betaFeedback

WorksMathematics

Where the Data End and the Mechanism Begins: Doubly Robust Causal Digital Twins for Treatment Strategies

Abstract

Digital twins are increasingly used to ask how a patient's health would evolve under alternative treatment strategies. When a twin couples a mechanistic model with learned components and is fitted to observational records, its answer mixes information from the data with assumptions from the mechanism. For sustained strategies that begin at a clinical landmark, we split the counterfactual mean into a part the data support, where patients follow the strategy with probability bounded away from zero, and a part they do not. On the supported part, a cross-fitted sequential doubly robust estimator that uses the twin as its outcome model is consistent whenever the decision model is, so twins that fit equally well but disagree as plug-in estimators agree after correction. On the unsupported part, we characterize the values compatible with the data and with a residual of size at most epsilon around the mechanism. This set is nested in epsilon, reduces to a point when the mechanism is exact and identifiable, and recovers assumption-free bounds as epsilon grows, so the penalty weight of a physics-informed loss becomes a sensitivity parameter. Knowledge constraints on how covariates enter the mechanism shrink the set, and when support grows over time, back-testing gives a data-based lower bound on epsilon. In simulations based on an energy-balance model of weight change under glucagon-like peptide-1 receptor agonist therapy, the correction reduced the spread of estimates across four twins from 0.66 to 0.03 percentage points of weight loss with conservative coverage; on unsupported patients, the exact mechanism was rejected in every replicate, the bounds covered the truth at every residual size not rejected, and a mechanism-free extrapolation was biased by 0.5 points.

Corresponding author: Cong Cao

Introduction

About half of patients who start a glucagon-like peptide-1 receptor agonist (GLP-1RA) for obesity stop within the first year [37, 4]. In randomized withdrawal trials, most of the weight lost is regained within a year of stopping [34, 39] (Aronne et al. 2024), and a nonlinear meta-regression of six trials estimated that 60% of lost weight returns at one year, with regain approaching a plateau at roughly three quarters of the loss [4]. In routine care, the average regain after discontinuation is much smaller, in part because many patients restart the drug or switch to another treatment [13]. A clinician who reviews a patient six months after initiation therefore faces a question that neither source answers directly: for this patient, what would happen to body weight under continuation, under discontinuation, and under a switch, and how much of that answer can be trusted?

Digital twins are designed to answer such questions by simulating a patient’s trajectory under alternative interventions [3] (National Academies of Sciences, Engineering, and Medicine 2024). In physiology, the most credible twins combine a mechanistic model with learned components, for example an energy-balance model of weight change [16, 15] augmented by a neural network that captures what the equations miss [32, 41]. Fitted to observational records, such a twin can be run forward under any strategy. The difficulty is that it returns an answer whether or not the data contain patients who followed that strategy. Where they do, the answer can be checked against what happened to comparable patients; where they do not, it is an extrapolation carried entirely by the mechanism and by the inductive biases of the learned component. The two kinds of answer look identical in the output.

Recent results show why this matters. Hybrid models that fit the data well can lose the causal validity of their mechanistic core as their learned components become more flexible [42]. Learned transport models with nearly identical predictive accuracy induced exposure mappings that implied different spillover effects [5], and low prediction error of nuisance functions did not identify the better causal estimator [6]. Observational data alone cannot certify that a twin is causally correct without assumptions such as no unmeasured confounding [10]. Hybrid mechanistic models improve counterfactual prediction out of distribution [28, 20], and partial mechanistic knowledge can support what-if predictions without controlled experiments under explicit structural assumptions [40], but these approaches target prediction accuracy and do not separate what the data identify from what the mechanism supplies. Neural methods for counterfactual outcomes over time handle time-varying confounding [35, 24, 12] and quantify predictive uncertainty [18], but they contain no mechanism and therefore no principled basis for extrapolation. Debiased estimation has been extended to functionals of neural-operator trajectories [19], without treatment strategies or confounding. In the literature on limited overlap, effects in regions without support are either excluded by trimming [11], extrapolated by flexible regression [31], or bounded under smoothness assumptions on the outcome function [27]; none use a mechanism as the basis for extrapolation.

This paper makes four contributions. First, for sustained strategies that begin at a landmark, we decompose the counterfactual mean into a supported part, which is nonparametrically identified, and an unsupported part, which is not (§ 2.3). Second, we show that a cross-fitted sequential doubly robust estimator that uses the twin as its outcome model is consistent for the supported part whenever the decision model is, and we derive its exact second-order remainder; as a consequence, twins that are observationally equivalent but disagree as plug-in estimators agree after correction (§ 4). Third, for the unsupported part we characterize the set of values compatible with the data and with a residual of size at most ϵ\epsilon around the mechanism, establish its monotonicity, validity, and limiting behaviour, and show that knowledge constraints on the parameter encoder shrink it (§ 5); because a physics-informed penalty weight controls the residual size, it becomes a sensitivity parameter with a causal interpretation. Fourth, when the set of supported histories grows over calendar time, we show that back-testing yields a lower bound on the residual size required by the data and falsifies smaller values (§ 6). Section 7 specifies and reports a simulation study based on published physiology and trial results.

We do not address the choice of adjustment set. Throughout, the adjustment set is the observed history itself. When the history is compressed into a learned representation, the requirements of Cao [7] apply in addition to those developed here; in particular, representation bias is first order and is not removed by the corrections below. The mechanism enters this paper only through the model for outcome dynamics.

Setup and estimands

Data structure

Time zero is treatment initiation. At a landmark τ\tau, for example six months, and at subsequent decision times τ=t0<t1<⋯<tK−1<tK=T\tau= t_0 < t_1 < \cdots< t_{K-1} < t_K = T, a treatment status Ak∈AA_k \in\mathcal{A} is recorded, where in the running example A={c,s,w}\mathcal{A} = \{c,s,w\} denotes continuing the current GLP-1RA, stopping, or switching to a different agent. We observe nn independent copies of

O=(H0,A0,L1,A1,…,LK−1,AK−1,Y),O = (H_0, A_0, L_1, A_1, \ldots, L_{K-1}, A_{K-1}, Y),

where H0H_0 is the landmark history, consisting of baseline covariates XX, the weight trajectory before and after initiation up to τ\tau, dose changes, and symptoms; LkL_k collects the information recorded in (tk−1,tk](t_{k-1},t_k], including weights measured at irregular visits and symptoms; and YY is the outcome at TT, for example the percentage of total weight loss relative to the weight at initiation. Write Hˉk=(H0,A0,L1,…,Ak−1,Lk)\bar{H}_k = (H_0,A_0,L_1,\ldots,A_{k-1},L_k) for the history available at tkt_k, with Hˉ0=H0\bar{H}_0 = H_0.

Strategies and estimands

A sustained strategy aˉ=(a,…,a)\bar{a} = (a,\ldots,a) assigns treatment status aa at every decision time from the landmark onwards; for a=sa=s it means stopping at τ\tau and not restarting before TT. Let YaˉY^{\bar{a}} denote the potential outcome under aˉ\bar{a}. The target is the counterfactual mean ψ(a)=E(Yaˉ)\psi(a) = \mathbb{E}(Y^{\bar{a}}) and contrasts such as ψ(s)−ψ(c)\psi(s) - \psi(c). Strategies are defined at the landmark, as in target trial emulation (Hernán and Robins, [17]); a patient who stops and later restarts has deviated from sˉ\bar{s}, and the restart is part of the decision process rather than of the strategy. This distinction is what separates the strategy estimand from the natural course after discontinuation reported in routine-care studies (Gasoyan et al., [13]).

For fixed aa, define the decision probabilities and the iterated conditional expectations

πk(hˉk)=P(Ak=a∣Hˉk=hˉk),QK=Y,Qk(hˉk)=E{Qk+1∣Hˉk=hˉk,Ak=a}.\pi_k(\bar{h}_k) = \mathbb{P}(A_k = a \mid\bar{H}_k = \bar{h}_k), \qquad Q_K = Y, \qquad Q_k(\bar{h}_k) = \mathbb{E}\{Q_{k+1} \mid\bar{H}_k = \bar{h}_k, A_k = a\}.

for k=K−1,…,0k = K - 1, \ldots, 0, where Qk+1Q_{k+1} is evaluated at Hˉk+1\bar{H}_{k+1} with Ak=aA_k = a. Each QkQ_k is a functional of the observed-data law wherever πk>0\pi_k > 0.

Identification and partial positivity

Assumption 1 (Sequential identification). (i) Consistency: Y=YaˉY = Y^{\bar{a}} and Lk=LkaˉL_k = L_k^{\bar{a}} when Aˉ=aˉ\bar{A} = \bar{a}. (ii) Sequential exchangeability: Yaˉ⊥ ⁣ ⁣ ⁣⊥Ak∣Hˉk,Aˉk−1=aˉk−1Y^{\bar{a}} \perp\!\!\!\perp A_k \mid\bar{H}_k, \bar{A}_{k-1} = \bar{a}_{k-1} for k=0,…,K−1k = 0, \ldots, K - 1.

We do not assume positivity everywhere. Call a history hˉk\bar{h}_k reachable from h0h_0 under aˉ\bar{a} if it has positive density under the law obtained by setting Aj=aA_j = a for j<kj < k.

Assumption 2 (Overlap gap). There is δ>0\delta> 0 such that, for every kk and every history reachable under aˉ\bar{a}, πk(hˉk)∈{0}∪[δ,1]\pi_k(\bar{h}_k) \in\{0\} \cup[\delta, 1]. Let S\mathcal{S} be the set of landmark histories h0h_0 from which πk≥δ\pi_k \ge\delta holds for every reachable continuation and every kk, and let pS=P(H0∈S)∈(0,1)p_{\mathcal{S}} = \mathbb{P}(H_0 \in\mathcal{S}) \in(0, 1).

The gap is a simplification. In practice πk\pi_k takes small positive values, and a threshold δ\delta is chosen so that histories with smaller probabilities are treated as unsupported, trading the variance of weighting against reliance on the mechanism (Remark 3). Histories fail to be supported for clinical reasons: a patient with a large early weight loss and no side effects may essentially never stop at six months, so that the strategy sˉ\bar{s} is unsupported for that patient type even though such patients are common.

