Why Corrections Need a Record
Data preparation sometimes identifies a value that is demonstrably wrong. A verified transcription mistake, for example, may be corrected before analysis. But changing a value without recording what it was, why it changed, and what evidence supports the change makes the prepared data difficult to check. Someone reviewing the analysis cannot tell whether a value was corrected from a source record, changed by a rule, or entered by mistake during preparation.
A documented correction leaves a trace from the value as first received to the value used for analysis. This trace supports review and helps another analyst reproduce the prepared data. It also makes it possible to revisit a decision if new evidence becomes available. The goal is not to make every data value look tidy; it is to ensure that each actual correction has a clear, supported explanation.
What to Record for Each Correction
A useful correction log is specific enough that someone who did not make the change can locate the affected value and understand the decision. A practical record includes the following fields:
- Record identifier: a stable identifier for the row or observation, not simply its current row number if rows might be sorted or filtered.
- Variable: the name of the field that changed.
- Original value: the value as received or first recorded. Preserve it rather than replacing it in the only copy.
- Corrected value: the value used after the documented correction.
- Reason: a concise explanation of the specific error being corrected.
- Evidence or source: the source used to verify the corrected value, such as the original paper form, an instrument record, or a confirmation from the person responsible for the measurement. Include enough detail to locate the evidence when possible.
- Date and responsible person or process: when the change was made and who made it, or which documented preparation process applied it.
These fields answer different questions. The reason explains why a value changed; the source explains what supports the replacement. “Value looked unusual” is not a sufficient reason to correct a value, and “checked” is not a sufficiently clear source description. If the evidence does not determine the right value, document the concern without inventing a replacement.
A correction log can be stored as a separate table linked to the data by a stable record identifier. The prepared data should also retain the original values, for example in a separate original-value field or an unchanged source file. This separation lets an analyst compare the received data with the analysis version and makes it harder for a correction to erase its own history.
Worked Example: Correcting a Verified Decimal-Place Error
Suppose a hypothetical measurement file contains a participant's recorded mass as 84.0 kg. The original measurement form for record P017 clearly reads 8.40 kg, and a second staff member verifies that the data-entry value has one decimal place in the wrong position. Show how to document the correction.
Identify the change. The affected record is P017, and the variable is mass_kg. The original value is 84.0 kg. The verified value is 8.40 kg, which can be stored numerically as 8.4 kg.
State the reason and evidence. The reason is “decimal-place transcription error.” The evidence is “original measurement form for P017, mass entry; reviewed by a second staff member.” The source description should point to where that form can be found in the study records, subject to the relevant privacy and access procedures.
Record the change. A correction-log row might read: record_id = P017; variable = mass_kg; original_value = 84.0; corrected_value = 8.4; reason = decimal-place transcription error; source = original measurement form for P017, verified by a second staff member; date = 2026-04-12; changed_by = data reviewer.
The arithmetic check is explicit: \(84.0 \div 10 = 8.40\), and \(8.40\) kg is represented as \(8.4\) kg when trailing zeros are not retained. The correction is justified by the form, not merely by the fact that 8.4 kg seems more plausible. The original value, 84.0 kg, remains available for audit.
From Log Entry to Reproducible Change
A log is most useful when it makes clear how a prepared value follows from the original data and documented evidence. For a simple correction, this relationship can be written as:
This rule separates two decisions: whether there is evidence for a correction, and what value that evidence supports. It also makes the preparation reproducible. Given the same original data and correction log, another analyst can apply the same documented replacements and compare the result. If a correction cannot be traced to its record, variable, and supporting evidence, the change is not fully reproducible.
Keep a clear distinction between an observed value and a value calculated from observations. If a value is created by a calculation, it is a derived value, not a correction to the original measurement. The calculation and inputs should be documented appropriately; do not label a new calculation as a correction simply because it replaces a field in an analysis file.
Worked Example: Recording a Transcription Correction
In a hypothetical study, record S042 has a laboratory result entered as 31.6. On review, the laboratory instrument’s saved output for the same record shows 13.6. The instrument output is linked to S042 by its record identifier. Assume the review confirms that 31.6 was a transcription error and that both values use the same units.
