Decentralizing centralized matching markets: Implications from early offers in university admissions

The matching literature commonly rules out that market design itself shapes agent preferences. Underlying this premise is the assumption that agents know their own preferences at the outset and that preferences do not change throughout the matching process. Under this assumption, a centralized matching market can often outperform a decentralized one. Using a quasi-experiment in Germany's university admissions, we provide evidence against this assumption. We study a centralized clearinghouse that implements the early stages of the university-proposing Gale-Shapley deferred-acceptance mechanism in real time, resembling a decentralized market with continuous offers, rejections, and acceptances. With data on the exact timing of every decision, we show that early offers are more likely to be accepted than (potential) later offers, despite early offers not being made by more desirable universities. Furthermore, early offers are only accepted after some time rather than immediately. These results and direct survey evidence are consistent with a model of information acquisition: it is costly for students to learn about universities and accepting a university that turns out to be inferior causes regret. We discuss and rule out some alternative hypotheses. Our findings motivate a hybrid mechanism that balances centralization and decentralization. By allowing sequential learning, it improves welfare, especially in markets with substantial learning costs.