Jon Heron posted on Thursday, May 12, 2016 - 5:14 am
we are currently experimenting with some non-linear growth models (specifically exponential decay) in which a Taylor series approximation is used to linearize the target function with respect to the random effects (here intercept, asymptote and half-life).
According the Preacher + Hancock's 2015 paper in Psych methods (http://dx.doi.org/10.1037/met0000028) these models are Population Averaged (PA) such that individual trajectories do not necessarily follow the same functional form as the mean-trend.
We have recently be exploring some strange outliers by exporting the factor scores and plugging values into our exponential target function. In doing so we have produced estimated trajectories that look nothing like those produced by the software.
If this is not merely a short-coming of factor scores and is due to the afore-mentioned PA issue, then how might Mplus be generating its estimated curves if it is not merely plugging values into the target function?