Assumptions test after negative binomial regression

[Thang Nguyen]

Dear all,

I have something to ask about assumptions test after negative binomial regression.

After finishing negative binomial regression, (nbreg y x1 x2 .....).

I would like to assess effects of the assumptions on regression model employed on the variance of the coefficient estimates.

Could you give me any suggestions ?

Thank you.

Best regards.

[George Ford]

I'd rewrite your inquiry, adding more detail. I have no idea what you are asking.

[Carlo Lazzaro]

Thang:
welcome to this forum.
As George receomended, more details about your query are needed to reply positively.
That said, usually researchers go -nbreg- following the results of an overdispersed -poisson- regression.
If this is the case, you may find https://journals.sagepub.com/doi/pdf...867X1401400406 useful.

[Thang Nguyen]

Dear George Ford,

After I found overdispersion in the Poisson regression, I ran a negative binomial regression.

Now, I want to explore the effects of the assumptions on the regression model employed on the variance of the coefficient estimates.

I don't know how to do that. I intend to assess the assumptions of the nbreg model, such as multicollinearity, robust standard errors, ..... and compare them with the Poisson regression results to verify changes in standard errors and model fit. I can do the Variance-Covariance Matrix.

Could you help me find the best way to approach this problem?

Thank you.

Best regards.

[Jeff Wooldridge]

Poisson regression is consistent for the beta parameters no matter what is the variance-mean relationship. NegBin is not. You’re better off using Poisson regression with robust standard errors. Then the only think to test is the functional form of the conditional mean function.

[George Ford]

See #3:

https://www.statalist.org/forums/forum/general-stata-discussion/general/1587040-why-do-poisson-and-negative-binomial-regressions-yield-the-same-result

And

https://www.statalist.org/forums/forum/general-stata-discussion/general/1567752-hausman-test-for-negative-binomial-fixed-effects-and-random-effects

Some interesting stuff here:

https://www3.nd.edu/~rwilliam/stats3/CountModels.pdf

In most instances, Poisson with robust standard errors is the way to go.

The data is inherently heteroskedastic. Why test it?

Multicollinearity is rarely worth bothering with. MC is about the X's, not the Y. You can compute the VIFs by regressing the X's on each other and making a computation with the R2. Or, just run the model using reg and estat vif. The Y variable doesn't influence the results (unless there are missing values).