Lawrence Lynn
8.9K posts
@PatientStormDoc
Crit/Care research physician. CoAuthor of “The Physician’s War” The Battle between Physicians & Pathological Consensus https://t.co/l8sZ0gKPTI
Why critical closing pressure (CCP) always felt wrong For years, I found CCP papers confusing — not because the idea of vessel collapse is wrong, but because CCP was often treated as something that exists all the time, as if it continuously governs perfusion in normal, open-flow states. That framing immediately caused problems and led to some very strange ideas. CCP was frequently presented as a downstream opposing pressure — something to subtract from MAP. Hence formulas like tissue perfusion pressure (TPP) = MAP − CCP. But a collapse threshold cannot be a back-pressure. CCP is only relevant once arterioles have collapsed. CCP was also used to explain autoregulation, even though autoregulation occurs in proximal arterioles at much higher pressures than those at which CCP becomes relevant in distal arterioles. CCP − Pms was then used to describe capillary perfusion, even though a collapse threshold cannot be an input pressure. This is simply a category error. CCP marks the point at which arterioles close. Above that point, flow is governed by ordinary arterial–venous pressure gradients and arteriolar tone — not by CCP. A waterfall analogy — developed for passive collapsible tubes (veins, zone-1 lung) — was exported into arteriolar physiology. But arterioles are not floppy veins. They possess active smooth-muscle tone, making them functionally open or closed, in a way veins are not. A true waterfall would make venous pressure irrelevant for organ blood flow — which it plainly is not. And here lies the central contradiction: If capillary perfusion were governed by CCP − Pms, how could there simultaneously be a “waterfall” in which venous pressure does not matter? The framework collapses under its own assumptions. From this came papers proposing that we might manipulate — even widen — waterfalls, which would in fact imply closing off more vascular beds rather than improving perfusion. In healthy physiology, there should be no collapse and no waterfall. None of this fits with clinical observation: • Raising MAP does not reliably restore perfusion • SVR does not reliably reflect tone • Oedema and venous congestion clearly impair organ blood flow • Monitors can reassure while tissues suffer The problem was never CCP. The problem was how it was framed. Active arteriolar closure was conflated with passive Starling-resistor behaviour. In reality, CCP is a conditional collapse threshold: vessels are either open or closed. It is not continuously “acting” when flow is normal. Once those ideas are separated, the contradictions disappear. CCP stops adding confusion and instead explains haemodynamic incoherence. CCP always felt wrong because it was being asked to do jobs it was never meant to do. Used correctly, it simplifies physiology — and supports genuinely personalised care. And that leads to the most important practical point: What matters clinically is restoring flow continuity across the macro-to-micro interface — reducing excessive tone, relieving external or venous constraint, and avoiding pressure strategies that worsen collapse rather than reopen closed beds. That is what CCP is actually useful for and why it finally makes sense. Read our full paper here 👇 @ThinkingCC @khaycock2 @EMNerd_ mdpi.com/2075-4426/16/2…
@yudapearl Here is a clearer “high school” explanation of Lynn’s Paradox (sequential RCT reversal). This is a real-world simulation, it models a structural flaw in current trial design. — A critical care researcher sees patients with severe infection who have fever and shortness of breath. She names this “Breathless Fever Syndrome” (BFS). It is a pragmatic synthetic syndrome, not a single disease with a common cause. To study treatment, she defines entry criteria: suspicion of infection + temperature >100.4°F + respiratory rate >20. A consensus group adopts these criteria for RCT enrollment. She receives @NIH funding and tests an anti-inflammatory drug (benefit in some infections, harm in others). She enrolls patients using these criteria. The first RCT shows benefit. The treatment and entry criteria are incorporated into guidelines. Over the next 4 years, patients treated under the same criteria in the same ICU do worse. A second RCT with the same protocol, same center, and same entry criteria shows harm. — Questions: 1Was this reversal predictable? 2What was missing in the RCT design? 3How could it be prevented? Answers: 1Yes 2A disease- or cause-specific diagnostic anchor 3Incorporate basic structural causal modeling into RCT design — Explanation: This is a “cause-agnostic RCT”. Enrollment is based on threshold criteria, not a causal disease. The entry criteria select a mixture of different infections, each with different treatment effects. The trial estimates only the weighted average: ATE = Σ pᵢ τᵢ Toy example of three infections (diseases): • Trial mixture (45%/35%/20%), effects (+35%/–25%/+10%) ->ATE = +9.0% • Second RCT, New mixture (25%/50%/25%) -> ATE = –1.25% Same criteria. Same setting. Opposite conclusions. The estimand changes because the underlying disease mixture changes. The criteria and treatment protocol stay exactly the same, but the real-world proportions of underlying infections shift. No amount of power, covariate adjustment, or statistical refinement can fix a design in which the enrollment gate is disease & cause-agnostic. This exact pattern has repeated for 40+ years in critical care. But causal tools which were not available when the critical care cause agnostic RCT became the standard now make this failure mode preventable! This is what your Book of Why II can teach. This is the mechanism behind Lynn’s Paradox and a central failure mode of modern critical care “synthetic syndrome” trials for four decades.
