Hey there, I would like to analyse asymmetric fixed effects in an interaction model. Asymmetric FE simply means splitting your main explanatory into positive and negative changes and test whether the effect of increasing a variable is the same as the effect of decreasing that variable in the opposite direction. This was implemented by Allison (2019) in Stata, R and SAS:

Allison, P. D. (2019). Asymmetric fixed-effects models for panel data. Socius, 5

The steps are
  1. Calculate difference of change scores for a variable
  2. Split them into positive and negative changes
  3. Generate a cummulative positive and negative sum
  4. Regress

Code:
use "https://statisticalhorizons.com/wp-content/uploads/wagerate.dta", clear
gen id = _n
reshape long lwage mar edu urb, i(id) j(year)
xtset id year

* FE interaction model (one 0/1-dummy and one continuous variable)
xtreg lwage mar##c.edu urb i.year, fe robust

* (1) differences
gen mardiff=d.mar
gen edudiff=d.edu
* (1*) initialize 0
replace mardiff=0 if mardiff==.
replace edudiff=0 if edudiff==.
* (2) split positive/negative
gen marpos=mardiff*(mardiff>0)
gen marneg=-mardiff*(mardiff<0)
gen edupos=edudiff*(edudiff>0)
gen eduneg=-edudiff*(edudiff<0)
* (3) cummulative sum
bysort id (year): gen marcumpos = sum(marpos)
bysort id (year): gen marcumneg = sum(marneg)
bysort id (year): gen educumpos = sum(edupos)
bysort id (year): gen educumneg = sum(eduneg)
* (4) regress ...
Do you have any thoughts on how to implement asymmetric FE in interaction model?
  • The first question that came to my mind is whether I should reduce pos/neg changes to one of my main explanatory variables? I analyse a continuous employment characteristic moderated by a binary employer trait. Usually I also show marginal effect plots (linear prediction) of my continuous variable at values of my varaible at groups. I also wonder how I could include post analysis with marginal effects into action.
Code:
* FE interaction model with (double?) asymmetric fixed effects
xtreg lwage marcumpos##marcumneg##educumpos##educumneg urb i.year, fe robust

* FE interaction model with (main) asymmetric fixed effects
xtreg lwage marcumpos##marcumneg##c.edu urb i.year, fe robust
  • Should I interaction positive and negative cummulative changes at all? Of do a split analysis?
Code:
* FE interaction model with positive asymmetric fixed effects
xtreg lwage marcumpos##c.edu urb i.year, fe robust

* FE interaction model with negative asymmetric fixed effects
xtreg lwage marcumneg##c.edu urb i.year, fe robust