The Lead That Closes From Behind

I.The First Known Thing — the gain that fades, and who moved

education economics · persistence and fadeout

A child given a good preschool programme arrives in school ahead. Measured against children who did not have the programme, the advantage on tests of early reading and arithmetic is large, often more than half a standard deviation. Then, over the first years of school, it shrinks. By the third or fourth grade the two groups are often indistinguishable on the tests, and the pattern is so regular across programmes, decades and countries that the field that studies it has a name for it: fadeout. Drew Bailey, Greg Duncan and their colleagues, in a review of 2017 and a longer treatment in 2020, define it plainly as a pattern of diminishing intervention impacts after the end of treatment, and they note that it is the usual case. The famous programmes are not exceptions. Head Start, the Perry Preschool, the Tennessee class-size experiment all show it. Three quarters or more of what a good teacher adds to a child's test score in one year is gone from the scores by the next.

the distinction the field insists on

The first thing the field learnt from looking closely is that a fading difference need not mean that anyone lost anything. In one mathematics programme the treated children finished the preschool year 0.63 standard deviations ahead; by the end of the first grade the gap was about forty per cent of that. But the treated children's scores had not fallen. Scores rose for both groups, and rose faster, on average, for the children who had not had the programme. The difference closed from behind. The field calls this control-group catch-up, distinguishes it sharply from loss of the skill in the treated, and observes that the two are indistinguishable in the treatment effect, which is a difference, and separable only when each group's absolute trajectory is followed. Most fadeout, on this evidence, is catch-up. The untreated children learn to read too; they merely learn it a year later, from their schools, and once they have, the year of advantage has no test on which to show.

the three conditions for a gain that stays

From this the reviewers derive a rule for which gains can persist. A skill must be malleable, or the programme cannot move it; it must be fundamental, feeding the adult outcomes or the later skills that feed them, or moving it does not matter; and it must be one that would not have developed anyway, or the untreated will arrive at it in their own time and the advantage will close. They call the three together the trifecta and note that the third is the hard one. Nearly every child eventually learns to count and to read. A programme that teaches those skills early therefore buys a lead on a state that everyone reaches, and a lead on a state that everyone reaches cannot, by construction, be a permanent difference in that state. The ceiling on how long the gain can be measured is fixed by how long the untreated take to arrive, which is a property of the schools and the children and not of the programme at all.

what survives the closing of the gap

The second thing the field learnt is stranger. In several of the same programmes in which the test-score advantage vanished within a few years, the treated children were later found, as adults, to have completed more schooling, earned more, and been arrested less. The measured effect on the thing the programme taught had gone; an effect on things it never taught remained. The reviewers offer two accounts and are careful about the evidence for each. One, which they call the institutional gateway or foot in the door, is that an early lead, however short-lived on the tests, can place a child on the other side of a door that closes — into a better classroom, a track, an expectation — so that the timing of the advantage does work the advantage itself does not; small interventions, they write, can have large and lasting effects if they are timed well, and it may not be essential to retain the immediate target of the intervention in order for someone to continue to benefit. The other, the sustaining-environments hypothesis, is that a lead persists only where what follows continues to invest in it; and here they report that the evidence is weak, that in meta-analysis the interaction between an early programme and the quality of the later schooling is approximately zero. What survives, in other words, is not the lead. It is whatever the interval of being ahead happened to buy while it lasted.

II.The Second Known Thing — how long a drug's effect lasts

drug discovery · resistance and durability of response

Every drug programme wants to know not only whether its compound works but for how long, and the field has two vocabularies for the answer. In cancer, where the untreated disease progresses and does not remit, the question is one of resistance: a response is measured within the treated patients, its duration is counted from the day it is documented, and its end is the tumour's doing — a resistant clone, a bypass pathway, a lost target. Durability here is a property of what the treated tumour does, and the field studies it by sequencing what comes back. In the far larger number of conditions that are not cancer, durability is a difference between arms. A treated group and a control group are followed on a symptom score, a function scale, a time to recovery, and the effect is the gap between them at each visit. The drug is durable if the gap is still there when the trial ends.

