Term structure of risk — a steeply falling hazard curve is a statement about heterogeneity, not about time
also called credit spread curve · front-loaded hazard · frailty and mixture models
Most relapses in this disease happen very early and then the risk falls away. In credit markets, a risk curve that falls like that is read as a sign that you are looking at two different kinds of borrower mixed together, not one borrower whose risk declines.
The same thing, for a scientist
Credit analysis reads the shape of a hazard curve, not only its level. A flat hazard means a memoryless failure process; a rising hazard means accumulating wear; a steeply *falling* hazard is the classic signature of a heterogeneous pool, because the weakest members default first and the survivors are progressively enriched for the strong. That inference is standard in credit and in reliability, and it is under-used in leukemia trial interpretation. In KMT2A-rearranged infant ALL the hazard is strongly front-loaded: two thirds of relapses occur within one year and 90% within two [E1], most high-risk infants never reach an intended transplant because of early events [E2], and post-immunotherapy lineage switch occurs at a median of 1.5 months with 81.4% inside six months [E3]. Read as a term structure, this says the population is a mixture rather than a homogeneous cohort with a uniform risk, which has direct consequences for how combinations should be dosed and for what a prognostic factor is supposed to predict.
Three readings of the same curve, and why they differ operationally
- Uniform hazard that happens to be high early. Implies continuous therapy for as long as tolerable, and prognostic factors that shift everyone's risk up or down.
- Declining hazard within individuals — each surviving month genuinely lowers the risk for that child. Implies the same treatment for everyone, tapering over time.
- Mixture of two populations — a high-hazard component that fails fast and a low-hazard component that mostly does not. Implies that the entire therapeutic question is identifying membership, and that intensifying therapy for the whole cohort delivers toxicity to the low-hazard component for no benefit.
These three are distinguishable by fitting the hazard function, which is a day of statistics on data the Interfant consortium already holds. The field routinely reports Kaplan-Meier curves and Cox models, both of which are compatible with all three readings and therefore silent on the question.
Why the third reading is the one a credit analyst would reach
Credit spreads on a pool of loans decline with seasoning for exactly this reason, and the inference to heterogeneity is so standard it has a name: survivorship-induced selection. Nothing about the inference is biological, which is what makes it a genuine import rather than a restatement — biology tends to reach for a mechanistic account of declining risk (immune reconstitution, clonal exhaustion, cumulative drug effect) where finance reaches first for composition.
The biological counterpart of a mixture is not mysterious: a fraction of cases carry a subclone or a cell state that is already resistant at diagnosis, and those cases fail fast. The graph already records a candidate for exactly that in multiply pretreated disease being menin-resistant before menin exposure, and in RAS pathway subclones.
Where the analogy does not carry
Loans in a pool are independent; patients on a protocol are not, because treatment intensity is adapted in response to early events, which distorts the observed hazard in ways a credit pool never experiences. Second, censoring by transplant is informative here and has no clean credit counterpart — patients are removed from the risk set precisely because they were doing well enough to be removed. Any hazard fitting must handle transplant as a time-dependent covariate or a competing risk, not as neutral censoring, and getting that wrong would manufacture a declining hazard out of nothing.
That artefact is the single most likely way this analogy produces a wrong answer, which is why it is named here and built into the analysis design.
| domain | finance |
|---|---|
| source concept | term structure of hazard rates, seasoning of credit pools, and frailty or mixture models as the standard reading of a declining hazard |
| methodological point | Kaplan-Meier curves and Cox models are compatible with all three readings of a front-loaded hazard and therefore cannot distinguish them. Fitting the hazard function, or a mixture-cure model, can. |
mapping
[
{
"bio": "relapse hazard over time after diagnosis",
"analog": "the default hazard curve of a loan pool"
},
{
"bio": "two thirds of relapses within one year, 90% within two",
"analog": "a steeply front-loaded, seasoning hazard curve"
},
{
"bio": "a subpopulation carrying pre-existing resistance at diagnosis",
"analog": "the weak tranche of the pool that defaults first"
},
{
"bio": "survivors at 24 months having a much lower forward risk",
"analog": "spread compression on seasoned loans, driven by composition not by improvement"
},
{
"bio": "a prognostic factor",
"analog": "a variable that should classify tranche membership rather than shift a uniform spread"
},
{
"bio": "continuous combination therapy for the full protocol duration",
"analog": "buying long-dated protection against a risk that has already passed"
},
{
"bio": "transplant removing patients from the risk set precisely because they were doing well",
"analog": "informative prepayment, which has no clean neutral-censoring counterpart"
}
]generated hypotheses
[ "hyp-front-loaded-hazard-implies-two-populations" ]
- If the population is a mixture, then every trial reporting an average benefit is averaging across two groups with different needs, and a therapy that helps the high-hazard component could be diluted into non-significance by the low-hazard one.
- Reading a declining hazard as a mixture when it is actually an artefact of informative censoring by transplant would produce a confident wrong conclusion and could justify withholding therapy from children who need it.
- Has anyone fitted the actual hazard function, rather than a Kaplan-Meier curve, to KMT2A-rearranged infant ALL relapse data? We could not find it, and Interfant holds the data to do it.
