Abstract

Data visualizations are now a primary medium for public communication, scientific reasoning, and decision-making. By transforming complex data into accessible graphical forms, visualizations are widely assumed to make information comprehensible to broad audiences. Yet the ability to accurately read, interpret, and critically evaluate a visualization, what researchers call visualization literacy, is neither uniform nor universal. It is a learned, multidimensional skill shaped by prior exposure, educational opportunity, and context. While the field has developed standardized instruments to measure visualization literacy, these tools were built under narrow assumptions: participants are typically paid, English-speaking, and recruited from Western online panels. When any of these conditions change, what our assessments reveal and what they conceal change with them. This dissertation examines what happens when those conditions are systematically varied. Across four complementary empirical studies, I investigate how standardized visualization literacy assessments behave when measurement conditions, cultural contexts, and agent types differ. First, I address the scalability barrier posed by long assessments by developing and validating Mini-VLAT, a 12-item short form of the 53-item Visualization Literacy Assessment Test (VLAT). Mini-VLAT preserves the psychometric properties of the full instrument while reducing completion time by nearly 80%, enabling visualization literacy assessment at scale. Second, I use regional adaptation as a methodological probe to examine what visualization literacy assessments reveal and conceal when standard assumptions may not hold. Through a Ghana-specific adaptation of Mini-VLAT deployed across three large-scale studies, I find that adapting dataset content while preserving chart types and task structures is insufficient to ensure psychometric equivalence. This contrasts with successful adaptations in other contexts, demonstrating that one size does not fit all. Third, I extend visualization literacy measurement beyond human readers by systematically benchmarking four leading visual language models against standardized assessments. While these models approach human performance on basic chart reading tasks, they fail substantially at detecting misleading visualizations, revealing specific gaps in AI visualization capabilities with direct implications for high-stakes analytical deployments. Finally, I develop and empirically examine a multidimensional framework for trust in data visualization. A large-scale empirical study demonstrates that visualization literacy is a significant predictor of trust calibration; viewers with higher Mini-VLAT scores perceive the same charts as clearer, and clarity is strongly tied to trust. Together, these four studies argue that visualization literacy assessments are psychometrically robust within the conditions for which they were designed but context-sensitive in ways that matter for research, practice, and equity. This dissertation advances a more honest and inclusive framework for understanding who can read the chart and under what conditions our measurements help us know.

Committee Chair

Chien-Ju Ho

Committee Members

Alvitta Ottley; Bum Chul Kwon; Jiaxin Huang; Roger Chamberlain

Degree

Doctor of Philosophy (PhD)

Author's Department

Computer Science & Engineering

Author's School

McKelvey School of Engineering

Document Type

Dissertation

Date of Award

4-29-2026

Language

English (en)

Download

Share

COinS