How to Check Your Data Before Submitting a Paper (2026 Guide)
Before submitting a paper, run four checks on the numbers. Every sample size in the text should match the tables, and percentages should sum correctly. Decimal places should be consistent, and missing data should be accounted for and explained. These are the errors reviewers find first, because they are the easiest to find.
Reviewers arrive with limited time and a reasonable heuristic. If the arithmetic in the tables does not hold, the analysis behind it probably was not checked either. That inference is unfair in individual cases and correct often enough that it shapes the review.
This guide covers why numeric errors dominate reviewer comments and which checks are worth running before submitting a paper. It ends with how to run them without reading every table by hand.
Why reviewers catch numbers, not theory
The N never matches
A study starts with 120 participants. Eight are excluded for incomplete responses, three more fail an attention check, and one condition loses two people to a technical fault. The methods section says 120, the analysis ran on 107, and one table still shows 109 because it was built before the last exclusion.
Nobody did anything wrong here. The count simply changed four times, and the manuscript records different moments in that sequence. A reviewer comparing the methods section against the tables sees a contradiction with no explanation.
The rounding drifts
Statistical output gives many decimal places. Different numbers get copied at different times, some rounded to two places and others to three. Percentages calculated from rounded subtotals sum to 99.8 rather than 100.
Individually these are trivial. Collectively they read as carelessness, and they are exactly what a careful reader notices while skimming a table.
The figure and the text disagree
A chart is exported early, the analysis is refined, and the text is updated. The figure keeps the old numbers. This is common enough that experienced reviewers check figures against prose as a matter of routine.
These three failures share a cause. Nothing went wrong in the analysis; the manuscript simply records several different moments in a process that kept moving. Catching that is a documentation problem, not a statistical one.
What this costs you
A revision round that was avoidable. Numeric inconsistencies rarely cause rejection on their own, but they generate a round of major revisions that adds months.
Reviewer attention spent in the wrong place. Every comment about a mismatched N is a comment not spent on your argument. You get a finite amount of engagement per review.
A correction after publication. The worst version is when nobody catches it before submitting a paper goes all the way through. Post-publication corrections are permanent and public, and they attach to your name rather than to the reviewer who missed them.
The checks worth running
Option 1: Reconcile every sample size
List every N and n that appears anywhere: methods, each table, each figure caption, the abstract. Trace each one to a specific filtering step. Any number you cannot trace is either an error or an undocumented exclusion.
Then check that the exclusions add up: starting sample, minus each documented exclusion, equals the analysed sample. Write that chain into the methods section. Reviewers rarely object to exclusions that are explained; they always object to ones that are not.
Option 2: Check the arithmetic in every table
Percentages within a category should sum to 100, allowing for a documented rounding note. Subgroup counts should sum to the group total. Means of subgroups, weighted by their sizes, should reconcile with the overall mean.
This is dull work and it is where most errors live. It is also the check reviewers are most likely to run themselves, which makes skipping it a poor gamble.
Option 3: Audit missing data and outliers
State how much data is missing, whether it is missing at random, and how you handled it. Listwise deletion, imputation and pairwise analysis produce different sample sizes, which is a frequent source of the mismatched N problem.
Do the same for excluded extreme values. Say how many were removed and on what rule. Consider reporting the key result both with and without them — see our guide to finding outliers without code. A result that survives that comparison is stronger, and one that does not is something you want to discover before a reviewer does.
Where all three hit the same ceiling
Each check is mechanical, and each requires holding the whole manuscript in your head at once. You are comparing a number in the methods against one in Table 3 against one in a figure caption. The document is one you have read so many times that you no longer see it.
That is the actual difficulty. Not that the checks are hard, but that self-checking a document you wrote is unreliable in a way no amount of care fixes.
How to run these checks with Powerdrill Bloom
Step 1: Upload your data
Upload the analysis dataset. Powerdrill Bloom profiles every column on arrival, so missing-value counts and group sizes are visible as facts rather than as things you need to remember.
Step 2: Describe the checks in natural language
Ask for the reconciliation directly. How many complete cases are there per group after applying these exclusions, do the category percentages sum to 100, and which columns have missing values. Then ask the comparison that matters — how the key result changes with and without the excluded cases.
Step 3: Export the chart, report, or deck
Take out a summary table of counts and missing values to check your manuscript against. A written note recording the exclusion chain for the methods section comes out of the same request.
Why this beats a manual read-through
| Manual check | Powerdrill Bloom | |
|---|---|---|
| Reconciling N across sections | Read and compare by hand | Counts returned per filter step |
| Missing-value audit | Column-by-column inspection | Profiled at upload |
| Result with and without exclusions | Rerun the analysis twice | Ask for both |
| After a revision | Repeat the entire pass | Re-ask on the new file |
The last row is the one that decides whether checks actually happen. An audit that takes an afternoon gets run once, before the first submission, and skipped before the resubmission. That is precisely when the numbers have just changed.
A pre-submission checklist
Trace every N to a filtering step. If you cannot say which rule produced a number, do not publish it.
Recompute the table totals. Percentages to 100, subgroups to the group total, and no exceptions without a rounding note stated on the table itself.
Regenerate figures after the final analysis. Never reuse a chart exported before the last revision. Our guide to making charts for a research paper covers the format requirements.
State the missing-data handling explicitly. Method and resulting sample size, in the methods section.
Check the abstract against the results. Abstracts are written early and updated last, so they carry the oldest numbers in the manuscript. They are also the part most readers see, which makes an error there disproportionately visible.
Have someone else read the tables. A colleague who has not seen the data will find in ten minutes what you have missed for weeks. Ask them to check arithmetic rather than argument, since that is the check you cannot perform on your own manuscript.
Conclusion
The checks that prevent revision rounds are unglamorous: reconcile the sample sizes, verify the arithmetic, account for missing data, and regenerate the figures last. None of them requires statistical sophistication, and all of them are easy to skip when a deadline is close.
The reason they get skipped is that they are expensive to repeat, and they need repeating every time the analysis changes. Try Powerdrill Bloom on the dataset before your next submission. See also our AI data cleaning page and the guide to cleaning and deduplicating data.
Frequently asked questions
What do reviewers check first in a paper?
Internal consistency of the numbers. Sample sizes across the methods, tables and abstract, whether percentages sum correctly, and whether figures match the text. These checks are quick and they signal how carefully the rest was prepared.
What should I check before submitting a paper?
State the starting sample, each exclusion with its rule and count, and the final analysed sample, so the chain adds up arithmetically. Explained exclusions are rarely questioned; unexplained changes in sample size always are.
Should I report results with and without outliers?
When the conclusion depends on a small number of extreme values, yes. Reporting both versions is stronger than quietly excluding them, and it prevents a reviewer discovering the sensitivity for you.
Why do my percentages not add up to 100?
Usually rounding each component independently. Either add a note that percentages may not sum to 100 due to rounding, or compute them so they do. Both are acceptable; silence is not.
How much missing data is too much?
There is no universal threshold, and the pattern matters more than the amount. Data missing at random is more tolerable than data missing systematically. State the amount, the pattern and your handling method, and let reviewers judge.