Comparing between Heckman selection model and PPML

[Imran Khan]

Dear all,

I need to investigate the allocation of U.S. foreign aid. More specifically, the rationale is to check the determinants of U.S aid allocation that a recipient country receives. For this purpose, I have a panel dataset of 100 foreign aid-recipient countries over 2000-2015. The dependent variable has large number of zeros i.e., most countries in the panel doesn’t receive U.S aid in all the years. Due to large amount of zero observations as well as the decision of US whether to allocate aid or not, if allocated, how much money to give, I intend to use either a Heckman sample selection model or Poisson Pseudo Maximum likelihood (PPML) Given this background, I have three questions:

First is there any statistical test in Stata to compare between Heckman sample selection model vs PPML?

Second, how to test the assumptions of the two models?

Third, I believe that the U.S decision is not sequential i.e., it does not sequentially first decide whether to allocate aid and then decide whether to give or not. Rather the two decisions are determined simultaneously. In this scenario of non-sequential decisions, will Heckman sample selection be appropriate? I am actually indifferent about testing the first step i.e., whether the aid is allocated or not as my main interest lies in testing how much amount is given.


Looking forward to your response.

Best regards,
Imran Khan.

[Joao Santos Silva]

Dear Imran Khan,

1 - Yes, there is a test, please search hpc.

2 - The assumptions of the sample selection model are very unlikely to hold; for PPML,try the RESET test.

3 - I think the sample selection model is inappropriate in either case. This is a model to deal with a particular data observability problem and you do not have that in your sample.

Best wishes,

Joao

[Imran Khan]

Dear Santos,

Many thanks for your reply.

1. I have found the hpc test and it seems to serve my purpose.

2. Is there still a way to test Heckman sample selection assumption as it has been widely used in the aid literature?

3. What if I want to investigate the decision of whether to give aid or not along with how much aid to give. Since PPML will only test the latter.

4. Will logged absolute amount of aid with zero values be an appropriate dependent variable in the case of PPML? Are they any other suitable tests to check the PPML estimator other than HPC and RESET?

Best regards,
Imran khan

[Imran Khan]

Dear Santos,

I tried to generate the following code for HPC test. Could you have a look at it and let me know if it makes some sense?

**PPML
ppml Aid GDP Edu Distance
predict fit1, xb

**Heckman selection
heckman Aid GDP Edu Distance, select(Aid GDP Edu Distance Democracy) 
predict fit2, xb

**HPC Test
hpc Aid GDP Edu Distance, a(fit1) b(fit2)

Best regards,
Imran Khan

[Joao Santos Silva]

Dear Imran Khan

I am afraid I that is not correct; please check the help file and read the paper on which the command is based.

Best wishes,

Joao

[Imran Khan]

Dear Joao,

Sorry for the delay in reply.

I just finished reading the paper and the helpfile. What I understood now is that I am doing some mistake in calculating the fitted values of the conditional expectation of the two competing models.

However, I couldn’t make the most out of the helpfile as I am not yet familiar with the codes on bootstrap.

I am wondering 1) If learning the bootstrap technique is essential for conducting the HPC test and 2) how can I calculate the fitted values of the conditional expectation of the two competing models in Stata?

Best regards,
Imran Khan.

[Joao Santos Silva]

Dear Imran Khan,

You do not need to bootstrap. See this tread for more information.

Joao