How to implement -xtdpdml- with interaction terms?

[Titir Bhattacharya]

Hi Statalist,

In my paper I validate a theoretical proposition through the following FE regression:

areg y x1##x2 i.round,a(id)

where y,x1,x2 are binary variables.

There is some concern that there might be reverse causality from y to x1.

Based on Leszczensky, L., & Wolbring, T. (2019) (available here) I am trying to implement the ML-SEM method through the -xtdpdml- command, as robustness check. I have gone through Moral-Benito,Allison,Williams,2016 guide on the command.

Based on my understanding, both x1 and x2 would be predetermined variables in my example. However, I'm not sure how to express the interaction

x1##x2

while using this command.
According to the guide, the syntax should be

xtdpdml depvar strictly_exogenous_variable, inv(time_invariant_variable) pre(predetermined variable)

Should the

x1##x2

go inside the predetermined variable?

Moreover, in my original specification, the time invariant variable was incorporated by absorbing individual level fixed effects through

,a(id)

. I'm not able to understand how I'm going to incorporate the fixed effects in this syntax, and whether I would need separate set of data for time invariant variables to include under

inv

.

I understand this question might seem confusing. It's partly because I myself am confused about how this command is working. i would appreciate it if you asked for further clarification if you need to help me out here.

Thanks,


Edited to add: I tried by adding interaction term in pre():

xtdpdml y, pre(L.x1##L.x2) tfix

and got the following error

factor variables not allowed
(error in option predetermined())

When I changed i into

xtdpdml y, pre(L.x1) tfix

, I got

no observations
r(2000);

[Tom Scott]

Titir Bhattacharya your variables have repeated measures (panel data) right? Inputting some data would help

[Richard Williams]

Stata’s sem command (and hence xtdpdml) does not support factor variable notation. Hence, you need to compute the interaction terms yourself before running xtdpdml.

Why don’t you get the interaction term working and then return if you have more Qs.

[Titir Bhattacharya]

Stata’s sem command (and hence xtdpdml) does not support factor variable notation. Hence, you need to compute the interaction terms yourself before running xtdpdml.

Why don’t you get the interaction term working and then return if you have more Qs.

Thanks Richard. I'll do that

[Titir Bhattacharya]

Titir Bhattacharya your variables have repeated measures (panel data) right? Inputting some data would help

Sure Tom, I'll do that

[Titir Bhattacharya]

Titir Bhattacharya your variables have repeated measures (panel data) right? Inputting some data would help

