Research Integrity for Surgeons

Ask a room of surgeons about research misconduct and everyone pictures the same thing: an invented dataset, a fabricated trial, a career ending in disgrace. That happens, and it is rare.

The far more common problem is quieter and nearly invisible from the inside. It is the outcome that was measured but not reported because it was not significant. The hypothesis written after the results were seen. The subgroup analysis that was the fourth one tried. The paper split into three because three publications count for more than one. Almost nobody doing these things believes they are behaving badly — and collectively they do more damage to the orthopaedic evidence base than fraud ever has.

This is the hub for our research series. It covers where integrity actually fails, and the practices that prevent it.

The spectrum

Integrity failures are not binary. They sit on a continuum, and the middle is where the volume is.

Misconduct is conventionally defined as fabrication, falsification and plagiarism — inventing data, altering it, or taking someone else's work. Those are career-ending and they are also, mercifully, uncommon.

The middle column is the one that should worry you, because it describes behaviour that is normal in many departments.

The practices worth naming

Selective outcome reporting. You measured eight outcomes. Two reached significance. The paper reports those two. Nothing was invented — but the reader cannot know what the other six showed, and the literature now contains a distorted signal. This is the single most damaging QRP in surgical research.

HARKing — Hypothesising After the Results are Known. Presenting a finding you discovered in the data as though it was what you set out to test. It converts an exploratory observation into an apparently confirmatory result, and it inflates the apparent strength of the evidence.

P-hacking. Trying analyses until one is significant — different cut-offs, different covariates, excluding outliers, testing subgroups. Each step may be individually defensible; the sequence is not, and it is rarely reported.

Salami slicing. Splitting one study into the smallest publishable units. It inflates a CV, fragments the evidence, and produces apparently independent papers that are the same patients counted repeatedly in later meta-analyses.

Gift and ghost authorship. Adding a department head who contributed nothing; omitting the trainee or statistician who did the work. See ICMJE Authorship and Disputes.

Undisclosed conflicts of interest. Especially relevant in orthopaedics, where industry relationships are legitimate, common and consequential — see Conflicts of Interest and Industry Relationships.

Why orthopaedics is exposed

Our specialty has features that make integrity harder rather than easier.

Device and implant research is close to industry. Funding, consultancy and royalty relationships are normal and often entirely proper — but they create pressure, and studies of a device by those with an interest in it are systematically more favourable. Disclosure is the minimum, not the whole answer.

Small single-centre series dominate. Underpowered studies with flexible outcomes are the ideal conditions for a spuriously significant finding.

Outcome measures are numerous and interchangeable. With a dozen validated scores available for the same joint, choosing the one that moved after the fact is easy and nearly undetectable. See PROMs Explained.

Registry data is powerful and easily over-interpreted. Large numbers produce statistical significance on effects too small to matter clinically, and observational confounding does not disappear because n is large.

Follow-up is long. Ten-year survivorship studies invite selective reporting of the analysis point that looks best.

The incentive problem, honestly

None of this is fixed by telling people to be more honest. The behaviour is a rational response to how surgeons are assessed.

Publications count for training applications, ARCP, fellowships and consultant interviews. Nothing counts negative or null results. Nobody has ever been appointed for a well-conducted study that found no difference. A trainee with three thin papers scores above one with a single rigorous negative study.

Being clear-eyed about that matters, because it tells you where the pressure will come from and when you are most likely to rationalise something. The moments of risk are predictable: near a deadline, when the primary outcome disappoints, and when someone senior suggests the data "probably shows" something.

What good practice actually looks like

Concrete, and mostly decided before you collect any data.

Register prospectively. Trials in a registry; increasingly, observational studies too. Registration is what makes selective reporting visible, and it protects you as much as the reader.

Write the protocol first, including the primary outcome, the analysis plan, and how you will handle missing data. A pre-specified analysis plan is the single most effective guard against p-hacking, because it removes the flexibility that makes it possible.

Name one primary outcome. One. Everything else is secondary and labelled as such.

Report everything you measured, including what did not reach significance. Null results are information.

Follow the reporting guideline for your design from the start — see CONSORT, PRISMA, STROBE.

Declare interests fully, including your department's and your co-authors'.

Involve a statistician early — before data collection, not after.

Keep your raw data, organised, with documented cleaning steps, for years.

Label exploratory work as exploratory. There is nothing wrong with hypothesis generation, provided you say that is what it is.

Correcting the record

You will, at some point, discover an error in something you have published. What you do next is the whole of your integrity.

• A correction for an error that does not affect the conclusions.
• A retraction where the conclusions no longer stand — whether from error or misconduct.

A self-reported correction costs you very little. A concealed one, later discovered by someone else, is a different matter entirely. Journals and readers are markedly more forgiving of authors who raise their own errors — see Retractions and What They Teach Us.

If you suspect misconduct

Difficult, and more common than people admit.

• Do not investigate it yourself. As with safeguarding and serious incidents, your role is to raise it, not to prove it.
• Do not confront the person. It rarely goes well and can prompt destruction of records.
• Raise it with the research integrity office at your institution — nearly every university and large hospital has one, or a named research governance lead.
• Document what you observed, factually, with dates.
• Do not put it in writing to third parties before taking advice; allegations carry legal risk.
• Expect it to be hard. Whistleblowers in research are frequently treated badly, and going through your institution's formal route is what protects you.

If you are a trainee and the person concerned is your supervisor, take advice first — from your indemnity provider, your association, or an independent senior colleague outside the department.

The rest of this series

• ICMJE Authorship and Disputes — who earns a place, and how to resolve it when they disagree
• How to Spot a Predatory Journal
• CONSORT, PRISMA, STROBE — which checklist, and when
• Preprints in Orthopaedic Research
• Open Access and APCs
• Retractions and What They Teach Us
• PROMs Explained
• How Orthopaedic Care Is Funded
• AI-Assisted Writing and Journal Policy

The summary

Fraud is rare and it is not your risk. Your risk is the ordinary, unremarkable set of practices that make a finding look stronger than it is — reporting what worked, framing an unexpected result as the hypothesis, splitting one study into three.

The defences are almost entirely front-loaded: register it, write the protocol, name one primary outcome, follow the reporting guideline, involve the statistician before you start. Do that and most of the ways to go wrong are closed off before you have collected a single data point.

Everything we operate from — every implant choice, every guideline, every consent conversation about risk — rests on this literature being roughly true. That is not an abstract argument. It is why it matters that your paper reports the six outcomes that did not reach significance.