Define the supported and unsupported counterfactual means

ψin(a)=E{Yaˉ∣H0∈S},ψout(a)=E{Yaˉ∣H0∉S}.\psi_{\mathrm{in}}(a) = \mathbb{E}\{Y^{\bar{a}} \mid H_0 \in\mathcal{S}\}, \qquad\psi_{\mathrm{out}}(a) = \mathbb{E}\{Y^{\bar{a}} \mid H_0 \notin\mathcal{S}\}.

Proposition 1 (Decomposition and identification). Under Assumptions 1 and 2:

(a) ψ(a)=pSψin(a)+(1−pS)ψout(a)\psi(a) = p_{\mathcal{S}}\psi_{\mathrm{in}}(a) + (1 - p_{\mathcal{S}})\psi_{\mathrm{out}}(a).

(b) ψin(a)=E{Q0(H0)∣H0∈S}\psi_{\mathrm{in}}(a) = \mathbb{E}\{Q_0(H_0) \mid H_0 \in\mathcal{S}\}, so the supported part is identified.

(c) If π0(h0)=0\pi_0(h_0) = 0 for every h0∉Sh_0 \notin\mathcal{S} and Y∈[ylo,yhi]Y \in[y_{\mathrm{lo}}, y_{\mathrm{hi}}], the sharp identified set for ψout(a)\psi_{\mathrm{out}}(a) under Assumptions 1 and 2 is [ylo,yhi][y_{\mathrm{lo}}, y_{\mathrm{hi}}].

Part (b) is the g-formula restricted to supported histories [33]. Part (c) states that the observed data place no constraint on ψout\psi_{\mathrm{out}} beyond the outcome range when the strategy is never initiated from the unsupported landmark histories, which is the situation in the running example and in the simulation. When positivity fails only at later decision times, the data carry information up to that time and the identified set is narrower; we do not characterize it here. Any informative statement about ψout\psi_{\mathrm{out}} must therefore come from assumptions, and the question is how to state them so that their strength is explicit.

Mechanism-informed causal digital twins

Mechanistic core

Let x(t)x(t) denote weight change from initiation and ua(t)∈[0,1]u_a(t) \in[0,1] the drug exposure implied by strategy aa after the landmark: uc(t)=1u_c(t)=1; us(t)=exp⁡{−(t−τ)/τd}u_s(t)=\exp\{-(t-\tau)/\tau_d\} for t≥τt \ge\tau, with washout time constant τd\tau_d; and uwu_w the exposure to the new agent. We use the linearized energy-balance model of Hall et al. [16] with the feedback control of intake described by Hall [15]. Changes in energy intake and expenditure are

ΔI(t)=−Pu(t)−k{u(t)}x(t),ΔE(t)=βEx(t),\Delta I(t)=-P u(t)-k\{u(t)\}x(t), \qquad\Delta E(t)=\beta_E x(t),

where PP is the persistent drug-induced reduction in intake, k(u)=koff−(koff−kon)uk(u)=k^{\mathrm{off}}-(k^{\mathrm{off}}-k^{\mathrm{on}})u is the strength of the feedback that raises appetite as weight falls, and βE\beta_E is the change in expenditure per unit weight change. With energy density ρ\rho, the mechanistic drift is

f(x,u;θ)=−1ρ[Pu+{βE+k(u)}x],θ=(P,kon,koff),(1)f(x,u;\theta)=-\frac{1}{\rho}\left[P u+\{\beta_E+k(u)\}x\right], \qquad\theta=(P,k^{\mathrm{on}},k^{\mathrm{off}}), \tag*{(1)}

with drug-specific values of (P,kon)(P,k^{\mathrm{on}}) when A\mathcal{A} contains more than one agent. Group-level fits of this model indicate that GLP-1RAs and bariatric surgery weaken the appetite feedback by roughly 40–70% relative to its baseline strength [15].

Twin

Definition 1 (Twin). A twin T=(g,r)T=(g,r) consists of an encoder gg that maps baseline covariates to mechanistic parameters, θi=g(Xi)\theta_i=g(X_i), and a residual r(x,u,z,t)r(x,u,z,t), where zz collects features of the landmark history. Under strategy aa, a twin generates post-landmark weight change for patient ii by

dxi(t)=[f{xi(t),ua(t);g(Xi)}+r{xi(t),ua(t),zi,t}]dt+σ dBi(t),t≥τ,dx_i(t)=\left[f\{x_i(t),u_a(t);g(X_i)\}+r\{x_i(t),u_a(t),z_i,t\}\right]dt+\sigma\,dB_i(t), \qquad t\ge\tau,

started from the state at τ\tau, with weights observed with measurement error at visit times. Under the decision process observed in the data, the same equation with uu given by the observed treatment path defines the twin’s law of the observed trajectories. A twin implies functions QkT(hˉk)Q_k^T(\bar{h}_k), obtained by conditioning the latent state on the weights observed up to tkt_k and simulating forward under aˉ\bar{a}.

The same twin also describes the pre-treatment window with u=0u = 0, where the feedback strength is koffk^{\mathrm{off}}. This is one route by which the mechanism carries information across strategies (Remark 2).

Knowledge constraints. Prior knowledge can restrict which covariates may affect which parameter. A mask M=(J1,…,Jq)M = (\mathcal{J}_1,\ldots,\mathcal{J}_q) requires the jjth component of gg to depend on XX only through XJjX_{\mathcal{J}_j}; for example, the drug effect PP may depend on glycaemic status and genetic risk for type 2 diabetes, whereas the off-drug feedback koffk^{\mathrm{off}} may depend on demographic characteristics and the pre-treatment weight trajectory. Let GM\mathcal{G}_M denote encoders that respect MM.

Residual size. For a reference measure μ\mu on (x,u,z,t)(x,u,z,t), let ∥r∥μ2=∫r2 dμ\lVert r\rVert_{\mu}^{2} = \int r^{2}\,\mathrm{d}\mu and Rϵ={r∈R:∥r∥μ≤ϵ}\mathcal{R}_{\epsilon} = \{r \in\mathcal{R} : \lVert r\rVert_{\mu} \le\epsilon\} for a function class R\mathcal{R}. The measure μ\mu must cover the states, exposures, and history features at which extrapolation is required, not only those seen in the data; otherwise the residual is unconstrained exactly where the answer depends on it. In the simulation, rr acts only off treatment and is linear in features, r=(1−u)(b0+bTz+bxxτ)r = (1-u)(b_0 + b^{\mathsf{T}}z + b_xx_{\tau}), where xτx_{\tau} is the observed weight change at the landmark, and μ\mu is the product of the empirical law of (z,xτ)(z,x_{\tau}) over all patients, supported or not, and the uniform law of uu on [0,1][0,1]; then ∥r∥μ2=13bTE(vvT)b\lVert r\rVert_{\mu}^{2} = \frac{1}{3}b^{\mathsf{T}}\mathbb{E}(vv^{\mathsf{T}})b with v=(1,z,xτ)v = (1,z,x_{\tau}).

Remark 1 (Physics-informed penalties and residual size). Physics-informed training minimizes a penalized loss Ln(T)+λ∥r∥μ2\mathcal{L}_n(\mathsf{T}) + \lambda\lVert r\rVert_{\mu}^{2}, where Ln\mathcal{L}_n measures fit to the observed trajectories [32, 8]; APHYNITY penalizes the norm of the learned component for a related purpose [41]. When the problem is convex in rr, Lagrangian duality gives a decreasing map ϵ↦λ(ϵ)\epsilon\mapsto\lambda(\epsilon) under which the penalized and constrained problems min⁡{Ln(T):∥r∥μ≤ϵ}\min\{\mathcal{L}_n(\mathsf{T}) : \lVert r\rVert_{\mu} \le\epsilon\} share solutions. For neural residuals the correspondence holds only locally, so we index all results by the realized residual size rather than by λ\lambda. Either way, the penalty weight sets how far the twin may depart from the mechanism, which § 5 turns into a statement about counterfactuals.

Definition 2 (Observationally equivalent twins). A twin is compatible with the observed-data law PP if, for every kk and every action with positive probability given the history, the twin reproduces the conditional law of Lk+1L_{k+1} given (Hˉk,Ak)(\bar{H}_k,A_k) and of YY given (HˉK−1,AK−1)(\bar{H}_{K-1},A_{K-1}), and reproduces the pre-treatment trajectories. Let

T(ϵ,M)={T=(g,r):g∈GM, r∈Rϵ, T compatible with P}.\mathcal{T}(\epsilon,M) = \{\mathsf{T} = (g,r) : g \in\mathcal{G}_M,\ r \in\mathcal{R}_{\epsilon},\ \mathsf{T}\ \text{compatible with }P\}.

Compatibility uses all observed treatment paths, not only those consistent with aˉ\bar{a}. Because the mechanistic parameters are shared across strategies, patients who continued treatment constrain PP and konk^{\mathrm{on}}, and the pre-treatment window constrains koffk^{\mathrm{off}}, even for patient types who never stop.

Remark 2 (Where the information for extrapolation comes from). Under (1), the trajectory after stopping is governed by the decay of Pu(t)P u(t) and by the return of the feedback from konk^{\mathrm{on}} to koffk^{\mathrm{off}}. For a patient who never stops, PP and konk^{\mathrm{on}} are informed by the on-treatment trajectory and koffk^{\mathrm{off}} by the pre-treatment trajectory and by the encoder, which is learned from patients with similar covariates who did stop. Strategy positivity fails because the decision depends on early response, a post-baseline quantity, while the covariates XX that enter the encoder may overlap well between supported and unsupported patients. The mechanism is what links the two.

