ijyliu
commented
3 years ago Main results here: https://github.com/ijyliu/ECMA-31330-Project/blob/main/Release/Defense_Spending.pdf
There are also some LASSO results here: https://github.com/ijyliu/ECMA-31330-Project/blob/main/Output/Regressions/LASSO_Country_Coeff_Values.csv
It kind of seems like they might all be noise, no clear patterns based on a clear inspection. That's probably because each individual LASSO only includes 25 observations/time periods. I made that restriction to make sure we include all of the other country variables for each country. But we can relax it to bump up the sample size.
Idea
We could look at spillovers in defense spending between countries using the LASSO (like the R&D spillovers model) and potentially also PCA/Factor methods.
A hypothesis is that spending in neighboring countries increases local spending. Yet another hypothesis is that ideological differences/alignment are somehow related.
Literature
In the literature, the main approach appears to be spatial autoregression: https://www.sciencedirect.com/science/article/pii/S0176268017301581 . I haven't seen anything using machine learning methods, even though I flipped through 10 Google Scholar pages. Though, here's a triple LASSO involving military spending but the question is different (it's looking at the fiscal multiplier): https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3819192
Method
Given our two hypotheses, I think we could provide an interesting generalization by not imposing/assuming that the model is fully spatial... I mean, the US and China and the UK and Russia aren't exactly neighbors, here there's an ideological driver. But we will still be using sparsity/trying to make a parsimonious model.
In addition to the general LASSO, we could also do a factor/PCA regression (maybe using stuff in VDem) and see if we can construct a geographic, and ideological factor. Maybe we can even see if these factors change over time- from the Cold War to 9/11 to US-China rivalry...
As yet another potential direction, we could go into textual data analysis: UN resolutions or NATO statements or something? Text data might allow for another look at an ideological dimension/other dimensions not easily ascertainable from the numbers. Here is UN voting data, which I think could be ideological: https://dataverse.harvard.edu/file.xhtml?fileId=4624867&version=28.0
Data
Here's all the defense spending data: https://www.sipri.org/databases/milex : I already downloaded it and put it on the Box. VDem has so much usable data that we could probably even get at heterogeneity in spillover effects- under what circumstances does a neighbor's civil war, etc. lead to more spending? Are effects stronger/weaker for ethnic conflicts, for example?
Problems/Robustness
I wonder if sparsity/dimension reduction is wise in this case. It seems like NATO or the US or USSR could have influence on the spending of a lot of countries. Then again, based off the patent example discussed in class, that's not a problem; we don't have issues if one country affects a lot of others, only if all countries are affected by a lot of countries.
For the LASSO: we don't really have a dependent variable? ie, we would be examining the relationship between a country's defense spending and that of all the other countries. Versus the R&D case, where there was something else effected by all the other firms. Maybe this means more focus should be placed on PCA/factors.