# causal time series analysis economics

I want to analyse the relationship between the level of globalization,and the level of income inequality between two specific countries. however, I'm quite lost as to which method is good to use (in spss). I'm certain I have to perform a time series analysis, but can someone advise which one is good to use? Kind regards

If you run a vector autoregression you could also follow up by testing for Granger causality. I don't know whether this is implemented directly in SPSS, but once you have an estimated model it is easy to calculate the statistic as it is basically just an F-test. In R you can do it directly via for example the package vars.

A vector autoregression is where you have a vector of outcomes for each country, $$y_t = (y_{China,t}, y_{Ghana,t}, y_{Britain,t},...) = a + B'X_t+\varepsilon_t$$ where $$X_t$$ is data, perhaps containing lagged values of $$y_t$$, and $$\varepsilon_t$$ is the error term.

The best known and most successful estimation framework for this setting is probably Arellano-Bond:

https://en.wikipedia.org/wiki/Arellano%E2%80%93Bond_estimator

which is available as a canned command in Stata and R. In other work, people put restrictions on the matrix $$B$$ to get identification and impose a causal interpretation on the regression: for example, some events precede others, and this means some of the terms in $$B$$ should'' be zero; this is popular in monetary economics, to see if stimulus has a causal impact on the economy. In general, causal identification in panel models is every difficult, because at some point everything is endogenous and it is hard to defend exclusion restrictions and assumptions of exogeneity.

Best of luck!

• Thank you! Do you know if the Arellano-Bond test can also (doable) be performed in spss? Apr 3 '20 at 15:12
• I have not used SPSS since elementary school, so I am not sure. A lot of Arellano-Bond is just getting the standard errors right, so SPSS might have canned panel data estimators that you can trick into doing the AB estimator, but maybe with wider confidence intervals than you want.
– user26098
Apr 3 '20 at 15:15
• Thanks I will look into that! I was also wondering if a sample of only 2 countries is sufficient for a panel analysis? Kind regards (and really thanks a lot for helping me out) Apr 3 '20 at 15:27
• There are two concerns: is the number of observational units $I$ large, and is the length of the panel $T$ large? If you want to estimate a model, you can have small $I$ but long $T$, and long $T$ but small $I$, but not both small.
– user26098
Apr 3 '20 at 15:32