Part 2 ended with a null result. This part is about why that null was the wrong thing to be looking for.
Every estimate so far collapses 44 years into a single number. That is standard practice, and it carries an assumption nobody usually states out loud: that the relationship being estimated is the same relationship in 1983 as in 2021. Our sample contains the end of the Cold War, the founding of the WTO, China's accession, a global financial crisis and a pandemic. Assuming stability across all of that is a strong claim to make silently.
So we tested it. We re-estimated the baseline on rolling ten-year windows — 1980–1989, then 1981–1990, and so on to 2014–2023 — each one a separate regression, each plotted at the year the window ends.
The coefficient does not sit still. In windows ending before 2015 it is small, unstable and mostly insignificant. In the window ending 2018 — the first containing only post-crisis observations — overall globalization comes in at 0.186 and clears the 1% level. By 2023 it is 0.183 and still significant. Social and political globalization follow the same path. Economic globalization is the holdout: significantly negative in the earliest window, and never decisively positive after.
A Chow test at 2008 confirms it formally, rejecting parameter stability for all four indices with F-statistics between 4.9 and 9.9. So the paper's conclusion is not 'globalization doesn't matter'. It is that a single long-run coefficient is the wrong object. It averages over periods that do not resemble each other, and the average is close to zero because the pieces point in different directions.
That, incidentally, explains the literature's disagreement better than any individual paper does. Yay and Aksoy report insignificant baseline results with data ending in 2010 — which, by our rolling estimates, is before the relationship turns. They were not wrong. They were looking at a different period, and nobody had established that the period was the variable.
There is a footnote to this that we only got to after submitting, and it complicates our own story.
A Chow test answers a specific question: is there a break at this date I have chosen? It does not answer where is the break strongest? Two tests do answer that — the QLR sup-Wald test, which searches every candidate year, and the Bai-Perron sequential procedure, which also decides how many breaks there are. We ran both.
They agree with each other and disagree with us. QLR puts the break at 1995 for all four headline indices, with sup-F statistics from 10.9 to 16.1 against a 5% critical value of 8.68. Bai-Perron selects 1995 for seven of nine indices, each with a single break. The strongest break of all is trade globalization, at 13.21 — notable, because trade sits inside the economic index that never reaches significance anywhere else in the paper.
The natural reaction is that 2008 was wrong. It is worth looking at the F-statistic path before concluding that. The curve peaks at 1995, drops well below the critical value through the early 2000s, and then rises to a second peak around 2008–09 that also clears the 5% line. Both dates are real breaks. 1995 is simply the bigger one.
The reading we take from that: the instability was building from the mid-1990s acceleration of globalization — the Uruguay Round, the founding of the WTO — and the financial crisis amplified a shift that was already underway rather than starting one. Our Chow test found something genuine. It just found the second-largest break, because we told it where to look.
Which is a fitting way for the project to end. The paper's contribution is not a coefficient. It is the observation that this literature has been arguing about the value of a parameter that does not hold still — and the small, slightly uncomfortable demonstration that when you stop imposing your assumptions on the data, it will happily tell you that you imposed the wrong one.