03 · What You Need to Know
What to Do After Discovering a Serious Research Error
First, preserve the original research record
Your first impulse may be to repair the spreadsheet, replace the figure, rerun the code, or delete the incorrect file. Make a corrected version if necessary, but do not destroy the original evidence of what happened.
Preserve relevant raw data, processed datasets, analysis scripts, laboratory records, notebooks, figure files, correspondence, version histories, protocol documents, and other records that may help reconstruct the error.
This serves two purposes. It allows you and others to understand how the problem arose, and it prevents a well-intentioned correction from looking like an undocumented alteration of the research record.
Use versioning or another appropriate method so that the original and corrected materials can be distinguished.
Verify that the apparent error is actually an error
Before sending alarming messages to every co-author and editor you have ever met, reproduce the problem.
Check the relevant data and code. Confirm that you are using the correct file version. Examine whether a result differs because of a genuine mistake, a legitimate methodological choice, a software version difference, or a misunderstanding about how the original analysis was conducted.
When the issue is technically complex, ask an appropriate collaborator or independent expert to verify it. A second set of eyes can be especially useful when the same person who created the analysis is trying to diagnose it.
Verification should be prompt, but it should not be theatrical. The objective is to establish the facts accurately.
Determine exactly what went wrong
"There is an error" is only the beginning of the diagnosis.
Was a variable miscoded? Were group labels reversed? Was the wrong dataset used? Did a formula omit several rows? Was an image placed in the wrong panel? Were observations incorrectly excluded? Did a transcription mistake affect one table? Was a methodological assumption violated?
Write down the specific error and reconstruct how it entered the workflow as far as the evidence allows.
This distinction matters because different errors have different consequences and may require different corrections.
Assess the scope rather than fixing only the first visible symptom
A single discovered error may be isolated, or it may indicate a systematic problem.
If one spreadsheet formula references the wrong cells, inspect similar formulas. If one sample is mislabeled, determine whether the labeling procedure could have affected others. If one figure uses the wrong image, review the figure-assembly workflow. If one analytical script contains a coding error, identify every output generated from that code.
ORI's guidance on honest error recommends considering the scope of potentially questioned conduct and examining the research record when determining what happened. Even when misconduct is not suspected, the same basic logic is useful for diagnosing an error accurately.
Recalculate the research using the corrected information
You cannot determine the seriousness of an error from its appearance alone.
A typo that looks embarrassing may have no effect on the analysis. A one-character coding mistake may reverse the principal conclusion.
Correct the error in a documented copy of the research workflow and rerun every affected analysis. Recreate affected tables, figures, estimates, confidence intervals, tests, models, or qualitative classifications as appropriate.
Then compare the corrected research with what was originally reported.
Ask what changes, not merely whether something changes
Assess the error at several levels.
| Level affected |
Question to ask |
| Underlying data |
Are the original observations themselves trustworthy, or does the error affect data provenance or validity? |
| Analysis |
Which calculations, classifications, models, or interpretations change after correction? |
| Results |
Do reported values, tables, figures, effect estimates, themes, or statistical conclusions change? |
| Interpretation |
Does the corrected evidence alter how the findings should be understood? |
| Main conclusion |
Does the study's central claim remain supported? |
| Downstream outputs |
Have the incorrect results already appeared in manuscripts, presentations, reports, datasets, repositories, theses, policy documents, or publications? |
This impact assessment should determine the response. The size of the original mistake is less important than what it does to the reliability of the research.
Do not quietly repair a consequential error
If an error materially affects research that others have seen, relied upon, reviewed, submitted, funded, or published, silently changing your private files is not enough.
The people responsible for the affected research need accurate information. Depending on the stage and context, that may include co-authors, supervisors, research leaders, institutional officials, ethics or regulatory bodies, funders, data repositories, conference organizers, journal editors, or other relevant parties.
Who must be informed depends on what the error affects and on applicable policies. Avoid broadcasting confidential information unnecessarily, but do not use confidentiality as a pretext for concealing a material problem.
If the manuscript has not been submitted, correct it before submission
An error discovered during ongoing research is usually easiest to manage.
Preserve the original records, correct the workflow, rerun the affected work, document what changed, and update the manuscript or report before it leaves the team.
If the error reveals a broader procedural weakness, fix that too. Correcting one cell while leaving the faulty process untouched merely schedules the sequel.
If the manuscript is under review, tell the journal when the error is material
If a submitted manuscript contains a consequential error, contact the journal through the appropriate author, usually the corresponding author, and explain the problem accurately.
The editor can determine whether revised files can be considered, whether review should pause, or whether withdrawal and later resubmission is more appropriate. Journal procedures differ, so do not assume that one response applies everywhere.
Continuing the review while knowingly allowing reviewers to evaluate materially incorrect results is difficult to reconcile with responsible research conduct.
If the work is published, the journal needs enough information to assess the correction
For a published article, contact the journal according to its correction policy. Explain what is wrong, how the error arose as far as you can establish, which parts of the article are affected, and what the corrected analysis shows.
Do not decide unilaterally that a correction, retraction, or another notice is required and then demand that exact label. Provide the evidence needed for the editor to apply the journal's policy.
ICMJE distinguishes among errors that can be corrected while leaving the paper's overall conclusions intact and errors serious enough to invalidate the principal results and conclusions. It also recognizes that honest errors can sometimes require retraction with republication when substantial changes are necessary but the corrected underlying science remains valid.
Retraction is not a confession of misconduct
Researchers sometimes resist necessary corrections because they fear that a retraction will be interpreted as evidence of fraud.
That fear is understandable, but the premise is wrong. Publications can become unreliable because of honest error. Editorial action concerns the reliability of the published record, while a misconduct finding concerns researcher conduct under a separate standard.
A researcher can therefore make an honest mistake serious enough to invalidate a publication without having committed misconduct.
This is the same reason a researcher can be cleared of misconduct while the research remains unreliable.
Do not label your own error misconduct before the facts are established
Discovering an alarming discrepancy can make researchers jump directly to moral conclusions about themselves or their collaborators.
Start with the evidence.
The current U.S. Public Health Service definition expressly excludes honest error from research misconduct. Whether conduct constitutes misconduct depends on the applicable definition, circumstances, state of mind, and evidence.
A serious mistake can still require major corrective action without becoming research misconduct rather than honest error.
Watch Out
Do not destroy, overwrite, backdate, or reconstruct records in a way that obscures what originally happened. Correct the research transparently while preserving the information needed to understand the original error.
If the error may involve someone else's misconduct, preserve the issue rather than privately adjudicating it
Sometimes investigating your own error reveals something more troubling. Perhaps a value you thought was a transcription mistake appears to have been deliberately altered by another researcher. Perhaps underlying data cannot be found because records were intentionally replaced.
If credible evidence raises a misconduct concern, preserve the relevant records and use the institution's established research integrity process. Do not secretly interrogate colleagues, alter files, or attempt to conduct a full misconduct investigation yourself.
Your immediate responsibility is to protect the evidence and research record while allowing the appropriate process to determine culpability.
Correct the process as well as the output
Once the immediate research record is addressed, ask why the error survived.
Was there no independent check of analysis code? Were filenames ambiguous? Did multiple researchers overwrite the same dataset? Was there no version control? Did nobody verify figures against source files? Were students inadequately trained? Did the team lack a reproducible workflow?
Some errors are genuinely isolated. Others expose a system that is waiting to produce the same mistake again.
Research integrity therefore requires learning from the error, not merely replacing the incorrect number.
