Decision making with visualizations: a cognitive framework across disciplines

The paper addresses how people make decisions using visualizations and argues for integrating decision-making frameworks into visualization cognition research. It identifies a gap: visualization studies often lack a unifying cognitive model that explains decision processes. The authors propose combining dual-process theories of decision making (Type 1: fast, low-effort; Type 2: slow, effortful, working-memory dependent) with established models of visualization comprehension (e.g., Pinker’s framework). The purpose is to present an integrated cognitive model for decision making with visualizations, review cross-domain evidence supporting dual-process mechanisms, and derive practical design recommendations. The study is important because visualizations inform high-stakes decisions (e.g., medical, weather, navigation), and a unified model could improve generalizability and design efficacy across disciplines.

The review summarizes two dominant decision-making traditions—rational, probabilistic models and heuristic/intuition-based approaches—and situates dual-process accounts as encompassing both. It contrasts dual-systems and dual-process conceptions, emphasizing Type 1 autonomy and minimal working-memory demands versus Type 2 controlled, capacity-limited processing. In visualization cognition, it contrasts perceptually focused frameworks with approaches emphasizing prior knowledge and cognitive fit. Pinker’s model of graph comprehension (from visual array to visual description, schema matching, message assembly, and conceptual message) is discussed, along with Cognitive Fit Theory and critiques of prior models (e.g., Patterson et al., 2014) that underemphasize working memory’s role in decision stages. The literature highlights interactions between bottom-up attention and top-down knowledge, the role of schema, and the need to integrate decision processes with visualization comprehension.

This is a selective, cross-domain review of empirical studies on complex decision making using static 2D computer-generated visualizations. The authors sampled representative studies across application areas such as meteorology/weather, healthcare risk communication, land-use/urban planning, finance, geospatial localization, statistics communication, social networks, and emergency management. The review focuses on identifying evidence for dual-process mechanisms in visualization decisions and distills four cross-domain findings: bottom-up attention effects, visual-spatial biases from encoding, cognitive fit (alignment/mismatch with tasks and schemas), and interactions of knowledge-driven processes with encoding. No original data were collected; methods synthesize prior experimental results (including eye-tracking, saliency modeling, individual differences, and task performance metrics).

• Evidence for Type 1 (bottom-up attention): Salient features in visualizations involuntarily capture attention and can help or hinder decisions. Examples include: icon arrays causing a foreground effect where viewers focus on icons and neglect base rates, leading to willingness to pay $125 more for improved tires versus text-only information; hurricane cone forecast saliency misinterpreted as storm growth, corrected by path-based ensemble displays but with path-intersection biases; weather map salience manipulations capturing gaze and aiding performance after training.
• Visual-spatial biases (Type 1): Encoding choices induce heuristics tied to visual-spatial conventions (e.g., containment within boundaries, deterministic construal of uncertainty, anchoring to initial placements, preference for high-quality/realistic images). Containment biases observed with bounded uncertainty regions (e.g., Google Maps blue dot) versus Gaussian fades; deterministic construal errors for temperature uncertainty bars; anchoring in aligning means with error bars by experts; high-fidelity images increase perceived scientific credibility but can impair task performance. Some biases (e.g., anecdotal evidence, side-effect aversion, risk aversion) can be reduced with appropriate visuals.
• Cognitive fit (Type 2): Mismatches between visualization, schema, and task require working-memory-driven transformations, increasing time and errors. When displays align with tasks/schemas, performance equalizes across working-memory capacity. Examples: network diagrams aligned with disconnection tasks eliminate WMC differences; schema priming (speed vs security) interacts with thickness/containment/distance encodings; choosing graphs vs tables aligned with spatial vs textual tasks speeds performance.
• Knowledge-driven processing (Type 1 and/or 2): Short-term training can override familiarity biases and improve visualization choice; however, some visual-spatial biases persist despite keys/instructions (e.g., error-bar misinterpretations). Individual differences (health literacy, graph literacy, numeracy) moderate benefits; natural frequency icon arrays improve comprehension among low numeracy/graph literacy. Time pressure manipulations (e.g., wildfire decisions at 5 s vs 30 s) show display effects under Type 1 that diminish with more time (Type 2).
• Model support: Evidence supports a decision stage influenced by both minimal and significant working-memory use, bottom-up attention pathways, schema matching, inference, and behavioral responses consistent with a dual-process framework.

The findings substantiate that decision making with visualizations involves both rapid, attention-driven processes and slower, working-memory-intensive reasoning. Bottom-up salience can guide attention toward task-relevant features when design aligns with goals but can also mislead via visual-spatial biases originating from encoding choices (e.g., containment, deterministic construal). Type 2 processes are recruited when cognitive fit is poor, requiring mental transformations to align the visualization with the viewer’s schema or task, which introduces time costs and potential errors, especially for users with lower working-memory capacity. Knowledge influences these dynamics: training and expertise can enhance performance, yet some visual-spatial biases remain resistant, likely because they arise early from attention capture. The integrated model bridges visualization comprehension and dual-process decision theories, offering explanatory power across domains and informing design strategies that either leverage Type 1 processing (by making critical relations perceptually evident) or reduce Type 2 burdens (by improving cognitive fit).

The paper contributes an integrated dual-process cognitive model for decision making with visualizations, supported by cross-domain empirical evidence. It identifies four generalizable findings: (1) bottom-up attention strongly shapes decisions; (2) encoding-induced visual-spatial biases can help or hinder; (3) better cognitive fit yields faster, more accurate decisions; and (4) knowledge-driven processes interact with encoding effects. Practical recommendations include directing salience to critical information, assessing saliency (e.g., via algorithms), aligning displays with users’ schemas and task demands, minimizing required mental transformations, using dual-task paradigms to assess working-memory demands, and accounting for individual differences (graph literacy, numeracy). Future research directions include formalizing schema–task–visualization alignment mechanisms, distinguishing when knowledge operates automatically versus via working memory, exploring expert performance under time/WM constraints, and systematically studying visual-spatial biases as a unique class of decision biases.

• The review is selective rather than systematic; it focuses on static 2D visualizations and may omit relevant interactive/3D or dynamic contexts.
• The integrated model is supported indirectly; few studies directly test all proposed pathways (e.g., explicit manipulation of schema matching mechanisms).
• Mapping knowledge-driven effects to Type 1 vs Type 2 processes remains unclear; more work is needed to determine when knowledge is automatic versus effortful.
• Predictive criteria for cognitive fit beyond broad categories (spatial vs textual) are under-specified; schema selection mechanisms are not fully characterized.
• Some evidence is domain-specific, limiting generalizability until further cross-domain validation.