A question.

I wanted to use asymmetric effects (see Allison 2019).

Basically: it divides a variable into two in panel data, making the changes in the positive direction one variable and the changes in the negative direction another variable.

Here is the Stata syntax for a two-wave panel data (revised from Allison's paper):

gen x=d.A

gen xpos=x*(x>0)

gen xneg=-x*(x<0)

bysort id (year): gen xpositive = sum(xpos)

bysort id (year): gen xnegative = sum(xneg)

now, when I use this without this division, namely, xtreg Y A, A is non-significant and positive.

But when I use this with this division, namely, xtreg Y xpositive xnegative, both are significant and NEGATIVE, which is very counterintuitive. I did t-tests for each waves but obviously it doesn't matter as this measures the difference.

what could be the reason?