Recipient Organization
PURDUE UNIVERSITY
(N/A)
WEST LAFAYETTE,IN 47907
Performing Department
(N/A)
Non Technical Summary
This proposal responds to the Economics, Markets, and Trade program priority of "Development of innovative empirical methods for addressing economic analysis using big data, machine learning, and natural language processing techniques." We combine principal-agent incentive design models, machine learning, and experimental economics to generate systematic knowledge about how behavioral nudge outcomes are affected by heterogeneity and contextual factors. Improving the methodology of behavioral nudging can support improved policy in the areas of health and nutrition, and sustainable agricultural, which are broad priorities for AFRI funding.Recent academic papers have suggested that publication bias has overstated the impact of nudges and that nudge outcomes are highly sensitive to heterogeneity and context. We know of no studies that attempt to capture this sensitivity in a systematic way. The objective of this grant proposal is to fill this knowledge gap.We propose that nudge design can be framed as a modified principal-agent problem where the principal chooses discrete nudge strategies rather than monetary incentives. By reframing nudges as an incentive problem, we have a systematic framework for evaluating the underlying mechanisms that drive uneven responses to nudges. However, given the huge variety of nudge architectures, we must also use the text processing capabilities of machine learning in combination with existing data on nudge interventions to reduce the dimensionality of nudge types and to identify important contextual factors. Experimental economics will be used to test whether the unevenness of responses to nudges can be attributed to failures of incentive compatibility across domains.
Animal Health Component
50%
Research Effort Categories
Basic
50%
Applied
50%
Developmental
(N/A)
Goals / Objectives
The main goal of this proposal is to combine reframed principal-agent models, machine learning, and experimental economics to generate systematic knowledge about how behavioral nudge outcomes are affected by heterogeneity and contextual factors. Moreover, reinforcement machine learning can be used to classify the optimal set of nudges from existing behavioral datasets, and these optimal nudges can be tested in the laboratory guided by the reframed principal-agent model. The predictions generated by the model, as well as the experimental results, can provide us with insights into the underlying mechanisms that generate uneven response rates to nudges. Our goal is to ultimately improve the efficacy and replicability of nudge interventions.
Project Methods
We will begin by developing a simple stylized principal-agent model, that abstracts away confounding details to uncover the key underlying economic forces that drive behavioral responses to nudges. We believe that a standard principal-agent model is an ideal model for this purpose because it offers a parsimonious framework for conceptualizing incentive design problems. Traditionally, the model has been used for designing pay-for-performance incentives. However, it can be easily modified to accommodate other forms of incentives, such as nudges. Rather than making contingent payments the choice variables for the principals, we only need to make the incentive choice set a set of N nudge strategies.Another advantage of the P-A model is that it includes a very precise criteria for evaluating incentives; namely, incentive compatibility. In essence, the agent's marginal utility from being exposed to the incentive must be greater than its effort cost from undertaking the behavioral change. Hence, any nudge that does not satisfy the incentive compatibility constraint of the agent is predicted to fail to induce a response. Because the nudge literature tends to focus more on psychological factors rather than economic factors, nudges are often implemented without consideration of incentive compatibility. Thus, as a first pass to understanding the unevenness of nudge reponses, we must rule out failures of incentive compatibility before moving toward more nuanced behavioral explanations that focus on deviations from rationality. The P-A model provides a concise and transparent framework for doing this.In order to introduce heterogeneity, we assume that there will be four types of agents as determined by exogenous variations in their payoff and effort cost functions. There will be a "high" type and a "low" type with respect to the utility function. There will also be "high" and "low" types with respect to effort cost/disutility. Thus, we get four types: "high-high," "high-low," "low-high," and "low-low" types. As a practical example, suppose that a nudge designer is designing nudges to support healthier diets. As a first pass, it might be convenient to classify agents into those that care about health (i.e., have "high" utility functions with respect to health) and those who do not care about health ("low" utility types). And then consumers also vary in their costs of search. Those who do not like to search have a "high" cost of effort. Those who will search have a "low" cost of effort. So high-high types might be those who care about health but do not like to search