Abstract. Behavioral economists have proposed that incentive contracts result in higher productivity when bonuses are “loss framed”—prepaid then clawed back if targets are unmet. We test this claim in a large-scale field experiment. Holding financial incentives fixed, we randomized the pre- or postpayment of sales bonuses at 294 car dealerships. Prepayment was estimated to reduce sales by 5%, generating a revenue loss of $45 million over 4 months. We document, both empirically and theoretically, that negative effects of loss framing can arise due to an increase in incentives for “gaming” behaviors. Based on these claims, we reassess the common wisdom regarding the desirability of loss framing.
Abstract. Five studies (N=3,526) reveal that willingness-to-pay (WTP) reflects perceived market price rather than personal valuation. Moreover, people tend to indicate WTP for the goods as if they were in the market for those goods, even when those goods are personally irrelevant (e.g., steak for vegetarians). We demonstrate that these market related characteristics of WTP explain some of the well documented anomalies such as preference reversals. Furthermore, merely highlighting that a WTP of $0 indicates not wanting to buy attenuates these anomalies. This research offers a novel interpretation for research using WTP anda simple intervention that shifts WTP closer to personal valuation.
Abstract. Are two heads better than one, or do too many cooks spoil the broth? Although researchers have generated a large number of nuanced answers to this question, they have had little success specifying the range of conditions for which a given answer applies. I argue that one of the keys to solving the puzzle is to better understand the underlying nature of the tasks being performed. Because no clear language exists to describe tasks in a way that allows for straightforward comparisons across studies, the role of task characteristics remains poorly understood. I will present an idea for developing a comprehensive, empirically grounded theory of group tasks using high-throughput, adaptive experiment design.
Abstract. An influential research tradition argues that the US media’s effects on voter preferences either are minimized by selective exposure or operate through subtle mechanisms. Using behavioral data from a media company on household-level television viewership, we recruited heavy consumers of Fox News to a field experiment in late 2020. We first document that heavy Fox News viewers are homogeneously conservative Republicans. During September 2020, we incentivized a randomized treatment group to instead watch CNN. At the end of the incentivized period, we find that changing the slant of participants’ media consumption had profound effects on: 1) factual beliefs. The treatment group was less likely to harbor misperceptions about COVID-19, vote-by-mail, and protests and policing, and more informed about Trump and Biden’s positions on them. 2) personal views and policy preferences. 3) issue importance. The treatment group saw COVID-19 as more important and efforts to quell protests as less important. 4) overall political views. The treatment group was much less approving of Donald Trump and Republicans generally. 5) attitudes towards the media. The treatment group evaluated Fox News more negatively and was more likely to state that it conceals negative information about Trump. However, many of these results decayed after the experiment ended and participants returned to their previous media diet. Our results are consistent with partisan coverage filtering impacting viewers’ opinions: partisan media filters which information it presents to an audience that cannot readily infer the contents of filtered information, resulting in changes in their factual beliefs, issue preferences, and evaluations of elected officials.
Abstract. We propose that decisions under risk often generate a consequential outcome that has been overlooked in the literature – error. Specifically, when prospects are tied to future states of the world (e.g., win $5 if horse #6 wins the race), risky decisions necessitate selecting a future state of the world. As a result, choosing among prospects generates error – the difference between the chosen and realized future state of the world. In 7 studies (N = 5,187), we find that people systematically value error and that they take more risk when their decisions generate error. These results suggest that error is a missing component in current theories of decision making under risk and that failing to account for error can result in biases in these theories. We believe people’s sensitivity to error has especially important implications for our understanding of real-world risk preferences because prospects in real-world decisions under risk are almost always tied to future states of the world.
Abstract. Across twelve studies using archival data, surveys, and experiments, we identify a novel factor that predicts gender differences in time stress and burnout. Across academic and professional settings, women are less likely to ask for more time when working under adjustable deadlines (Studies 1-4a). Women’s discomfort in asking for more time on adjustable deadlines uniquely predicts time stress and burnout, controlling for marital status, industry, tenure, and delegation preferences (Study 1). Women are less likely to ask for more time to complete their tasks because they hold stronger beliefs that they will be penalized for these requests and worry more about burdening others (Studies 1-2d). We find no evidence that women are judged more harshly than men (Studies 3-5). We also document a simple organizational intervention: formal processes for requesting deadline extensions reduce gender differences in asking for more time (Studies 6a-7).
Abstract. Hypotheses involving interactions, where one variable modifies the association between another two, are common in experimental and observational social science. Interactions are typically tested relying on models that assume effects are linear, e.g., with a regression like y=ax+bz+cxz. In the real world few effects are linear, invalidating inferences about interactions. For instance, in realistic situations, the false-positive rate for an interaction can be 100%. The sign of the interaction can be wrong with certainty. And results from probing the interaction estimates, where one computes the effect of interest for different levels of the moderator, can be entirely disconnected from reality. This paper proposes a revised toolbox for studying interactions which is curvilinear-robust, giving correct answers 'even' when effects aren't linear. It's applicable to most study designs, while introducing minor modifications to current analytical and reporting practices. The presentation combines statistical intuition, examples of published results, and simulations.
Abstract. In several countries, individuals have multiple opportunities to join the organ donor registry. In Ontario, Canada, for example, one can register when visiting a branch of ServiceOntario, the agency of the Ministry of Government and Consumer Services, which provides a single point of contact for most government services in the province such as driver’s license or health card issuance and renewal. However, registration rates are still below the desired level in Ontario. In collaboration with ServiceOntario, the organ and tissue donation agency of Ontario (TGLN), and the Behavioral Insight Unit at the provincial Treasury Ministry, we conducted a natural field experiment to study the effects of providing absolute and relative performance feedback to customer service representatives (CSRs). We ran the experiment in three waves over two years (2017-2019) with all government-owned 80 branches in the Province to either a baseline condition consisting of an email sent to CSRs to encourage them to ask clients if they would like to join the organ donor registry, or to one of two performance feedback interventions. The first intervention treatment added to the baseline e-mail information about the number of new registry members (per hundred transactions) that the CSR had signed up in the previous six months. The second treatment included, in addition to the CSR’s individual signup performance, the average signup rate in the region the branch belonged to and the performance at the 80th percentile in the region. Overall, providing performance feedback (no matter if only individual signup performance or also average and 80th percentile reference performances) resulted in a significant increase in daily signups compared to encouragement alone. The effects were particularly strong in the few weeks immediately following receipt of a treatment email, but a substantial effect persisted for several months.
Behavioral science and experimentation have the potential to inform policy, but discrepancies in intervention effectiveness across evaluations could make policymakers question the scalability of empirical findings. In this talk, I will offer insights that contribute to explain some of these discrepancies by leveraging data from two pre-registered randomized controlled trials (RCTs) testing the impact of nudges on COVID-19 vaccine uptake. In February-April, 2021, we delivered interventions via text to people one day (first RCT; N=187,134) and eight days (second RCT; N=149,720) after they received notification of vaccine eligibility from their healthcare provider. I will share key lessons gleaned from this research focusing on three questions. First, how do our interventions affect vaccinations across different outcome measures? Second, how does the efficacy of our interventions vary with people’ baseline motivation to act? Third, do hypothetical responses match behavior in the field? Altogether, this work advances our understanding of nudges’ impact and limits, helps reconcile divergent effects documented across evaluations, and provides implications for evaluating the scalability of RCT results.
