You can very well fit regression equation between the studied variables to develop a model.It will be very interesting.

if you don't have a placebo/untreated control group, then this is impssible to say.

If you took a group of persons with autism, measured Glc, divided them with a median split into a high & low Glc group, waited, and then measured Glc again, then I would say the time*Glc interaction (where both groups come toward the middle Glc) would very likely be a regression toward the mean.  To rule out regression toward the mean, you need a control group.  I’m not sure I understand your method perfectly, James.  You mention a placebo (so random assignment where only half got the drug) and you also mention a control group of neuro-typical persons.  Skipping the neuro-typical persons (for now), if you divided both your drug intervention and placebo group (of those with autism) into high/low Glc, waited with both groups, and measured both groups again, you might have some evidence for an effect of the drug.  Do you have a three way interaction such that the group with the drug has Glc come to the middle but the control group remains unchanged pre to post (or doesn’t regress toward the mean as much)?  If so, that seems like genuine evidence!  Adding in a neuro-typical group, you would be able to see if the drug brings Glc to the middle for only those with autism or if it works more generally.  Hope this is helpful. ~ Kevin

Thanks all. I think a little bit of methodological clarification is in order.

Participants (ASD and controls) underwent two MRS scans separated by at least a week. Before each scan they received an acute dose of either citalopram or placebo in a randomised order. Participant and experimenter were blind to drug assignment. This is a crossover design, so all participants got placebo and drug. The order was randomised and counterbalanced, so half got citalopram at T1 then placebo at T2 and the other half got placebo at T1 and citalopram at T2.

I have a correlation between placebo Glx and the difference between citalopram Glx and placebo Glx.

Those who have the highest Glx under placebo have the greatest decrease at citalopram and vice versa. This is true in both ASD and controls (and the whole cohort). So I'm wondering if this is a simple case of RMT or if it suggests the normalising effect of citalopram.

Thanks.

I think your result might be about your calculations rather than about the drug or regression toward the mean.  Let’s say you have any two variables, X and Y.  Then you define a new variable as X-Y.  Make up completely random data for X and Y (e.g., within Excel using the RAND() function).  Calculate the X-Y column.  You’ll find a strong positive correlation between X and X-Y, and a strong negative correlation between Y and X-Y.  You’ll also get correlations with a percent change variable rather than difference (i.e., (X-Y)/X), albeit somewhat weaker.