through a store in order to find healthy foods. Thus a nudge in the form of placing healthy foods at eye level or providing easy-to-interpret health information will likely convert these high-high types into responders. At the other end of the spectrum are low-low types. These folks do not care about healthy eating and also have low search costs. So a nudge that places healthy food in easy-to-find locations and junk foods in difficult-to-find places will not deter these types from finding junk food and bypassing the healthy foods. The in-between cases are the high-low for which a nudge is not necessary as they would find healthy foods even without the nudge, and low-high types for which a nudge may work because they might grab the first item they see even if they are not explicitly worried about healthy eating.Partitioning agents into the above four types also allows us to examine the impact of context on nudge effectiveness. For example, there might be venues, such as Whole Foods-style supermarkets, where the majority of consumers care about health, so that would be a context in which there are mostly high-high and high-low types. In this case, nudges that impact the cost of search might be most effective. In conventional supermarkets, we might get the whole range of types from high-high to low-low, so nudges will have highly mixed response rates.We will use the above model to generate predictions under three scenarios that account for heterogeneity and context:The impact of nudge interventions when the principal is implementing nudges for the full range of agent types. This scenario is analogous to a nudge policy that seeks to nudge a general population toward a behavioral goal.The impact of nudge interventions when the principal is implementing nudges for a targeted range of agent types that are favorable to nudge interventions (e.g. high-high and high-low types only). This would be analogous to a nudge policy that targets a specific set of consumers who care about health but may have different motivations to seek out information about health.The impact of nudges when the principal is implementing nudges for a challenging subset of consumers (e.g., low-high and low-low types only). This would be analogous to a situation where a nudge practitioner is targeting a low-income population that is more concerned with meeting daily caloric needs than health considerations. This population might also be income- and time-constrained which make health lower on their priority list.The three scenarios above should yield varying predictions about the ease/difficulty of satisfying the agents' incentive compatibility constraints, and thus, we anticipate differences in predicted response rates across different scenarios. This information would yield insight into how nudge outcomes vary across context and heterogeneity.Next the model will be used as a blueprint for designing economic experiments. The primary purpose of these first-round experiments is to test the basic reframed principal-agent model in an environment stripped of complicating factors. This allows for a clean test of the predictions and can provide clear insights into the underlying mechanism driving response rates, which we hypothesize to be the ability to satisfy agents' incentive compatibility conditions. Given that we will test theory and contextual factors, we need a high degree of control over the experimental design to mimic the theoretical model. Thus, we will rely more on laboratory experiments than field experiments. It is well known among experimental economists that laboratory experiments afford a greater degree of control over design elements.We will also apply machine learning text processing capabilities to analyze the various nudges that are observed in existing nudge datasets. Given the large range of nudges, unsupervised learning combined with the rich set of feature variables in these datasets can be used to subgroup similar types of nudges in order to reduce the dimensionality problem of having to deal with a huge range of nudges. The different broad classes of nudges can then replace the abstract nudges that we specified initially. Unsupervised learning can also be used to cluster and segment consumers. In the abstract model, the four types of consumers (agents) are rather artificial and distributed across the population with equal probability. But we can be more specific about identifying types and distributions using unsupervised learning and actual nudge datasets. We can also extract information on how often nudges are targeted toward specific populations or whether nudges in practice are mostly untargeted toward general populations. Reinforcement learning can help us solve for the optimal set of nudge strategies. In practice, nudge architecture appears to be implemented in a non-systematic manner or based on some fixed behavioral principle, such as creating default choices. However, in a principal-agent approach, one typically solves for the optimal incentive strategies. Reinforcement learning can be used for combinatorial optimization in order to find the optimal nudge strategies. We can also relax the restriction that monetary incentives must be held constant and use reinforcement learning to find the optimal combination of nudge and monetary incentives.