It's not a report. It's a case.

Advocacy, not fraud.

Let's be precise about what a manufacturer's submission is, because both naïve extremes are wrong. It is not fraud — the manufacturer is not lying, faking data, or miscalculating. Outright deception would be caught, and the consequences (regulatory, reputational, legal) would be catastrophic; no serious company plays that game. But it is also not a neutral, disinterested report. It sits in between, in a category that's easy to miss if you've only been taught to check for errors: it is advocacy.

Here's how advocacy works without a single lie. Any real economic analysis contains dozens of points where the methodology permits a range of honest, defensible choices — which comparator to use, which endpoint, how long a time horizon, how to extrapolate beyond the data, which source to take a utility value from, which population to analyse. At each of these forks, more than one answer is legitimately arguable. And at each one, the manufacturer — rationally, legally, and without any individual choice being wrong — selects the answer that happens to favour their drug. Each pick, examined alone, is defensible: "this comparator is used in some settings," "this extrapolation fits the data acceptably," "this subgroup was pre-specified." No error to point to.

The tell isn't in any single choice. It's in the pattern. Think of it as the testimony of a witness called by one side in court: every sentence they say may be true, and yet the whole testimony is shaped, selected, and sequenced to persuade. Your job as the assessor is not to catch a lie — there may be none — but to recognise that you are reading a case, and to work out what the impartial truth underneath it looks like. And the way you do that is by following the levers.

Follow the levers.

So how do you actually read a case, rather than a report? You go through it lever by lever — every point where a choice was made — and at each one you ask the same four questions:

1. What choice was made here? Identify the fork: the comparator, the horizon, the extrapolation, whichever it is.
2. Was it justified, or merely permissible? There's a world of difference between "this is the right choice, and here's the clinical and methodological reason" and "this is within the range of things you're allowed to do." Advocacy lives in the gap between justified and permissible.
3. Which way does it push the result? Does this choice lower the ICER (favour the drug) or raise it (against the drug)? Almost every choice pushes one way or the other.
4. What would an equally honest opposite choice do? If you flipped this to the impartial best answer, how much would the ICER move?

Notice what this is not. It's not checking the arithmetic — the arithmetic is fine. It's not hunting for a mistake — there isn't one. It's tracing direction: at each fork, which way did they turn, and were they entitled to? One lever, by itself, tells you little — any single defensible choice is, well, defensible. The power of the method comes from doing it across every lever and then stepping back to look at the pattern of directions. Because a case reveals itself not in any one turn, but in whether all the turns went the same way.

The map of levers.

Here's where this final module pays off everything that came before, because the levers are the course. Every place a manufacturer can tilt an analysis is a topic you've already learned to build — and now you learn to detect. The main levers, roughly in order of power:

• The comparator (Module 1's PICO) — the single most powerful lever. Compare the drug to an older, weaker, or cheaper treatment instead of the true current standard of care, and its relative benefit inflates automatically. The drug didn't get better; the yardstick got worse. Changing the comparator alone can flip a verdict.
• The population — restrict the analysis to the subgroup where the drug performed best, and average results improve.
• The endpoint (Module 6) — lean on a surrogate that moved nicely rather than the hard outcome that matters, borrowing optimism the hard data might not support.
• The time horizon (Module 5) — stretch it so a slowly-accumulating benefit keeps adding up, or shorten it to hide a late-emerging cost.
• Survival extrapolation (Module 8) — beyond the trial data, fit an optimistic tail that keeps patients alive and benefiting far past what was observed.
• Model structure, utilities, data sources, discounting, perspective (Modules 5, 6, 8) — each a fork where a favourable but defensible input can be chosen.
• The sensitivity analysis (Module 9) — the meta-lever: a narrow analysis that varies trivial parameters (stable by definition) while never seriously testing the real uncertainties (the comparator, the extrapolation), producing a reassuring "robust" ICER that proves nothing.

You know each of these intimately, because you spent this course learning to construct them. Reading a manufacturer's case is that same knowledge, run in reverse: where you once asked "how do I model this properly?", you now ask "how might this have been tilted, and which way?"

Follow the levers, live.

Here's a manufacturer's submission with an ICER of £18,000 — under the £30,000 threshold, so "fund." Each lever below is currently set to the manufacturer's choice — defensible, but favourable. Flip any lever to the impartial choice and watch the ICER move. The lesson is in what it takes to change the verdict. (An illustrative additive model — real levers interact; here they're additive to show the mechanism.)

Watch what it took. No single lever was the smoking gun — flip any one and the verdict mostly survives, exactly as the manufacturer would argue. But the £18,000 was never one number; it was five favourable choices stacked on top of each other, each defensible alone, together carrying the drug from "reject" to "fund." That's how advocacy works: not one lie, but a direction — every fork turned the same way. And here's the signature you're looking for: five independent choices, and all five happen to favour the drug. Flip them to impartial and the case dissolves. You never found an error. You found a slope.

Now you.

Direction betrays intent.

Now the deep point, the one that separates an expert assessor from a checklist-follower. You will rarely be able to say "this choice is wrong." Each one is defensible; the manufacturer will defend each one, correctly. So if your critique is "I don't like this comparator," you'll lose the argument — they'll show it's been used elsewhere, and move on.

