Is it normal to have a negative pseudo R-Squared (McFadden’s pseudo R-squared) in tobit regression?

[Kevin Traen]

Dear stata users,

First of all: thanks for reading my question.
I have run a tobit model and found a negative pseudo R-squared value (-0,342).
Is this normal and how do I need to interpret this?

Kind regards

[Eric de Souza]

Always helps to see your results for the tobit estimation

[Marcos Almeida]

Welcome to the Stata Forum / Statalist,

There is this very helpful FAQ that enlightens the matter: https://www.stata.com/support/faqs/s...ics/pseudo-r2/

[Kevin Traen]

This is a screenshot. Thanks in advance!

[Eric de Souza]

Screen shots are not accepted by Stata unless they are in png format (see the FAQ).. In any case Marcos has given the answer above in #3. Textbooks say that the pseudo R2 has no meaning or utility because, although it is bounded above by +1 it is not bounded below and can be negative.

[Nick Cox]

The FAQ is helpful but in detail misleading in an irrelevant way as whether the normal density at 0, presumed to be the mean, is more than 1 is contingent on the standard deviation used!

I guess that the point of the pseudo-R-squared in question is limited in the following senses, presumed to be evident on reflection:

* If the log likelihood is what is being maximised, it belongs in measures of fit.

* You should only compare different models fitted to the same dataset! Whether one model is better than another for the same data is easier to judge than whether one model is good or poor.

I like phrases like badness of fit (used by Joseph B. Kruskal in 1969 and perhaps others earlier) and misfit (used by Bowley around 1920).