Causal Support: Modeling Causal Inferences with Visualizations.

Area of Science:

• Human-Computer Interaction
• Cognitive Science
• Statistics

Background:

• Visual analytics (VA) software often leaves analysts' causal inference models implicit.
• This lack of explicit models raises concerns about the statistical validity of visual insights.
• There is a need for formal evaluation of causal inferences made from visualizations.

Purpose of the Study:

• To formally evaluate the quality of causal inferences derived from visualizations.
• To establish causal support, a Bayesian cognitive model, as a normative benchmark for causal inferences.
• To assess how chart users detect treatment effects and confounding relationships.

Main Methods:

• Adopted causal support, a Bayesian cognitive model, as a normative benchmark.
• Conducted two experiments with crowdworkers to assess causal inference detection.
• Evaluated sensitivity to sample size and the impact of interactive cross-filtering.

Main Results:

• Chart users' causal inferences were insensitive to sample size, deviating from the causal support benchmark.
• Interactive cross-filtering showed potential for improving sensitivity but did not guarantee reliable performance.
• Users performed similarly with common visualizations and textual contingency tables on average.

Conclusions:

• Causal support is a valuable framework for evaluating inferences in visual analytics.
• Analysts' mental models need to be made more explicit within VA software.
• Improvements are needed to enhance the statistical validity of visual causal inferences.