the conditions that mend themselves

A great many of those conditions mend themselves. Influenza, otitis media, acute low back pain, the sprained ankle, the uncomplicated fracture, a first episode of depression, Bell's palsy, the common cold: in each the untreated course runs, over days or weeks or months, to a state most patients reach without help. Drugs for them are tested against placebo on the same scales as drugs for anything else, and the effects that appear are real and often useful, and they close. Oseltamivir, taken early in influenza, shortens the illness by something under a day; on the seventh day both groups are well. A steroid course given in the first days of Bell's palsy had, in the largest trial, 83 per cent of patients fully recovered at three months against 64 per cent without it, and 94 against 82 at nine months: a gap of nineteen points that had narrowed to thirteen as the untreated face mended. The field records these numbers, calls the early gap the effect and the later one its durability, and does not, as a rule, decompose the change between them.

the tools the field has, and what they were built for

It is not that the field is unaware that placebo arms improve. Every trial in these conditions plots the placebo trajectory, and the shape of it — the placebo response, regression to the mean, the natural history — is discussed at length, mostly as a nuisance that makes effects hard to see. There are designs that come close to the question. The delayed-start design, used in Parkinson's disease to ask whether a drug modifies the disease rather than its symptoms, gives the control group the drug later and asks whether the early group's lead survives once both are treated; a lead that survives is read as disease modification. Lead-time bias, in the screening literature, names the illusion that earlier diagnosis lengthens survival when it has only moved the clock. Cure-fraction models separate the patients who will never have the event from those who will merely have it later. Each of these is a tool for one indication or one design, and none of them asks the general question: whether, for a condition that mends itself, the time over which any drug's effect can be measured is fixed by the condition rather than by the drug, and whether what is left after that time is predictable from the lead alone.

where durability goes unread

The consequence is that two drugs whose effects both fade by the same visit are read the same way, though one may have lost its effect and the other may simply have been caught up with; that a drug whose effect fades is judged less durable than one tested in a condition that does not mend, though the comparison is between conditions and not between drugs; and that the question of what a short lead is worth — whether the day of influenza saved, or the two months of a working face, or the earlier remission of a depression, bought anything that persists after the untreated have arrived — is asked, when it is asked, one indication at a time, as a clinical debate about that indication, and not as a quantity that could be computed for all of them from the same two numbers.

III.The Bridge

no preschool study cites a placebo arm; no trial cites a preschool

The halls do not meet. Fadeout belongs to education economics and developmental science, and its literature is about children, schools and test scores; durability of response belongs to clinical pharmacology and trial methodology, and its literature is about tumours, symptom scales and hazard ratios. No review spans them, and the education hall's own examples — reading, counting, class size — have no pharmacological content at all. Two neighbours should be set aside. It is not the prefix refutability of an earlier specimen, in which a deadline converts a claim that could only be confirmed into one that can be refuted; there the deadline is imposed by the trial, whereas here the closing is done by the untreated course and would happen with no deadline set. And it is not the redress separation of another, in which a remedy fails because the injury is by now sustained by parties it does not reach; there the injury persists and the remedy is what falls short, whereas here the injury itself resolves and it is the comparison that falls short. What the halls share is a rule about the measured durability of any advance on a state that the comparator will reach on its own.

X: the ceiling belongs to the comparator, not to the intervention

Call it convergence fadeout. Where an intervention advances a system to a state that the untreated course reaches on its own, the measured effect is the difference between two trajectories, and it is bounded by the comparator's arrival rather than by anything the intervention did. Four things follow. The effect decays at the rate the comparator converges, so that a fading difference is indistinguishable from a lost state until the treated arm is measured absolutely, and the one is routinely read as the other. The ceiling on how long the effect can be measured is a property of the untreated course — the time by which the untreated arrive — and is therefore computable, from natural history alone, before the intervention exists; and every intervention in that condition, whatever its mechanism, is measured against the same ceiling. Durable separation on the state itself is possible only where the untreated course does not reach it, which is the education hall's third condition, that the skill would not have developed anyway. And what survives convergence is not the lead but what the lead bought: whatever irreversible good or harm accrued to one arm and not the other during the interval in which they differed, so that the residual is predicted by the length of the lead and the hazard of the irreversible during it, with no term for the intervention itself. Read into the trial hall, each has a consequence the hall does not state. A condition that mends itself has a convergence horizon, and the durability of any drug tested in it saturates there; a drug whose effect closes at the horizon has not been shown to lose anything, and a drug whose effect closes before it has. The Bell's palsy figures separate into a tempo component, the months of a working face gained, which must close, and a quantum component, the twelve or so patients in a hundred whom the untreated course never brings to full recovery, which is the only part of the effect that can last. And the value of a lead in a condition that mends is the integral, over the lead, of what goes wrong irreversibly in the meantime — the complication of the influenza, the job lost to the depression, the contracture of the face — a quantity the condition sets and the drug only multiplies by the time it saves.

