Multiple missing values in Panel Data

[Muhammad Yousaf Malik]

Hi,
I am trying to run the xtreg command for the panel data. I have attached the summary table, where the dependent variable is missing 616 values. The dependent variable is Chinese outward foreign direct investment (OFDI), showing multiple missing values over the years in multiple countries. In terms of data, we can say that the dependent variable has missing values, but in reality, it means that China only invested in particular countries in specific years. I have run the xtreg using original regression and also using the imputed values. I adopted the method of imputation by following the article https://phantran.net/multiple-imputa...lues-in-stata/.

Though the imputed results remain consistent with the original results, the imputed results show most variables to be significant. I would be grateful if you could guide me to use which results in my paper (imputed or original). Also, can we run the Granger Causality using the imputed dataset since the original dataset is unbalanced?

[Carlo Lazzaro]

Muhammad:
before worrying about missing data, I would check via -xttest0- whether your results shows evidence of a panel-wise effect.

[Muhammad Yousaf Malik]

Hi,
Please find the xttest0 attached.

[Carlo Lazzaro]

Muhammad:
as the -xttest0- outcome rejects the null, your data show evidence of a panel-wise effect.
That said, I would consider the option to stick with your original unbalanced panel dataset, as missingness is explained by the fact that

...China only invested in particular countries in specific years

.

[Muhammad Yousaf Malik]

Hi Carlo,
Thanks for your prompt replies. I would be grateful if you could also provide a solution to carry out the Granger Causality test on the original data set. Considering some replies in another post, converting the unbalanced panel data into balanced panel data significantly reduces the number of observations. I am hoping for your reply.

[Carlo Lazzaro]

Muhamnad:
1) see -help vargranger-;
2) I'm afraid but I do not follow your statement about the reduction in the number of observations due to the coversion of an unbalanced panel into its balanced counterpart.

[Muhammad Yousaf Malik]

Hi Carlo,
Thanks for the reply. From the original dataset (OECD + non-OECD), I take a sample of OECD countries. After running the -xtreg-, I also carry the -xttest0-, where null is not rejected, thus suggesting to use of a pooled regression. I carry out the pooled regression for these OECD countries' panel data by adopting the method shown in this article https://www.projectguru.in/pooled-pa...ression-stata/.

In some of your old posts regarding the pool regression for the panel data, you commented that 'pooled OLS seldom outperforms -xtreg- when you deal with a panel dataset.' (https://statalist.org/forums/forum/g...nel-regression). I have attached the result for both the -xtreg- and pooled regression, where both show similar findings. Which result should I choose to present in the paper? Also, for your statement ''pooled OLS seldom outperforms -xtreg- when you deal with a panel dataset.' is there any reference paper that I can add in the methodological part to highlight why we still used the -xtreg-, even the LM test showed to use pooled regression. I am hoping for your reply.

[Carlo Lazzaro]

Muhammad:
1) I meant that starting off with a pooled OLS when dealing with a panel dataset should not be researcher's first choice as, if there's evidence that a panel-wise efffect exists, -xt- commands do it better;
2) conversely, in your case you are forced to switch to a pooled OLS as your dataset does not support the evidence of a panel-wise effect.
As an aside, it seems that you're using a -wide- format in your -regress-. It's much better to -reshape- and adopt the -long-layout.