Since 1932, when Justice Louis Brandeis remarked that in a federal system states can serve as "laboratories" of democracy, political decentralization has been thought to stimulate policy experimentation. We reexamine the political economy behind this belief, using a simple model of voting in centralized and decentralized democracies. We find that the electoral logic suggests the opposite conclusion: centralization usually leads to "too much" policy experimentation, compared to the social optimum, while decentralization leads to "too little." Three effects of centralization — an "informational externality," a "risk-seeking" effect, and a "risk-conserving" effect—account for the different outcomes.