The supported part: doubly robust twins

Fix SS; in practice it is estimated by thresholding estimated decision probabilities on a split of the data that is not used for estimation. Divide the estimation sample into JJ folds. For observations in fold jj, let π^k\hat{\pi}_k and a twin T^\hat{T} be fitted on the other folds, set Q^k=QkT^\hat{Q}_k=Q_k^{\hat{T}} and Q^K=Y\hat{Q}_K=Y, and define the weights W^−1=1\hat{W}_{-1}=1 and W^k=W^k−11(Ak=a)/π^k(Hˉk)\hat{W}_k=\hat{W}_{k-1}\mathbb{1}(A_k=a)/\hat{\pi}_k(\bar{H}_k). The estimator of the supported part is

ψ^in(a)=∑i=1n1(H0i∈S)ϕ^(Oi)∑i=1n1(H0i∈S),ϕ^(O)=Q^0(H0)+∑k=0K−1W^k{Q^k+1(Hˉk+1)−Q^k(Hˉk)}.(2)\hat{\psi}_{\mathrm{in}}(a)=\frac{\sum_{i=1}^{n}\mathbb{1}(H_{0i}\in S)\hat{\phi}(O_i)}{\sum_{i=1}^{n}\mathbb{1}(H_{0i}\in S)},\qquad\hat{\phi}(O)=\hat{Q}_0(H_0)+\sum_{k=0}^{K-1}\hat{W}_k\{\hat{Q}_{k+1}(\bar{H}_{k+1})-\hat{Q}_k(\bar{H}_k)\}. \tag*{(2)}

This is the sequential augmented inverse-probability-weighted estimator ([2, 36, 25]) with the iterated regressions supplied by the twin rather than by separate regression fits. The corresponding influence function is

Din(O)=1(H0∈S)pS[∑k=0K−1Wk{Qk+1(Hˉk+1)−Qk(Hˉk)}+Q0(H0)−ψin(a)],D_{\mathrm{in}}(O)=\frac{\mathbb{1}(H_0\in S)}{p_S}\left[\sum_{k=0}^{K-1}W_k\{Q_{k+1}(\bar{H}_{k+1})-Q_k(\bar{H}_k)\}+Q_0(H_0)-\psi_{\mathrm{in}}(a)\right],

with WkW_k the weights at the true decision probabilities.

Lemma 1 (Exact remainder). For fixed π^k\hat{\pi}_k bounded below on supported histories and fixed Q^k\hat{Q}_k,

E{1(H0∈S)ϕ^(O)}−E{1(H0∈S)Q0(H0)}=∑k=0K−1E[1(H0∈S)W^k−1π^k(Hˉk)−πk(Hˉk)π^k(Hˉk){Q^k(Hˉk)−Qk(Hˉk)}].\mathbb{E}\{\mathbb{1}(H_0\in S)\hat{\phi}(O)\}-\mathbb{E}\{\mathbb{1}(H_0\in S)Q_0(H_0)\} = \sum_{k=0}^{K-1}\mathbb{E}\left[\mathbb{1}(H_0\in S)\hat{W}_{k-1}\frac{\hat{\pi}_k(\bar{H}_k)-\pi_k(\bar{H}_k)}{\hat{\pi}_k(\bar{H}_k)}\{\hat{Q}_k(\bar{H}_k)-Q_k(\bar{H}_k)\}\right].

Each term is the product of the error in one decision model and the error in the twin’s regression at the same time point, so the bias vanishes if, at every kk, either the decision model or the twin is correct.

Theorem 1 (Doubly robust twins on the supported part). Suppose Assumptions 1 and 2 hold, π^k≥δ/2\hat{\pi}_k\ge\delta/2 on supported histories with probability tending to one, and π^k→πˉk\hat{\pi}_k\to\bar{\pi}_k and Q^k→Qˉk\hat{Q}_k\to\bar{Q}_k in L2(P)L_2(P) on supported histories for some limits πˉk\bar{\pi}_k and Qˉk\bar{Q}_k.

(a) If, for each kk, either πˉk=πk\bar{\pi}_k=\pi_k or Qˉk=Qk\bar{Q}_k=Q_k, then ψ^in(a)→ψin(a)\hat{\psi}_{\mathrm{in}}(a)\to\psi_{\mathrm{in}}(a) in probability.

(b) If πˉk=πk\bar{\pi}_k=\pi_k and Qˉk=Qk\bar{Q}_k=Q_k for all kk and ∑k∥π^k−πk∥∥Q^k−Qk∥=oP(n−1/2)\sum_k\lVert\hat{\pi}_k-\pi_k\rVert\lVert\hat{Q}_k-Q_k\rVert=o_P(n^{-1/2}), with norms over supported histories, then n{ψ^in(a)−ψin(a)}→N{0,Var⁡(Din)}\sqrt{n}\{\hat{\psi}_{\mathrm{in}}(a)-\psi_{\mathrm{in}}(a)\}\to N\{0,\operatorname{Var}(D_{\mathrm{in}})\}.

(c) Let T^(1)\hat{T}^{(1)} and T^(2)\hat{T}^{(2)} be two twin-fitting procedures. If both satisfy the conditions of (b), then ψ^in(1)(a)−ψ^in(2)(a)=oP(n−1/2)\hat{\psi}_{\mathrm{in}}^{(1)}(a)-\hat{\psi}_{\mathrm{in}}^{(2)}(a)=o_P(n^{-1/2}). If both are misspecified, with limits Qˉ(1)≠Qˉ(2)\bar{Q}^{(1)}\ne\bar{Q}^{(2)}, but the deci-

sion models are consistent, both estimators converge to ψin(a)\psi_{\mathrm{in}}(a), whereas the plug-in estimators converge to E{Q‾0(m)(H0)∣H0∈S}\mathbb{E}\{\overline{Q}^{(m)}_0(H_0)\mid H_0\in\mathcal{S}\}, m=1,2m=1,2, which differ in general.

Part (c) is the formal version of an empirical phenomenon: twins that fit the observed trajectories equally well, and that would pass any check based on prediction error [6], can imply different counterfactuals, as learned transport models did in Cao [5]. On supported histories the correction removes this disagreement to first order. Under misspecification of the twin, consistency of ψ^in\widehat{\psi}_{\mathrm{in}} rests on the decision models, and the rate is that of π^k\widehat{\pi}_k; with correctly specified parametric decision models it is n\sqrt{n}.

Remark 3 (Choice of threshold). A larger δ\delta shrinks S\mathcal{S}, stabilizes the weights, and moves more of ψ(a)\psi(a) into the part carried by the mechanism. Reporting ψ^in\widehat{\psi}_{\mathrm{in}}, the bounds of § 5, and pSp_{\mathcal{S}} over a grid of δ\delta makes this trade-off visible. We do not study asymptotics with δn→0\delta_n\to0.

The unsupported part: mechanism-indexed bounds

For a twin T\mathrm{T}, let ψout(T)=E{Q0T(H0)∣H0∉S}\psi_{\mathrm{out}}(\mathrm{T})=\mathbb{E}\{Q^{\mathrm{T}}_0(H_0)\mid H_0\notin\mathcal{S}\} be the value it implies on unsupported histories, and define

B(ϵ,M)=[inf⁡T∈T(ϵ,M)ψout(T), sup⁡T∈T(ϵ,M)ψout(T)].B(\epsilon,M)=\left[\inf_{\mathrm{T}\in\mathcal{T}(\epsilon,M)}\psi_{\mathrm{out}}(\mathrm{T}),\ \sup_{\mathrm{T}\in\mathcal{T}(\epsilon,M)}\psi_{\mathrm{out}}(\mathrm{T})\right].

Assumption 3 (Mechanism adequacy at level ϵ∗\epsilon^*). The potential trajectories under every strategy are generated, given the landmark history, by a twin T∗=(g∗,r∗)\mathrm{T}^*=(g^*,r^*) with g∗∈GM∗g^*\in\mathcal{G}_{M^*} and ∥r∗∥μ≤ϵ∗\lVert r^*\rVert_{\mu}\le\epsilon^*.

The level ϵ∗\epsilon^* is unknown. Assumption 3 with a small ϵ∗\epsilon^* is a strong statement about physiology; with a large ϵ∗\epsilon^* it is weak.

Proposition 2 (Properties of the bounds). Suppose Assumptions 1 and 2 hold.

(a) Monotonicity. If ϵ≤ϵ′\epsilon\le\epsilon' then B(ϵ,M)⊆B(ϵ′,M)B(\epsilon,M)\subseteq B(\epsilon',M); if GM⊆GM′\mathcal{G}_M\subseteq\mathcal{G}_{M'} then B(ϵ,M)⊆B(ϵ,M′)B(\epsilon,M)\subseteq B(\epsilon,M').

(b) Validity. Under Assumption 3, ψout(a)∈B(ϵ,M)\psi_{\mathrm{out}}(a)\in B(\epsilon,M) for every ϵ≥ϵ∗\epsilon\ge\epsilon^* and every MM with GM∗⊆GM\mathcal{G}_{M^*}\subseteq\mathcal{G}_M.

(c) Exact mechanism. Suppose the parameters are identified at ϵ=0\epsilon=0: any two twins in T(0,M)\mathcal{T}(0,M) have encoders that agree almost surely on the support of X among supported patients and in the pre-treatment window, and the support of X among unsupported patients is contained in that set. Then B(0,M)B(0,M) is a single point if T(0,M)\mathcal{T}(0,M) is non-empty. If T(0,M)\mathcal{T}(0,M) is empty, the exact mechanism with mask MM is refuted by the observed data.

(d) No mechanism. Let I0∗betheidentifiedsetfor∗\mathcal{I}_0 *be the identified set for*\psi_{\mathrm{out}}(a)$ under Assumptions 1 and 2 alone. Suppose the residual class is rich: for every c∈I0c\in\mathcal{I}_0 and every η>0\eta>0 there is a twin, compatible with PP and with finite residual size, whose out-of-support value lies within η\eta of cc. Then ⋃ϵB(ϵ,M)\bigcup_{\epsilon}B(\epsilon,M) is dense in I0\mathcal{I}_0, and B(ϵ,M)⊆I‾0B(\epsilon,M)\subseteq\overline{\mathcal{I}}_0 for every ϵ\epsilon.

Part (a) says that a stronger trust in the mechanism (smaller ϵ\epsilon) and stronger knowledge about how covariates enter it (a smaller mask) both narrow the bounds. Part (b) says that the narrowing is valid only if the trust is warranted. Parts (c) and (d) give the two ends of the scale: point identification by the mechanism, and the assumption-free set of Proposition 1(c). Reporting B(ϵ,M)B(\epsilon,M) as a function of ϵ\epsilon shows how strongly a conclusion about unsupported patients depends on the mechanism; this is the role that sensitivity parameters play in analyses of unmeasured confounding, here attached to the physics-informed penalty. Part (c) also supplies a falsification test: if no twin with small residual is compatible with the data, the mechanism with that residual size is rejected, in the spirit of Cornish et al. [10].

Combining with Proposition 1(a), the counterfactual mean satisfies

ψ(a)∈pSψin(a)+(1−pS)B(ϵ,M)for all ϵ≥ϵ∗.\psi(a) \in p_S\psi_{\mathrm{in}}(a) + (1-p_S)B(\epsilon,M) \quad\text{for all } \epsilon\ge\epsilon^*.

