The Design Theory /Causal Inference Divide Explained by the Position of U

research methods design
,
1

Over on X there is commonly an argument ongoing between CI and DT. I have struggled to understand the divide and have extensively reviewed the history. Below I present my thesis for its existence as well as a means to connect the siloed in the interest of subject/participant safety and patient care.

Copied from X ————

“Below I present the “Spectrum of U” as a unifying framework, aimed at bringing the intellectual giants @f2harrell & @yudapearl into closer alignment to help us critical care.

Pearl’s formulation,
Y = f(X, U),
when translated into clinical medicine, states that a patient’s outcome (Y) results from what we do (X) (give a drug) acting on the patient’s underlying relevant biological state including the effects of other external actions (U). Pearl’s core claim is that U is not metaphysical speculation but a real, causally operative substrate. Even when unobserved, it represents the actual biological configuration upon which the intervention acts. In Pearl’s framework, U is explicitly treated as real rather than metaphysical.

However Pearl’s framework is incomplete. In reality, U is not binary but spectrum from the real (and even substantially knowable) to the functionally metaphysical.

The difference is a function of the data defining U and the domain knowledge. Together they can move U to the Left. (left = more concrete, knowable, and reproducible.)

However, causal interpretability depends not only on the functional reality of U, but on whether U can be meaningfully reproduced across the population to which a causal claim is applied. For a trial result to be transportable, the underlying causal world must be sufficiently stable that it can be approximated again, in another cohort, another hospital, or at another time. The value to the public of internal validity is severely limited without external transportability.

In stable, well-characterized diseases, U is relatively structured and constrained. Although incompletely known, it can be approximated and reconstituted across different populations. The causal world is coherent enough that we can plausibly say: this intervention acts on the same kind of system again and again.

Here we again see the value of domain knowledge to move the design such that U moves to the left. A range of combined techniques from cSM, OS, to RCT may be effective here.

However critical illness represents a teaching field by its effect on U’s position on the spectrum. By its nature, critical illness is a state of global physiological destabilization: cascading failure of interacting systems, and deep dependency on context, timing, and iatrogenic perturbation. The causal background is no longer localized or reproducible. It becomes a moving configuration rather than a stable substrate.

As complexity increases, the ability to reproduce U across populations progressively erodes. At some point, U remains biologically real but ceases to be reconstructable. When a causal world cannot be reliably reproduced, it begins to function as a metaphysical equivalent, a causal state that exists but cannot be coherently instantiated again.

This is the setting (the rightward positioned U) in which randomization becomes especially valuable in practice. It protects against systematic bias, but it cannot restore causal coherence.

But the public benefit of randomization ultimately depends on a functionally reproducible U or a U that is rendered a reproducible equivalent by adjustment

Under this framework, given the goal is a more leftward U, synthetic syndromes can now be recognized causing the opposite.

When labels such as sepsis & ARDS are applied they aggregate patients whose U configurations are not merely complex but fundamentally divergent

Here, U is no longer just unknowable, it is operationally irreconstructible. The intervention is applied not to a reproducible biological system but to a disease mix whose internal causal structure cannot be reassembled. In this sense, U becomes “functionally metaphysical” despite its biological reality. Neither DT nor CI can deliver meaningful, transportable causal guidance when U is metaphysically equivalent

The goal is to move U leftward using data, domain knowledge, and causal structural modeling (cSM), & then deploy randomization as a synergistic technology

“…They are approaching the same equation from opposite sides of U. DT treats U as unknowable and therefore “blows it up,” using randomization as nuclear weaponry to balance ignorance and isolate treatment effects.

In contrast CI treats U as partially knowable and seeks to move it leftward, pulling structure out of latency using domain knowledge and causal modeling. Both are correct, but incomplete alone.

The risk of CI alone is a false sense they have pulled U into the “light of the Left”.

The risk of design theory is that they trust to much in their strength (internal validation by blowing up U) and fail to focus on transportability.

The tragedy of the critical care synthetic-syndrome era was that the DT community trusted too much in randomization, so wide, incoherent entry gates, made U functionally metaphysical, so randomization balanced a trial artifact rather than a real causal substrate.

However, my fundamental thesis is that there is a deeper lesson. The “U Left-Right spectrum” indicates that the problem cannot be solved by fixing gates. Rather we need to make formal causal modeling part of trial design so that the U we balance actually corresponds to biology.

1 Like
2

Continuing the discussion from X after pushback from professor Elias Bareninboim.

Critical care urgently needs cSM, but Pearl’s framework cannot simply be imported verbatim. It must be reframed in a way that reflects how ICU trials are actually conceived, gated, & executed.

Pearl correctly teaches that causal inference allows valid conclusions even when background some variables remain unobserved. However, in critical care the dominant problem is not that these variables “cannot” be observed, but that they “were not” observed despite being biologically meaningful & often clinically accessible.

The central problem in critical care is not that decisive background variables are inherently unmeasurable; it is that many of them are in fact biologically meaningful & knowable, yet are excluded from the model for reasons of convenience, consensus, or disciplinary inertia. This renders Pearl’s formulation incomplete for ICU practice.

In critical care’s consensus based, cause-agnostic research, trial entry is defined by “synthetic syndromes” rather than coherent causal entities. As a result, large portions of decisive biological structure are relegated to U not because they are unknowable, but because they are inconvenient for trial logistics or incompatible with inherited taxonomies.

If the present concept of U were introduced into the prevailing critical care paradigm without reinterpretation, it would likely be assimilated in a manner that reinforces, rather than corrects, existing epistemic practices. Specifically, U would be systematically construed as an inherently aggregated substrate inside a “synthetic syndrome”rather than as a heterogeneous mix of unmodeled yet potentially knowable causal structure.

This interpretation is already reflected in the widespread invocation of the critical care term-of-art “heterogeneous syndrome,” a neologistic construct that purports to describe diversity while instead functioning as a semantic placeholder & used as a trail gate. A source of unexamined causal incoherence.

To make Pearl’s theory operational for intensivists, U must therefore be partitioned into:

Uₖ -causally relevant elements not in the model but potentially knowable through present domain knowledge, diagnostics, and cSM. (e.g., pathogen-specific biology, dynamic physiological relationships.)

Uᵤ -causally relevant elements that remain genuinely inaccessible or irreducible under current scientific constraints.

In critical care, and particularly in synthetic syndromes such as sepsis and ARDS, a disproportionately large share of clinically decisive causality lies in Uₖ. Yet conventional trial design treats this region as if it were Uᵤ, relying on randomization and consensus thresholds to neutralize its effects.

This creates what I define as a rightward U: a causal landscape in which biological structure that should guide intervention design is absent from the model, rendering the trial formally compliant with the consensus but causally incoherent. In this state, causal inference is not merely weakened it becomes structurally unsafe.

Therefore, translating cSM into critical care requires more than importing Pearl’s notation; it requires exposing where the discipline embraces “functional inaccessibility” (shifting U rightward) by consensus. The solution is not simply to “shrink U,” but to shift U leftward, progressively migrating biologically coherent, knowable causality into the observed model before the trial is ever allowed to operate. This will require abandoning 50 year old consensus science which has never worked but is so deeply embedded that it defines the present state of the pathological science itself

3

One reason SCM has not penetrated into clinical care as a synergistic tool is the lack of “disclosure rigor”. Design theory has extensive rigor defined by CONSORT but with a few weak blind areas of disclosure. But the same blind areas are present in SCM.

In some ways CI methodology elevates its blind areas into a strength, as if CI represents a magical tool to see past the unknowns and discover the counterfactual.

Critical care medicine is highly complex and relationally dynamic. We have seen even the best of DT stumble there. CI emerging with a blind U and undifferentiated S will fare no better and almost certainly, much worse.

The best role for CI at present in clinical medicine would be to explicate pathophysiologic relationships and pathways in the design of RCT and OS. But the mutually shared blind portions explain why this is not a simple matter of combining them. CI adds little, and certainly not enough. in its present poorly disclosed form.

The portions of CI which are needed ant present only go up to “Rung 2”. ie P (Y | do(x)) . The value of Rung 3 would seem to decline as a function of complexity of the relevant clinical state. As causal systems become increasingly heterogeneous and biologically opaque, counterfactual utility diminishes and risks becoming a formal abstraction detached from clinical reality. Unless rung three can be enabled by processor CI layering onto EMR data its utility in complex clinical science seems low.

But for rung 2 there is the value of design disclosure. The design can be built using SCM as a scaffold but not in its present form. Additional explication of the symbols is required.

Here is a summary from X.

“Thanks very much for this clarification. Let me gently push back, purely for the purpose of sharpening the discussion.

I understand and accept that, in SCM, the theory does not formally equate “unknown” and “omitted.” But in practice the analyst cannot avoid that equivalence, because both are symbolically merged into the same U, with no formal requirement for disclosure. In critical care design theory we have learned, to our sorrow, that a distinction not disclosed is a distinction functionally erased.

Our goals are slightly different. Your goal is to preserve and clarify a formally coherent overarching framework. My goal is to bring a formal tool into a raw domain that has been harmed for decades by tolerance of excessive, undisclosed design discretion, and to force consensus based flawed use of design discretion into the open.

I am not critiquing Pearl’s math. I am critiquing the reality in which that math will live, if it is migrated wholesale into critical care science.

That is why I’m proposing what you might call a “disclosure extension” to SCM: an added layer of “symbolic causal modeling” that makes the content of the symbols themselves explicit. In critical care, if SCM is imported without such an extension, the worst offender will not actually be U, but S, which is closely related to the opacity of U.

In causal notation, S=1 is treated as a single selector or gate. But that gate can represent radically different kinds of things:
-a definitive diagnostic test for a real disease,
-a specific biological condition (eg a mutation),
-a threshold defining a population,
-a surrogate marker, or

-a consensus-defined “synthetic syndrome” (e.g. sepsis, ARDS, CAP).

Formally, while these are all given the same symbolic status; epistemically, they are nowhere near equivalent. So U collapses ignorance and omission, and S collapses causal coherence and causal fiction.

In my upcoming paper (aimed at clinicians) I discuss this need for formalization if S=1 disclosure extensively, precisely because in a concrete domain an equation should be able to stand on its own, with minimal additional prose, and still convey the crucial distinctions.

One simple step in that direction is to make the symbolism itself carry more of that burden. For example, we might explicitly define a subset:

S_syn= 1

to denote inclusion based on a synthetic symbolic gate constructed from a threshold rule T that aggregates a heterogeneous set of underlying diseases D1, D2,..Di that do not share known causal unity. In words: S_syn=1 means “this is a synthetic syndrome defined by a threshold rule over a mixture of diseases,” not a coherent disease entity.

What I am proposing, then, is not a replacement for SCM but a symbolic overlay that forces disclosure of what S and U actually mean in a given domain, beyond their basic overarching mathematical role.

I am proposing this as a means to improve and pre and post audit critical care RCT and OS and as a requirement for grant applications critical care.

Clinicians and Trialists will probably have to build that symbolic disclosure layer themselves, and the act of doing so may be the best means to promptly reform the science…”