- If the cohort is a mixture, what identifies membership at diagnosis — RAS subclone burden, breakpoint class, methylome, MRD kinetics — and can any of them be measured before therapy starts?
- Does the frontline blinatumomab result work by shrinking the high-hazard component or by lowering the hazard in both, and can the published curves distinguish those?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
In KMT2A-rearranged infant ALL, 3-year event-free survival is below 40% and most relapses occur during treatment, with two thirds occurring within one year and 90% within two years of diagnosis.
Blinatumomab Added to Chemotherapy in Infant Lymphoblastic Leukemia · van der Sluis IM, de Lorenzo P, Kotecha RS, et al. · New England Journal of Medicine 2023What this evidence cannot show
This figure appears as background context in the paper rather than as a hazard analysis the authors performed. It should be treated as a well-sourced summary statement, not as a fitted hazard function, and the fitting is precisely what is being proposed.
Evidence that argues against this (1)
- WHERE THE ANALOGY BREAKS, AND IT IS THE CENTRAL INFERENCE. The entry's key move is that a steeply falling hazard is 'the classic signature of a heterogeneous pool'. In survival analysis this is a frailty argument, and the frailty is not identified from the marginal hazard alone: the same observed decline is reproduced by different assumed frailty distributions with materially different implications. Cécilia-Joseph and colleagues computed the asymptotic bias induced by omitted covariates under gamma, log-normal and binary frailty and showed the attenuation grows strongly with follow-up duration, concluding explicitly that 'the possibility of interpreting the long-term decrease in treatment effects as a bias induced by heterogeneity and trial duration' has to be considered. The corollary for this entry is that a front-loaded hazard is *consistent with* a mixture but does not establish one; a homogeneous population with a time-varying treatment intensity, or a cure-fraction model, or simply a protocol whose most intensive phase is early, will generate the same curve. `hyp-front-loaded-hazard-implies-two-populations` therefore needs an external, non-survival discriminator — a molecular or clinical covariate that splits the pool prospectively — and cannot be established from the shape of the curve. Cécilia-Joseph E et al. 2015
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
In Interfant-06, high-risk infants had 6-year event-free survival of 20.9% with an intention to transplant and only 46% actually received a transplant because many had early events; overall 6-year EFS was 46.1% and OS 58.2% in 651 infants, with KMT2A rearrangement the strongest prognostic factor.
Outcome of Infants Younger Than 1 Year With Acute Lymphoblastic Leukemia Treated With the Interfant-06 Protocol: Results From an International Phase III Randomized Study · Pieters R, De Lorenzo P, Ancliffe P, et al. · Journal of Clinical Oncology 2019What this evidence cannot show
Prognostic analysis is a Cox model, which assumes proportional hazards and therefore cannot itself detect the mixture structure this analogy proposes.
Evidence that argues against this (1)
- The Interfant-06 pattern the entry reads as pool heterogeneity has a mundane competing explanation that the credit analogy hides: the intended consolidation is scheduled at a fixed point, so patients who event early cannot reach it, and 'only 46% received a transplant because many had early events' is a statement about the protocol's timetable interacting with a front-loaded hazard, not necessarily about two biological populations. That interaction is the immortal-time structure Suissa describes, in which exposure (transplant) can only be acquired by surviving to the landmark, and it both inflates the apparent benefit of consolidation and makes the cohort look retrospectively as if it contained a fragile and a robust subpopulation. Any term-structure reading of this dataset must be done on landmark-conditioned hazards. Suissa S 2008
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
In Project EVOLVE, post-immunotherapy lineage switch occurred at a median of 1.5 months (1 month after blinatumomab, 2.8 months after CAR T cells), with 81.4% of cases inside six months and 64.3% of the 70-case cohort carrying a KMT2A rearrangement.
Project EVOLVE: an international analysis of postimmunotherapy lineage switch, an emergent form of resistance · Project EVOLVE consortium · Blood 2025What this evidence cannot show
A case series of events, not a cohort denominator, so it describes the timing of switch among switchers rather than a population hazard.
Evidence that argues against this (1)
- The very early, tightly clustered timing of post-immunotherapy lineage switch is read here as evidence of a pre-existing weak subpopulation defaulting first. An equally good reading is that the timing simply tracks the therapy: switch follows blinatumomab at a median of one month and CAR T at 2.8 months, i.e. it follows the exposure that causes it, and the MLL/AF4 lineage-switch data attribute the phenotype to chromatin-modifier-driven reprogramming rather than to selection of a pre-formed myeloid pool. A hazard that is front-loaded because the causal exposure is front-loaded carries no information about pool heterogeneity at all. Tirtakusuma R et al. 2022
the falling-hazard-implies-mixture inference is not identified from marginal survival data; frailty-distribution choice changes the answer and the bias grows with follow-up
reads the front-loaded relapse hazard as a seasoning credit pool
- related to →Lineage switch as antigen escape from CD19-directed therapy in KMT2A-rearranged leukemia
lineage switch is a second, even more front-loaded failure mode
generated by reading a declining hazard as heterogeneity