Computation and inference

The endpoints of B(ϵ,M)B(\epsilon,M) are values of optimization problems over twins, as identified sets for counterfactuals in structural dynamic models are Kalouptsidi et al. [22]. We approximate them with extremal twins (Algorithm 1): for each sign, the twin’s out-of-support value is pushed up or down subject to a fit constraint that admits near-minimizers of the empirical loss, in the spirit of set estimation with near-minimizers Chernozhukov et al. [9]. The fit threshold is set once, by the best fit in the richest residual class considered, so that the estimated sets are nested in ϵ\epsilon as in Proposition 2(a); if no twin with residual size ϵ\epsilon attains the threshold, the estimated set is empty and the mechanism with that residual size is rejected, the sample counterpart of Proposition 2(c). Confidence intervals for the endpoints can be obtained by the bootstrap and combined into an interval for the partially identified parameter Imbens and Manski [21].

In implementation, compatibility is imposed through the mean trajectory: a twin is compatible if it minimizes the population loss L(θ)\mathcal{L}(\theta), the expected squared error of predicted weights along the observed treatment paths, over the class Θϵ\Theta_\epsilon of twin parameters with residual size at most ϵ\epsilon. Write Θ0(ϵ)=arg⁡min⁡ΘϵL\Theta_0(\epsilon)=\arg\min_{\Theta_\epsilon}\mathcal{L} and Θ^(ϵ)={θ∈Θϵ:Ln(θ)≤inf⁡ΘϵLn+cn}\widehat{\Theta}(\epsilon)=\{\theta\in\Theta_\epsilon:\mathcal{L}_n(\theta)\le\inf_{\Theta_\epsilon}\mathcal{L}_n+c_n\}.

Proposition 3 (Consistency of extremal twins). Suppose that Θϵ\Theta_\epsilon is compact, that sup⁡Θϵ∣Ln−L∣=OP(an)\sup_{\Theta_\epsilon}|\mathcal{L}_n-\mathcal{L}|=O_P(a_n) with an→0a_n\to0, that θ↦ψout(θ)\theta\mapsto\psi_{\mathrm{out}}(\theta) is continuous on Θϵ\Theta_\epsilon and its sample analogue converges to it uniformly in probability, and that cn→0c_n\to0 with cn/an→∞c_n/a_n\to\infty. Then Θ^(ϵ)\widehat{\Theta}(\epsilon) converges to Θ0(ϵ)\Theta_0(\epsilon) in Hausdorff distance in probability, and the endpoints of {ψ^out(θ):θ∈Θ^(ϵ)}\{\widehat{\psi}_{\mathrm{out}}(\theta):\theta\in\widehat{\Theta}(\epsilon)\} converge in probability to min and max of ψout\psi_{\mathrm{out}} over Θ0(ϵ)\Theta_0(\epsilon).

Under Assumption 3, the true twin minimizes L\mathcal{L} whenever ϵ≥ϵ∗\epsilon\ge\epsilon^*, so Θ0(ϵ)\Theta_0(\epsilon) is the set of compatible twins in this mean sense, and the limits are the endpoints of B(ϵ,M)B(\epsilon,M). In the simulation, ana_n is of order n−1/2n^{-1/2} and we use the relative slack cn=22/Nc_n=2\sqrt{2/N}, with NN the number of weight measurements.

Algorithm 1 Extremal twins for B(ϵ,M)B(\epsilon, M)

Require: increasing grid ϵ1<⋯<ϵJ\epsilon_1 < \cdots< \epsilon_J, mask MM, relative slack cn>0c_n > 0

  1. Fit the richest twin T^rich\widehat{T}_{\mathrm{rich}} and set the threshold τn=(1+cn)Ln(T^rich)\tau_n = (1+c_n)\mathcal{L}_n(\widehat{T}_{\mathrm{rich}})

  2. for j=1,…,Jj = 1,\ldots,J do

  3. T^j=arg min{Ln(T):g∈GM, ∥r∥μ≤ϵj}\widehat{T}_j = \mathop{\mathrm{arg\,min}}\{\mathcal{L}_n(T): g \in\mathcal{G}_M,\ \lVert r\rVert_\mu\le\epsilon_j\}

  4. if Ln(T^j)>τn\mathcal{L}_n(\widehat{T}_j) > \tau_n then

  5. report B(ϵj,M)=∅B(\epsilon_j,M) = \varnothing and continue

  6. end if

  7. for s∈{−1,+1}s \in\{-1,+1\} do

  8. T^j,s=arg max{sψ^out(T):g∈GM, ∥r∥μ≤ϵj, Ln(T)≤τn}\widehat{T}_{j,s} = \mathop{\mathrm{arg\,max}}\{s\widehat{\psi}_{\mathrm{out}}(T): g \in\mathcal{G}_M,\ \lVert r\rVert_\mu\le\epsilon_j,\ \mathcal{L}_n(T) \le\tau_n\}, started at T^j\widehat{T}_j

  9. end for

  10. Record the endpoints, taking the envelope with those at ϵj−1\epsilon_{j-1}

  11. end for

Calibrating the residual size over time

The size ϵ\epsilon is a sensitivity parameter, but data can sometimes say which values are too small. Treatment practice changes over calendar time, for example through drug shortages or changes in insurance coverage, and patient types for whom a strategy was unsupported in an early era may follow it in a later one.

Let e∈{1,2}e \in\{1,2\} index two eras with observed-data laws P(1)P^{(1)} and P(2)P^{(2)} and supported sets S1⊆S2\mathcal{S}_1 \subseteq\mathcal{S}_2, and let V=S2∖S1\mathcal{V} = \mathcal{S}_2 \setminus\mathcal{S}_1 be the back-test region, unsupported in era 1 and supported in era 2.

Assumption 4 (Invariant dynamics). The law of the potential trajectories under each strategy, given the landmark history, is the same in both eras. The eras may differ in the distribution of histories, in the decision process, and in the visit process.

Let BV(1)(ϵ,M)B^{(1)}_{\mathcal{V}}(\epsilon,M) be the bounds computed from twins compatible with P(1)P^{(1)}, with the out-of-support value averaged over the era-2 distribution of histories in V\mathcal{V}, and let ψV=E(2){Q0(2)(H0)∣H0∈V}\psi_{\mathcal{V}} = \mathbb{E}^{(2)}\{Q^{(2)}_0(H_0)\mid H_0 \in\mathcal{V}\}, which is identified in era 2. Define

ϵcal=inf⁡{ϵ≥0:ψV∈BV(1)(ϵ,M)}.\epsilon_{\mathrm{cal}} = \inf\{\epsilon\ge0 : \psi_{\mathcal{V}} \in B^{(1)}_{\mathcal{V}}(\epsilon,M)\}.

Proposition 4 (Back-testing). Suppose Assumptions 1 and 2 hold in both eras and Assumptions 3 and 4 hold, with GM∗⊆GM\mathcal{G}_{M^*} \subseteq\mathcal{G}_M.

(a) ϵcal≤ϵ∗\epsilon_{\mathrm{cal}} \le\epsilon^*.

(b) Every ϵ<ϵcal\epsilon< \epsilon_{\mathrm{cal}} is inconsistent with Assumption 3 at level ϵ\epsilon.

(c) If, in addition, extrapolation to the remaining unsupported set U=S2c\mathcal{U} = \mathcal{S}_2^c requires no larger residual than extrapolation to V\mathcal{V}, in the sense that inf⁡{ϵ:ψoutU(a)∈BU(2)(ϵ,M)}≤ϵcal\inf\{\epsilon: \psi_{\mathrm{out}}^{\mathcal{U}}(a) \in B_{\mathcal{U}}^{(2)}(\epsilon,M)\} \le\epsilon_{\mathrm{cal}}, then ψoutU(a)∈BU(2)(ϵ,M)\psi_{\mathrm{out}}^{\mathcal{U}}(a) \in B_{\mathcal{U}}^{(2)}(\epsilon,M) for every ϵ>ϵcal\epsilon> \epsilon_{\mathrm{cal}}, and also at ϵ=ϵcal\epsilon= \epsilon_{\mathrm{cal}} if the infimum is attained.

Back-testing therefore gives a lower bound on the residual size that the data require, not an upper bound. Parts (a) and (b) need only invariance; part (c) needs an additional assumption that the back-test region is at least as hard to extrapolate to as the target region, which cannot be checked. In the simulation design below it holds by construction.

In practice ψV\psi_V is estimated by (2) in era 2 and the bounds by Algorithm 1 in era 1. For each ϵ\epsilon in a grid ϵ1<⋯<ϵJ\epsilon_1 < \cdots< \epsilon_J, consider the null hypothesis H0(ϵ):ψV∈Bν(1)(ϵ,M)H_0(\epsilon): \psi_V \in B_\nu^{(1)}(\epsilon, M), rejected when the Wald interval ψ^V±z1−αse^\hat{\psi}_V \pm z_{1-\alpha}\widehat{\mathrm{se}} does not intersect the estimated bounds, and let ϵ^L\hat{\epsilon}_L be the smallest grid value at which H0H_0 is not rejected.

Corollary 1 (Lower confidence bound). Under the conditions of Proposition 4, Theorem 1(b) applied in era 2, and Proposition 3 applied in era 1, let ϵj∗\epsilon_{j^*} be the smallest grid value with ϵj∗≥ϵcal\epsilon_{j^*} \ge\epsilon_{\mathrm{cal}}, and suppose ψV\psi_V lies in the interior of Bν(1)(ϵj∗,M)B_\nu^{(1)}(\epsilon_{j^*}, M). Then lim inf⁡nP(ϵ^L≤ϵj∗)≥1−α\liminf_n \mathbb{P}(\hat{\epsilon}_L \le\epsilon_{j^*}) \ge1-\alpha.

Combined with Proposition 4(a), ϵ^L\hat{\epsilon}_L is an asymptotic lower confidence bound for the grid level just above ϵcal\epsilon_{\mathrm{cal}}, and hence for nothing larger than ϵ∗\epsilon^* rounded up to the grid.

Simulation study {#sec:7}

The study follows the ADEMP structure ([29]); the code is a single script described in Appendix § 13. All results below use ω=1\omega= 1, the setting in which the mechanism omits part of the dynamics.

