An ABA data outlier review begins by preserving the observation and verifying its source, definition, unit, exposure, timing, access, health, context, observer, and system state. Correct a documented error with an audit trail. Retain a valid unusual value and investigate what it may reveal. Show any summary with and without the point transparently, keeping automatic deletion out of the review.

Treat “outlier” as a review prompt

An unusual point can reflect a data-entry error, a change in observation opportunity, a real event, an altered setting, a health or access condition, or ordinary variation in a small sample. Define what drew attention to the point: visual distance from the recent series, a predeclared statistical flag, a rule violation, or a clinical event. The label should open an investigation and should not become a synonym for bad data.

Statistical rules are method-dependent. A box-plot rule uses quartiles and IQR, and NIST’s box-plot guidance describes that structure. Quartile algorithms can differ, especially in small samples. Record the software, version, method, comparison cohort, and cutoff if a formal flag is used. A point can be clinically important without crossing a statistical fence, and a flagged value can still be accurate.

Freeze the record before investigating

Preserve the original value, source record, author, collection and entry times, observation duration, graph version, and any calculations that used it. Give the ABA data outlier review an investigation ID and record who noticed the point, when, and through which rule. Limit access to people who need the information.

If an error is confirmed, append the corrected value with the correction author, timestamp, reason, source evidence, and affected reports. Do not delete the original entry or quietly rerun a graph. The team should be able to reproduce the earlier summary and see exactly why it changed.

Verify definition, exposure, and arithmetic

Check whether the same operational definition, counting method, opportunity rule, observation length, and inclusion rule applied. Recalculate the value from the primary tally or event record. Review unit conversions, percentage denominators, device settings, duplicate events, delayed entry, and data imported from another system.

A high count during a longer observation may be ordinary when expressed as a rate. A high percentage across two opportunities may be less stable than a similar percentage across twenty. Preserve both count and exposure. Never convert a missing observation into zero, and never assume a session had the standard duration when the source shows an early end or extension.

Review context through the right roles

Look for documented changes in setting, people present, schedule, routine, task or activity, communication access, health, sleep report, medication record when within scope, treatment implementation, staffing, equipment, or safety conditions. Separate direct observations from later hypotheses. Ask Cleo and relevant supporters what was different using an accessible method.

The qualified clinician interprets clinical significance and decides whether additional assessment or a treatment review is appropriate. Medical questions go to an appropriate health professional. Privacy, payer, legal, or safety determinations stay with their authorized owners. The data reviewer can surface the discrepancy and coordinate evidence without claiming authority outside the measurement review.

Work the fictional investigation

Cleo’s session counts are 2, 3, 3, 4, and 17. The raw mean is (2 + 3 + 3 + 4 + 17) / 5 = 5.8 events per observation, and the median is 3. If the fifth value were removed, the four-point mean would be 12 / 4 = 3.0. That sensitivity result shows the point’s influence; it does not justify deletion.

The source log confirms that the first four observations lasted 30 minutes each and the fifth lasted 180 minutes during a disrupted routine. Their rates are 4, 6, 6, 8, and 5.7 events per hour after rounding. The count of 17 is accurate and remains visible. The comparison report uses rate because exposure differs, while retaining the count, duration, disruption note, and the fact that conditions may still be incomparable for other reasons.

This example also shows why a count-only graph can mislead. The longer session increases the opportunity for events. Rate corrects for duration mathematically, yet it cannot remove effects of fatigue, changed activities, staff transitions, or the disrupted routine.

Handle missingness and neighboring points

Inspect the observations before and after the unusual value. A single spike surrounded by the prior level poses a different descriptive question from a persistent shift. Keep missing sessions, cancellations, and delayed observations visible because they affect whether the point is isolated and whether the comparison window is representative.

Do not widen or narrow the comparison cohort after seeing the result. Define the phase, dates, setting, and eligible observations first. If the observation falls under a new definition or exposure rule, show separate summaries and explain the break. When a valid point is retained in a very small sample, prefer raw display and cautious wording over a precise threshold claim.

Graph and report the result transparently

Keep the original point on the raw graph. Add a concise annotation tied to the investigation record, such as “180-minute observation; disrupted routine; count verified.” If a rate graph is added, retain the count and exposure table nearby. Avoid changing axis limits merely to make the point appear smaller.

A report could say: “One count of 17 occurred across 180 minutes; the other four counts ranged from 2 to 4 across 30 minutes each. Source review confirmed the value. Rates ranged from 4 to 8 events per hour, with 5.7 per hour for the long observation. The point was retained, and the changed duration and routine limit direct comparison.” This states what was checked and what remains uncertain.

Keep interpretation bounded

An outlier investigation can identify a correction, exposure difference, documented context, or need for more observation. It cannot prove what caused the value, whether an intervention was effective, or whether the measured outcome reflects benefit or harm. NIST outlier guidance advises investigating unusual points before deletion. Single-case design and visual-analysis sources provide broader interpretation context.

Review the outcome’s importance with Cleo. Keep AAC continuously available under ASHA’s AAC guidance, and preserve access to food, water, bathroom use, mobility, prescribed health care, rest, relationships, and emergency help. A statistical flag must never become a reason to withhold support.

Clinician review checklist and limitations

Before closing the investigation, confirm:

  • the original value, graph, source, exposure, and audit history are preserved;
  • the operational definition, tally, unit, duration, denominator, and arithmetic reproduce;
  • missing, zero, canceled, extended, and shortened observations remain distinct;
  • the comparison cohort and any statistical flag method were declared;
  • direct context evidence is separated from hypotheses and routed by authority;
  • graphs show the point, denominator or exposure, and sensitivity view transparently;
  • Cleo’s accessible feedback and meaningful-outcome perspective are recorded; and
  • corrections, decisions, open questions, owner, and review date are attributable.

Outlier rules are sensitive to sample size, distribution, quartile convention, and software. Rate adjustment addresses exposure but may leave major contextual differences. A valid unusual point can carry important process information, and a corrected error can affect earlier decisions. Reopen the review when source records, definitions, denominators, context, or software change, and keep every affected report linked to the outcome.

Related resources

Sources