The power isn't in any single lever. It's in the distribution of directions. Ask yourself: if these choices were being made impartially — genuinely trying to model the truth — which way would they fall? Randomly. Some would happen to favour the drug, some would happen to go against it, scattering both ways, because an honest modeller isn't steering toward a result. So when you open a submission and find that every debatable choice, all of them, happens to favour the drug — comparator, endpoint, horizon, extrapolation, utilities, all tilting the same way — you are looking at something with the probability of ten coin-flips all landing heads. Not impossible for any single flip. Astronomically unlikely as a set. That improbability is your finding.

And here's why it's such a strong finding: you don't have to prove any individual choice is wrong. You concede that each is defensible — and then point out that their unanimous direction is not something impartial modelling produces. You're not attacking the assumptions; you're attacking the improbability of their alignment. That's an argument with no good rebuttal, because the manufacturer can defend each choice individually but cannot explain why all of them, independently, happened to help. Direction betrays intent — even when every step is clean.

The right response: rebuild, don't reject.

So what do you do with a case you've read this way? The immature move is to write "the analysis is optimistic and biased — we reject it." This is useless, for a simple reason: every manufacturer submission is optimistic and biased. Saying so is a truism, not an assessment. A decision-maker who's told "it's biased" still has no idea what the drug is actually worth. You've labelled the problem without solving it.

The expert move is harder and far more valuable: rebuild the base case impartially. Take the manufacturer's model — which is competently built, remember — and flip each lever to its impartial, best-justified setting: the true standard of care as comparator, the hard endpoint, the cautious extrapolation, the full uncertainty range. Then re-run it. Now you have two numbers to put in front of the committee: the manufacturer's £18,000, and your impartial reconstruction's £47,000. The difference between them is your assessment. It quantifies exactly how much of the claimed cost-effectiveness was real and how much was constructed.

And crucially, sometimes the benefit survives the reconstruction — you flip every lever and the ICER barely moves, which tells you the case was genuinely solid, and you can say so with confidence. Other times it collapses, and you can show precisely why. Either way, you've done something a mere accusation never could: you've replaced "I distrust this" with "here is the impartial number, and here is how it was reached." That's not hostility to the manufacturer — it's the highest form of respect for the process: instead of believing the case or rejecting it, you've measured how much truth remains once the advocacy is subtracted. Reading the levers tells you how it was built; rebuilding tells you what it's really worth.

What's the strongest and most useful thing to do?

Why this matters for HTA

Critically appraising a manufacturer's submission is the single most common thing many HTA practitioners actually do — and doing it well is a specific discipline, not general scepticism:

• Read for direction, not for errors. The arithmetic will be right and each choice defensible; that's the nature of competent advocacy. Your job isn't to find a mistake but to trace, at every fork, which way the choice tilts — and to notice when every fork tilts the same way. A submission where all the debatable calls happen to favour the drug is showing you its construction, no error required.
• Make the cumulative case, not a single objection. One challenged assumption is easily defended and easily survived — the stack holds. The argument with no rebuttal is the pattern: independent choices whose unanimous alignment impartial modelling wouldn't produce. Direction across many levers, not doubt about one, is what carries weight.
• Rebuild, then compare — never just reject. Your deliverable is not "this is biased" (true of everything, useful for nothing) but an impartial reconstruction: flip each lever to its best-justified setting, re-run, and present both numbers. The gap is the assessment. Sometimes the benefit survives and you say so; sometimes it dissolves and you show why. Either way you give the decision-maker a defensible figure, which is the entire point of the exercise.

A manufacturer's submission is a witness for one side: every sentence may be true, and the whole thing is still built to persuade. Your task is not to catch it lying — it usually isn't — but to reconstruct the impartial testimony underneath. Read the levers to see how the case was built; rebuild the base case to learn what it was really worth.

Reading the manufacturer's case, in one breath.

• A manufacturer's economic submission is advocacy, not fraud and not a neutral report: rarely any error, but at every fork where the method permits a range of honest choices, the favourable one is taken. Each is defensible alone; together they tilt the result. It's a witness for one side — every sentence true, the whole built to persuade.
• You read it by following the levers: at each choice, ask what was chosen, whether it's justified or merely permissible, which way it pushes the ICER, and what an impartial opposite would do. You're tracing direction, not hunting errors — the arithmetic is fine.
• The levers are the course: comparator (the most powerful — a weaker yardstick inflates relative benefit), population, endpoint (surrogate vs hard), time horizon, survival extrapolation (optimistic tail), utilities, discounting, and the sensitivity analysis that varies trivia while dodging the real uncertainties.
• Direction betrays intent: any single favourable choice is defensible, but all choices favouring the drug has the probability of ten heads in a row — impartial modelling scatters both ways. You don't prove any one wrong; you show their alignment is what an honest model wouldn't produce. Then the expert response: rebuild the base case impartially and re-run, and put both numbers before the committee. The gap is your assessment — not "it's biased" (true of everything), but "here's the impartial figure, and here's how much of the claim survived."

Don't look for the lie — there usually isn't one. Look at the slope of the whole thing: which way every choice leaned, and whether an honest model would ever lean that consistently. Then rebuild it straight, and read off what the technology was actually worth.

You can now take a manufacturer's case apart and rebuild an impartial number from it — the analytical core of the job. But an assessment that no one reads or understands changes nothing. The number has to become a recommendation: written so a committee grasps it, trusts it, and can act on it. Turning analysis into a decision-ready HTA report is the next lesson.