Aims. (i) To verify that, on supported histories, twins that disagree as plug-in estimators agree after correction (Theorem 1); (ii) to evaluate the validity and width of B(ϵ,M)B(\epsilon, M) on unsupported histories as functions of ϵ\epsilon and of the mask, and the rejection of too small ϵ\epsilon (Proposition 2); (iii) to evaluate back-testing (Proposition 4 and Corollary 1); and (iv) to compare with plug-in twins, the mechanistic core alone, and a mechanism-free g-computation estimator.

Data-generating mechanism. Baseline covariates are age, sex (60% female), body mass index, glycated haemoglobin, and a genetic risk score for type 2 diabetes, standardized except sex; weight at initiation is max⁡{70,100+15⋅BMIz}\max\{70, 100 + 15 \cdot\mathrm{BMI}_z\} kg. Gastrointestinal symptoms SS at the landmark are binary with probability expit⁡(−0.8+0.3⋅female)\operatorname{expit}(-0.8 + 0.3 \cdot\mathrm{female}), and a cost shock CC, recorded in the landmark history, has probability 0.10 in era 1 and 0.35 in era 2. Weight change follows the energy-balance model of § 3 on a weekly grid. The energy density and the expenditure slope are sex-specific values of the linearized NIH Body Weight Planner model, ρ=9,383\rho= 9{,}383 and 9,916 kcal/kg and βE=28\beta_E = 28 and 24 kcal/day per kg for men and women ([14]). Individual parameters are

Pi=P0exp⁡(0.15BMIi+0.10femalei−0.15HbA1ci−0.10PRSi+0.15Si),kioff=95exp⁡(0.10agei−0.08femalei),kion=(1−wi)kioff,wi=0.4+0.3expit⁡(0.8femalei−0.6HbA1ci),\begin{aligned} P_i &= P_0 \exp(0.15\mathrm{BMI}_i + 0.10\mathrm{female}_i - 0.15\mathrm{HbA1c}_i - 0.10\mathrm{PRS}_i + 0.15S_i), \\ k_i^{\mathrm{off}} &= 95\exp(0.10\mathrm{age}_i - 0.08\mathrm{female}_i), \qquad k_i^{\mathrm{on}} = (1-w_i)k_i^{\mathrm{off}}, \\ w_i &= 0.4 + 0.3\operatorname{expit}(0.8\mathrm{female}_i - 0.6\mathrm{HbA1c}_i), \end{aligned}

so that the baseline feedback strength matches the value estimated by Hall [15] and treatment weakens it by 40–70%. These coefficients define the true mask M∗M^*: PP depends on sex, BMI,

HbA1c, genetic risk, and symptoms; ww on sex and HbA1c; and koffk^{\mathrm{off}} on age and sex. Drug exposure is u=1u=1 on treatment and decays with time constant τr\tau_r after stopping. Variation in the parameters is explained by the covariates, so that the encoder can represent it; unexplained variation enters through the intake noise below. The design does not include random effects in the parameters, which would require updating individual parameters from early weights. We set P0=800P_0=800 kcal/day, which gives mean total weight loss of 10.3% at 26 weeks and 14.4% at 78 weeks under continuation, close to the 15% reported at 68 weeks in STEP 1 ([38]), and τr=14\tau_r=14 weeks, which, together with the residual described next, gives a median regain of 59% of the lost weight 52 weeks after stopping, close to the 60% estimated by [4]. Intake fluctuates weekly with standard deviation 100 kcal/day, and weights are recorded at weeks 4, 8, 12, 16, 20, and 26 and every 8 weeks thereafter up to week 78, with measurement error of standard deviation 1 kg.

The true dynamics include a residual that the mechanistic core omits: after a patient first stops, intake stays lower by an amount proportional to the weight lost by the landmark, r∗=ωγxτ(1−u)r^*=\omega\gamma x_\tau(1-u) with γ=40\gamma=40 kcal/day per kg, a stand-in for behavioural changes learned during treatment. Without it, the core predicts that all lost weight is eventually regained; with it, regain levels off below the pre-treatment weight, as in the meta-regression of [4]. This residual lies in the twin’s residual class, so Assumption 3 holds with ϵ∗=∥r∗∥μ\epsilon^*=\lVert r^*\rVert_\mu, about 250 kcal/day and computed in each replicate. The dial ω∈{0,1}\omega\in\{0,1\} makes the mechanism exact (ω=0\omega=0) or not (ω=1\omega=1).

At the landmark τ=26\tau=26 weeks, the probability of stopping is expit⁡{−1.5+1.2S+1.0C−0.15(early loss−8)}\operatorname{expit}\{-1.5+1.2S+1.0C-0.15(\text{early loss}-8)\}, with early loss the observed percentage weight loss. In era 1 it is zero for patients with early loss of at least 12% and no symptoms, who make up 10% of the cohort and form the unsupported set for the stopping strategy. In era 2, a cost shock triggers stopping with probability 0.5 among these patients, so that those with a cost shock form the back-test region VV and those without one remain unsupported. Because the cost shock is independent of the dynamics, the condition of Proposition 4(c) holds in this design. Every 8 weeks after the landmark, patients on treatment stop with probability expit⁡(−3+C+0.5S)\operatorname{expit}(-3+C+0.5S), and patients off treatment restart with probability min⁡{0.9,expit⁡(−3.5+0.15 regain)}\min\{0.9,\operatorname{expit}(-3.5+0.15\,\text{regain})\}, where regain is the weight regained in kilograms since stopping. Restarting in response to regain creates time-varying confounding, and its frequency is comparable to the roughly 20% of patients who restarted within a year in routine care ([13]). Each era has n=2,000n=2{,}000 patients and each setting uses 200 replicates.

Estimands. The counterfactual means of the percentage of total weight loss at 78 weeks under continuation, ψ(c)\psi(c), and under stopping at the landmark without restarting, ψ(s)\psi(s), with the supported and unsupported parts of ψ(s)\psi(s). True values are computed by simulating 60,000 patients under each strategy.

Methods. The twin has log-linear and logit-linear encoders for PP, ww, and koffk^{\mathrm{off}}, an estimated recovery time constant, and the residual (1−u)(b0+b⊤z+bxxτ)(1-u)(b_0+b^\top z+b_xx_\tau) of § 3, with zz the five covariates and SS. Decision models are logistic regressions. The methods are: (M1) the mechanistic core without residual, as a plug-in estimator and with the correction (2); (M2) hybrid twins with penalty weights λ∈{10−7,10−5}\lambda\in\{10^{-7}, 10^{-5}\} and an unrestricted encoder, and (M3) a hybrid twin with the true mask, each as a plug-in estimator and with the correction; (M4) iterated linear regressions among strategy followers, a mechanism-free g-computation estimator that extrapolates linearly to unsupported patients; and (M5) the bounds B(ϵ,M)B(\epsilon, M) of Algorithm 1 on unsupported histories, with unrestricted and true masks, over ϵ∈{0,0.1,0.25,0.5,1,2}×ϵ∗\epsilon\in\{0, 0.1, 0.25, 0.5, 1, 2\} \times\epsilon^*. Corrections use two-fold cross-fitting with the same folds for all twins. Back-testing compares the era-2 corrected estimate on V\mathcal{V} with era-1 bounds evaluated on the era-2 histories in V\mathcal{V}, and reports ϵ^L\hat{\epsilon}_{\mathrm{L}} of Corollary 1 with α=0.025\alpha= 0.025.

Performance measures. On supported histories: bias, mean estimated standard error, and coverage of 95% intervals for each twin, with and without correction, and the standard deviation across twins within a replicate before and after correction. On unsupported histories: coverage of ψout(s)\psi_{\mathrm{out}}(s) by B(ϵ,M)B(\epsilon, M), its width, and the share of replicates in which the set is empty. For back-testing: the distribution of ϵ^L/ϵ∗\hat{\epsilon}_{\mathrm{L}}/\epsilon^* and the share of replicates with ϵ^L≤ϵ∗\hat{\epsilon}_{\mathrm{L}} \le\epsilon^*. Monte Carlo standard errors are given in the table notes.

Results

All results are based on 200 replicates with n=2,000n = 2{,}000 patients per era. On average, 10.6% of patients were unsupported for the stopping strategy, and the true residual size was ϵ∗≈249\epsilon^* \approx249 kcal/day.

Supported histories. Table 1 shows that the four twins, all fitted to the same data, gave plug-in estimates whose biases ranged from −0.06-0.06 to −1.37-1.37 percentage points of total weight loss. The mechanistic core and the strongly penalized hybrid were most biased because they cannot represent the intake reduction retained after stopping and therefore predict too much regain. After the correction (2), all biases were at most 0.04 points and the within-replicate standard deviation across twins fell from 0.656 to 0.029, the behaviour described in Theorem 1(c). The two misspecified twins kept a small bias (0.042 and 0.030, about a third of their standard errors), as expected when consistency rests on the decision models alone (Theorem 1(a)). Coverage of nominal 95% intervals was 0.975 to 0.995, so the variance estimator was conservative; in the first 50 replicates, the empirical standard deviations of the corrected estimates were 10–20% below the mean estimated standard errors. This is consistent with the known effect of treating estimated weights as known in inverse-probability-weighted estimators (Lunceford and Davidian, [26]). For continuation, which is supported for every patient, the corrected estimate had bias −0.014-0.014 and coverage 0.950. The mechanism-free iterated regression estimator was also nearly unbiased on supported histories (bias −0.056-0.056): where the data support a strategy, the mechanism is not needed.

Table 1: Supported histories, strategy “stop at the landmark”: bias of plug-in and corrected estimates of ψin(s)\psi_{\mathrm{in}}(s) (true value 4.03 percentage points of total weight loss at 78 weeks), ω=1\omega= 1, 200 replicates.

TwinPlug-in biasCorrected biasMean SECoverage
Mechanistic core, no residual−1.3690.0420.1160.995
Hybrid, λ=10−7\lambda= 10^{-7}−0.0760.0050.0750.975
Hybrid, λ=10−5\lambda= 10^{-5}−0.9690.0300.0990.990
Hybrid with true mask, λ=10−7\lambda= 10^{-7}−0.0640.0060.0750.980
SD across the four twins0.6560.029

SE, estimated standard error; SD, mean within-replicate standard deviation of the four estimates. Monte Carlo standard errors are about 0.005–0.008 for the corrected biases and 0.005–0.011 for the coverages.

Table 1.

