One or more outcomes in a single model?
Posted: Tue Apr 04, 2006 8:24 am
Dear all,
I have a general question regarding the number of outcomes (endogenous latent variables) in PLS modelling. I came across the advice that each LV should be connected to at least two other LV for identification purposes. Because of this, I built a modell that contained a number of predictors (exogenous LV) and two outcomes, in this case behavior and satisfaction.
However, a colleague argued that reviewers might object to this, because the measurement models of the predictors would be optimized for predicting both outcomes instead of just one. Thus, there should be more error variance ("noise") in the predictors for each outcome. R² is in fact a little lower (delta = -.07).
I'd still prefer a more stable model with less "capitalization on chance", but a lower R² over an instable model with maximum R². This seems quite important to me because I have a second sample and would like to cross-validate the results.
My questions are:
1) Would it be better to build a single model for each of the outcomes?
2) Do you know good references on this topic?
3) Do you have experiences with comments from reviewers? ;-)
Of course, I'm also very interested in your ideas about this topic in general. I appreciate all your comments, and thanks in advance!
Best wishes,
Joachim Schroer
I have a general question regarding the number of outcomes (endogenous latent variables) in PLS modelling. I came across the advice that each LV should be connected to at least two other LV for identification purposes. Because of this, I built a modell that contained a number of predictors (exogenous LV) and two outcomes, in this case behavior and satisfaction.
However, a colleague argued that reviewers might object to this, because the measurement models of the predictors would be optimized for predicting both outcomes instead of just one. Thus, there should be more error variance ("noise") in the predictors for each outcome. R² is in fact a little lower (delta = -.07).
I'd still prefer a more stable model with less "capitalization on chance", but a lower R² over an instable model with maximum R². This seems quite important to me because I have a second sample and would like to cross-validate the results.
My questions are:
1) Would it be better to build a single model for each of the outcomes?
2) Do you know good references on this topic?
3) Do you have experiences with comments from reviewers? ;-)
Of course, I'm also very interested in your ideas about this topic in general. I appreciate all your comments, and thanks in advance!
Best wishes,
Joachim Schroer