Vuong test- no result

[Soyoung Han]

Hello, I have two questions. I'm running zip regressions, and I added the vuong option to see which model fits better (zip vs. poisson). The code is zip gsp expired inflation, inflate(_cons) vuong.

1. However, I get no vuong statistics (as the picture shows). Can anyone guess why this might be happening?
2. What does it actually mean to just include a constant in the inflate option?

Thank you!

[Joao Santos Silva]

Dear Soyoung Han,

1 - The Voung test that was implemented by Stata was not valid; see here for details.
2 - It means that the probability of zero-inflation does not depend on the regressors.

Best wishes,

Joao

[Soyoung Han]

Hello! Thank you for your reply! Does that mean I cannot use vuong test at all in Stata? even with zinb?

[Soyoung Han]

Do you have any suggestion on how to choose between poisson, negative binomial, zero-inflated poisson or zero-inflated negative binomial regressions?

[Joao Santos Silva]

Can you tell us a bit more about what you are modelling?

[Weiwen Ng]

...

1 - The Voung test that was implemented by Stata was not valid; see here for details.
...

As a side note, Joao's linked presentation raises some concerns about Stata's estimation routines where the likelihood is known to have multiple local maxima. Joao pointed out that the Heckman and zero-inflated negative binomial likelihoods are both prone to local maxima and (if I read correctly) not globally concave. Joao cites one author who commented that in estimating a Heckman model, ideally one should conduct a grid search over plausible values of rho (which controls the correlation between the error terms of the selection equation and the main equation - apologies if I have misstated the issue). He also believes that Stata does not conduct a grid search when estimating Heckman models.

Local maxima are a known issue in estimating latent class analysis models, and presumably in finite mixture models also (LCA models are, I believe, a subset of FMMs). I am aware that when fitting LCA models, MPlus starts with widely-varied parameter start values, and saves all results. I have commented elsewhere that Stata does not do this by default. There is an option to do so, however, although Stata doesn't explicitly enable you to save the results of all random starts; I have written elsewhere that you can make Stata do this, but the result appears to be quite a bit slower than some alternative software, and it did involve some program writing on my part.

If Joao is correct, then perhaps it would be a good addition to Stata 16 to add options to randomly vary important starting parameters in the relevant commands, to enable us to automatically save the results and start values, to inspect the saved runs, and to reproduce the highest consistently-converged log likelihood.

[Soyoung Han]

Joao Santos Silva I'm trying to get the impact (a slope parameter estimation) of a status of a trade program on the imports. I'm using 4 methods I specified above because I have excess zeros in my data. The equation goes like import = beta0 + beta1*status + beta2*covariates + error term. I want to figure out which method fits the data the best.

[Joao Santos Silva]

Soyoung Han: if you are looking at imports (which is not a count) you should stick to Poisson regression; the other estimators are only suitable for count data.

Weiwen Ng: I think you read correctly ; indeed it would be great if Stata could improve these procedures.

Best wishes,

Joao

[Soyoung Han]

Joao Santos Silva Thank you for your comment! However, my monthly import data has over-dispersion and import data is in integers, so I thought I could use negative binomial regression as well.

[Joao Santos Silva]

Soyoung Han: in what units are imports measured?

[Soyoung Han]

Joao Santos Silva They are either in USD or EUR, but are in integers.

[Joao Santos Silva]

Thanks, Soyoung Han. They may be integers, but are not counts and you can change their scale (for example converting to pounds), so overdispersion is not well defined in this context. As I said before, Poisson is the only approach valid in this context.

Best wishes.

Joao

[Soyoung Han]

Joao Santos Silva Thank you so much for your comments. After reading your comments and reviewing more literatures, I decided to use ppml for import value data. Additionally, I collected import quantity data to carry out other regressions I mentioned above.

I would like to include multiple fixed effects in the ppml regression, such as country pair fixed effect. So far, my code looks like:

egen id = group(exporter importer product)
egen eximid = group(exporter importer)
xtset id
xi: xtpqml import status inflation i,eximid, i(id) fe vce(exporterproductid)

However, it basically stopped. Do you have any suggestion on how I can deal with this problem? I am having the same problem with xtpoisson.

xtpoisson quantity status i.productid, fe vce(robust)

It basically stopped.

Thank you so much!

[Joao Santos Silva]

That is probably because you have perfect predictors. Try using ppml with all the dummies.

Best wishes,

Joao