Unsupported histories. The true value on unsupported histories was 5.67 points, and the mechanism-free estimator, which extrapolates linearly, was biased by −0.49-0.49 points while reporting a single number. Table 2 shows the bounds. For ϵ≤0.25ϵ∗\epsilon\le0.25\epsilon^*, no twin met the fit threshold in any replicate, so the estimated set was empty: the data rejected the exact mechanism and every residual smaller than a quarter of the true one, the sample counterpart of Proposition 2(c). For every residual size that was not rejected, the bounds covered the true value in at least 99.5% of replicates. Their width increased with ϵ\epsilon, from 2.23 to 3.02 points with an unrestricted encoder, as Proposition 2(a) requires, and the true mask narrowed them by 25% at ϵ=0.5ϵ∗\epsilon= 0.5\epsilon^* and by about 10% at larger values.

Table 2: Unsupported histories, strategy “stop at the landmark”: mechanism-indexed bounds B(ϵ,M)B(\epsilon, M) for ψout(s)\psi_{\mathrm{out}}(s) (true value 5.67), ω=1\omega= 1, 200 replicates.

ϵ/ϵ∗\epsilon/\epsilon^*Unrestricted encoderTrue mask M∗M^*
CoverageWidthEmptyCoverageWidthEmpty
00.000–1.0000.000–1.000
0.100.000–1.0000.000–1.000
0.250.000–1.0000.000–1.000
0.501.0002.2340.0000.9951.6830.000
1.001.0002.8490.0001.0002.5510.000
2.001.0003.0160.0001.0002.7400.000

Coverage, share of replicates in which the estimated set contains the true value (an empty set does not); width in percentage points, averaged over non-empty sets; empty, share of replicates in which no twin met the fit threshold. Mechanism-free iterated regression: bias −0.49.

Table 2.

Back-testing. The calibrated level was ϵ^L=0.5ϵ∗\hat{\epsilon}_{L} = 0.5\epsilon^* in every replicate, so it never exceeded ϵ∗\epsilon^*, as Proposition 4(a) requires. In this design, however, back-testing added nothing to the fit-based rejection: because the retained intake reduction acts on all patients who stop, the era-1 data already determine its size, and ϵ^L\hat{\epsilon}_{L} coincides with the smallest residual size that the era-1 data do not reject. Back-testing is informative when the omitted mechanism acts mainly on patients for whom the strategy is unsupported, so that a small residual fits the era-1 data but extrapolates wrongly. The present design does not include this case.

Discussion

A digital twin that combines a mechanism with learned components returns an answer for every strategy, and this paper separates the parts of that answer by their source. Where patients followed the strategy, the twin serves as an outcome model inside a doubly robust estimator, and the conclusion does not depend on which of several equally well-fitting twins is used. Where they did not, the twin’s answer is reported as a set indexed by how far the twin may depart from the mechanism, and the physics-informed penalty that controls this departure becomes an explicit sensitivity parameter. When practice changes over time, back-testing tells which departures are too small to be consistent with the data.

The simulation illustrates both parts of this division. Four twins that fit the same data disagreed by more than a percentage point of weight loss as plug-in estimators, and the correction brought them to within a few hundredths of a point of each other and of the truth. On patients whom the data did not cover, a mechanism-free extrapolation was off by half a point and gave no sign of it, whereas the bounds were empty when the residual was too small to fit the data and covered the truth once it was not; knowledge constraints on the encoder narrowed them. Two findings qualify these results. The variance estimator of the correction was conservative, so coverage exceeded the nominal level; and in the present design back-testing could not improve on fit-based rejection, because the omitted mechanism also acted on supported patients.

The framework complements our earlier work. Cao [5] showed that learned models with the same predictive accuracy can imply different causal conclusions; Theorem 1 shows when such disagreement disappears and Proposition 2 quantifies it when it does not. Cao [6] showed that prediction error is a poor guide to causal performance; accordingly, no result here relies on the predictive accuracy of the twin. Cao [7] studied when a learned representation is an adequate adjustment set; here the adjustment set is the observed history, and the mechanism enters only through the model for outcome dynamics.

The work has limitations. Sequential exchangeability given the observed history is assumed; reasons for stopping such as cost and insurance coverage may be incompletely recorded, and sensitivity analysis for unmeasured confounding of the decision process is a natural extension. The overlap gap of Assumption 2 simplifies the treatment of histories with small but positive decision probabilities. Individual mechanistic parameters may be weakly identified from sparse clinical weights, a difficulty already noted for energy-balance models ([15]); the population estimands studied here are less sensitive to this than individual forecasts. Back-testing relies on invariant dynamics across eras, and its coverage guarantee on the remaining unsupported set relies on an assumption that cannot be checked. Finally, the framework has not yet been applied to clinical data.

Data and Code Availability

No data were analyzed in this version. The simulation code and instructions for reproducing § 7.1 are available at https://github.com/congcao/causal-digital-twins.

Funding

The author received no specific funding for this work.

Competing Interests

The author declares no competing interests.

Proofs

Proposition 1. Part (a) is the law of total expectation. For (b), fix h0∈Sh_0 \in\mathcal{S}. By Assumption 2, πk≥δ\pi_k \ge\delta along every continuation reachable under aˉ\bar{a}, so each QkQ_k is well defined on those histories. Backward induction using consistency and sequential exchangeability [33] gives E(Yaˉ∣Hˉk=hˉk,Aˉk−1=aˉk−1)=Qk(hˉk)\mathbb{E}(Y^{\bar{a}} \mid\bar{H}_k = \bar{h}_k, \bar{A}_{k-1} = \bar{a}_{k-1}) = Q_k(\bar{h}_k) for reachable hˉk\bar{h}_k; at k=0k = 0 this is E(Yaˉ∣H0=h0)=Q0(h0)\mathbb{E}(Y^{\bar{a}} \mid H_0 = h_0) = Q_0(h_0), and averaging over H0∈SH_0 \in\mathcal{S} gives the result. For (c), let N={h0∉S:π0(h0)=0}\mathcal{N} = \{h_0 \notin\mathcal{S} : \pi_0(h_0) = 0\}, which has positive probability. On N\mathcal{N} the observed-data law contains no information on the conditional law of YaˉY^{\bar{a}}, because A0≠aA_0 \ne a almost surely. Given any law P†P^\dagger compatible with the observed data and the assumptions, modify the conditional law of the potential trajectories under aˉ\bar{a} given H0H_0 on N\mathcal{N} only, independently of the observed treatments, keeping YaˉY^{\bar{a}} within [ylo,yhi][y_{\mathrm{lo}}, y_{\mathrm{hi}}]. The observed-data law and consistency are unaffected, because potential outcomes under aˉ\bar{a} are never observed on N\mathcal{N}. Exchangeability at k=0k = 0 is preserved because the modified law does not depend on A0A_0, and exchangeability at k≥1k \ge1 conditions on A0=aA_0 = a, an event of probability zero on N\mathcal{N}. Under the assumption of part (c), N=Sc\mathcal{N} = \mathcal{S}^c. Taking the modified law of YaˉY^{\bar{a}} given H0H_0 on N\mathcal{N} to be a point mass at any c∈[ylo,yhi]c \in[y_{\mathrm{lo}}, y_{\mathrm{hi}}] gives ψout(a)=c\psi_{\mathrm{out}}(a) = c, so every point of the interval is attained; conversely ψout(a)∈[ylo,yhi]\psi_{\mathrm{out}}(a) \in[y_{\mathrm{lo}}, y_{\mathrm{hi}}] because YaˉY^{\bar{a}} takes values there. Hence the interval is the sharp identified set.

Lemma 1. Write Δk=Q^k−Qk\Delta_k = \widehat{Q}_k - Q_k, with ΔK=0\Delta_K = 0, and Q~k(hˉk)=E{Q^k+1∣Hˉk=hˉk,Ak=a}\widetilde{Q}_k(\bar{h}_k) = \mathbb{E}\{\widehat{Q}_{k+1} \mid\bar{H}_k = \bar{h}_k, A_k = a\}, so that Q~k−Qk=E{Δk+1∣Hˉk,Ak=a}\widetilde{Q}_k - Q_k = \mathbb{E}\{\Delta_{k+1} \mid\bar{H}_k, A_k = a\}. All expectations below include the factor 1(H0∈S)\mathbb{1}(H_0 \in\mathcal{S}), which is a function of H0H_0 and is omitted from the notation. Because W^k\widehat{W}_k is non-zero only if Ak=aA_k = a, and the nuisance estimates are fixed,

E{W^k(Q^k+1−Q^k)}=E{W^k(Q~k−Q^k)}=E{W^k−1πkπ^k(Q~k−Q^k)}.\mathbb{E}\{\widehat{W}_k(\widehat{Q}_{k+1} - \widehat{Q}_k)\} = \mathbb{E}\{\widehat{W}_k(\widetilde{Q}_k - \widehat{Q}_k)\} = \mathbb{E}\left\{\widehat{W}_{k-1}\frac{\pi_k}{\widehat{\pi}_k}(\widetilde{Q}_k - \widehat{Q}_k)\right\}.

where the second equality integrates over AkA_k given Hˉk\bar{H}_k. Writing Q~k−Q^k=(Q~k−Qk)−Δk\widetilde{Q}_k - \widehat{Q}_k = (\widetilde{Q}_k - Q_k) - \Delta_k and using E{W^k−1(πk/π^k)(Q~k−Qk)}=E(W^kΔk+1)\mathbb{E}\{\widehat{W}_{k-1}(\pi_k/\widehat{\pi}_k)(\widetilde{Q}_k - Q_k)\} = \mathbb{E}(\widehat{W}_k\Delta_{k+1}),

E(ϕ^)−E(Q0)=E(Δ0)+∑k=0K−1E(W^kΔk+1)−∑k=0K−1E(W^k−1πkπ^kΔk).\mathbb{E}(\widehat{\phi}) - \mathbb{E}(Q_0) = \mathbb{E}(\Delta_0) + \sum_{k=0}^{K-1}\mathbb{E}(\widehat{W}_k\Delta_{k+1}) - \sum_{k=0}^{K-1}\mathbb{E}\left(\widehat{W}_{k-1}\frac{\pi_k}{\widehat{\pi}_k}\Delta_k\right).

