Someone has quoted you a number. An external control arm saves you 40% and a year, they said, and the slide had a nice arrow on it. The instinct behind that number is sound. Recruiting a randomised comparator arm is genuinely slow and genuinely expensive, and if you can stand up a credible comparator from data that already exists, of course you save something.
Then look at what Amgen actually did with blinatumomab.
They ran it two ways. BLAST, a single-arm confirmatory study in MRD-positive ALL (NCT01207388), needed roughly 39 months and 75 sites across 11 countries to enrol 116 patients [1]. TOWER, a randomised trial against standard-of-care chemotherapy in relapsed or refractory ALL (NCT02013167), needed about 24 months and 117 sites to randomise 405 patients [2]. Same drug, same sponsor. The design without a comparator arm was the one that took longer in calendar time.
That is the whole problem with the number on the slide. The saving from an external control arm is not a percentage you can carry from one programme to the next. It is arithmetic you run on your own asset, and this post gives you the variables and the sums rather than a headline. So stop repeating the figure you were quoted. You are about to build the two columns it hides.
Give the number on the slide a name. Call it the brochure delta: the fixed savings percentage that travels from pitch to pitch as though it were a feature of the external control arm itself, like a spec on a datasheet.
It isn't one. The saving is the gap between two costs that both belong to your specific programme. On one side, what a comparator arm would cost you to recruit and run. On the other, what an external control would cost you to build and defend. Both of those turn on your disease and your population as much as on your endpoints and your data. Change the asset and you change both numbers, which means you change the gap.
The number is not portable.
Let's be honest about what the brochure figure actually is. It is someone else's subtraction, done on someone else's asset, sold to you as though it were yours. Sometimes the sign is even right. But you would not accept a valuation of your lead programme computed from a competitor's cap table, and you should not accept a savings estimate computed from a competitor's trial. If you are sizing a raise or a go/no-go on this decision, you need the two columns, not the percentage.
Take the randomised side first, because it is the better-measured one. Four variables set it.
Sample size: your effect size and variance assumptions fix how many patients the comparator arm needs, and that headcount is the multiplier on everything below it. I'm deliberately not naming a number here. The right one depends on your endpoint and your assumed treatment effect, and anyone quoting a universal figure for "the comparator arm you'll need" is selling the brochure delta again.
Per-patient cost: what each of those patients costs varies enormously by therapeutic area. Stergiopoulos and colleagues put oncology at roughly $87,300 a patient, hospital-acquired and ventilator-associated pneumonia at $89,600, and endocrine studies at $57,700 (Stergiopoulos et al. 2018) [3]. Pick your area; the spread is the point.
Screen-failure rate: this is the quiet one, and the single biggest driver of per-patient cost variation. In trials without decentralised elements, screen-failure rates run around 29.9% to 31.5%, and screen failures alone account for about 11% of total trial cost (DiMasi et al. 2023) [4]. In fact, every patient you randomise is hiding roughly half a patient you screened, consented and paid for and then could not use.
Recruitment rate, sites and activation time: a comparator arm is a recruitment operation long before it is a headcount, and finding the patients is where most of the money goes. Watch how long the machine takes just to switch on. Site and contract activation runs anywhere from 6 to 24 months depending on the country (Crow et al. 2018) [5]. Even the faster end of that range is not fast: Lai and colleagues measured average site contract execution at 7.9 months in the US and 8.7 months outside it (Lai et al. 2021) [6]. And in one analysis of ALS trials, median contract execution was 105 days and median IRB approval 125 days before a single patient ever walked through the door (Atassi et al. 2013) [7]. None of that has yet produced a data point.
And all of that machinery costs money every day it runs. A Phase II or III trial burns roughly $40,000 a day on average, and a Phase III alone about $55,716 a day (Smith, DiMasi and Getz 2024) [8]. Delay carries a second price on top of the running cost: the Tufts Center for the Study of Drug Development puts a day of delayed launch at around $500,000 in lost sales [8]. When you are staring at your runway, those are the numbers that empty it.