Oh my goodness how did @JClinEpi allow this? The authors are pretending that point estimates are true values! For this study the subjects should not have been given point estimates & you really can't answer the survey questions without Bayesian posterior probs. #Statistics
It’s a deep observation - ‘not a causal entity’! even though everything is caused, not everything is causing! 😊 made-up constructs are not causes. that’s a big hurdle in teaching causal graphs in epidemiology. the best way to design causal inference, experimental or otherwise, is to start with the study unit: real-world entity to which intervention is applied and from which response is recorded
There are two instantiations of the paradox which, unrecognized, have plagued critical care science for decades. First “RCT-to-clinical protocol effect reversal”. Suppose investigators conduct a RCT to test a treatment for patients with severe infection, which experts label “sepsis.” A consensus panel defines disease and cause agnostic triage criteria for sepsis, and patients meeting those criteria are enrolled in the trial. The treatment protocol is tested in what appears to be a well-designed &! powered RCT that meets all methodological standards & reports strong evidence of benefit. The treatment protocol is then converted into clinical guidelines. However, when implemented in practice, the protocol surprisingly results in patient harm, even when applied with the same treatment timing, the same hospital system, the same caregivers, & using the same clinical entry criteria as the trial. This is Lynn’s paradox. The paradox arises because the RCT was designed without causal modeling. Sepsis is not a single disease or causal entity it was initially derived as similarity heuristic & then migrated to causal disease equivalent status in the 80s. In practice, more than 50 different infectious diseases/causes can satisfy the consensus criteria used to define sepsis. If the treatment benefits some of these diseases/causes but harms others, the overall result of the RCT reflects only the average effect across that specific tested mixture of diseases. The trial appears positive only because the particular mixture of diseases among enrolled participants happened, by chance, to favor those that benefited from treatment. Unfortunately, this mixture is not stable. When the same criteria are applied in routine clinical practice, the proportions of diseases will shift. The result can be clinical guidelines derived from a “positive” RCT that produces net harm. This phenomenon also appears as RCT reversal, in which sequential trials of the same intervention in the same setting and the same every criteria, equivalent covariates, and treatment timing produce opposite conclusions. This is Lynn’s Paradox of sequential RCTs. The mechanism can be illustrated with a simple toy model. Imagine sequential RCTs enrolling patients who meet the latest consensus “sepsis criteria” but some participants have one infectious diseases (d1) and other participants have another infectious disease (d2). (both diseases meet the trial entry criteria). The treatment effect benefits d1 to the same degree that it harms d2 If the first RCT happens to enroll more participants with d1, the trial will show benefit. If a subsequent RCT enrolls a higher proportion of d2 the result will reverse and show harm, even though the hospital, caregivers entry criteria, trial design, power, and covariate mix and treatment protocol remains equivalent This is difficult to understand only because no one trained in SCM would think that disease & cause agnostic entry criteria for an RCT is acceptable. How could covariates be modeled as they may be different for each disease/cause and the disease/cause mix is not stable. Yet this is standard structure in critical care syndrome science Lynn’s paradox was the reason the ventilator protocols failed during the pandemic. Severe COVID pneumonia met the disease agnostic criteria for adult respiratory distress syndrome (ARDS) so the guideline committee thought they had strong evidence for a particular ventilator protocol (high PEEP) and produced guidelines with “strong” evidence based recommendations for this ventilator protocol But COVID pneumonia did not exist when the source RCT was performed it only met the criteria used in the trials. But meeting the syndrome criteria establishes causal equivalence in syndrome science so they thought they had high level RCT evidence for the protocol. It failed & has now been abandoned. That was Lynn’s Paradox Requiring SCM interrogation as part of RCT design would prevent the paradox
Finally published: the COBRRA trial, the first randomized head-to-head comparison of major DOACs — something the companies would never have done themselves, as a direct confrontation goes beyond their commercial interests. This was driven by independent investigators. The results somewhat support the idea of apixaban being the “safer” DOAC. That said, the list of study limitations is long, and for me, labeling something as “safe” is not enough — after all, even placebo is “safe” when it comes to bleeding. nejm.org/doi/full/10.10…
Which Paradox explains decades of sequential RCT reversal and is correctable by causal modeling? (Hint: it’s not Simpson or Berkson). @yudapearl Ans: In standard cause-agnostic randomized trials (CARs), enrolling a synthetic syndrome, the treatment effect observed in the trial can be expressed as: ATE_t = Σ_i w_{i,t} τ_i where i = underlying causal disease w_{i,t} = proportion of disease i in trial t τ_i = treatment effect for disease i. If the disease mixture differs across trials, the weighted average treatment effect can change sign: Σ_i w_{i,1} τ_i > 0 Σ_i w_{i,2} τ_i < 0. Sequential trials can therefore produce opposite conclusions even when the treatment effects within each disease remain constant. This is “Lynn’s Paradox”. Should be in “Book of Why II” (Call the paradox what you will. I’m just seeking reform.)
Lynn’s Paradox (with deference to Simpson) “A phenomenon in which sequential cause-agnostic randomized trials (CARs) testing a treatment for a disease-agnostic syndrome yield opposing conclusions across studies because enrolled participants represent a shifting mixture of diseases with opposing treatment effects.”
The official journal of the Canadian Paediatric Society has just acknowledged that more than 100 of its case reports are fabricated. Incredible reporting from @RetractionWatch: retractionwatch.com/2026/03/03/can…