Shifting the index in the first sum and using ΔK=0\Delta_K = 0 and W^−1=1\widehat{W}_{-1} = 1 gives ∑k=0K−1E{W^k−1(1−πk/π^k)Δk}\sum_{k=0}^{K-1} \mathbb{E}\{\widehat{W}_{k-1}(1-\pi_k/\widehat{\pi}_k)\Delta_k\}, which is the stated expression.

Theorem 1. Write s=I(H0∈S)s = \mathbb{I}(H_0 \in\mathcal{S}), φ\varphi for φ^\widehat{\varphi} evaluated at the true nuisances, so that Din=s(φ−ψin)/pSD_{\mathrm{in}} = s(\varphi-\psi_{\mathrm{in}})/p_S, and P\mathbb{P} for expectation over a new observation with the nuisance estimates held fixed. With JJ folds of sizes njn_j, the estimator is ψ^in=Pn(sφ^)/Pn(s)\widehat{\psi}_{\mathrm{in}} = \mathbb{P}_n(s\widehat{\varphi})/\mathbb{P}_n(s), so

ψ^in−ψin=Pn{s(φ^−ψin)}Pn(s),\widehat{\psi}_{\mathrm{in}}-\psi_{\mathrm{in}}=\frac{\mathbb{P}_n\{s(\widehat{\varphi}-\psi_{\mathrm{in}})\}}{\mathbb{P}_n(s)},
Pn{s(φ^−ψin)}=(Pn−P){s(φ−ψin)}+∑jnjn(Pn,j−P){s(φ^j−φ)}+∑jnjnRj,\mathbb{P}_n\{s(\widehat{\varphi}-\psi_{\mathrm{in}})\}=(\mathbb{P}_n-\mathbb{P})\{s(\varphi-\psi_{\mathrm{in}})\}+\sum_j\frac{n_j}{n}(\mathbb{P}_{n,j}-\mathbb{P})\{s(\widehat{\varphi}_j-\varphi)\}+\sum_j\frac{n_j}{n}R_j,

where φ^j\widehat{\varphi}_j uses nuisances fitted outside fold jj, Pn,j\mathbb{P}_{n,j} is the empirical measure of fold jj, and Rj=P{s(φ^j−ψin)}=P{s(φ^j−Q0)}R_j=\mathbb{P}\{s(\widehat{\varphi}_j-\psi_{\mathrm{in}})\}=\mathbb{P}\{s(\widehat{\varphi}_j-Q_0)\} because E(sQ0)=pSψin\mathbb{E}(sQ_0)=p_S\psi_{\mathrm{in}}; RjR_j is the remainder of Lemma 1 for the fold-jj nuisances. Conditionally on the data outside fold jj, the middle term has mean zero and variance at most ∥s(φ^j−φ)∥2/nj\|s(\widehat{\varphi}_j-\varphi)\|^2/n_j, so it is OP(∥φ^j−φ∥/n)O_P(\|\widehat{\varphi}_j-\varphi\|/\sqrt{n}) by Chebyshev’s inequality ([23]). Since YY is bounded and π^k,πk≥δ/2\widehat{\pi}_k,\pi_k\ge\delta/2 on supported histories, the weights satisfy W^k≤(2/δ)k+1\widehat{W}_k\le(2/\delta)^{k+1} and ∥φ^j−φ∥≤C∑k(∥Q^k−Qk∥+∥π^k−πk∥)\|\widehat{\varphi}_j-\varphi\|\le C\sum_k(\|\widehat{Q}_k-Q_k\|+\|\widehat{\pi}_k-\pi_k\|) for a constant CC depending on δ\delta, KK, and the outcome range; the middle term is therefore oP(n−1/2)o_P(n^{-1/2}) under the L2L_2 convergence in (b) and oP(1)o_P(1) under the convergence to limits in (a). By the Cauchy–Schwarz inequality and W^k−1≤(2/δ)k\widehat{W}_{k-1}\le(2/\delta)^k,

∣Rj∣≤∑k=0K−1(2/δ)k+1∥π^k−πk∥∥Q^k−Qk∥≤Cδ,K∑k=0K−1∥π^k−πk∥∥Q^k−Qk∥,Cδ,K=(2/δ)K,|R_j|\le\sum_{k=0}^{K-1}(2/\delta)^{k+1}\|\widehat{\pi}_k-\pi_k\|\|\widehat{Q}_k-Q_k\|\le C_{\delta,K}\sum_{k=0}^{K-1}\|\widehat{\pi}_k-\pi_k\|\|\widehat{Q}_k-Q_k\|,\qquad C_{\delta,K}=(2/\delta)^K,

with norms in L2L_2 over supported histories along which the strategy has been followed. Finally, Pn(s)=pS+OP(n−1/2)\mathbb{P}_n(s)=p_S+O_P(n^{-1/2}) and the numerator is OP(n−1/2)O_P(n^{-1/2}) in case (b), so the ratio equals the numerator divided by pSp_S up to oP(n−1/2)o_P(n^{-1/2}). For (a), each term of the bound converges to ∥π‾k−πk∥∥Q‾k−Qk∥=0\|\overline{\pi}_k-\pi_k\|\|\overline{Q}_k-Q_k\|=0, and the other terms are oP(1)o_P(1). For (b), ψ^in−ψin=(Pn−P)Din+oP(n−1/2)\widehat{\psi}_{\mathrm{in}}-\psi_{\mathrm{in}}=(\mathbb{P}_n-\mathbb{P})D_{\mathrm{in}}+o_P(n^{-1/2}) and the central limit theorem applies. Part (c) follows by applying (b) to each procedure, which yields the same leading term (Pn−P)Din(\mathbb{P}_n-\mathbb{P})D_{\mathrm{in}}, and, under misspecification of the twins, by applying (a) with π‾k=πk\overline{\pi}_k=\pi_k; the plug-in estimator Pn(sQ^0)/Pn(s)\mathbb{P}_n(s\widehat{Q}_0)/\mathbb{P}_n(s) converges to E{Q‾0(m)(H0)∣H0∈S}\mathbb{E}\{\overline{Q}^{(m)}_0(H_0)\mid H_0\in\mathcal{S}\}.

Proposition 2. (a) Rϵ⊆Rϵ′\mathcal{R}_{\epsilon}\subseteq\mathcal{R}_{\epsilon'} and GM⊆GM′\mathcal{G}_M\subseteq\mathcal{G}_{M'} imply T(ϵ,M)⊆T(ϵ′,M′)\mathcal{T}(\epsilon,M)\subseteq\mathcal{T}(\epsilon',M'), and the infimum and supremum over a larger set are weakly more extreme. (b) Under Assumptions 1 and 3, T∗\mathcal{T}^* generates both the potential and the observed trajectories, so it is compatible with PP; it belongs to T(ϵ,M)\mathcal{T}(\epsilon,M) whenever ϵ≥ϵ∗\epsilon\ge\epsilon^* and GM∗⊆GM\mathcal{G}_{M^*}\subseteq\mathcal{G}_M, and ψout(T∗)=ψout(a)\psi_{\mathrm{out}}(\mathcal{T}^*)=\psi_{\mathrm{out}}(a). (c) At ϵ=0\epsilon=0 every twin is determined by its encoder. By assumption, compatible encoders agree almost surely on the support of XX among unsupported patients, so they imply the same post-landmark law under a‾\overline{a} for those patients and the same ψout(T)\psi_{\mathrm{out}}(\mathcal{T}). If no twin is compatible, the observed data are inconsistent with Assumption 3 at ϵ∗=0\epsilon^*=0 with mask MM. (d) Every compatible twin, combined with the

observed decision process, defines a law that reproduces PP and satisfies Assumptions 1 and 2, because the twin generates potential trajectories from H0H_0 without reference to later decisions, so that sequential exchangeability holds by construction. Hence ψout(T)∈I0\psi_{\mathrm{out}}(\mathrm{T}) \in\mathcal{I}_0 for every compatible twin, and B(ϵ,M)⊆I0B(\epsilon, M) \subseteq\mathcal{I}_0. Conversely, for c∈I0c \in\mathcal{I}_0 and η>0\eta> 0, richness gives a compatible twin with finite residual size ϵc,η\epsilon_{c,\eta} and ∣ψout(T)−c∣<η\lvert\psi_{\mathrm{out}}(\mathrm{T}) - c\rvert< \eta; this twin belongs to T(ϵ,M)\mathcal{T}(\epsilon, M) for ϵ≥ϵc,η\epsilon\ge\epsilon_{c,\eta}, so ⋃ϵB(ϵ,M)\bigcup_{\epsilon} B(\epsilon, M) meets every neighbourhood of every point of I0\mathcal{I}_0.

Proposition 4. (a) By Assumption 4, T∗\mathrm{T}^* generates the potential trajectories in both eras, so it is compatible with P(1)P^{(1)} and belongs to the era-1 set at every ϵ≥ϵ∗\epsilon\ge\epsilon^*. Under the era-2 identification assumptions, ψV=E(2){QT∗(2)(H0)∣H0∈V}\psi_V = \mathbb{E}^{(2)}\{Q_{\mathrm{T}^*}^{(2)}(H_0) \mid H_0 \in\mathcal{V}\}, which is the value implied by T∗\mathrm{T}^* and therefore lies in BV(1)(ϵ∗,M)B_V^{(1)}(\epsilon^*, M). Hence ϵ∗\epsilon^* belongs to the set whose infimum defines ϵcal\epsilon_{\mathrm{cal}}. (b) By monotonicity, ψV∉BV(1)(ϵ,M)\psi_V \notin B_V^{(1)}(\epsilon, M) for every ϵ<ϵcal\epsilon< \epsilon_{\mathrm{cal}}, and by validity this is impossible if Assumption 3 held at level ϵ\epsilon. (c) Let ϵU\epsilon_U denote the infimum in the stated condition, so ϵU≤ϵcal\epsilon_U \le\epsilon_{\mathrm{cal}}. For ϵ>ϵcal\epsilon> \epsilon_{\mathrm{cal}} there is ϵ′∈[ϵU,ϵ)\epsilon' \in[\epsilon_U,\epsilon) with ψoutU(a)∈BU(2)(ϵ′,M)\psi_{\mathrm{out}}^U(a) \in B_U^{(2)}(\epsilon', M), and monotonicity gives membership at ϵ\epsilon. If the infimum ϵU\epsilon_U is attained, the same argument applies at ϵ=ϵcal\epsilon= \epsilon_{\mathrm{cal}}.

