Taking Truncation to Task: A Task-Based Exploration of Axis Truncation in Bar Charts

Paper Title

Taking Truncation to Task: A Task-Based Exploration of Axis Truncation in Bar Charts

Publication Info

• Topic area: Effects of axis truncation in bar charts on task performance.
• Keywords: Axis truncation, bar charts, data visualization, task performance, ratio heuristic, data labeling, chart readability, experimental design.

Background and Problem

• Problem / challenge: Existing guidelines broadly discourage axis truncation in bar charts, labeling it as misleading, particularly for ratio judgments. However, these guidelines lack nuance and fail to consider potential benefits of truncation for other tasks.
• Significance: Misapplied guidelines may hinder chart readability and communication in critical contexts like healthcare or emergency response.
• Motivation and related work: Prior research has focused on the negative effects of truncation, particularly its impact on ratio judgments, but has not explored its effects across a broader range of tasks. This paper addresses this gap by investigating the nuanced effects of truncation on various bar chart tasks.

Solution

• Proposed approach: A task-based exploration of axis truncation effects across seven bar chart tasks, using controlled experiments to evaluate performance differences in accuracy and speed.
• Novelty:
  1. Demonstrates that truncation can improve performance for certain tasks (e.g., Filter) while increasing error for others (e.g., Ratio).
  2. Shows that the effects of truncation depend on the degree of truncation and the data range.
  3. Introduces data labeling as a mitigation strategy to reduce truncation-related errors.
• Procedure and key techniques:
  • Conducted three experiments with controlled stimuli to measure task performance under truncated and untruncated conditions.
  • Experiment 1: Compared performance across seven tasks for truncated vs. untruncated charts.
  • Experiment 2: Analyzed the interaction between truncation effects and data range (small, medium, large).
  • Experiment 3: Tested data labeling as a mitigation strategy for truncation effects.

Results

• Concrete findings:
  • Truncation increased error for Ratio tasks but reduced error for Filter tasks.
  • Truncated charts improved speed for Retrieve Value and Make Comparison tasks without reducing accuracy.
  • Data labeling significantly reduced errors for both Ratio and Filter tasks and mitigated differences between truncated and untruncated charts.
• Advantage over baselines:
  • Truncated charts showed faster task completion for Retrieve Value and Make Comparison tasks compared to untruncated charts.
  • Data labeling neutralized the negative effects of truncation on Ratio tasks and enhanced overall accuracy.
• Experiments / evaluation:
  • Experiment 1: N=62, binary truncation design, seven tasks.
  • Experiment 2: N=57, varied data ranges (small, medium, large), seven tasks.
  • Experiment 3: N=60, binary labeling design, two tasks (Filter, Ratio).
  • Metrics: Accuracy (error rates), speed (duration), and demographic/literacy correlations.
• Limitations and future work:
  • Results may not generalize to real-world settings with time pressure or lower attention.
  • Focused on binary truncation; future work should explore subtle truncation levels and alternative designs (e.g., aspect ratios, logarithmic scales).
  • Did not assess other metrics like decision-making, aesthetics, or trust.

Summary

This paper investigates the nuanced effects of axis truncation in bar charts across seven tasks, demonstrating that truncation can improve accuracy and speed for some tasks (e.g., Filter, Retrieve Value) while increasing error for others (e.g., Ratio). The effects of truncation depend on the degree of truncation and the data range. Data labeling emerged as an effective strategy to mitigate truncation-related errors, particularly for Ratio tasks. These findings challenge the universal rejection of truncation, advocating for a more nuanced, task-specific approach to bar chart design.