Hello. I am running a fixed effects models with an interaction of a continuous and categorical variable. I hypothesize a moderation effect of some categories that alleviates a negative effect of my continuous variable. The outcomes is life satisfaction (thus ordered dependent in principle).

My base model setup looks promising: My interaction terms are stat. sig. (as a first indication). Thus I check the marginal effect of my main regressor which is also significant. I further check this ME at representative values of my categorical variable (this is stat. sig.) and finally the difference of these AME of continous at representative of the categorical variable.


Code:
*******************************************************************************
*** life satisfaction panel data
* https://www.diw.de/documents/dokumentenarchiv/17/diw_01.c.412698.de/soep_lebensz_en.zip
use soep_lebensz_en.dta
xtset id year
*******************************************************************************

*******************************************************************************
// BASE INTERACTION MODEL
* run FE regression with interaction of continuous and "categorical" variable
set more off
xtreg satisf_org  c.education##c.education##no_kids, fe

* average marginal effect of education is stat. sig.
margins, dydx(education)
* AME of education at representative values stat. sig.
margins, dydx(education) at(no_kids = (0 1 2 3)) post

* difference of AME is not significant (in this example) but for my data
lincom _b[1bn._at] - _b[2._at]
*******************************************************************************
Next I want to do a robustness check with a nonlinear ordered model. I noticed the brand new community command feologit from:

Baetschmann, G., Ballantyne, A., Staub, K. E., & Winkelmann, R. (2020). feologit: A new command for fitting fixed-effects ordered logit models. The Stata Journal, 20(2), 253-275.

Code:
*******************************************************************************
// ROBUSTNESS CHECK: fixed effects ordered logit with "blow up cluster" (BUC)
feologit satisf_org  c.education##c.education##no_kids, group(id)

* I don't know exactly what this means.
logitmarg, dydx(education)

* I can also calculate this (but it's subject to special assumptions).
margins, dydx(education)
margins, dydx(education) at(no_kids = (0 1 2 3))  post
lincom _b[1bn._at] - _b[2._at]
*******************************************************************************
I know from "help feologit postestimation" that "predict and margins contain assumptions about the fixed effects and will produce inconsistent estimates when these assumptions do not hold." As chapter 4.3 from the Journal says "estimates for MEs cannot generally be obtained".

What can you recommend for conducting the robustness check and proceeding? Kind regards