Visualizing uncertainty and qualifier for Bayesian decision-making
MetadataShow full item record
Probabilistic decision-making is ubiquitously present in medical diagnosis, intelligence analysis, weather forecasting, and other command and control domains. People often rely on incomplete and uncertain information to make high-stake decisions in such complex domains. For example medical diagnostic tests have varying levels of sensitivity and specificity, indicating the chances of missing a diagnoses or falsely diagnosing an illness. An informed diagnostic decision should combine prior knowledge about a disease (e.g. base rate) with new evidence such as test result. A history of previous research shows that people generally do not function well at combining and reasoning about such probabilities. The goal of this research was to develop and to evaluate a novel visualization method for representing uncertainty and its qualifiers in single- and multi-step Bayesian reasoning. The visualization method was instantiated in an interactive software tool, which visualizes prior and posterior probabilities with colors and shapes. Qualifiers of uncertainty (e.g. range of uncertainty, overlapping uncertainty) are further represented with graphical variables such as color saturation and semi-transparency. An experiment tested the effectiveness of the visual representation by asking 60 participants to answer 28 questions regarding six Bayesian reasoning scenarios. The results show significant performance advantage in participants' response accuracy for the two improved display formats augmented by the visualization method in static screenshots or in interactive software. The performance advantage was present across all domains tested and among participants with diverse levels of prior statistical knowledge. It is also evident that the visualization method can reduce participants' subjective workload and mitigate the adverse effect of increasing complexity in the Bayesian reasoning scenarios. The development of a novel visualization method and evaluation scheme, and the knowledge gained from the experiment, has demonstrated a good potential in contributing to future implementations of visualized Bayesian decision-support for practitioners in medical diagnosis and possibly other command and control domains. The visualization method and the software could also benefit the general public by promoting broader understanding and applications of Bayesian decision-making. This study provides a necessary step towards a range of future researches in visualized Bayesian reasoning and in uncertainty visualization for decision-making.