Data Management
Use quality sampling to inspect large data batches
Define what a sample can tell you and keep serious defects visible during batch review.
IGS Editorial · 3 min read · Updated 17 November 2026

When a data batch is too large to inspect manually in full, sampling can help identify quality problems. It does not automatically prove that every unchecked record is correct. The method, sample size and consequences of missed errors need to be understood.
Start with the purpose of inspection. You may be checking whether a new process is ready, estimating common error patterns or verifying a contractual acceptance rule. These purposes can require different approaches and, in some cases, specialist statistical advice.
Define the batch and defect
State which records belong to the batch and what counts as an error. A misspelled optional note differs from a wrong account identifier. Separate defect categories by their effect on the intended use.
Use a reference source or approved rule for each check. Reviewers cannot assess accuracy consistently if they disagree about the authoritative value. Resolve ambiguous definitions before interpreting the results.
Avoid convenient-only selection
Selecting the first rows or easiest records may miss patterns elsewhere in the batch. Choose a method that suits the question and record how items were selected. Consider different sources, operators or time periods when they may have different risks.
If you deliberately oversample high-risk work, explain that the sample is not a simple estimate of the whole batch. It can still be valuable for finding problems, but the result should not be presented as though every record had the same selection chance.
Calibrate the reviewers
Have reviewers check common examples and discuss disagreements. Record edge cases and the accepted interpretation. Otherwise apparent quality differences may reflect reviewer judgement rather than differences in the data.
Keep a record of both the defect and its source. A value may be entered correctly from a flawed source document. That requires a different corrective action from an entry error.
Decide what findings trigger
Agree whether certain defects require a wider review, batch hold or source investigation. A serious error should not disappear inside a favourable average. Define who can authorise release and what evidence they need.
For example, an incorrect entity reference may justify checking related records even if only one instance appears. The response should reflect the consequence and likely pattern, not merely the count observed.
Report uncertainty honestly
Describe the sample, method, defect definitions and limitations. Avoid claiming a precise batch-wide quality level unless the design supports that inference. Distinguish an exploratory review from a formal acceptance sampling plan.
Use findings to improve the process and then check whether the change worked. Repeated inspection without cause analysis creates a permanent correction loop. Sampling is most useful when it helps the team identify where errors arise and choose proportionate controls for future work.
Respond to a critical defect in a sample
Suppose a reviewer finds that one record is linked to the wrong account. Even if every other sampled field is accurate, the consequence may justify holding related work and investigating the cause. An average quality score should not decide the response by itself.
Identify whether the problem could affect a source batch, transformation rule or operator sequence. Expand the review in a reasoned way and record why that scope was chosen. Avoid assuming that the single observed instance is either isolated or universal without further evidence.
After correction, validate the affected records and check the control intended to prevent recurrence. Explain the original sample's limits in the report. If a formal acceptance decision requires a statistical plan, obtain appropriate expertise rather than improvising confidence from a convenient sample. Sampling is a tool for informed judgement. Its value depends on defect definitions, selection and response rules that reflect the actual use and consequences of the data.