Look hard at what dominates this column. The headcount, taken in isolation, is almost a side issue; what runs up the bill is how hard your particular population is to recruit. Which is exactly the condition under which an external control starts to look attractive.
Free download
The External Control Arm Design Checklist
The 12 design points regulators probe first, in one checklist you can run against your protocol before database lock.
Get the checklist →Now the side the slide never itemises. It has three variables too, and one of them does not behave the way the pitch implies.
Does the data already exist, and does it fit? An external control is only cheap if a natural-history cohort or patient-level data from a prior trial already exists, actually matches your inclusion criteria, your endpoints and your era of standard of care, and clears the bar of what we've elsewhere called regulatory-grade RWE. If it does, you are mostly buying access and analysis. If it doesn't, and you have to build the cohort yourself, the floor is brutal. The CINRG Duchenne natural-history registry (NCT00468832) took something like 14 years and 21 sites to mature into longitudinal data on 551 patients rich enough to anchor a comparator [9]. That is the real price of "we'll just build our own registry."
Data access and the statistical build: here I have to be straight about a gap in the evidence. There is no defensible published figure for what it costs to build an external control arm: the licence for the data, the epidemiological and statistical labour to curate it, match it and analyse it. I went looking. The numbers that circulate in vendor decks did not survive being chased back to a primary source, so I will not repeat them dressed up as an industry estimate. Get your own quotes, from more than one group, and treat the line item as real even though nobody has benchmarked it in public.
The regulator's conservatism premium the line that runs backwards: here is the variable the brochure delta quietly assumes away. The selling point of a single-arm design is "fewer patients." The regulators do not agree that it automatically means fewer patients. ICH E10, final since 2001, is explicit that the external control estimate should be made conservatively, "possibly leading to a larger sample size than would be needed in a placebo-controlled trial" [10]. The EMA's 2024 reflection paper on single-arm trials makes the same point in its own words: "uncertainty with respect to bias may outweigh any gains in precision compared to a randomised controlled design" [11] (we drew ten lessons for drug developers from that paper). Read that again. The design you chose to shrink your patient count can be sent back to you bigger. For what the FDA specifically wants to see inside an externally controlled protocol, we have covered its draft guidance separately; this section is about the economics, not the checklist.
I watched this happen. Years ago I worked on a rare-disease programme that pencilled an external control arm into the budget as the cheap, fast option, more or less on the strength of a number like the one on that slide. The comparator data existed, which felt like a win. Then the cohort needed cleaning and curating we had not costed, the statistical plan grew a conservatism margin to survive the agency's questions, and the treated arm crept upward to carry it. The saving did not vanish. But by the time we had priced the build and the margin honestly, most of it had.
So put all three into the column: data access, build labour, and a conservatism premium that can add patients rather than remove them. The "fewer patients" line is often simply false.
So how do you actually work out whether an external control arm will save you anything? You do not need software to do this. You need four steps and your own numbers.
Size the comparator arm you would actually need. Start from your endpoint, your effect size and your variance. If you are going the external-control route, do not assume this number shrinks. Add the conservatism premium from the section above; the regulator may want it larger.
Cost the randomised comparator arm. Take that sample size, divide it by one minus your screen-failure rate to get the patients you must actually screen, and multiply by your per-patient cost. Then add the machine: site-activation months plus accrual months, multiplied by your per-day running cost. That is the column the brochure delta is subtracting from.
Cost the external control arm. Add the data-access or licensing quote, the statistical and epidemiological build time (real quotes, and if you have to build the cohort from scratch, the multi-year build), and the conservatism premium you carried down from step one.
Subtract, and read the sign. The number you get, positive or negative, large or small, is your delta. It is specific to this asset, defensible to a board and to an agency, and unrelated to any percentage on a slide.
