Experimental researchers in political science frequently face the following inference problem: which of several treatment arms produces the greatest return, where returns may be expressed in terms of campaign donations, votes for a candidate, or some other support for a political cause? Multi-arm trials are typically conducted using a static design in which fixed proportions of study participants are allocated to each arm. However, a growing statistical literature suggests that adaptive experimental designs may be far more efficient in finding the most effective treatment arm. An important class of adaptive designs uses randomized probability matching to dynamically allocate subjects to treatment arms. We review the operating characteristics of randomized probability matching and suggest that it has many potential applications in political science. We discuss the design and analysis of original experiments using this approach and compare their efficiency to traditional static designs.
The colloquia are organised by the Centre for Experimental Social Sciences.