Do not rely on caption text alone to correct a confidence-interval chart

Change the visual encoding instead of trying to repair a confidence-interval chart with more caption text. For example, do not keep the same 95% confidence-interval bars and add extra text about prediction intervals; replace the interval encoding or rescale the axis, then make the caption match the plot.

The visual marks dominate effect-size perception in these charts. Extra explanation in the caption does not undo the effect of narrow confidence-interval bars enough to bring judgments close to those produced by outcome-uncertainty displays.

Mechanism: Readers anchor on the displayed intervals, so added text about other statistics does not sufficiently change how large or reliable the treatment looks.

Evidence: In the first experiment, adding extra caption text about additional interval information reduced the bias only somewhat; participants still judged the same confidence-interval plots as showing a more effective treatment than prediction-interval plots (Hofman et al., 2020).

Notes: Extra text helped less than changing the plot itself.

• User Goal: Help readers judge whether a treatment is effective.
• Task: Compare two summarized conditions from a plotted result.
• Chart Setting: The figure currently uses confidence-interval bars, and the proposed repair is to add explanatory text.
• Component: The caption or surrounding text can be edited.
• Audience: Lay readers or other nonexpert consumers of the result.
• Success Criterion: The revision materially changes the judgment readers form from the figure.

Break it when: The plot itself already encodes outcome uncertainty and the caption is only reinforcing what is shown. Why: The failed repair in the study was adding extra text while leaving a confidence-interval visualization unchanged.

Sacrifice: You give up the convenience of a text-only revision. Risk: Changing the plot may require more space or rework than editing the caption. Mitigation: If confidence-interval bars must remain, change the axis as well rather than only lengthening the caption.

Mistake: Add definitions or extra text about prediction intervals while leaving the confidence-interval marks unchanged. Why it fails: The chart still visually emphasizes inferential uncertainty, so readers continue to overestimate treatment effectiveness.

Failure Sign: The proposed revision changes words but not marks. Quick Check: Compare the before and after figures; if the intervals and axis are unchanged, do not expect the bias to disappear. Stronger Test: Compare the caption-only revision against a plot revision such as a prediction-interval version or a rescaled confidence-interval version before publication.

• Replace the confidence-interval encoding with a prediction-interval encoding when readers must judge likely individual outcomes.
• If confidence intervals must stay, rescale the axis instead of only adding caption text.
• After changing the plot, rewrite the caption so it matches the displayed uncertainty.