Does graphical editing require interactive software?

No. Paper or static graphs can support it, although linked interactive views can make investigation easier.

Is every outlier an error?

No. An outlier can be a valid and important observation.

Is editing the same as imputation?

No. Editing identifies and handles quality issues; imputation fills missing or unusable values through a specified estimation process.

AI Glossary / Statistics and probability / Beginner

Graphical data editing

Definition

Graphical data editing is the use of graphs to identify potentially erroneous, inconsistent, or otherwise unusual data for investigation and possible correction. In interactive systems, selecting a point can reveal its underlying record and highlight related observations in other displays. The graph helps locate cases requiring attention; it does not establish that an unusual value is wrong or authorize changing valid observations simply to make a dataset look smoother.

What it means

Tables can conceal patterns that become obvious visually: an isolated point, an abrupt break in a time series, a cluster with different units, or a relationship that suddenly changes. Graphical editing brings these patterns into a data-quality workflow.

It can support record-level review or investigation of influential contributions to aggregate results. For example, a current-versus-previous-period plot may reveal a large change that deserves checking against the original report.

The essential distinction is between detecting a suspicious pattern and deciding what it means. A genuine business expansion can look like an error, while a consistent unit-conversion mistake may produce a smooth, plausible-looking graph. Evidence beyond appearance is needed to resolve the case.

How it works

Choose displays suited to the data and intended checks: histograms for distributions, scatterplots for relationships, time plots for changes, and linked views for investigating groups. Label units and transformations clearly and preserve identifiers connecting plotted points to source records.

Flag unusual observations, inspect relevant context, and consult source evidence or a knowledgeable data owner where needed. Record a decision: retain as valid, correct a confirmed error, or mark unresolved according to the project's documented procedure. Missing values requiring estimation should remain distinguishable from verified corrections.

Preserve original data and an audit trail of changes, reasons, and reviewers. Recalculate affected results and reapply checks after a correction. Review priorities can reflect likely impact, but small systematic errors and less visually prominent groups still require suitable checks. Graphical methods complement logical validation and other quality controls.

Three practical examples

Survey processing: inspect a plot of reported sales against employee count, then investigate an isolated record for possible unit or transcription errors.
Sensor monitoring: examine time-series discontinuities and compare device-maintenance logs before treating a sudden jump as an invalid measurement.
AI preparation: use linked distributions and record views to investigate unusual training examples, protecting legitimate rare cases from automatic deletion merely because they look different.

Think of it this way

Think of a dashboard warning light. It directs attention to a possible problem but does not tell the mechanic to replace a particular component without diagnosis. A graph likewise helps an editor find a record worth examining. The responsible action comes from the investigation, not from the visual signal alone.

Worked example

Consider an original example of five stores with employee counts 10, 12, 14, 16, and 18, and reported sales of 100, 120, 140, 160, and 1,800 thousand currency units. A scatterplot of sales against employees places the fifth store far above the first four.

The reported sales total is 2,320 thousand, and the mean is 464 thousand. Sales per employee are 10, 10, 10, 10, and 100 thousand. The fifth ratio is a strong reason to investigate, but neither the plot nor the ratio proves a mistake.

Suppose the original signed report confirms sales of 180 thousand and the entry log identifies an added zero. Correcting that verified transcription error gives total sales of 700 thousand and a mean of 140 thousand. The correction reduces the total by 1,620 thousand. Independent summation verifies both totals and means.

If the source instead confirmed 1,800 thousand, the proper decision would be to retain the observation and explain its context, not force it onto the visual relationship. Deleting the fifth record would leave a mean of 130 thousand among the other four, which is a different analysis from correcting its value. The example shows how review evidence determines the appropriate action.

Illustrative case study

In a fictional business survey, editors use linked scatterplots and historical charts to review unusual reports. One selected point reveals a change from reporting amounts in thousands to reporting whole currency units. Supporting records confirm the conversion, so the team makes a documented correction and checks all affected totals.

Another unusual point reflects a genuine acquisition and is retained with a note. The editors also review whether the form's unit instructions caused repeated confusion. The fictional case illustrates how graphical editing can improve both current records and future collection procedures without treating every outlier as a defect.

Common misconceptions and mistakes

Common mistakes include deleting every point beyond a boxplot whisker, moving values toward a fitted line without evidence, and using a graph's appearance as proof that all remaining data are correct.

Overplotting can hide duplicate or dense records. Axis limits, log transformations, and aggregation can conceal patterns or exaggerate their importance. A change in a time series may reflect a definition change rather than a data-entry error.

Excessive editing can introduce bias by forcing observations to match an analyst's expectations. In AI datasets, this can remove rare but important situations and make evaluation misleading. Review rules should distinguish legitimate extremes, confirmed mistakes, and unresolved anomalies, rather than merging them into one deletion category.

Business use

Use graphical editing where visual relationships can guide targeted investigation and source evidence is available. Track the number of flagged records, confirmed errors, retained unusual cases, and the impact of edits on important outputs. Keep access to detailed records appropriate to their sensitivity. The business benefit comes from better-supported data and fewer recurring collection errors, not simply a tidier chart.

Check your understanding

Answer: A point lies far from the main cloud, but its source documentation confirms the recorded value.

Should it be changed to match nearby points?

Answer: No. Retain the valid observation and investigate whether its context affects the analysis. A visual pattern is a review signal, not a replacement value.

Related terms

Data editing
Outlier detection
data validation
micro-editing
macro-editing
data provenance
audit trail
imputation

Learn next

Learn how scatterplots, distributions, and time-series displays reveal different issues. Then practice separating detection, verification, correction, and imputation, with a reproducible record of every decision.

Time series

Sources

OECD, Glossary of Statistical Terms, Graphical data editing, printed page 235: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf — definition attributed to UNECE's statistical-data-editing glossary. Statistics Canada, Quality Guidelines, Editing (archived methodological guidance): https://www150.statcan.gc.ca/n1/pub/12-539-x/2009001/editing-verification-eng.htm — graphical detection, over-editing, verification, and audit principles. The stores and survey case above are original fictional teaching examples.

Previous: Graph semantics Next: Graphical degree sequence