You Have to Publish and You Hate Research: The Minimum Viable Paper

Your program wants a scholarly project. Your portfolio has an empty box before the next review. Your college wants a research output before fellowship, and you have no interest whatsoever in research. That is not laziness and it does not need dressing up: you are good at your job, your hours go to patients, and the paper is an administrative requirement somebody else decided for you. The practical problem is not motivation, it is that nobody has told you where your part starts and stops. Without that line there are only two bad exits: hand the whole thing to someone else and put your name on work you cannot defend, or try to learn statistics in a month you do not have. There is a third, and it starts by naming the three decisions that are yours.

A minimum viable paper is not a bad paper

Minimum viable here means the smallest paper that honestly answers a real question using data your service already generates. A well described case series, a retrospective comparison of two protocols, a validation of a questionnaire in your population, an agreement study between two raters. None of those needs funding, a trial, or two years. They need a tight question, a design that holds up, and a discussion that promises no more than it shows.

What it is not: slicing one study into four papers, digging through the dataset until something falls below .05 and writing the hypothesis afterwards, or buying a byline on a manuscript you have not read. That last one has its own market, and it pays to know how a paper mill differs from a statistical consultant: the difference is not the price or who types, it is who makes the decisions and who can defend them in front of a reviewer.

The advantage of the minimum viable route is that it concentrates your hours on the part nobody else can do. With a clean data package, six to eight hours of your own time spread over several weeks is usually enough to get a paper ready for an indexed journal. The rest is execution and waiting, and neither of those has to come out of your diary.

The question is yours, and no statistician can write it for you

A statistician knows which test belongs to which design. What a statistician cannot know is which uncertainty on your ward costs decisions every week. That asymmetry is why fully outsourced papers read like nothing in particular: someone outside ended up picking the question the data could answer rather than the one worth answering. A weak question is not repaired later by a better model.

A question earns a paper if it passes two filters. First, the answer would change something you do; if you picture the result going either way and you shrug at both, it is not a question, it is an exercise. Second, it is not already settled in the guideline you open every week. Those two filters take five minutes and kill most of the ideas you will have, which is exactly their job.

Whatever survives gets written in one line with four slots: in [which patients], is [which comparison] associated with [which outcome, measured how] at [what time point]? An example from a community mental health clinic: in adults referred with moderate depression between 2023 and 2025, do those started on a stepped-care pathway differ from those started on standard follow-up in PHQ-9 score at three months? Writing that sentence takes twenty minutes and it determines everything downstream.

Notice what is missing from it: the name of a statistical test. The test follows from the design and the type of variable, and that deduction is delegable; it is already worked out in the guide on how to choose a statistical test. Arriving with the test decided and the question blurry is the reverse order, and it is what produces flawless analyses that answer nothing.

The design is yours as well, and that is where papers die

Clinical papers are almost never rejected over the test. They are rejected over a comparison that does not hold: the patients on the second pathway were the more unwell ones, precisely because that is why they were put on it. It is called confounding by indication, and no analysis repairs it if severity was never recorded. You are the one who knows that trap exists, not the person running the numbers: it is clinical knowledge wearing a methodological costume.

So the design gets written before anyone looks at the data, on a single page: who is in, who is out, which date window, which primary outcome (one, not five) and what happens to missing values. That page is the only thing separating a confirmatory analysis from a fishing trip, and it is the first thing anyone helping you will ask for. The full template is in the data analysis plan, and for a clinical paper it fits on one page. Prespecify it, date it, and keep it.

One more check that takes five minutes and saves months. With 60 patients per arm, α = .05 two-tailed and 80% power, your study can detect from d = 0.52 upwards; with a standard deviation of 5.6 points on the PHQ-9, that is 2.9 points. With 20 per arm the smallest detectable effect rises to d = 0.91, which is 5.1 points. That matters, because the change usually treated as clinically meaningful on the PHQ-9 is around 5 points: the first design can see something a patient would notice, the second can only see what is already unmissable. Run the numbers yourself in the clinical sample size calculator before you commit to anything.

If the number comes out badly you can still change the question, widen the date window, bring in a second site, or turn the work into something descriptive and honest. Once the data are collected that conversation no longer exists, and all that is left is writing carefully about what the design was able to see.

What the number means in clinic: nobody else can write that

Say the analysis comes back like this: the stepped-care group scores 1.4 points lower on the PHQ-9, with a standard deviation of 5.6 and 150 patients per group. In figures, d = 0.25, t(298) = 2.17, p = .031, 95% CI [0.02, 0.48]. Statistically significant, and anyone who can drive the software gets that far. The question only you can answer is the next one: does 1.4 points on that scale change anyone's week?

The conventional answer is no. The threshold usually cited for a meaningful change on the PHQ-9 sits around 5 points, and 1.4 is under a third of it. The interval does not rescue the finding either: its optimistic end, d = 0.48, is 2.7 points, still well short. You have an effect that is real, reasonably precisely estimated, and clinically irrelevant. All three at once, with no contradiction.

With that in hand the discussion sentence writes itself and survives review: the difference was statistically significant but smaller than the change usually regarded as clinically meaningful on this measure, so it does not support redesigning the pathway on its own. A reviewer cannot ask for more, and a reader knows what to do with your paper. Nobody outside the service would have written that sentence, because it takes knowing what one point on that scale means to the person reporting it.