Compare the values. The value first entered is 31.6; the verified instrument value is 13.6. The numerical difference is \(31.6 - 13.6 = 18.0\) units. This calculation describes the size of the discrepancy; it does not by itself justify changing the value.
Establish the support. The supporting source is the saved instrument output linked to S042. The reason is “transcription discrepancy; entered value does not match instrument output.” The record link and source location allow a reviewer to confirm that the comparison concerns the same observation.
Enter the correction. The log records record_id = S042; variable = lab_result; original_value = 31.6; corrected_value = 13.6; reason = transcription discrepancy; source = instrument output linked to S042; date and reviewer as applicable.
The prepared field may use 13.6, but the received value 31.6 should remain preserved. If the source could not be matched confidently to S042, the evidence would not support this replacement; the discrepancy should instead be flagged for review rather than resolved by guesswork.
When the Evidence Is Incomplete
Not every flagged value can be corrected. A value may appear inconsistent with another field, or a source document may be unreadable or unavailable. These observations justify review, but they do not necessarily establish the true value. Document what was checked and what remains uncertain. Do not silently replace a value with a typical value, an average, or the value that seems most plausible.
If a correction is made in stages, keep enough information to reconstruct the sequence. For example, record the value as received and the final corrected value, and describe any intermediate step that matters to the decision. If multiple sources disagree, document that disagreement rather than presenting one source as decisive without justification. A transparent record can show that a question was investigated even when the answer remains unresolved.
Worked Example: Documenting an Unresolved Discrepancy
A hypothetical data file lists a visit duration of 15 minutes for record V108. A note attached to the record appears to say either 15 or 75 minutes, but the handwriting is unclear, and no electronic time record is available. The value is unusual in context, but there is no verified source that settles the reading.
Do not manufacture a correction. Neither the unusual value nor the ambiguous note proves that 15 is wrong or that 75 is right. There is no supported corrected value to enter.
Document the review. Record V108 and the visit_duration variable as flagged for review; retain 15 as the value received; describe the concern as “handwritten note is ambiguous”; identify the note as the source reviewed; and record the review date and reviewer. State that the discrepancy remains unresolved.
This entry preserves the evidence and the decision not to change the value. If a clear source is later found, a new, dated correction record can cite it and provide the supported replacement. Until then, the log should not imply that an unverified value has been corrected.
Checking the Correction Trail
Before using corrected data, review the log as a set of linked claims: a particular value in a particular record was changed for a stated reason, on the basis of identified evidence. A short audit can check whether those claims are complete and internally consistent.
Confirm that the record identifier and variable identify one specific field in the original data.
Check that both are recorded clearly and that the corrected value in the prepared data matches the log.
Ensure the reason describes the error and the source is specific enough to verify the replacement.
Starting from the retained original value, apply the documented replacement and confirm that the resulting prepared value is the same.
The central check is that the recorded correction can be independently followed. A list of corrected values without original values cannot show what changed. A list of reasons without sources cannot show why the replacements are trustworthy. A source without a record identifier may not establish which observation it supports. Good documentation connects all of these pieces.
Common Mistakes and Why They Matter
- Overwriting the original. Replacing 84.0 with 8.4 in the only copy removes the evidence of what was received. Preserve the original in an unchanged source file or a separate field.
- Using plausibility as proof. A value that seems unlikely may still be valid. Correct it only when evidence supports the replacement; otherwise, document the concern.
- Writing vague reasons. “Fixed” does not explain the error. “Decimal-place transcription error, verified against the original form” states both the issue and the basis for the change.
- Failing to identify the source. “Checked with records” may not let another person locate the evidence. Name the relevant document or system and link it to the observation where possible.
- Confusing corrections with new calculations. A calculated value should be traceable to its inputs and calculation. It is not an observed value that has been corrected.
Check Your Understanding
Use the ideas in this tutorial to answer the following questions.
- What information should a correction record contain so another analyst can locate and review a changed value?
- A value looks implausible, but there is no source record available. Why is that not enough to justify replacing it?
- A form shows 6.20, while the data file contains 62.0. If the form is verified as the source for that record, what should be preserved and what should be documented?
- Why is a correction log row that says only “fixed typo” inadequate?
- A calculated value replaces a field in an analysis file. What distinction should be made between this calculation and a correction to an observed value?