Generally, in longitudinal analysis, three data points are the minimum required to fit a line while guarding against reversions to the mean. The lead author is currently assessing data with a third longitudinal timepoint of the ADOS severity measure. That said, the research article tried to guard against this problem by defining the minimum amount of change to be considered a reliable change i.e. a Reliable Change Index statistic (RCI; Jacobson and Truax 1991), which is, in Latex: RCI_{{Z\,SCORE}} = \frac{{({\text{ADOS}}\,{\text{CSS}}{\mkern 1mu} {\text{Time}}{\mkern 1mu} 3 - {\text{ADOS}}\,{\text{CSS}}{\mkern 1mu} {\text{Time}}{\mkern 1mu} 1)}}{{\sqrt {2\left( {SD\sqrt {1 - r_{{xy}} } } \right)^{2} } }} which takes into account test-retest reliability correlations.
Roughly speaking, this meant the the change groups ended up changing dX>+2 or dX<-2 on a 0-10 ADOS severity scale (4 is usually the clinical threshold for ASD diagnosis), while the no change group was defined roughly as between -2<dX<+2.
Furthermore, in data I did help analyze which is currently included in a journal submission, the change groups defined here exhibited altered longitudinal trajectories of white matter fractional anistropy and diffusivity, suggesting biological relevance one might not expect with a statistical reversion to the mean.
> suggesting biological relevance one might not expect with a statistical reversion to the mean.
Just going to note here that regression to the mean is not 'just' statistical. It happens any time two variables are not perfectly correlated. The imperfect correlation could be due just as easily to 'biological relevance' as it could be to some sort of more purely statistical random error like rater error, so pointing out biology is no evidence. (As in the classic example of the flu patient going to the doctor and recovering afterwards; they really were biologically sick and really did biologically recover.)
