Cumulative Knowledge in the Social Sciences: The Case of Improving Voters’ Information

Abstract

Cumulative knowledge requires (at least) two conditions to be met: unbiasedness and comparability. Research designs should be unbiased so that researchers obtain correct estimates of an underlying quantity. Empirical specifications, the actual regression run, should permit comparability so that researchers measure the same quantity across distinct studies. The first condition is covered by the causal revolution, the second is the object of this paper. Using the example of interventions providing additional information to voters, we show that unbiasedness does not imply comparability. Any two studies that employ the commonly used specification to analyze the electoral consequences of informational campaigns estimates different estimands. This holds true even after removing all external validity issue, all statistical noise, and all sources of bias. We highlight conditions to restore comparability and describe specifications that satisfy them.