Conflicts of Interest Don't Work the Way You Think

The popular picture of a conflicted study is a researcher at a keyboard, quietly changing a number that didn't say what the sponsor wanted. It is a satisfying picture. It has a villain, a moment of corruption, a line that was crossed. It is also, for the most part, wrong — and its wrongness is the most protective thing about it. While everyone watches the spreadsheet for tampering, the real influence has already happened, upstream, in the open, in rooms where no rule was broken and no number was touched. By the time the data exists, it is usually honest. The shaping was done long before anyone ran an analysis.

This matters because we have built our entire defence against the wrong threat. The disclosure slide, the competing-interests footnote, the declaration at the foot of the paper — all of it is engineered to catch the keyboard villain. None of it catches the thing that actually moves results, because that thing isn't fraud. It's a sequence of design decisions, each one defensible on its own, each one tilting gently in the same direction. You cannot find it by reading harder. You find it by knowing where to look.

The result is decided before the first patient is enrolled

A study's most important conclusions are settled at the design stage, not the analysis stage. This is the part that gets missed, because design feels like neutral plumbing — recruit some people, give them the thing, measure what happens. In fact, the plumbing is where the argument is won.

Start with the comparator. 'The drug works' is an empty sentence until you finish it: works compared with what. Test something against nothing — against a placebo — where a perfectly good standard already exists, and you have engineered a flattering result before a single dose is given. Test it against a real rival, but at a dose chosen to underperform, or at the wrong moment in the illness, and you have done the same thing more subtly. The comparator is one of the quietest levers in all of medicine precisely because it is set early, looks like an administrative detail, and determines the answer. Nobody fakes anything. They simply choose the opponent.

Then the endpoint. Hard outcomes — does the patient live longer, function better, suffer less — are slow, expensive, and stubborn. Surrogate outcomes — the blood marker, the score on a scale, the shadow on a scan — are fast, cheap, and obliging. A study built to move a surrogate will, often enough, move it, and the surrogate will be reported as if it were the thing that matters. Sometimes the surrogate genuinely tracks patient benefit. The history of medicine is a long catalogue of confident occasions when it did not. Choosing what to measure is choosing what you are allowed to claim, and that choice is made on a whiteboard, not in the data.

Then the population. Who gets enrolled is who gets to generate your result. Recruit the people most likely to respond and least likely to be harmed — the younger, the fitter, the ones without the complicating second and third diagnoses — and the trial will look cleaner and safer than the drug will ever look in the clinic, where the patients are older, frailer, and on five other things. This isn't deception. Tight eligibility criteria are good science in their own way; they reduce noise. But they also quietly pre-select the answer, and the gap between the trial population and the real one is decided in the protocol, long before recruitment opens.

None of these moves requires a bad actor. Each is a legitimate methodological choice that a thoughtful, honest team could defend at a podium. That is exactly why they are so much more powerful than fraud: they survive scrutiny, because there is nothing hidden to find.

Then the literature gets shaped by what doesn't appear

Influence operates a second time after the data is in — not by altering studies, but by deciding which ones the world ever sees, and how loudly.

The largest force here is the simplest: the study you are reading exists partly because it was positive. Its duller siblings — same question, flatter answer — are far more likely to have been quietly shelved, never written up, never submitted, never indexed. The published record isn't a neutral sample of what was tried; it's the survivors, and survival correlates with telling a good story. Any single encouraging result therefore arrives pre-filtered by an editorial process that flatters in aggregate, and the filtering is invisible from inside the paper. You can read the whole thing, cover to cover, and never see the negative version that ran alongside it and went nowhere.

A second force multiplies the positive. One study's worth of encouraging data can be carved into several papers, each reporting a slice, each citing the others, until a modest signal acquires the bulk and self-reference of a literature. To a reader doing a quick scan, three papers feel like more evidence than one. Often they are one, wearing three coats.

A third force shapes the connective tissue. Review articles, commentaries, the summarising pieces that busy clinicians actually read instead of the primary studies — these can function less as scholarship than as infrastructure, framing a field's questions and emphases in a particular direction while breaking no rules of honesty. The individual claims may all be true. The selection, the weighting, the choice of what to foreground and what to leave as a footnote — that is where the steering happens, and it leaves no fingerprints.

Disclosure was built to catch a different crime

Here is the uncomfortable part. The disclosure slide, which we treat as the immune system against all of this, is almost perfectly designed to miss it.

A declaration of interest tells you that a relationship exists. It does not tell you what that relationship shaped — which comparator it favoured, which endpoint it preferred, which patients it would rather enrol. It announces the existence of a financial tie and stops, as though naming the connection were the same as tracing its effect. But the effect lives in the design decisions, and those decisions are not on the slide. You are handed the fact of a relationship and left to do nothing useful with it, because the mechanism by which it mattered is precisely the part that isn't disclosed.

It can be worse than useless. There is a real possibility that disclosure backfires — that an author who openly declares an interest is read as more candid, more trustworthy, having come clean, while an identical claim from an undeclared source draws more suspicion. The honest act of disclosing can purchase a small licence to be believed. The slide, meant to lower your guard's threshold, can quietly raise it. A defence that sometimes increases trust in the conflicted claim is not much of a defence.

And underneath both problems sits the deepest one: most conflicted authors are not lying, and do not feel conflicted at all. They believe their conclusions. The comparator looked reasonable to them. The endpoint seemed sensible. The enthusiasm is sincere. Influence at the design stage doesn't need anyone to act in bad faith — it works through good faith, through a thousand small judgement calls each made honestly and each, on average, leaning the same way. This is why the keyboard-villain model is so dangerous to hold. It tells you to look for guilt, and there usually isn't any. There is just a slope, and everyone walking down it convinced they're on level ground.

Read for the decisions, not the declaration

If the influence is structural, the defence has to be structural too. The move is not to scan the disclosure footnote and feel reassured or alarmed by it. The move is to interrogate the design.

Ask who chose the question, and who chose the opponent. What was this tested against, and would a fair fight have used a different comparator, a different dose, a different moment in the disease? Ask what was actually measured — a hard outcome that matters to a patient, or a surrogate that was easier to move — and treat a surrogate result as a hypothesis about benefit rather than a demonstration of it. Ask who was enrolled, and who was quietly left out, and whether the people in the study resemble anyone you'd apply it to.

Then look for the siblings. A single encouraging study is a data point, not a verdict, and the most important context for it is often the trials you can't see in the paper — the ones registered and never published, the flatter results that didn't make the journey. Knowing that registries exist, that protocols are filed before results are known, that the published literature is the visible tip of a larger and duller iceberg, is half the defence by itself. It reframes any one positive paper as a survivor, to be held loosely until something independent agrees with it.

And use the funding source the right way. Not as a verdict — 'sponsored, therefore false' is lazy, and it would discard a great deal of perfectly sound science. Use it as context for the design choices. A funding relationship doesn't tell you the result is wrong. It tells you which way to expect the small decisions to have leaned, and where to check whether they did. That is a far more useful instrument than indignation.

What this means

The reassuring story about conflicts of interest is that they are a problem of dishonesty, caught by honesty — declared, footnoted, and thereby handled. The truer story is that the most consequential influence is entirely compatible with integrity. It is spent early, in the open, on choices that look like methodology and quietly function as conclusions. The data, when it finally arrives, is clean. It was never the point of attack. The slide that declares the relationship is real and worth reading, but it tells you the least interesting thing — that money was in the room — and almost nothing about the only thing that matters, which is what the money did. To find that, stop reading the declaration and start reading the design. That's where it went.