Conducting an Experiment at Multiple Sites with Small Subject Pools: How is Raven Score Effective as a Covariate?

It is necessary to extract some patterns of human behavior from events observed in subject experiments in order to choose particular parameters for computer experiments on trades of information. Many experimental sites are, however, faced with practical problems due to their small subject pools. In conducting a subject experiment at multiple sites for avoiding those problems, homogeneity of subjects’ behavior is important for integrating the data collected at separate sites. Ideally, preliminary test results on the homogeneity are needed in planning the experiment. This paper clarifies a condition under which subjects’ behavior in a bandit experiment with the context of weighted voting can be homogenized across different subject pools in five universities in Japan by covariate adjustment with their cognitive ability scores on Raven’s APM test. Among experimental sites located in different regions, we could not obtain homogeneity of subjects’ behavior by covariate adjustment with their cognitive ability scores in addition to their attribute information, although it was not difficult to obtain the homogeneity across the experimental sites located in the same region.