Five statistical problems that get nursing papers rejected

You have a finished manuscript, eight months of work on your ward behind it and a list of journals ranked in case the first one says no. You also have a suspicion you have not quite put into words: that the numbers are the weak part. You measured before and after, there was nobody to compare against, the patients are whoever came through your ward, and you ran the analysis with what you remembered from biostatistics plus a couple of tutorials. The decision in front of you is not whether the work is worth anything — it is, often more than some of what does get published — but what can still be changed and what cannot, because that is what decides whether next week goes into redoing tables or into rewriting the discussion. What follows are the five statistical reasons a nursing paper gets rejected, ordered by how much each one costs to deal with.

The pre-post design with nothing to compare against

It is the commonest design in nursing because it is the one clinical practice allows: you train the team, you change a protocol, you measure before and you measure after. The problem is not that it is a bad design, it is that it isolates nothing. Between the baseline and the six-month measurement came your intervention, yes, and also staff turnover, the hospital's safety campaign, the September change of residents, and the fact that whoever scores very badly on a first measurement tends to score somewhat better on the second even if you do nothing at all. That last one has a name — regression to the mean — and it is the alternative explanation that most often ends up written into a reviewer's report.

If there is still time, there are ways out that do not cost a randomized trial: a contemporaneous comparison ward even if you did not allocate it at random, an interrupted time series if you have enough points before and after (a series, not two measurements), or a stepped rollout across wards. For that family of non-randomized evaluations there is also a reporting guideline of its own, the TREND statement, created by the CDC's HIV Prevention Research Synthesis project, with a 22-item checklist (checked in September 2026). Attaching it does not fix the design, but it tells the editor you know which category you are playing in.

And if a comparison group is no longer possible, which is the usual situation once the work is done, what is left is the verb. «The intervention reduced anxiety» is not supported by your design. «After the intervention, the mean score dropped 4.2 points, 95% CI [1.8, 6.6]» is supported, and the discussion then lists the alternative explanations one by one instead of waiting for somebody else to list them. An honest manuscript about a weak design has somewhere to go; one that pretends the design is strong does not, and gets rejected in a bad tone on top of it.

One ward, 41 patients and a conclusion about the profession

The problem is almost never n. Nobody rejects a paper for having 41 patients; they reject it because the abstract says «in hospitalized patients» and the data say «on the trauma ward of one public tertiary hospital, morning shift, between March and October». The distance between those two sentences is what the reviewer flags, and it flags itself: read the title, then look at Table 1.

There are also two technical consequences worth locating. The first is power: with 41 cases, a non-significant result does not say the intervention does not work, it says your study could not tell. The confidence interval shows which effects are still compatible with your data, and it is usually an uncomfortably wide range that includes clinically relevant values; that gets written down, not hidden, and how to write it is in what to do when your results are not significant. The second is dependence between observations: if patients were cared for by five different professionals, or if you measured the same patient several times, your data have structure and a flat analysis swallows it without warning you.

The way out is not getting more patients, which at this stage you are not going to get. It is matching the claim to the sampling, giving intervals instead of verdicts, and describing your ward in enough detail — nurse-to-patient ratio, case mix, recording system, hospital size — for a reader to decide whether it resembles theirs. And sometimes it is changing format: the same work that does not fit as an original article fits as a brief report or as an improvement report, with faster review and without pretending to be something it is not.

The validated scale you never checked in your own sample

The pattern is recognisable at a glance: «scale X was used, validated in the Spanish population, with a Cronbach's alpha of .89». That .89 belongs to the validation study, with its sample, its language, its context and sometimes its decade. Reliability is not a property of the questionnaire that travels attached to it like the number of items: it is a property of the scores you obtained, in the people you measured, under your administration conditions.

This is not one reviewer's pet peeve, it is written into the standards. The APA quantitative reporting standards (JARS-Quant, Appelbaum et al., 2018) ask you to «estimate and report values of reliability coefficients for the scores analyzed (i.e., the researcher's sample), if possible», and add that if you report coefficients from other sources, these should be directly compared with the characteristics of your own sample; the document is explicit that the term reliability refers to test scores in particular samples and not to tests or testing instruments (checked in September 2026). In health sciences, the equivalent reference for measurement properties is COSMIN (COnsensus-based Standards for the selection of health Measurement INstruments), focused on health measurement instruments and in particular on patient-reported outcome measures (checked in September 2026).

