Which regression model should I use? Mlogit is not possible due to violation of the IIA

[Kim Schwartau]

hello everyone,
I am unsure which type of regression to use. I have a dataset sorted by firm and a dependent variable, that is called "engagement" - the variable has multiple possible categories: no engagement; customer, supplier or partner or a combination of any two or three of the outcomes. Here are two examples:

Firm: BOSTON SCIENTIFIC CORP
Engagement: customers; suppliers

Firm: GOODYEAR TIRE & RUBBER CO
Engagement: partners

The initial idea was an mlogit (the engagements are not naturally ordered) where "no engagement" was the base outcome, where I had 7 categories (as there are 7 different outcome combinations of the Enagement variable), but as some categories overlapped, the IIA assumption was violated.
I have now separated the outcomes in the following way:

gen customers = strpos(Engagement, "customers") > 0
gen partners = strpos(Engagement, "partner") > 0
gen suppliers = strpos(Engagement, "suppliers") > 0

gen number_partners = customers + partners + suppliers

I thought of using a nested model or a mixed logit model, but I am not sure how to actually integrate the engagement variable as my dependent variable, as I would like to understand in the end which category and which amount of partners, and maybe also which combination of partners has the highest rrr for my independent variable (which is binary).

I appreciate any advice or code, thanks a lot
Kim

[George Ford]

here's a thought. I suppose it depends on what you're after.

https://stats.stackexchange.com/questions/250438/is-it-possible-to-do-multinomial-logit-regression-where-multiple-outcomes-can-oc

[Kim Schwartau]

thanks a lot George Ford , the example you sent matches my case & data structure.

I am still unsure how to fit the model using a random effect, do you have any advice?

[George Ford]

what exactly is your hypothesis?

[Kim Schwartau]

George Ford "Pleading for net zero has a larger impact on including suppliers and other parties in the engagement than not engaging." - it refers to my variable where firms engage with one or more of the above-mentioned parties (customes, suppliers, partners), my data structure is almost the same as in the example you sent me with the fries and burgers etc., exept that my identifier are CUSIPs and there are no other characteristics or categories i am testing for

[Joseph Coveney]

"Pleading for net zero has a larger impact on including suppliers and other parties in the engagement than not engaging."

I don't follow your hypothesis statement in light of how your variables are described in #1.

How does "not engaging" or "no engagement" differ from just where all three outcome variables are zero?

Or is "no engagement" a separate predictor variable such that despite enforcing a "no engagement" policy the company or firm might still have one or more engagements with customers, suppliers and partners?

For clarity in the description of your problem, you might want to show your data structure a little more explicitly.

[George Ford]

so engagement = f(pleading), but you want to know how the effects differ by the type of engagement?

[Kim Schwartau]

sure, sorry, so my dependent variable is "engagement" and has the following outcomes:


- no engagement
- suppliers
- customers
- partners
- or a combination of suppliers customers and partners

Here are some examples:

Firm: BOSTON SCIENTIFIC CORP
Engagement: customers; suppliers

Firm: GOOGLE
Engagement: no engagement

Firm: GOODYEAR TIRE & RUBBER CO
Engagement: partners

Firm: AES CORP (THE)
Engagement: customers, partners, suppliers



My independent variable is "pledge for net zero" which is binary and either 1 or 0. I would like to find out which outcome category is the most probable, based on the literature I am assuming firms that plead for net zero are more likely to include their suppliers in their engagement than firms that do not. In my mlogit I thought of using "no engagement" as a base outcome, that's why I originally included it. But as described before the mlogit model was rejected. I would like to understand which combination of engagement parties or which engagement party is most affected by the pledge for net zero. I could exclude the "no engagement" outcome of my engagement variable.

Joseph Coveney, but by stating "How does "not engaging" or "no engagement" differ from just where all three outcome variables are zero?" you are right I did not differentiate that in what I stated above. I no engagement is another outcome of my variable engagement, that was not considered in the coding I described above.

I hope I was able to describe my issue, thanks for your advice!

[George Ford]

Start here, I think.

eststo e1: reg customers plead
eststo e2: reg suppliers plead
eststo e3: reg partners plead
suest e1 e2 e3 , robust

eststo e4: reg customers plead suppliers partners
eststo e5: reg suppliers plead partners customers
eststo e6: reg partners plead customers suppliers
suest e4 e5 e6 , robust

[Kim Schwartau]

George Ford thanks a lot.
Is it still ok to choose a linear regression model even if all my variables included are binary?

[George Ford]

sure, especially when you have only dummy regressors.

you can do logit and get margins, and they'll be the same (or so close that it doesn't matter)