Eligibility criteria are a cohort tax

A protocol’s eligibility section is written by committee, and the committee is right every time. The medical monitor wants the comorbidity exclusion because of a signal in the phase 1 data. The statistician wants the biomarker cut because it sharpens the effect size. Regulatory wants the washout because a precedent asked for one. Safety wants the organ-function floor. Each argument is individually sound.

Nobody in that room is charged for the patients their criterion removes. That is the structural problem, and it is why eligibility is the most reliable place to find a trial that will not enrol on time.

The arithmetic is multiplicative, not additive

Intuition says criteria add up. They multiply. If a criterion retains some fraction of the screened population, the eligible pool after a series of criteria is the product of those fractions, not their sum. The consequence is that the marginal cost of a criterion rises with every criterion already in the protocol — the tenth one is applied to an already-thinned population, so it removes fewer people in absolute terms but a larger share of what you had left.

There is a second effect the multiplication hides. Criteria are not independent. Age floors, organ-function thresholds and performance-status requirements all correlate — they are largely the same underlying patients. So the true joint retention is usually better than the naive product, which is a genuine argument against panicking at the arithmetic above. But correlation cuts the other way too: a biomarker requirement and a prior-treatment exclusion can be anti-correlated, and the population that satisfies both may be far rarer than either alone suggests. Neither direction can be reasoned out from a spreadsheet. It has to be measured against real cohort data.

What kind of criterion is it?

Not all criteria tax the same population in the same way, and it helps to score a protocol along axes rather than as a single count. The eight we score are:

• Laboratory and biomarker thresholds — the cuts that require a test result to fall in a range.
• Comorbidity and safety exclusions — conditions that rule a patient out for their own protection.
• Prior and concomitant treatment, and washout — what they have had, what they are on, and how long they must come off it.
• Performance status and general fitness — ECOG, Karnofsky, and their equivalents.
• Consent, compliance, contraception and lifestyle — the requirements about behaviour rather than biology.
• Age, sex, BMI and other demographics.
• Diagnosis confirmation and disease severity — how firmly the disease must be established, and how bad it must be.
• Reproductive and cycle status — where the indication makes it relevant.

The axes matter because they have different remedies. A washout period is a scheduling tax — those patients exist and can be enrolled later, at a cost in time. A biomarker cut is a prevalence tax — those patients do not exist in your catchment at all, and no amount of extra sites in the same geography fixes it. A consent-and-compliance criterion is often a site tax, falling much harder on some centres than others. Counting criteria without distinguishing these tells you a protocol is long, which you knew.

Strict compared to what?

An absolute criterion count is close to meaningless. Twenty-two criteria is unremarkable in one oncology setting and extraordinary in another. The only useful statement is relative: this protocol sits at the Nth percentile of strictness against comparable trials in this indication and phase.

That comparison is only honest if both sides are scored the same way. It is easy — and tempting — to score your own protocol with a rich, structured extraction of its criteria, and the peer trials with a cruder pass over registry text, because that is what is available. The result is an asymmetry that makes your protocol look systematically stricter or looser than its peers depending on which classifier is more sensitive. One classifier, both sides, and where an extractor fills a gap on your side only, say so in the output.

Pricing a criterion before it goes in

The practical fix is procedural, not analytical. Make the cost visible at the moment of the decision, in the protocol review where the criterion is proposed. For each candidate criterion, three things:

1. What does it retain? A measured or cited retention estimate, with its basis — a real-world cohort, a published cohort characterisation, or an explicit “unknown”. “Unknown” is a legitimate answer and a useful one; it tells you the criterion is being added blind.
2. What does it buy? The specific risk it removes or the specific effect-size gain it delivers. If nobody can state this in a sentence, that is informative.
3. Which tax is it? Scheduling, prevalence, or site. This determines whether more sites, more time, or a different geography is even a possible remedy.

Run that on the ten most restrictive criteria and you will usually find two or three that cost a great deal of cohort and buy very little — typically inherited from a previous protocol in the programme, where they were load-bearing and now are not. Those are the cheapest weeks you will ever recover from a timeline.

And when the answer comes back that the criterion is genuinely necessary, you have still gained something: a protocol whose enrolment assumption is now defensible, with a named reason for the tax you chose to pay.