have would be most appreciated. Thanks! |
|
|
The latent predictor has mean zero, right? So the f and f-square variables shouldn't be so highly correlated. Check the factor scores for them. If that doesn't solve the problem, please send input, output, data and license number to Support and we will look into it. |
|
|
Thanks Bengt - I just checked the Tech4 output and the f and f-square variables are indeed uncorrelated. But why then should including the f-square as a predictor wipe out the effect of the f variable? |
|
|
Why don't you send the input and data to Support so we can look at it. |
|
|
will do - thanks!!!!! |
|
|
I forgot that Cov(f, f*f) = 0 under the assumption (which we make) that f is normal. In principle, uncorrelated predictors should not harm each other but in real data you cannot be sure. A more complex model may dilute significance. |
|
Back to top |