|
|
Fixing non-significant slopes - or not? |
|
Message/Author |
|
C. Lechner posted on Saturday, September 21, 2013 - 5:09 am
|
|
|
Dear Linda and Bengt, I have a quadratic growth model in which growth factors (i s q) are regressed on a set of covariates. In an unconditional model, all growth factors have significant means but neither s nor q have significant variance. When fixing the variance of q to zero, however, the variance of s is highly significant. Yet, as often happens, covariates have a significant effect on both s and q when added to the model (the residual variance of q is not fixed to zero in this model). My question is whether you would recommend fixing the variance of q to zero and regressing the covariates only on i and s - or not. I found a lot of conflicting recommendations in the literature concerning this question. I know that the significance test of the variance components is not very reliable, and that covariates can still have significant effects on insignificant variance components. But I lack a clear rationale favoring one way to proceed over the other. Thanks in advance! |
|
|
This is discussed in the Snijders and Bosker book. The recommend not fixing s and q at zero. |
|
Back to top |
|
|