IV.The Question That Becomes Possible

neither discipline would ask this alone

The trial hall asks whether a drug's effect is still there at the last visit and reads its absence as the drug wearing off. The education hall has no drugs, but it has established that a gain on a state everyone reaches must close, that it usually closes from behind, that the closing time belongs to the untreated, and that what outlasts the closing is what the lead bought while it lasted. Neither hall asks whether, across the conditions that mend themselves, the time at which drug effects fade is set by the condition rather than the drug, and whether what remains is predicted by the lead without reference to what the drug was.

the convergence fadeout hypothesis

For conditions whose untreated course reaches the state a treatment produces, can a convergence horizon be assigned to each condition from natural-history data alone, such that the treatment–control separation of every drug tested in that condition, whatever its mechanism, closes at that horizon and not before; and is the separation that remains beyond the horizon predicted by each drug's lead time multiplied by the hazard of irreversible sequelae during the lead, with no term for the drug — so that two drugs with equal lead in the same condition leave equal residue?

why it would matter

Because durability is read as a property of the drug, and if in a large class of conditions it is a property of the condition, then drugs are being compared on a quantity they do not own, and a programme choosing an indication is choosing, without knowing it, the longest its effect will ever be measurable. If the horizon rule holds, that number can be read from natural-history cohorts before a molecule is made, and an indication in which no drug can look durable can be recognised as such in advance. If the residue rule holds, the worth of a lead in a self-mending condition can be computed from the condition's own hazard of irreversible harm, and the argument over whether accelerating a recovery is worth anything — which the field now conducts separately for influenza, for depression, for back pain — becomes a single calculation with different inputs. And if neither holds, then drugs whose effects fade at the same visit differ in ways the visit does not show, and the field will have reason to measure its treated arms absolutely, which it seldom does.

V.How It Could Be Proven — or Broken

a candidate must be breakable, or it is merely poetry

An intercognix candidate earns nothing until it names its own test. This one needs no new trial; it needs the serial visit data of trials already published, and natural-history cohorts the epidemiologists already keep.

the honest status

Note the narrowness. The claim concerns conditions whose untreated course reaches the treated state, and says nothing about the conditions that do not — the metastatic cancers, the neurodegenerations, the chronic diseases that only worsen — in which durability really is a property of the treated arm and its resistance, and where the whole of the field's oncology vocabulary applies. It does not say that drugs for self-mending conditions are worthless; a day of influenza and two months of a face are worth having, and the residue rule is a way of pricing them, not of dismissing them. It does not claim that the treated arm never loses its state; it claims that loss and catch-up are different events with the same signature in a difference, and that the field seldom looks at the signal that would tell them apart. It is worth saying what does not transfer. A child's test score is a scale on which both groups keep climbing for years, so that catch-up has room to happen; a recovery scale has a ceiling at recovered, and both arms reach it, which makes the convergence complete rather than partial and the residue on the state itself exactly zero. The education hall's adult outcomes are decades away and its lead is a year; the clinic's irreversible outcomes may be weeks away and its lead a day, and whether a hazard integrated over a day is ever large enough to measure is an empirical question the education hall cannot answer. The whole thing rests on one computation that a single group with the published trials and the natural-history cohorts could make in a season: whether the time at which drug effects fade in self-mending conditions is set by the condition, and whether what remains is set by the lead. If it is not, this is an elaborate way of saying that placebo arms improve, and should be said so. Until that computation is made, this remains exactly what its stamp says: a candidate. A structural echo between two halls of the library, waiting for someone with the tools of both to walk the corridor.