Bayesian Analysis Symposium at Psychonomics

Symposium: Practical Benefits of Bayesian Data Analysis

Friday Nov. 19, 2010, 9:45am-11:55am.
Annual Meeting of the Psychonomic Society, St. Louis, Missouri

Please note that the times listed in the official Program are wrong.
Correct times are shown below.
9:45 - 9:50. Introduction.
John K. Kruschke, Indiana University, Bloomington.
(Please note: Symposium attendees may also be interested in the optional Bayesian tutorial on Thursday November 18 at the conference of the Society for Computers in Psychology.)

9:50 - 10:10. Practical Advantages and Applications of Bayesian Hypothesis Tests in Experimental Psychology.
Eric-Jan Wagenmakers, University of Amsterdam.
Experimental psychologists can profit greatly from the adoption of Bayesian hypothesis tests. Through a series of practical examples, I illustrate how Bayesian hypothesis tests allow researchers to quantify evidence in favor of the null hypothesis; how they allow researchers to monitor the evidence as the data accumulate and stop whenever they feel so inclined; how they put a premium on parsimony, such that a high-N study will not automatically lead to a “significant” result; and, finally, how they relate to concepts that are intuitive and relevant. Throughout this presentation, I emphasize the recent software developments that make Bayesian hypothesis testing feasible, easy, and fun.

10:15 - 10:35. Examples of Using Flexible Psychological Models in the Bayesian Analysis of Data.
Michael D. Lee, University of California, Irvine.
Bayesian methods allow for a more flexible and mature approach to analyzing data than do traditional methods. The flexibility comes because it is straightforward in a Bayesian setting to make realistic assumptions about the complexities of experimental data, including pervasive issues like individual differences. The maturity comes because generic descriptive statistical models can easily be replaced by domain-specific models of psychological processes, with meaningful psychological parameters replacing default statistical ones. We give two case studies that make these general points, coming from the memory retention and category-learning literatures. The memory retention example focuses on the form of the forgetting curve. The category-learning example focuses on the role of selective attention in learning. In both cases, a Bayesian analysis reveals more information in the data than the traditional analysis can manage and allows for stronger and more general claims to be drawn.

10:40 - 11:00. A Hierarchical Bayesian Dual-Process Model Reveals That Recognition Memory May Be Mediated by a Single Process.
Jeffrey N. Rouder & Michael S. Pratte, University of Missouri, Columbia.
The dual-process signal detection (DPSD) model of Yonelinas has proved pivotal in assessing processes underlying recognition memory. In conventional analysis, data are averaged over people or items to form hit and false alarm rates. We show how this averaging may distort parameter estimates and threaten prior conclusions of separate recollection and familiarity processes. We develop a Bayesian hierarchical DPSD model that posits variation in multiple processes across conditions, individuals, and items. This model yields simultaneous estimates of recollection and familiarity effects across conditions, people, and items. Analysis across a number of confidence-rating recognition memory tasks reveals no strong evidence for separate processes. Recollection and familiarity estimates covary strongly across conditions, people, and items, indicating that much of the variation in confidence-rating recognition memory can be accounted for by a single mnemonic process.

11:05 - 11:25. A Hierarchical Bayesian Framework for Series of Response Times.
Peter F. Craigmile, Mario Peruggia & Trisha Van Zandt, Ohio State University. (Presented by Trisha Van Zandt.)
Response time (RT) data arise from reactions to a succession of stimuli under varying experimental conditions over time. Because of the sequential nature of the experiments, there are trends (due to learning, fatigue, fluctuations in attentional state, etc.) and serial dependencies in the data. The data also exhibit extreme observations that can be attributed to lapses, intrusions from outside the experiment, and errors occurring during the experiment. Any adequate analysis should account for these features and quantify them accurately. We demonstrate how simple Bayesian hierarchical models can be built for several RT sequences, differentiating between subject-specific and condition-specific effects.

11:30 - 11:50. Multiple Comparisons and Power Make Sense in Bayesian Analysis.
John K. Kruschke, Indiana University, Bloomington.
In experiments with multiple conditions, Bayesian methods encourage thorough data analysis and discovery, including numerous multiple comparisons, because Bayesian analysis provides rational estimates of individual and group parameters without being affected by which comparisons the analyst might intend to conduct. Bayesian analysis produces a complete distribution of credible combinations of parameter values. From this distribution, simulated data reveal the probability of achieving any research goal. Bayesian analysis thereby provides straightforward estimates of statistical power and replication probability, even for the complex experimental designs and goals of real research. These points are illustrated with actual analyses of choice and response time data from experiments in human learning.

This page URL: