Lawrence Lynn

@fluidloading @bentatoo31 @Denis_Faust @Procto_Log The story of pathological consensus and shortcutting the RCT method of Hill/Fischer with the Petty/Bone science. The lumping paradigm. The Physician’s War: The Story of the Hidden Battle between Physicians and a Science Based on Pathological Consensus a.co/d/8OtJlVk

Discussion of subgroup gates at datamethods. It would be great to have input from CI exerts. @soboleffspaces @eliasbareinboim discourse.datamethods.org/t/collider-in-…

🚨 HELL YES! EXPEL ILHAN OMAR FROM CONGRESS NOW and STRIP her phony citizenship for the immigration fraud she’s been pulling for YEARS! Rep. Fine is 100% RIGHT: President Trump has the power to boot this radical Islamist straight back to Somalia where she belongs! No more anti-American Squad members hating on the U.S. from inside our own government. DEPORT THE FRAUD! Clean house! @ericbolling @RepFine

Come contribute to the discussion of causal inference (CI) at Vanderbilt Datamethods. discourse.datamethods.org/t/the-growing-…

It is frustrating to have to reduce complex structural pathologic methodology to elementary explanations, but the degree of institutional conditioning around the RCT paradigm requires a return to the most simple explanations of the most basic principles which were unknowingly violated, institutionalized and remain hidden within complex mathematical embellishments.

There is a profound resistance to structural causal modeling and structural thinking within parts of modern statistics, a striking anomaly in science. The result has been that simple structural mistakes cause decades of waste, reversals, and harmful RCT-derived guidelines. Yet despite repeated failures, the underlying framework remains unquestioned because the RCT itself is an institutional icon beyond structural critique.

This simple graphic explains ~4 decades of RCT failure and reversed guidelines for harm in critical care. @CritCareReviews @MarionKCampbell @yudapearl

The Causality-Enlightened RCT Trialist

The most elegantly and definitively provable error of the science during COVID was the tragic early ventilator and high-PEEP guidelines. This came from a flawed, 1980s-era syndrome gated RCT model with decades of failure but sustained by elite consensus. After the loss, the offending ventilator guidelines were abandoned but causal modified RCT which generated the error was not. Without reform, it will happen again.

IN THE ROOM: DR. JAY BHATTACHARYA ON COVID, SCIENCE, AND THE FIGHT AGAINST TECHNOCRACY Former UK Prime Minister @trussliz speaks with NIH Director @DrJBhattacharya about the lab leak theory, the failures of the Covid response, and the deeper crisis inside modern science. They also discuss peer review, censorship, lockdowns, scientific authority, pharmaceutical incentives, and why Bhattacharya believes the West now needs a second scientific revolution rooted in openness, replication, and freedom. Timestamps: 3:00 - Did Covid come from a lab leak 7:00 - The broken incentives inside science 14:00 - Big pharma creates problems to sell solutions 19:20 - We must trust science, not "agencies 24:47 - Censorship has no place in science 27:10 - The role of the elites 31:28 - Can truth and freedom still win

@DrJBhattacharya @yudapearl The COVID ventilator-guideline failure was driven by an mis-specified selection rule (S): trials estimated a mixture estimand (E3), not a disease-level effect. The guidelines have been abandoned but the pathological methodology that created them has not. Like many major failures of perceived “well reasoned science”, this is not reducible to standard explanations, rather there is a simple but elegant apical error in the science which caused the failure. The NIH should investigate this apical error as it was one of the most important failure modes of the pandemic response.

You see the point? Let me make it explicit. The tragic failure of world wide “strong,” evidence-based ventilator guidelines during COVID can only be understood by integrating physiologic insight, historical perspective, and a Pearlean causal framework. This implies an urgent need to elevate the cognitive framework of critical care science. That deserves repeating: “We must elevate the cognitive framework of critical care science to protect the public!” Yet this mode of reasoning is extraordinarily difficult to convey within a “Babylonian” black box, cause agnostic RCT paradigm…so we must find a way to accomplish it soon.

There was a parallel competition of epistemological frameworks in critical care science (CCS). In the 1980s, CCS was moving toward a physiologic, relational, time-series paradigm grounded in mechanistic coherence. This trajectory was displaced by a “black box” approach that collapsed the physiologic envelope into cause-agnostic randomized trials (CARs). These trials estimate a third-layer (mixture) estimand, E3 = Σ_i π_i(S=1) · E2(i) = Σ_i P(D = i | S = 1) · [ Σ_xi τ_i(xi) · P(xi | D = i, S = 1) ] which is internally valid but not anchored to a stable biological data-generating process, precluding cumulative knowledge. The result was decades of lost opportunity to map the physiological “universe.” “Babylonian prediction” without causal structure, left CCS reliant on opaque models consistently generating the cause mixture paradox. When faced with new, pathologically complex disease, this framework lacked the causal grounding and flexibility required for adaptive treatment.

Why should I care about counterfactuals? At decision time all I observe are background covariates C, I don’t see outcomes nor natural treatment choices. What role do counterfactual quantities actually play for decision making? @yudapearl @eliasbareinboim @f2harrell

@barttels2 This is dad toughening up the weaker baby in that same nest site last year. His brother, the stronger baby, already flew to nearby limbs. I wonder where those babies are now. They hang around in limbs for awhile and then…they are gone.

@barttels2 Thanks. This is the non-zoomed distance from my open 3rd story window. I should put a web cam up. She had wonderful babies there last year. Don’t miss the sound on that other video.

Winning the best mother award. Large hailstorm a few days ago. Courageous Red Shouldered Hawk Mother protecting her eggs laid in Beech tree near the house. ~Dime size hail

I understand that you could not have anticipated that investigators (and later census groups) would simply guess thier own (disease and mechanism agnostic) threshold gates but this new trial design was introduced in 1987 one year before your causal graph teachings. This has expanded broadly in both RCT and OS, so yes without the required attributes of the selection gate S, causal inference is incomplete “right out of the gate.”

Yes. The answer is incomplete until the structure of the selection variable S is made explicit by a formal set of required gate attributes. When S is a synthetic gate (which is allowed at present) that mixes distinct causal systems, the estimand E[Y1 - Y0 | S=1] is mathematically precise but does not correspond to a stable causal target (i.e., it is not functionally causal).

I see 465 replies to the question "what does imply causation", and I am wondering: Why is it still a question? Is the answer provided by modern causal analysis incomplete? IOW, Is there a situation where this question is not given a definitive answer? @2Philosophical_ #bookofwhy @smueller

Well I wish causal inference was off the hook but alas CI has no formal restrictions on the gate (S) either. Same problem, equivalent bad result. CI handles the covariate mass from synthetic S (a synthetic data generating process) by modeling it, but this assumes a single underlying causal system, which is precisely what a synthetic S violates.

Not my fight. It's the doctor's dilemma: serve the patient or serve the system. My friend, Prof. Sam Shortt, wrote a book, The Doctor Dilemma, honestly exploring the conflict between two ethical dimensions (the individual patient and the society). In his analysis, the dilemma breaks down into two competing loyalties: 1. Patient Historically, the physician's primary duty is fiduciary: to act solely in the best interest of the person in front of them. 2. the "greatest good for the greatest number" As healthcare has grown to population delivery, the doctor is increasingly a "steward" of collective resources. Shortt concludes that the modern doctor became "doubly bound": they cannot serve the patient without being irresponsible to the system, nor serve the system without distancing from the patient. The change is profound. If you treat a patient you need αιτιολογία, if you treat a population of patients you only need a pyramid of evidence, with its golden calf - meta-analysis of RCTs = an average of the averages αιτία is gone. Not my life, not my fight.

@DrJBhattacharya There was a stunning discovery during the recent investigation of the structure of clinical trials. Neither inference (CI) nor decision theory (DT) explicitly specifies the required attributes of the selection rule S. Bradford Hill did, but those restrictive rules were lost after about 4 decades as funding for trials ramped up in the 1990s. One possible reason is pragmatic: imposing strict constraints on S would likely reduce the number of admissible (funded) trials and ~everyone wanted to do them. In the absence of such constraints, investigators began define S using arbitrary thresholds, scores, or consensus syndromes, thereby constructing synthetic data-generating processes (SDGPs) which generates a third layer estimand which looks robust but is an artifact of the trial and potentially harmful to the public to transport. Because these S (selection gate) constructions are unconstrained, SDGPs can generate an essentially unlimited number of trials, each internally valid but anchored to a different, gate-defined mixture. This elimination of the restriction on the gate (S) generated the present golden age of trial expansion where performance of the trail itself is all that’s needed for success. This creates a structural asymmetry: 1. The number of possible SDGP-based trials is unbounded 2. The number of true, coherent biological DGPs is finite 3. Buildable knowlege comes from the study of biological DPGs. The study of SDPGs is discardable research. As a result, the trial ecosystem in some fields can become dominated by SDGP studies. These are well funded discardable research which do not advance the knowledge in the field.

One naturally wonders why a Turing Award winning framework, as articulated by @yudapearl, has no explicit constraints on the admissible S. Or are they provided and I missed it? I suspect that no one anticipated that investigators would routinely construct selection rules for RCT and OS that generate synthetic data generating processes SDGPs as they do today. If such pathological derivations of S had been anticipated, where the selection rule does not correspond to a coherent data-generating process, then explicit constraints on admissible S might have been formalized. Hopefully a formalization of S will make it into the Book Of Why II .

@dylanarmbruste3 Yes. Along with the introduction of the “cause agnostic RCT” (CAR). If you find anyone else doing this work I would like to know. This is new stuff so if you have any questions I am happy to answer. pubmed.ncbi.nlm.nih.gov/41992366/

@PatientStormDoc Is it your definition?

“RCTs are causal by design” is a popular refrain. It’s succinct and persuasive… but is it true? Functionally no. Not in the sense that clinicians actually require. Critical care physicians have learned this the hard way over 4 decades not to trust that common slogan. Physicians don’t just care about “technical causality”. They care about “functional causality”. But what is functional causality? A causal effect is functionally causal if it supports reliable and durable guidelines applicable to real patients, because the estimated effect reflects a mathematically robust, stable, transportable biological (eg physiological) response to intervention because the target during the trial was, and in the population is, a real (non-synthetic) biological data generating process (DGP). This is “guideline-level causality” (GLC) The structural gap is that an RCT identifies: P(Y |do(X), S = 1) But clinicians need something closer to: P(Y |do(X)) So these are only “functionally equivalent” if: The trial selection S corresponds to a biologically (pathophysiologically) coherent target captured by S and this is where the slogan fails because CONSORT does not assess this requirement. This is true because S=1 is ambiguous and as broad as the statistician and trialist desires under CONSORT. S can select a synthetic data generating process (SDGP) such as a the latest biologically cause and disease agnostic consensus derived threshold triage. Bottom line: “RCTs are (technically) causal by design, but even well constructed RCT, as defined by CONSORT they are not necessarily clinically useful or safe to transport by design.” @f2harrell @yudapearl @soboleffspaces @eliasbareinboim