There is a claim you hear constantly in economic debate, and it sounds obvious the first time: once capital, firms and workers can move freely across borders, governments have to compete for them. Competing means lower taxes. Lower taxes mean less revenue. Less revenue means the welfare state has to shrink. Globalization, on this account, quietly dismantles social protection whether voters want it or not.
There is an equally common claim that says the opposite. Open economies are more exposed to shocks from abroad — a collapse in export demand, a financial crisis transmitted through integrated capital markets, an industry wiped out by import competition. Citizens who face that exposure demand insurance against it, and the state that provides it is the welfare state. On this account globalization makes social protection more necessary, not less.
These are the efficiency hypothesis and the compensation hypothesis, and the uncomfortable thing about them is that both are reasonable. Rodrik's classic 1998 result is that more open economies have bigger governments, which is compensation. The tax-competition literature — Devereux, Lockwood and Redoano on corporate rates, Egger and co-authors on who ends up carrying the tax burden — is squarely efficiency. Neither camp is doing bad economics. They are looking at the same world and finding different things in it.
So what does the evidence say? Mostly: it depends on who is asking. Brady, Beckfield and Seeleib-Kaiser looked at 17 affluent democracies and found effects that were real but small, inconsistent across indicators, and smaller than domestic political factors. Dreher found no systematic effect on taxation and spending. Meinhard and Potrafke found a clear positive effect, driven especially by social globalization. Yay and Aksoy found nothing overall, but something once you split countries by welfare regime.
That last paper is the one we took as our starting point. Yay and Aksoy (2018) run the question across welfare regimes — social democratic, conservative, liberal, Mediterranean, post-communist — using the KOF globalization indices, and report that the regime you belong to changes the answer. It is a careful paper. It also stops in 2010.
Ours asks a simple follow-up: does that hold when you extend the sample to 2023?
The thirteen years we added are not quiet ones. They contain the sovereign debt crisis, a decade of austerity across southern Europe, the sharpest peacetime expansion of social spending in living memory during COVID, and a broad political turn against trade openness. If the relationship between globalization and welfare spending is a stable structural fact, none of that should matter much. If it is contingent on the period you happen to look at, extending the sample should show it.
The setup is deliberately close to the original so the comparison means something. Thirty-two OECD countries, annual data from 1980 to 2023, dependent variable social security transfers as a share of GDP from the Comparative Political Data Set. Globalization is measured by the KOF indices, which run 0–100 and come in an overall version plus economic, social and political sub-indices — so we can ask not just whether globalization matters but which kind.
Every specification is a two-way fixed effects panel: country effects absorb the fact that Denmark is permanently different from Japan in ways we cannot observe, year effects absorb whatever hit every country at once. Controls are the ones Yay and Aksoy use, for the same reasons — GDP per capita for Wagner's Law, inflation, the budget deficit and government debt for fiscal room, and population and the dependency ratio for demographic pressure. Everything on the right-hand side enters lagged one year.
One thing worth saying before any results: none of this is causal, and we do not pretend otherwise. Globalization is not assigned to countries at random, there is no non-globalized control group, and the least globalized economies mostly lack the statistical capacity to serve as one anyway. What we have is within-country variation over four decades, and what we can report is association. That is a real limitation, not a ritual disclaimer — it shapes how much weight any of the numbers can carry.
In Part 2 we run the thing and get an answer that is, at first glance, disappointing: almost nothing is significant. The interesting part is what happens when you refuse to accept the one result that is.