* Example generated by -dataex-. To install: ssc install dataex
clear
input str8 id float(round y x1 x2)
"IN010001" 2 . . .
"IN010001" 3 1 1 1
"IN010001" 4 1 1 0
"IN010001" 5 1 1 0
"IN010002" 2 . . .
"IN010002" 3 1 1 1
"IN010002" 4 1 1 1
"IN010002" 5 0 1 0
"IN010003" 2 . . .
"IN010003" 3 1 1 1
"IN010003" 4 1 0 0
"IN010003" 5 1 0 1
"IN010004" 2 . . .
"IN010004" 3 1 1 0
"IN010004" 4 1 1 0
"IN010004" 5 1 1 0
"IN010005" 2 . . .
"IN010005" 3 0 0 1
"IN010005" 4 1 1 0
"IN010005" 5 1 1 0
"IN010006" 2 . . .
"IN010006" 3 0 1 1
"IN010007" 2 . . .
"IN010007" 3 0 1 1
"IN010007" 4 0 1 0
"IN010007" 5 0 1 0
"IN010008" 2 . . .
"IN010008" 3 1 1 0
"IN010008" 4 0 0 0
"IN010008" 5 0 0 1
"IN010009" 2 . . .
"IN010009" 3 1 1 1
"IN010009" 4 1 1 0
"IN010009" 5 1 0 1
"IN010010" 2 . . .
"IN010010" 3 1 1 1
"IN010010" 4 1 1 0
"IN010010" 5 1 1 0
"IN010011" 2 . . .
"IN010011" 3 1 1 1
"IN010011" 4 0 1 0
"IN010011" 5 1 0 0
"IN010012" 2 . . .
"IN010012" 3 1 1 1
"IN010012" 4 1 1 0
"IN010012" 5 1 0 0
"IN010013" 2 . . .
"IN010013" 3 1 0 0
"IN010013" 4 1 1 1
"IN010013" 5 1 1 0
"IN010014" 2 . . .
"IN010014" 3 1 1 0
"IN010014" 4 1 1 0
"IN010014" 5 1 1 1
"IN010015" 2 . . .
"IN010015" 3 0 1 1
"IN010015" 4 1 1 0
"IN010015" 5 1 1 1
"IN010016" 2 . . .
"IN010016" 3 0 1 0
"IN010016" 4 1 1 0
"IN010016" 5 1 1 0
"IN010017" 2 . . .
"IN010017" 3 1 1 0
"IN010017" 4 1 1 0
"IN010017" 5 1 0 0
"IN010018" 2 . . .
"IN010018" 3 1 1 0
"IN010018" 4 1 1 0
"IN010018" 5 1 1 0
"IN010019" 2 . . .
"IN010019" 3 1 1 1
"IN010019" 4 1 0 0
"IN010019" 5 1 1 1
"IN010020" 2 . . .
"IN010020" 3 1 0 1
"IN010020" 4 1 0 0
"IN010020" 5 0 0 0
"IN010021" 2 . . .
"IN010021" 3 1 1 0
"IN010021" 4 1 1 0
"IN010021" 5 1 1 0
"IN010023" 2 . . .
"IN010023" 3 1 1 0
"IN010023" 4 1 1 0
"IN010023" 5 1 1 0
"IN010024" 2 . . .
"IN010024" 3 1 1 1
"IN010024" 4 1 1 0
"IN010024" 5 1 1 1
"IN010025" 2 . . .
"IN010025" 3 0 1 0
"IN010025" 4 0 0 0
"IN010025" 5 0 0 1
"IN010026" 2 . . .
"IN010026" 3 1 1 0
"IN010026" 4 1 1 1
"IN010026" 5 1 1 1
"IN010027" 2 . . .
"IN010027" 3 0 1 1
end

[Richard Williams]

A couple of comments:

• I assume round is supposed to be the time var? Or did you not include the time var in your extract?
• When round = 2, y, x1, and x2 have missing values. If that is true for the entire data set you might as well drop cases where round = 2.

This is the code I ran.

encode id, gen(idnum)
xtset idnum round
drop if round == 2 // All other vars are missing when round == 2
gen x1x2 = x1*x2
xtdpdml y, pre(L.x1 L.x2 L.x1x2) tfix // Won't converge
xtdpdml y, pre(L.x1 L.x2) tfix // Converges

I don't know if that is getting close to what you want or not. But, by dropping round = 2, at least you can get results and tweak the model from here. You may want to add ylag(0) or do other tweaks.

Also, at least in the extract, y is a 0/1 dichotomy. For your purposes, perhaps that is ok. But, xtdpdml wants y to be continuous. You might want to look at this thread, especially Allison's post in #5.

https://www.statalist.org/forums/for...int-estimation

[Tom Scott]

You could consider running - xtdpdml y, pre(L.x1) tfix - at various levels of x2 as opposed to including an interaction

[Titir Bhattacharya]

A couple of comments:

• I assume round is supposed to be the time var? Or did you not include the time var in your extract?
• When round = 2, y, x1, and x2 have missing values. If that is true for the entire data set you might as well drop cases where round = 2.

This is the code I ran.

encode id, gen(idnum)
xtset idnum round
drop if round == 2 // All other vars are missing when round == 2
gen x1x2 = x1*x2
xtdpdml y, pre(L.x1 L.x2 L.x1x2) tfix // Won't converge
xtdpdml y, pre(L.x1 L.x2) tfix // Converges

I don't know if that is getting close to what you want or not. But, by dropping round = 2, at least you can get results and tweak the model from here. You may want to add ylag(0) or do other tweaks.

Also, at least in the extract, y is a 0/1 dichotomy. For your purposes, perhaps that is ok. But, xtdpdml wants y to be continuous. You might want to look at this thread, especially Allison's post in #5.

https://www.statalist.org/forums/for...int-estimation

Thanks a lot Richard for your detailed explanation. It really helped clear things up and get me started.

Also, thanks for pointing me towards the relevant thread.

[Titir Bhattacharya]

You could consider running - xtdpdml y, pre(L.x1) tfix - at various levels of x2 as opposed to including an interaction

Thanks Tom for the suggestion.