Short answer
HR data input quality control tests whether each source is fit for the decision before it enters analysis. Define the intended use, population, period and fields; check missing and duplicate records; preserve provenance and transformations; distinguish employee accounts from verified events; and publish unresolved limits beside the result. Clean formatting alone does not make an input accurate or representative.
Begin with a data-use card
A quality check needs a purpose. The same field may be adequate for sending an invitation and inadequate for comparing teams.
Copy this card before opening the dataset:
| Field | What to specify |
|---|---|
| Decision | The action, test or question the data may inform |
| Population | Who or what should be represented |
| Period | Event dates and the reporting cut-off |
| Unit | Employee, invitation, response, case, shift or another defined record |
| Critical fields | Values required for this use |
| Source owner | Person or system responsible for the source |
| Quality threshold | What must be checked before use |
| Excluded use | Decisions the data cannot support |
| Escalation | Who resolves or accepts a material quality issue |
The UK Government Data Quality Framework describes quality as fitness for purpose and separates completeness, uniqueness, consistency, timeliness, validity and accuracy. It also recommends treating issues at source and documenting quality through the data lifecycle.
Run nine checks before analysis
1. Purpose and applicability
Ask whether the source measures the concept in the decision. Course completion records attendance or completion under a defined rule. They do not by themselves establish that someone can perform a task in a different setting.
The US Government Accountability Office's data reliability guide assesses accuracy, completeness and applicability for the intended purpose. It uses a risk-based approach, so a low-consequence local exploration does not need the same assurance as a decision affecting employment or a large investment.
2. Population and coverage
List eligible, invited, reached, contributing and excluded populations. Check whether a missing group shares a route, shift, site or access constraint. A high overall participation figure can still hide an absent population.
3. Definitions
Maintain a short dictionary for every critical field:
- exact name and definition;
- valid values and units;
- effective date and version;
- source system or collection question;
- owner and correction route.
If one site records “handover delayed” when the next shift starts and another records it after 24 hours, combining their counts creates a false comparison.
4. Completeness
Measure both missing records and missing critical fields. State the denominator. A dataset can have every expected row and still lack the field needed for the decision.
5. Uniqueness
Define what makes a record unique. Check repeat submissions, reissued invitations, merged employee IDs and copied extracts. Keep a decision log for duplicates instead of deleting them silently.
6. Validity and consistency
Test expected formats, ranges and relationships. Then compare the same definition across systems, sites and periods. A valid date can still refer to the wrong event; a permitted category can still be interpreted differently.
7. Accuracy and corroboration
Trace a risk-based sample back to an appropriate source. The US Office of Personnel Management's current federal workforce data-quality page distinguishes accuracy, completeness and timeliness and describes agency review, authoritative comparisons and validation checks. Its operational examples are specific to US federal data, but the distinction applies well to an HR quality log.
An employee response is source material. A follow-up may clarify the event, date or meaning. It does not convert a self-reported account into independent verification. When the decision requires it, compare the account with a permitted document, observation or another relevant perspective.
8. Timeliness
Record the event date, capture date, processing date and availability date. “Updated today” can describe a file refresh even when the underlying event is months old.
9. Provenance and transformation
Keep the path from source to output:
- original record or account;
- collection question and context;
- corrections or exclusions;
- joins and deduplication;
- translation, coding or classification;
- summary or metric;
- human review and release status.
Never let a generated summary replace the source. Label employee wording, analyst interpretation and machine-assisted fields separately.
The qualitative HR data guide provides a fuller record schema for source material, derived fields, access and retention after the input checks are complete.
Use this quality issue log
| Issue | Affected records | Decision risk | Immediate treatment | Root cause owner | Recheck |
|---|---|---|---|---|---|
| Missing evening-shift invitations | 24 | Coverage | Exclude workforce-wide claim | Invitation owner | Next campaign |
| Duplicate invitation IDs | 4 | Inflated participation | Resolve against source register | Data operations | Before release |
| Definition changed mid-period | All three sites | Invalid trend | Split the periods | Measure owner | After definition review |
| Summary lacks source link | 7 themes | Weak traceability | Hold themes from decision pack | Analysis lead | Before meeting |
A correction should propagate to derived tables and reports. Preserve the original issue, treatment, owner and date so the same failure can be prevented upstream.
Fictional example: an onboarding input file
Pinebrook Services is fictional. HR wants to decide whether to test a new first-week handover across four sites. The eligible cohort contains 120 new starters.
- Invitations were delivered to 96 people: 96 / 120 = 80.0% delivery coverage.
- The file contains 72 submissions.
- After deduplication, 68 unique submissions remain; four duplicates were linked to reissued invitations.
- Three of those 68 submissions fall outside the agreed first-week window.
- The usable file contains 65 unique, in-window submissions: 65 / 120 = 54.2% of the eligible cohort.
Neither percentage measures onboarding quality. Evening-shift delivery was lower than day-shift delivery, so the 65 accounts do not support a claim about all new starters. HR reports the coverage gap, reviews the invitation route and restricts the findings to participating groups.
Within the usable accounts, several people describe not knowing who approves a customer exception. Analysts retain the question, source passage, site and date, then check the current handover document and ask the process owner. They find two sites using an old document. That supports a targeted document test; it does not prove the document caused every reported difficulty.
Protect employment records while correcting them
Quality control should not expand access or collect extra identity by default. The UK Information Commissioner's Office guidance on keeping employment records covers lawful collection, minimisation, accuracy, retention, security, transparency and worker access. It was under review for changes in UK law when checked, so employers should confirm the current requirements for their programme.
Give people an appropriate route to correct inaccurate personal information. Restrict raw and identifiable material to authorised roles, and make sure an analytic correction reaches any decision record that used the faulty input.
Where guided conversations fit
Guided follow-up can improve specificity by asking for the event, example, condition or missing detail. It cannot promise that every answer is complete, comparable or true. Managers receive a prepared brief rather than raw employee responses, and HR grouped material requires at least five respondents before display.
Review the Lontra product overview for the employee conversation and manager-brief workflow. If it is used as an input source, record the campaign purpose, question, population, coverage, collection dates, access and review status beside other data sources.
Good input quality control does not certify a dataset once and move on. It makes every material limitation visible at the point where a person decides whether the data are usable.
