An ABA summary table audit traces each cell to locked source data and verifies the cohort, analysis unit, formula, weighting, missingness, exclusions, rounding, stratification, privacy, accessibility, and version. Recalculate a sample independently and preserve corrections. A table becomes review-ready when its values and construction choices are reproducible from the underlying records.

Freeze the table and its purpose

Begin the ABA summary table audit with a release candidate that has a version, owner, extraction time, intended audience, clinical question, eligible cohort, maturity cutoff, and planned review date. Lock the source extract or preserve a reproducible query and its parameters. Reviewers need to know exactly which records and software state produced the table.

Define what every row and column represents. A row may be a client, phase, week, session, target, site, or service line. A percentage column needs its numerator, denominator, inclusion rule, and missing-state treatment. A mean needs its analysis unit and weighting. A range or IQR needs its source values and calculation convention.

Trace each cell to evidence

Build a data dictionary with field name, plain-language label, definition, unit, source system, source field, transformation, allowed missing states, and responsible owner. Each displayed value should link back to a source record or a documented calculation. Derived fields need formulas that a second person can inspect.

Label source times and versions. Clinical data, scheduling records, treatment-integrity measures, access notes, and client feedback may update on different schedules. A quarterly table that mixes a Monday clinical extract with a Friday schedule extract can be legitimate only when the as-of dates and limitations are visible.

Reconcile the cohort and denominator

Start from the declared eligible population, then account for included, excluded, incomplete, late, duplicate, and unresolved records. The row total should reconcile to the cohort. Avoid completed-case denominators that silently omit clients or sessions still awaiting data.

For every percentage, independently count its numerator and denominator from the locked extract. Check whether denominators differ by row because of opportunity count, exposure, missing observations, or eligibility. Preserve zeros as valid values and blanks as their actual missing states. A hidden filter or collapsed group can remove records while leaving formulas apparently intact.

Inspect formulas, weighting, and rounding

Review spreadsheet ranges, named formulas, joins, grouping, pivot settings, weights, date boundaries, phase rules, and unit conversions. Look for copied formulas that stop one row early, mixed absolute and relative references, text values treated as zero, and rounding applied before aggregation.

Recalculate high-impact values and a spread of ordinary rows from primary records. A practical sample can include the largest and smallest values, each phase or setting, rows with missing data, a row near a decision threshold, and randomly selected remaining rows. A small table may permit reproduction of every row. Preserve full precision in the working file and state display rounding.

Validate labels and time structure

Compare client, phase, setting, staff, service, procedure, payer, and date labels with the authoritative source. A correct number under the wrong phase is still a material error. Check whether phase boundaries, service transitions, observer changes, and definition revisions are reflected in both data and labels.

Order time-based rows by actual date or session sequence and inspect gaps. A monthly average can hide a midmonth phase change. A client summary can combine two settings with different exposure. Split incompatible records or label the mixed scope deliberately.

Work the fictional release audit

Esme audits an 18-row quarterly summary. Three rows use the wrong denominator, one hidden filter excludes an eligible record, and two rows carry the wrong phase label. The six affected rows are distinct, so the first-pass error count is 6 of 18, or 33.3% after rounding 0.333 to one decimal place.

The audit retains all 18 rows in the denominator. The team restores the filtered record, repairs the three denominator calculations, corrects the phase labels, and appends a correction log with affected row IDs. A second reviewer then reproduces 18 of 18 rows, or 100%, from the locked source and data dictionary. That second-pass result does not erase the 6-of-18 release finding.

The audit record should distinguish six affected rows from six errors. A row may contain several errors, and one hidden filter may affect multiple calculations. Report row-level impact, error count by type, and any decisions or reports that used the earlier version.

Protect privacy and accessibility

Use the minimum identifiable detail needed for the review, store extracts in approved locations, restrict roles, and control downloads. Small cells, rare services, exact dates, or narrative notes may identify a person even after names are removed. Apply the organization’s approved privacy rule and route legal or security questions to qualified owners.

Make the table usable with clear headers, units, row labels, sufficient contrast, accessible reading order, and text descriptions for visual encodings. Avoid conveying phase or status by color alone. During review with Esme or another AAC user, keep communication tools continuously available under ASHA’s AAC guidance.

Write the release note

A review-ready note can say: “This table includes 18 eligible quarterly rows from the locked August 15 extract. Each row was reproduced after correction using the attached data dictionary. The first-pass audit identified six affected rows: three denominator errors, one hidden-filter exclusion, and two phase-label errors. Current display values are rounded to one decimal place; working calculations retain full precision.”

State missingness, exclusions, analysis unit, weighting, range or quartile convention, software version, and unresolved issues. If the table feeds a graph, confirm that the graph uses the same corrected version and labels. Archive the release candidate, correction log, released table, and reviewer sign-off.

Keep interpretation within scope

Reproducibility means another qualified reviewer can obtain the same displayed result from the declared records and rules. It does not validate the operational definition, prove treatment integrity, demonstrate benefit, or establish cause. A perfectly calculated mean can still summarize an unsuitable outcome or incompatible observations.

NIST descriptive guidance distinguishes location, spread, and shape. Its outlier guidance supports investigating unusual values before deletion, and its box-plot page shows why quartile conventions must be declared. Single-case design and visual-analysis sources provide clinical research context.

The BACB ethics hub and CASP public summary provide professional context. The BCBA Test Content Outline is examination content and does not serve as a release protocol. Qualified clinicians remain responsible for clinical interpretation and for asking whether the outcome is meaningful to the person.

Clinician release checklist and limitations

Before releasing the table, confirm:

  • the version, purpose, cohort, cutoff, source extract, query, and owner are preserved;
  • every row reconciles to included, excluded, incomplete, duplicate, and unresolved states;
  • numerators, denominators, formulas, weights, units, missingness, and rounding reproduce;
  • phase, setting, service, person, and date labels match authoritative records;
  • unusual values and corrected records retain their source and audit history;
  • a second reviewer tested the defined sample or complete table independently;
  • privacy controls, accessible reading order, labels, and export behavior were checked; and
  • the release note states construction choices, limitations, open issues, and next review.

An audit is bounded by the source data and tests performed. Reproducing 18 rows cannot detect an inaccurate operational definition, an undocumented event, or a source-system defect shared by both reviewers. Spreadsheet and query software can change behavior across versions. Reopen the audit when data, definitions, filters, formulas, labels, cohort rules, or software change, and link downstream graphs or decisions to the exact released version.

Related resources

Sources