In practice it is ten minutes: you compute the coefficient in your own data, you put it in the methods next to the original one and, if it has dropped, you say so. An alpha falling from .89 to .62 in your population is not a mistake of yours to be covered up; it is information about the instrument in that context and it usually yields one of the more interesting discussion paragraphs in the manuscript. Which coefficient applies and how it is written up is in how to report Cronbach's alpha in APA 7.

The quality improvement project presented as research

This is the fastest rejection of the five, and it often comes from the editor without ever reaching review. An improvement project — rolling out a checklist, lowering the fall rate, shortening time to first mobilization — is an intervention on a system: it runs in cycles, it is adjusted as it goes, and its data were collected to manage a service, not to test a hypothesis. Forced into the mold of an experimental study, all of its virtues look like defects: there is no prior hypothesis, the protocol changed halfway, the measurement was taken when it could be taken.

The specific guideline exists and editors know it: SQUIRE 2.0 (Standards for QUality Improvement Reporting Excellence), published in 2015 simultaneously in seven journals and cataloged by the EQUATOR Network as the reporting guideline for quality improvement in health care (checked in September 2026). Changing the label changes the structure of the methods section, changes the journal you submit to, and changes both what is required of you and what is allowed: under SQUIRE, a protocol that evolved in cycles stops being a defect and becomes something you describe.

There is a side effect that gets forgotten and arrives late and expensive: the ethics committee. If the work was done as service improvement and you now present it as research on patients, the journal will ask for the approval number and its date, and that conversation with your committee can take months or end with the conclusion that it cannot be published as it stands. It is worth having it before you write, not once the journal's email has arrived. And it is worth saying the obvious: calling it quality improvement does not demote the work, it places it.

The effect that appears nowhere

The result says «the experimental group improved significantly (p = .03)» and the sentence ends there. It does not say how much, in what units, or with what precision. In nursing this weighs more than in other disciplines because the scales have no units that mean anything on their own: dropping 4.2 points on an anxiety scale running from 0 to 60 tells nobody anything, not even you, until you put it next to something.

Here too the standard is explicit. JARS-Quant asks, in the results section, for «effect-size estimates and confidence intervals on those estimates that correspond to each inferential test conducted, when possible» (checked in September 2026). That means a d, a g, a mean difference in the original unit, an odds ratio, whichever fits, with its interval, and for every test you ran, not only the one that came out well. If you present six tests and three effect sizes, the reviewer's question will not be technical, it will be about why you chose those three. How the commonest one is computed and written is in how to report Cohen's d in APA 7.

And then there is the step almost nobody takes, which is precisely the one that turns a result into a nursing finding: comparing your effect against the minimal clinically important difference for that instrument, where one has been published. A statistically significant change that falls below that threshold is still a result, but it is a different result, and saying it before the reviewer does puts you on the good side of the discussion.

Which of the five can be fixed by writing

Split the list into two columns and next week plans itself. In the homework column go two things: reliability in your sample and effect sizes with their intervals. They come out of data you already have, they need no new collection and no permissions, and they close in one afternoon with the file open. In the limits column go the other three: the missing comparison group, the single-ward sample and the nature of the project. Those are not corrected; they are declared, and declaring them means lowering every claim to what the data support and choosing the journal that publishes that kind of work.

Order matters because every round of review is measured in months, not weeks. If you submit with the two pieces of homework undone, the first letter will most likely ask for them — they are the easiest things to see from outside — and you will have spent those months on something that was sitting on your own computer. And if you have not declared the limits either, the letter will ask for both at once, which is when an editor starts calculating whether a second round with you is worth it compared with the next manuscript in the queue.

None of this is an exotic analysis or a high-impact-journal demand: it is the methodological part of the craft, the same part that holds up the question and the search. If you are at the earlier stage — framing the question, searching the evidence, placing your study on a level — that is in the guide to evidence-based nursing research.