Run it back through Amgen's blinatumomab pair and you can see why the sign is never guaranteed. BLAST removed the comparator arm and still needed 39 months and 75 sites for 116 patients, because the eligible population was the binding constraint on both designs. Taking the comparator arm away removed a cost. It did not remove the recruitment problem. Do this sum before you commit the design in your clinical development plan or your raise deck, because the answer is yours to defend, not the vendor's.
The sign of your delta is mostly decided before you cost a single line item, by one question: how hard is your target population to recruit at all?
Where it is real, and large: in narrow and ultra-rare diseases, a randomised comparator is often impossible to recruit at all, and sometimes impossible to justify ethically. Up to 30% of rare-disease trials are discontinued early, primarily because they cannot accrue patients (Khachatryan et al. 2023) [12]. That is the structural reason external controls exist at all, and it is where they earn their reputation. The eflornithine approval in high-risk neuroblastoma is the cleanest recent example: the FDA accepted an externally controlled design, citing concerns about the feasibility and the long projected time to complete a new randomised trial, with a comparator built from 270 patients' data from a completed cooperative-group trial, matched three-to-one against 90 treated children (Duke et al. 2024) [13]. The selumetinib approval in NF1 plexiform neurofibroma leaned on 50 matched external patients against 50 treated (Gross 2021) [14]. The pathway is real when the population forces it.
Where it shrinks or reverses: in a common disease where a comparator arm recruits quickly and cheaply, the data-access, build and conservatism costs can swallow the whole saving, and occasionally exceed it. And rare does not automatically mean external control. In spinal muscular atrophy, an ultra-rare and devastating condition with every ethical argument for avoiding a placebo, Biogen and Ionis still ran randomised, sham-controlled trials in ENDEAR (NCT02193074) and CHERISH (NCT02292537) [15][16]. They had the archetypal case for an external control and chose randomisation anyway, because the population, though small, could be recruited.
Locate your asset on that axis first. If your population makes a randomised comparator genuinely hard to fill, the delta is probably large and probably in your favour. If it doesn't, do the arithmetic before you assume anything.
Here is the honest objection, put as well as I can put it. In rare disease the direction is obvious. You cannot recruit a randomised comparator, so the external control is faster and cheaper, end of discussion, and sitting down to do per-variable sums is analysis paralysis while the runway burns. Just build the thing.
Grant the first half. Where recruitment is genuinely infeasible, do not agonise over the direction of the saving; it is in your favour and you know it. That said, the arithmetic was never about direction. It is about magnitude and risk. You still need the number to size the raise, to decide whether to license a cohort or build one, and to price the conservatism premium before the agency prices it for you. And the external-control path is not automatically the fast one, even in rare disease. The EMA refused marketing authorisation for defibrotide sodium in March 2013 after its external control arm shrank from 86 to 32 patients at interim analysis, then granted it in July 2013 once an additional external comparator was supplied (Khachatryan et al. 2023) [12]. That is a four-month delay bought by a comparator problem the sponsor discovered late. One case is not a pattern. Still, it is enough to show the external-control column carries a downside the brochure delta never prices in.
What I would genuinely like to end on is a published, defensible figure for what it actually costs to build an external control arm, benchmarked by disease area the way trial costs now are. It does not exist yet. Until it does, the honest move is unglamorous. Build your own two columns. Get three real quotes for the data and the build. Size the comparator arm both ways, with the premium in. And treat anyone who hands you a portable savings percentage the way you would treat a sticker price with no mention of tax, delivery or the trade-in on your old car: technically accurate, and still not the number you will actually pay.
If you want that modelling done properly and at pace, with the per-asset numbers rather than the brochure ones, it is the kind of work our team and the InovaSight platform exist to do. Either way, run the columns before you commit the design.
Get the monthly digest
The 5 things evidence leads need to know each month: regulatory moves, RWE developments and what they mean in practice. No pitch, one email a month.
Journal titles are given as descriptors, not verbatim, except where a document's official title is quoted. PMIDs and NCT IDs are the verification anchors.