In this section we describe the basic experimental setup common to all experiments we ran in this study. Auditors were recruited via the employee networks of the Center for Microfinance (CMF), with the goal of recruiting reliable, capable individuals who would be able to conduct the audits effectively.
The audit team was led by a full-time audit manager, who had previously worked managing a financial product sales team for an international bank. This employee, along with a principal investigator, provided intensive introductory training on life insurance. Each auditor was subsequently trained in the specific scripts they were to follow when meeting with the agents. Each auditor's script was customized to match the auditor's true-life situation (number of children, place of residence, etc.). However, auditors were given uniform and consistent language to use when asking about insurance products and seeking recommendations. Auditors memorized the scripts, as they would be unable to use notes in their meetings with the agents.
Following each interview, auditors completed an exit interview form immediately, which was entered and checked for consistency. The auditors and their manager were told neither the purpose of the study nor the specific hypotheses we sought to test. Auditors were instructed not to lie during any of the sessions. The audit process was designed to mimic customer behavior as much as possible, and allow our auditors to act naturally. The audits scripts were written by a former life insurance salesperson, with the goal of representing typical transactions.
Following pilots, we ran a series of experiments to understand under what circumstances advice might improve. In each experiment, treatments were randomly assigned to auditors, and auditors to agents. Note that because the randomizations were done independently, this means that each auditor did not necessarily do an equivalent number of treatment and control audits for any given intervention of interest. Since we were identifying agents as the experiment proceeded, we randomized in daily batches. To ensure treatment fidelity, auditors were assigned to use only one particular treatment script on a given day.
Life insurance agents were identified via a number of different sources, most of which were websites with national listings of life insurance agents. Contact procedures were identical across the treatments. While some agents were visited more than once, care was taken to ensure that no auditor visited the same agent twice, and to space any repeat visits at least four weeks apart, both to minimize the burden on the agents and to reduce the chance that the agent would learn of the study. At the experiments' conclusion, auditors were offered a bonus which they could use towards purchasing a life insurance plan of their own choosing.
THE THREE EXPERIMENTS
Below we describe the audit counts from our three experiments, disaggregated by treatment combinations. The first column provides the total number of audits for each treatment combination, the second column provides the total number of auditors involved for each treatment combination, and the third column provides the number of distinct agents visited for each treatment combination. The fourth column indicates the mean of the main dependent variable, by treatment assignment, for each experiment. a) Since agents may have been visited by more than one auditor, the number of agents visited is less than the total number of audits.