The test is whether removal affects reading

What it is

A task-based removal test deletes an element and compares performance on predefined tasks to determine whether the element is irrelevant load, necessary scaffolding, or redundancy with both cost and benefit. “Reading” includes accuracy, speed, consequence of error, comprehension, memory, confidence calibration, and accessibility—not merely whether the result looks cleaner. The test converts Tufte's erasure heuristic into a measurable hypothesis, but one test cannot prove universal uselessness.

Why it happens

Deletion creates a counterfactual version for estimating an element's marginal contribution, but elements interact: removing a legend may be harmless only when direct labels remain. A cue redundant under normal viewing can become a recovery path for low vision, grayscale output, reduced size, or anomalous data. “No significant average decline” is not equivalence; inadequate power, an easy task, or omission of rare high-consequence states may conceal harm.

Studying it

State the task, users, device, and acceptable non-inferiority margin before randomising complete, removal, and necessary factorial-ablation variants. Alongside accuracy and time, measure error type, conclusion recall, confidence calibration, anomaly detection, and assistive-technology success. Use intervals or equivalence/non-inferiority tests rather than treating nonsignificance as no effect. Control learning, order, and layout reflow, and correct multiple comparisons when testing many elements.

Where it stops holding

Evidence transfers only to tested tasks and states. Motion, interactive help, responsive layout, and live updates require runtime prototypes; static screenshots cannot establish their function. Legal, ethical, provenance, safety-warning, and uncertainty requirements may impose independent completeness constraints even when a short task metric does not decline. A chart serving distinct audiences must test them separately rather than prune for an average user.

Applying it

• State each candidate's function, beneficiary task, possible failure condition, and removal hypothesis, then create controls with the same data and principal layout.
• Remove single elements first, then test combinations that may substitute for or depend on one another. Record accuracy, time, error consequence, memory, and access failures.
• Set non-inferiority margins and a sample rationale for critical tasks. Retain low-cost elements or gather more evidence when estimates remain uncertain.
• Log test conditions, version, and decision in the chart record. Re-test when data, audience, device, or task changes.

Related

• Same group: U1.14.1 Non-data elements consume attention while carrying no information · U1.14.2 Over-pruning damages readability
• Adjacent: U2.17.2 One question, several viable charts — they differ in reading cost · U1.01.1 Visual channels have a stable empirical accuracy ranking
• Search terms: removal test · ablation study · non-inferiority · task-based evaluation