Experimental Design
We run a selection contest with two agents who compete by performing a two-period estimation task. Agents will be paid based on their performance, which is measured by the accuracy in the estimation task (i.e. the difference between the estimate and the correct solution). Furthermore, a principal selects one of the two agents as the winner of the contest. This agent will thereby get a selection bonus as well as an additional payoff by playing a third part of the estimation task that is also paid based on accuracy. The principal will be paid based on the selected agent’s accuracy in the third period, which incentivizes the principal to select the best-performing agent. With this basic framework, we follow the seminal work by Falk and Zimmermann (2017) on consistency as a signal of skills in selection contests with asymmetric information. We build upon their work to find out who benefits from selection-by-consistency by analyzing the selection of specific types of individuals and the role of consistency across different contexts.
To this end, we vary the stability of the decision environment by manipulating the level of information agents receive. First, in the ChangingDifficulty treatment, we provide agents with an easier task in the second period, to reflect the idea that leaders may still stick to their initial decisions, even when they realize there were wrong, because they may benefit from that. Second, in the ChangingSolution treatment, we change the solution of the task in the second period, to capture the idea of a changing or volatile environment in which a leader may be required to be flexible and to adjust to a new situation.
We also vary the selection mechanism, introducing selection based on luck and selection based on merit for two main reasons. The first is to distinguish between those who behave consistently for strategic reason from those who behave consistently for intrinsic reasons or those who are consistent as a byproduct of being highly able. Here, we eliminate the strategic reasons by replacing principals with a computer programme that selects agents randomly or based on their accuracy of estimates. The second reason is to inform our subsequent analyses on the perceptions of inequality where we intend to find out how selection-by-consistency is perceived compared to selection by merit and luck, as prominent considerations for fairness views regarding the sources of inequality in the literature.
Accordingly, we will have five conditions to which participants will be randomly assigned (by a computer):
a) Principal-Agent treatments
• Main treatment (perfect stability in terms of the solution and the level of difficulty)
• ChangingDifficulty treatment (second estimate becomes easier)
• ChangingSolution treatment (correct answer may change)
b) Agents treatments
• RandomSelection treatment (selection by a uniformly-mixing computer)
• PerformanceSelection treatment (selection by a computer choosing the more accurate agent)
In case of selection effects due to differences in task ability, such as gender-specific ability to solve the task, we could consider further control conditions without principal.
We collect multiple observations per individual across three periods. The longitudinal nature of the data allows us to learn more about the potential dynamics in our experiment, such as learning effects. For example, it is possible that some agents learn how to exploit the option to behave strategically over time, while the first estimates from the first period may only inform about the strategic behavior of agents with a very good understanding of the strategic context. Additionally, in each period, we collect multiple observations from each principal who makes selection choices for five pairs of agents, while one selection choice is randomly determined as relevant to the payments.
To learn more about the potential implications of selection-by-consistency and to identify the determinants of consistent behavior, we collect multi-faceted information on individuals and their characteristics prior to the start of the consistency experiment, including dishonesty, risk attitude, competitiveness, overconfidence, and pro-sociality. In each case, we use incentivized elicitation methods to capture these traits. This also applies to a measure of strategic ability. To complement the information at the individual level, we ask subjects further questions about their socio-demographic background, inclinations, and personality traits, such as Big Five.
We expect strategic consistency to play a role in the Principal–Agent Treatments. Since strategic motives are plausibly related to several traits, we expect that those could be predictive of consistent behavior in our experiment. First, as the success of strategic consistency is uncertain and depends on a variety of different factors, including the behavior of other individuals in the experiment, risk-tolerant individuals are more likely to employ strategic consistency. Second, since winning a selection contest over another person is more attractive to agents with preferences for competition, more competitive individuals are more likely to employ strategic consistency. Third, as the (expected) success of strategic consistency depends on the (expected) lack of strategic ability of competitors, overconfident individuals are more likely to employ strategic consistency. Fourth, strategic consistency could imply a need to deviate from the best possible estimate in the absence of any strategic incentive; accordingly, dishonest individuals are more likely to employ strategic consistency. Fifth, because strategic consistency is inefficient from a social perspective when individual payoffs are increased at the expense of overall payoffs, more pro-social individuals are less likely to employ strategic consistency.
Since all these characteristics and preferences are potentially gender-specific, we expect there could be a gender gap in consistent behavior. The same could occur if intrinsic consistency is found to be gender-specific. As we suspect that consistency varies with gender for a variety of reasons, we will balance treatment assignment using a stratified randomization by a computer.
In addition to the determinants of strategic consistency, we investigate the consequences for selection outcomes and hence inequalities, based on the notion that the more successful the consistency strategy is, the more likely inequalities along the lines of our above predictions could occur. If for examples male agents benefit strongly from strategic consistency, we expect that they also have a higher likelihood of being selected, all else being constant.