The Final Checklist

The study is done, the sections are written, the interpretation is calibrated. One task remains, and it's the one most students skip: the consistency pass. Graders and reviewers rarely re-run your analysis, because they usually can't. But they can check quickly whether your own document agrees with itself. Does the sample size in the abstract match the method? Do the degrees of freedom fit the design? Is every figure referenced, every citation listed? An internal contradiction suggests that something deeper may be wrong. It is also the easiest kind of error to catch, and an easy way to lose marks.

Do the numbers match everywhere?

A single result appears in several places: the abstract, the results sentence, maybe a table, maybe the discussion. Every one of those copies must be identical. An abstract that says r = .28 while the Results report r = .31 tells the reader that at least one number was typed from memory, and now they doubt all of them. The same goes for sample sizes, means, and percentages. Pick the true value once, then make every mention agree.

Check your degrees of freedom

Degrees of freedom are a check on your design. For an independent two-group t-test, df = N − 2; for a one-way ANOVA, the denominator df = N − k; for a correlation, df = N − 2. So if you report two groups of 60 but write t(116), the arithmetic doesn't add up: 60 + 60 − 2 = 118, not 116. A grader who knows the formula reads your df straight back to your n, and a mismatch signals either a typo or a misunderstanding. Check that every reported df is consistent with the sample size and design you described.

The cross-reference checks

Three quick checks catch most of the rest. Check that every table and figure is referenced in the text ("as Figure 1 shows…") and that every one you reference actually exists. An unreferenced figure is clutter, and a reference to a missing one confuses the reader. Check that every citation is in the reference list and every reference is cited. A name in the text that's missing from the list, or the reverse, is the classic reference-manager slip. Then check that every statistic is complete: test, degrees of freedom, the statistic's value, an exact p, and an effect size. Floor p at p < .001, and never write p = .000 (see Reporting Statistics in APA Style).

Two of those sweeps are automated now

The p-value check and the mean-plausibility check are mechanical, so software can do them, and journals increasingly run that software on a manuscript before a human reads it. statcheck reads a manuscript, pulls out every result written in APA form, recomputes p from the test statistic and its degrees of freedom, and flags the pairs that disagree. Several psychology journals, Psychological Science among them, now run it as part of peer review. A large survey run with statcheck looked at more than 250,000 p-values in roughly 16,700 articles. Of the articles reporting null-hypothesis tests, 49.6% contained at least one inconsistent p, and 12.9% contained one far enough off to flip the significance decision. Just under a tenth of all reported p-values did not match their own statistic.

You can run that check on yourself without installing anything. Paste a result into the APA results formatter; it recomputes p the same way and says so when the value you typed and the value your statistic implies disagree.

GRIM does the equivalent job for means. When a measure is a whole number for each person (items recalled, a sum of 1-to-7 ratings), the mean of n people has to be some whole number divided by n. Most decimals are therefore unreachable. At n = 20 the possible means step by 1/20, so a reported 4.32 could not have come from that sample: 4.32 × 20 = 86.4, and a total of 86.4 items is impossible. The test works only when the sample is small relative to the number of decimals reported. The mock paper below gives recall means of 28.4 and 24.1 from groups of 60, and both are perfectly reachable (1,704 and 1,446 items), so GRIM has nothing to say about them. That is the usual result. It means the means are possible, not that the check failed to work.

Neither tool says anything about honesty, and both find far more typos than misconduct. Fraud and self-correction covers using them on someone else's paper. On your own paper, they do the arithmetic half of the final pass in a couple of minutes. That leaves your attention for the half no software can do: checking that each sentence you wrote is one your data support.

Find the six inconsistencies

Here is a one-page paper in miniature, with an abstract, method, results, a table, a figure, and references. Six internal inconsistencies are hidden in it: numbers that disagree, degrees of freedom that don't fit, a display that's never referenced, a citation that's missing. Read like a grader and click each problem. Some things you click will turn out to be fine. That's part of the job too.