The mirror case turns up just as often and confuses people more. A difference of 5.5 points, above the threshold, with 8 patients per group: d = 0.98, t(14) = 1.96, p = .070, 95% CI [-0.06, 2.02]. The correct reading is not that there is no effect, it is that this design could never have seen one, and you defend it by showing the interval next to the smallest detectable effect you calculated before you started. Same calculation as the previous section, used here to avoid binning a result that deserves a bigger study.

What you can hand over entirely, and why it is not cheating

The whole execution is delegable: cleaning the dataset, checking assumptions, running the models, computing effect sizes with their confidence intervals, building the APA 7 tables, applying the journal's format, tidying the reference list and drafting the technical half of the reviewer response. None of that requires knowing your patients, and all of it would cost you weeks learning something you will not touch again for three years.

The ethical test is not who touched the keyboard, it is the ICMJE authorship criteria: substantial contribution to the design or to the analysis and interpretation, involvement in drafting or critical revision, approval of the final version, and accountability for the work. Someone who designs the analysis with you, runs it and revises the draft usually meets all four and belongs on the author list; someone who only formats tables belongs in the acknowledgments. What is never acceptable is a contribution that appears nowhere at all. The full argument is in whether hiring a statistician is ethical.

The line you must not cross has a thirty-second test. If you cannot explain to a first-year trainee, in one sentence each, what you compared, why that test fits that design, and what the number means in clinic, you are not in a position to sign. You do not need to derive the formula or write the code. You need to hold the paper up in a departmental meeting and answer a reviewer without forwarding the email to someone else.

That test also protects you from something specific and expensive: the reviewer who asks for a sensitivity analysis, an adjustment for covariates or a mixed model three months later, when the whole thing has faded from memory. If you understood the division of labor from the start, you know who to call and what to ask for; that situation is worked through in what to do when a reviewer asks for an analysis you cannot run.

The handover pack, and the hour that saves you three weeks

The difference between a paper that goes out in six weeks and one that dies in March is almost always an hour of preparation nobody did. Whoever runs the analysis cannot guess what the column VAR00012 holds, or whether 999 is a missing value or a very unwell patient, so they ask, they wait, they ask again, and the weeks disappear there. This is what you settle before the first meeting.

The dataset and the dictionary. One row per patient, one column per variable, no colored cells, no total rows, no comments inside the cells, no two facts in the same column. Alongside it, a one-page document giving each column's name, what it measures, in which unit, what values are allowed and how a missing value is coded. It is the single biggest time-saver and almost nobody brings it.

The question, the primary outcome and the decision. The one-line sentence with its four slots, the primary outcome flagged as such (one, not five candidates) and what you would do differently in clinic depending on the answer. Add the threshold of clinical relevance you intend to use and where it comes from. You hold that number, and it is what turns a results table into a conclusion.

The journal, and what you demand back. Name the target journal, because table format, word limits and the reporting guideline (STROBE for observational work, CONSORT for trials) change the job from day one. And always ask for the same things back: the assumptions checked, effect sizes with intervals, the exact APA 7 sentences for the results section, and the output or code that reproduces every number. With those, you can defend a paper you did not run yourself.

Frequently asked questions

Can you publish a paper if you hate research?

Yes, and in clinical work it is the norm. The route is the minimum viable paper: a tight question answered with data your service already generates, a design written on one page before anyone looks at the numbers, and the execution handed over. With a clean data package your own time is around six to eight hours spread over several weeks. What does not work is trying to learn statistics and publish at the same time.

Is it ethical to pay someone to run the statistics for my paper?

Yes, provided the contribution is recorded. The standard is the ICMJE one: whoever designs the analysis with you, runs it and revises the draft meets all four authorship criteria and belongs on the byline; whoever only formats tables belongs in the acknowledgments. What is not acceptable is someone else choosing the question, interpreting for you, and writing a manuscript you sign without being able to defend it. That is the line between a consultant and a paper mill.

What should I bring to a statistician so I do not waste time?

Five things: the dataset with one row per patient and one column per variable; a one-page dictionary explaining each column, its unit and the missing-value code; the question in a single sentence; the primary outcome flagged (only one); and the target journal. Add the threshold of clinical relevance you plan to use. That hour of preparation decides whether the paper takes six weeks or six months.

How many patients do I need to publish a clinical study?

It depends on the smallest difference you want to be able to see, not on a magic number. With 60 per arm, α = .05 two-tailed and 80% power, you detect from d = 0.52 upwards, which with a standard deviation of 5.6 points on the PHQ-9 is 2.9 points. With 20 per arm the smallest detectable effect rises to d = 0.91, or 5.1 points, roughly the change usually called clinically meaningful. Do that calculation before you collect, not after.

What is the minimum I have to understand about my own analysis?

Three sentences: what you compared with what, why that test fits that design, and what the number means in clinic. If you can say them in a departmental meeting without reading, you can sign the paper. You do not need to derive the formula or write the code; you need to answer a reviewer without forwarding the email to somebody else.

Is a retrospective study using routine clinical records good enough?

It is, and it is the fastest route to an honest publication, provided you declare the design for what it is and deal with confounding by indication. If the patients on the second pathway were the more unwell ones, you need severity recorded or the paper falls apart in review. State the date window, the inclusion criteria and the primary outcome before looking at the data, and write it up against STROBE.

Keep the division of labor: the question, the design and the clinical meaning are yours and are never handed over; the execution, the tables and the formatting belong to whoever does them quickly and well. If the data are already collected and a deadline is closing in, or you have a question and no idea whether it survives a design, that is precisely what I do in my statistical consulting: you bring the clinical judgment, I bring the analysis, and you sign a paper you can defend in any meeting.