What to check before you hit submit

Design and verbs. Look in the abstract for the word that describes what you actually did — one-group pre-post, quasi-experimental, cross-sectional, interrupted time series — and if it only appears on page seven, move it up. Then look for the verbs: «reduced», «improved», «caused», «was due to». Each one is a causal claim and it has to be supported by the design, not by the intention you had when you did the work.

Scales and effects. For every instrument, a reliability coefficient computed in your sample, not the one from the validation paper. For every p, an effect size with its interval. It is a boring count to do, which is why it is the one most often left undone, and also why it is the first thing spotted by anyone who reviews manuscripts regularly.

Reporting guideline. Decide which one applies and attach the completed checklist: CONSORT if you randomized, STROBE if it is observational — cohort, case-control or cross-sectional, published in 2007 simultaneously across several journals — TREND if you evaluated an intervention without randomizing, and SQUIRE 2.0 if it is quality improvement. All four are cataloged by the EQUATOR Network (checked in September 2026), so there is no guessing about which one exists.

The first two blocks can be checked by anyone with the manuscript printed out and an hour to spare. What cannot be seen by eye is the fine count: whether each p matches its degrees of freedom, whether each effect size matches the F or the t you reported yourself, whether the references you cite exist as you cite them. That is what the Previum sample report does: it recomputes the values from your own numbers with code, not with a model, and checks the references one by one against eight databases. It is not peer review: it is your own assessment of your own manuscript before you send it, and what it still cannot do for you is decide whether your ward resembles anybody else's.

Several of these problems never reach a reviewer: they end in a desk rejection, with the editor returning the nursing paper in days. Before you submit, run a desk-reject check on your manuscript.

Frequently asked questions

Can a pre-post study with no control group be published in nursing?

Yes, and it is published every month. What changes is not the possibility, it is the journal and the claim: with that design you cannot write that the intervention caused the change, but you can describe the observed change, give its interval and discuss the alternative explanations seriously. Name the design in the abstract, not on page seven, and attach whichever reporting guideline applies. The rejection comes when the text promises a trial and the methods deliver something else.

I have 38 patients from a single ward. Is it worth submitting?

It depends on what you want to claim, not on the 38. If the conclusion is about what happened on your ward, with the context described and the intervals in place, it is publishable and useful to anyone with a similar ward. If the conclusion is about «hospitalized patients», there is too much ambition and not enough sample. Read your own conclusion before submitting and ask yourself who it actually applies to.

The scale is already validated. Do I have to compute alpha again?

Yes, in your sample. Reliability refers to the scores obtained in a particular sample, not to the questionnaire, and that is how the APA quantitative reporting standards word it. Compute the coefficient in your own data, put it next to the one from the validation study and comment on the difference if there is one. It takes ten minutes and removes one of the most automatic reviewer comments there is.

My work is a quality improvement project. Does that count as research?

It is not a yes or a no, it is a label you choose and it drags consequences behind it. If you present it as quality improvement, you have SQUIRE 2.0 as a reporting guideline, journals that publish exactly that, and a coherent story. If you present it as research on patients, you will be asked for ethics committee approval dated before data collection, and that is where a lot of good projects get stuck. Decide the label before writing the methods, not afterwards.

It did not come out significant. Do I bin it?

No, but change what you write. With a small sample, a non-significant test does not say there is no effect, it says your study could not see it; the confidence interval shows which effects remain compatible with your data, and that sentence is perfectly publishable. What does not fly is hunting through subgroups until one lights up and then telling the story as if it had been planned from the start.

Is this something a tool catches, or do I need a person?

It splits. The counting — whether each p matches its degrees of freedom, whether effect sizes match the F or the t, whether the references exist — is done better by a machine, because it does not get tired on page 20. Deciding whether your design supports the verb you wrote, or whether the work should be reframed as an improvement report, is judgment and it takes a person who has seen many cases. Start with the counting, which is cheap, and save the judgment for what is left.

Keep the split: reliability in your sample and effect sizes are homework; the missing comparison group and the single-ward sample are limits to be declared, not defects to be covered. Do it in that order, attach whichever reporting guideline applies and bring the verbs down to what the design supports. If the manuscript is already written and you cannot tell which of the five is going to cost you the rejection, that is exactly what we look at in the nursing service.