Kvam, P. D., Romeu, R. J., Turner, B. M., Vassileva, J., & Busemeyer, J. R. (2019). Testing the factor structure underlying behavior using joint cognitive models: Impulsivity in delay discounting and Cambridge gambling tasks. psyarxiv.com/4hw73/
Kvam, P.D., Pleskac, T.J., Yu, S., & Busemeyer, J R. (2015). Interference effects of choice on confidence: Quantum characteristics of evidence accumulation. Proceedings of the National Academy of Sciences, 112(34), 10645--10650. Link, PDF
Kvam, P.D., Cesario, J., Schossau, J., Eisthen, H., & Hintze, A. (2015) Computational evolution of decision-making strategies. In D.C. Noelle, R. Dale, A.S. Warlaumont, J. Yoshimi, T. Matlock, C.D. Jennings, & P.P. Maglio (Eds.), Proceedings of the 37th Annual Meeting of the Cognitive Science Society (pp. 1225--1230). Austin, TX: Cognitive Science Society. Link, PDF
Our work pursues an understanding of cognitive and neural processes, especially those related to decision-making. As part of this effort, we develop computational models and other tools that can be applied to a variety of problems spanning topic areas in psychology and other sciences.
Approaching problems from a mathematical and statistical understanding of behavior allows us to construct more complete models of cognition, leveraging different sources of data to gain insight into shared underlying processes. As such, our research is often interdisciplinary, involving tools and collaborators from other areas like economics, biology, computer science, physics, and statistics. We focus on three main approaches to generating new insights about cognition and psychology: models of behavior, models of neural and evolutionary processes, and joint models that put together multiple sources of data.
Understanding the root causes of behavior requires us to consider not only the descriptive fit of a model or theory, but also its suitability as a generative model of the data. When I think about what makes for a satisfying theory, I want to know why and how people produce different behaviors, rather than simply finding the best statistical description of phenomena.
Computational models embody this approach to studying cognition by using a generative model to predict responses alongside process measures like response times, mouse trajectories, or neural data. The models I work with not only describe what information is considered and what responses people give, but how a person’s beliefs or preferences change over time. A good example of both considerations comes from our research into how people assign prices to different risky prospects. In this work, we showed that the distributions of prices people assign to different items possess extreme skew that depends on both the item attributes and the type of price (e.g., buying vs selling price) and tended to be unreliable when elicited multiple times. Furthermore, the mean prices people assigned to items shifted as an interaction between time and price type. These findings violate several assumptions of existing price models, which suggest that prices should be reliable, normally distributed, and static across time. To address this deficit in models of pricing, we developed a theory that explains price judgments as the output of a dynamic anchoring and adjustment process. This generative model not only predicts the properties of the choice, response time, and mouse trajectory data, but it posits an explanation for why we find skew, unreliability, and temporal shifts in prices.
Price judgments, like many of the decisions and judgments we make in the real world, require a decision maker to select between many possible options. Despite the prevalence of multi-alternative and continuous selection tasks in the real world -- from pricing to spatial navigation, confidence, or estimation tasks –- models of the decision process have focused primarily on selections between 2-3 alternatives. Much of our cognitive modeling work has striven to extend both our theoretical and empirical understanding of decisions that offer many possible response alternatives. Our previous and ongoing work investigates and models phenomena in perception as well as higher-level cognitive processes, including judgments of color, orientation, magnitude, price, confidence.
A critical consideration when implementing models of cognition is whether a proposed explanation is plausible given what we know about human physiology and ancestry. We know that the nervous system is composed of neural circuitry that has evolved over generations as well as shifted over lifetimes with new learning. Connectionist networks and evolutionary algorithms allow researchers to explore how particular models might be instantiated in the brain.
Evolutionary models can also be used to constrain or inform our psychological theories. Several important classes of decision models, from heuristics to Bayesian cognition, make implicit or explicit reference to evolution as the generator of decision strategies. Exploring the process of evolution computationally using artificial organisms allows us to directly manipulate evolutionary environments and examine what conditions can lead to the evolution of the complex cognitive processes humans enjoy. For example, my current work suggests that having a variety of fitness-relevant tasks that can yield rewards if performed well –- such as memory recall, decision making, spatial navigation, and foraging –- leads to the evolution of complex, general-purpose cognitive machinery like working memory, which does not evolve in solitary tasks
Being able to relate different types of empirical data to one another relies on effective methods for connecting observations to a common set of underlying parameters. Hierarchical Bayesian methods offer an avenue for doing so while also accounting for group and individual differences (multilevel structure). We have been adapting these methods to connect behavior across tasks and relate it to self-report measures. We do so by loading cognitive model parameters onto latent traits that are thought to underlie performance on many tasks, such as impulsivity or working memory, effectively creating a structural equation model with an embedded cognitive model (composing the "measurement" part of the model) predicting behavior on its corresponding task. We have shown that this approach lets us better estimate individual differences as well as better estimate cognitive model parameters. In turn, we can leverage the model parameters and trait estimates to better characterize clinical populations and predict important outcome measures like drug use and risky or externalizing behaviors.
The joint modeling work takes advantage of the interesting data on alcohol and drug misuse, much of which was collected by our collaborators Jasmin Vassileva and Peter Finn, who are supported by grants from the National Institutes of Health, R01AA13650 (National Institute on Alcohol Abuse and Alcoholism) and R01DA021421 (National Institute of Drug Abuse).