Experimenting with sequential allocation procedures

In experiments that consider the use of subjects, a crucial part is deciding which treatment to allocate to which subject – in other words, constructing the treatment allocation procedure. In a classical experiment, this treatment allocation procedure often simply constitutes randomly assigning subjects to a number of different treatments. Subsequently, when all outcomes have been observed, the resulting data is used to conduct an analysis that is specified a priori. Practically, however, the subjects often arrive at an experiment one-by-one. This allows the data generating process to be viewed differently: instead of considering the subjects in a batch, intermediate data from previous interactions with other subjects can be used to influence the decisions of the treatment allocation in future interactions. A heavily researched formalization that helps developing strategies for sequentially allocating subjects is the multi-armed bandit problem. In this thesis, several methods are developed to expedite the use of sequential allocation procedures by (social) scientists in field experiments. This is done by building upon the extensive literature of the multi-armed bandit problem. The thesis also introduces and shows many (empirical) examples of the usefulness and applicability of sequential allocation procedures in practice.