Multi Level Tobit, Ordered Probit

[Theocharis Iosifidis]

Dear Stata Forum Members,

I would like to kindly raise the following question: I am currently working on my Thesis which examines macroeconomic and financial variables that affect the choice of payment method in M&A deals. There are three potential dependent variables cash payment = 2, hybrid payment = 1, stock payment = 0. For that reason, and based on the existing bibliography I decided to employ an ordered probit regression and a multilevel tobit regression (metobit). My question is the following: I find a specific variable (interest rate of banks to non-financial corporations) to be insignificant when using the ordered probit approach, and significant when I am using the metobit approach. I know that in order to provide a solid answer someone needs to see the data that I use etc, however my simple question is whether is it possible that Ordered Probit and Multilevel Tobit can lead to a different result with respect to significance levels.

Thanks in advance,

Theocharis

[Joseph Coveney]

I'm curious: I'd never before heard of trying to fit tobit models to categorical outcomes, and to those that don't appear to be subject to censoring. Is there a source that recommends this?

I'll hazard an answer to your question: why not? Why would you expect different models to obtain the same result, except by chance?

[Theocharis Iosifidis]

Hi Joseph,

Thanks for your reply. My benchmark paper from Faccio and Masulis (2005) https://onlinelibrary.wiley.com/doi/...7f-1wyFBujb7WM is using both an Ordered Probit and a Tobit model for his output... Basically I am using the ordered probit to classify the payment method as I explained 0,1, or 2 and the Tobit model to examine the percentage of cash that the bidder paid for the acquisition with censoring between 0% and 100%

[Joseph Coveney]

. . . I am using the ordered probit to classify the payment method as I explained 0,1, or 2 and the Tobit model to examine the percentage of cash that the bidder paid for the acquisition with censoring between 0% and 100%

That you have separate outcome variables wasn't in your original post.

I don't know what the article has to say about it, not the least because it's behind a paywall, but again my first take on an answer would be: why not? Even if the underlying phenomena are approximately contemporaneous, perhaps even somehow related, why would you expect that the result for one model fitted to one measure would be the same as for a different statistical model fitted to another measure?

And as others have reminded us, the difference between statistically significant and not statistically significant isn't itself necessarily statistically significant.

[Tom Scott]

As an aside, this seems like a situation where a stereotype logistic regression might be better than an ordered logistic regression: https://www.stata.com/manuals13/rslogit.pdf

See also https://digitalcommons.wayne.edu/cgi...&context=jmasm