Here is the baseline, run separately for each of the four KOF indices on 996 country-year observations. Overall globalization: 0.088, t of 1.39. Economic globalization: −0.020, t of −0.48. Political globalization: 0.046, t of 1.40. Three coefficients, three failures to reject. On the headline question — does globalization move social security transfers — the answer is no.
The fourth is different. Social globalization comes in at 0.129 with a t of 2.63, significant at the 1% level. That is the one result in the baseline that looks like a finding, and it lines up with Meinhard and Potrafke, who also singled out social globalization. The tempting move is to write it up as the paper's contribution.
We spent most of our robustness work trying to break it instead.
Start with the standard errors, because they turn out to matter more than we expected. We cluster by country, which handles heteroskedasticity and serial correlation within a country but not correlation across countries. A Pesaran CD test rejects cross-sectional independence decisively — unsurprising for 32 open economies sharing business cycles, and doubly so for a sample containing both the financial crisis and a pandemic. Driscoll-Kraay standard errors are built for exactly that.
Switch to them and the picture changes: overall globalization becomes significant (p = 0.007), political globalization becomes significant (p = 0.011), and social globalization stays significant. So depending on which defensible error structure you choose, you report either one significant index or three. We lead with the country-clustered results because they are the more conservative of the two, and we say plainly that the inference is sensitive to the choice. A paper that only showed you the Driscoll-Kraay column would be telling you something true and misleading at once.
Then the test that actually did the damage. Everything so far assumes globalization moves first and welfare spending responds. Why should it? A generous welfare state is itself a feature of a country's political economy, and might plausibly shape how open that country becomes. So we ran the specification backwards: KOF indices on the left, lagged social security transfers on the right, same controls, same fixed effects.
For overall, economic and political globalization, nothing. Lagged transfers do not predict them. But for social globalization the coefficient is 0.225 with a t of 2.15 — welfare spending in one year predicts social globalization in the next. That is the wrong direction for the story we would have liked to tell.
Rather than stop at 'reverse causality may be present', we decomposed social globalization into its six sub-components: interpersonal, informational and cultural, each split into de facto and de jure. Only one is significant. De jure interpersonal globalization comes in at 0.355. Its de facto counterpart is 0.044 and nowhere near significance.
That distinction is the whole point. De facto interpersonal globalization measures realised flows — migrants who actually arrived, tourists who actually came. De jure measures institutional openness: visa freedom, telephone and mobile subscriptions, the density of airports with international connections. So the channel running from welfare spending to social globalization is not people moving. It is that more generous welfare states tend to be more institutionally open countries in the first place.
This brushes against the welfare magnet hypothesis — Agersnap, Jensen and Kleven estimate a migration elasticity with respect to benefits of about 1.3 using a Danish reform — but it does not support it. The welfare magnet mechanism runs through realised migration, which is precisely the component our data says is insignificant. We flag the connection and decline to claim it.
So the honest summary of the baseline is: no robust relationship, one apparent exception, and the exception does not survive being pushed on. It is a thinner result than we would have liked. It is also, we would argue, the correct one — and the specification curve on the project page makes the point sharper still: across all 64 possible combinations of our six controls, social globalization is positive in every single one, and we still discount it, because a robust sign is not the same as a credible direction of causation.
Part 3 is where the paper stops being a null result. Not because anything becomes significant, but because we stopped asking whether the relationship exists and started asking when.