Proposition 3. Let L∗=min⁡θ∈ΘL\mathcal{L}^* = \min_{\theta\in\Theta} \mathcal{L}. On the event sup⁡∣Ln−L∣≤cn/2\sup\lvert\mathcal{L}_n - \mathcal{L}\rvert\le c_n/2, whose probability tends to one because cn/an→∞c_n/a_n \to\infty, every θ∈Θ0(ϵ)\theta\in\Theta_0(\epsilon) satisfies Ln(θ)≤L∗+cn/2≤inf⁡Ln+cn\mathcal{L}_n(\theta) \le\mathcal{L}^* + c_n/2 \le\inf\mathcal{L}_n + c_n, so Θ0(ϵ)⊆Θ^(ϵ)\Theta_0(\epsilon) \subseteq\widehat{\Theta}(\epsilon); and every θ∈Θ^(ϵ)\theta\in\widehat{\Theta}(\epsilon) satisfies L(θ)≤L∗+2cn\mathcal{L}(\theta) \le\mathcal{L}^* + 2c_n. By compactness and continuity of L\mathcal{L}, for every η>0\eta> 0 there is κ>0\kappa> 0 such that L(θ)≤L∗+κ\mathcal{L}(\theta) \le\mathcal{L}^* + \kappa implies dist⁡{θ,Θ0(ϵ)}<η\operatorname{dist}\{\theta, \Theta_0(\epsilon)\} < \eta; since cn→0c_n \to0, Θ^(ϵ)\widehat{\Theta}(\epsilon) lies in the η\eta-neighbourhood of Θ0(ϵ)\Theta_0(\epsilon) with probability tending to one. This is Hausdorff convergence (compare [9 Theorem 3.1]. Uniform continuity of ψout\psi_{\mathrm{out}} on the compact set and uniform convergence of its sample analogue then give convergence of the endpoints.

Corollary 1. At ϵj∗\epsilon_j^* the null hypothesis holds by Proposition 4 and monotonicity. By Proposition 3 the estimated bounds at ϵj∗\epsilon_j^* converge to their population values, which contain ψV\psi_V in their interior, and by Theorem 1(b) the Wald interval covers ψV\psi_V with probability tending to 1−α1-\alpha or more. On the intersection of these events the interval intersects the estimated bounds, H0(ϵj∗)H_0(\epsilon_j^*) is not rejected, and ϵ^L≤ϵj∗\widehat{\epsilon}_L \le\epsilon_j^*.

Simulation details

The simulation is implemented in a single Python script, twin_sim.py, using NumPy, SciPy, pandas, and scikit-learn. All constants of Section 7 are set at the top of the script, and python twin_sim.py --calibrate prints the calibration targets reported there. Twins are fitted by nonlinear least squares with the penalty of Remark 1. For the extremal twins of Algorithm 1, the constraint ∥r∥μ<ϵ\lVert r\rVert_\mu< \epsilon is imposed by writing the residual coefficients as b=ϵv/(1+vTGv)1/2b = \epsilon v/(1+v^\mathsf{T}Gv)^{1/2}, with GG the Gram matrix of § 3, and each endpoint is found by continuation: an added least-squares term pulls the out-of-support value toward successive targets, and the last target reached within the fit threshold, with relative slack cn=22/Nc_n = 2\sqrt{2/N}, is recorded. Each replicate runs on one core. In the first 50 replicates, with a four-point grid of residual sizes, the median time per replicate was 13 minutes (interquartile range 11–16); the 200 replicates reported in § 7.1 used a six-point grid and ran with 30 parallel jobs on a 32-core virtual machine.

References

  1. [2]Heejung Bang and James M. Robins. Doubly robust estimation in missing data and causal inference models. Biometrics, 61(4):962–973, 2005.pubmed.ncbi.nlm.nih.gov/16401269
  2. [3]Bergthor Björnsson, Carl Borrebaeck, Nils Elander, et al. Digital twins to personalize medicine. Genome Medicine, 12:4, 2020.DOI
  3. [4]Brajan Budini, Steven Luo, Martin Tam, Isabel Stead, Andrew Lee, Angelica Akrami, Antonio Vidal-Puig, and Adrian Park. Trajectory of weight regain after cessation of GLP-1 receptor agonists: A systematic review and nonlinear meta-regression. eClinicalMedicine, 93:103796, 2026. doi: 10.1016/j.eclinm.2026.103796.DOI
  4. [5]Cong Cao. Causal inference under interference with learned exposure mappings, 2026a. arXiv:2608.19224.arxiv.org/abs/2608.19224
  5. [6]Cong Cao. When prediction error is not enough: Evaluating nuisance-function prediction for causal estimation, 2026b. arXiv:2609.00071.arxiv.org/abs/2609.00071
  6. [7]Cong Cao. When reconstruction is not enough: Doubly robust counterfactual and interventional mediation analysis with learned multimodal representations, 2026c. Manuscript; code archived at https://doi.org/10.5281/zenodo.23524452.
  7. [8]Cong Cao, Ramit Debnath, and R. Michael Alvarez. Physics-based machine learning for predicting urban air pollution using decadal time series data. Environmental Research Communications, 7:051009, 2025.
  8. [9]Victor Chernozhukov, Han Hong, and Elie Tamer. Estimation and confidence regions for parameter sets in econometric models. Econometrica, 75(5):1243–1284, 2007.DOI
  9. [10]Rob Cornish, Muhammad Faaiz Taufiq, Arnaud Doucet, and Chris Holmes. Causal falsification of digital twins. Journal of Machine Learning Research, 27(152):1–52, 2026.
  10. [11]Richard K. Crump, V. Joseph Hotz, Guido W. Imbens, and Oscar A. Mitnik. Dealing with limited overlap in estimation of average treatment effects. Biometrika, 96(1):187–199, 2009.DOI
  11. [12]Dennis Frauen, Tobias Hatt, Valentýn Melnychuk, and Stefan Feuerriegel. Estimating average causal effects from patient trajectories. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023.
  12. [13]Hamlet Gasoyan, Rebecca Schulte, Christopher B. Boyer, Nicholas J. Casacchia, W. Scott Butsch, Phuc Le, Ali Aminian, Marcio L. Griebeler, Bartolome Burguera, and Michael B. Rothberg. Obesity treatments and weight changes in clinical practice after discontinuation of semaglutide or tirzepatide. Diabetes, Obesity and Metabolism, 28(6):4795–4805, 2026. doi: 10.1111/dom.70660.DOI
  13. [14]Juen Guo, Danielle C. Brager, and Kevin D. Hall. Simulating long-term human weight-loss dynamics in response to calorie restriction. The American Journal of Clinical Nutrition, 107(4):558–565, 2018. doi: 10.1093/ajcn/nqx080.DOI
  14. [15]Kevin D. Hall. Physiology of the weight-loss plateau in response to diet restriction, GLP-1 receptor agonism, and bariatric surgery. Obesity, 32(6):1163–1168, 2024. doi: 10.1002/oby.24027.DOI
  15. [16]Kevin D. Hall, Gary Sacks, Dhruva Chandramohan, et al. Quantification of the effect of energy imbalance on bodyweight. The Lancet, 378(9793):826–837, 2011.DOI
  16. [17]Miguel A. Hernán and James M. Robins. Using big data to emulate a target trial when a randomized trial is not available. American Journal of Epidemiology, 183(8):758–764, 2016.DOI
  17. [18]Konstantin Hess, Valentýn Melnychuk, Dennis Frauen, and Stefan Feuerriegel. Bayesian neural controlled differential equations for treatment effect estimation. In International Conference on Learning Representations, 2024.arxiv.org/abs/2310.17463
  18. [19]Konstantin Hess, Dennis Frauen, Niki Kilbertus, and Stefan Feuerriegel. Debiased neural operators for estimating functionals, 2026. arXiv:2604.19296.
  19. [20]Samuel Holt, Tennison Liu, and Mihaela van der Schaar. Automatically learning hybrid digital twins of dynamical systems. In Advances in Neural Information Processing Systems, volume 37, 2024.arxiv.org/abs/2410.23691
  20. [21]Guido W. Imbens and Charles F. Manski. Confidence intervals for partially identified parameters. Econometrica, 72(6):1845–1857, 2004.
  21. [22]Myrto Kalouptsidi, Yuichi Kitamura, Lucas Lima, and Eduardo Souza-Rodrigues. Partial identification and inference for dynamic models and counterfactuals. Discussion Paper 14402, Centre for Economic Policy Research, 2020.
  22. [23]Edward H. Kennedy. Semiparametric theory and empirical processes in causal inference. In Hua He, Pan Wu, and Ding-Geng Chen, editors, Statistical Causal Inferences and Their Applications in Public Health Research, pages 141–167. Springer, 2016.arxiv.org/abs/1510.04740
  23. [24]Rui Li, Stephanie Hu, Mingyu Lu, Yuria Utsumi, Pritwhish Chakraborty, Daby M. Sow, Piyush Madan, Jun Li, Mohamed Ghalwash, Zach Shahn, and Li-wei Lehman. G-net: A recurrent network approach to G-computation for counterfactual prediction under a dynamic treatment regime. In Proceedings of Machine Learning for Health, volume 158 of Proceedings of Machine Learning Research, 2021.arxiv.org/abs/2003.10551
  24. [25]Alexander R. Luedtke, Oleg Sofrygin, Mark J. van der Laan, and Marco Carone. Sequential double robustness in right-censored longitudinal models, 2017. arXiv:1705.02459.arxiv.org/abs/1705.02459
  25. [26]Jared K. Lunceford and Marie Davidian. Stratification and weighting via the propensity score in estimation of causal treatment effects: A comparative study. Statistics in Medicine, 23(19):2937–2960, 2004.
  26. [27]Yuanzhe Ma, Yian Huang, and Hongseok Namkoong. A sensitivity approach to causal inference under limited overlap, 2025. arXiv:2511.22003.arxiv.org/abs/2511.22003
  27. [28]Tomer Meir, Ori Linial, Danny Eytan, and Uri Shalit. Structured hybrid mechanistic models for robust estimation of time-dependent intervention outcomes, 2026. arXiv:2602.11350.arxiv.org/abs/2602.11350
  28. [29]Tim P. Morris, Ian R. White, and Michael J. Crowther. Using simulation studies to evaluate statistical methods. Statistics in Medicine, 38(11):2074–2102, 2019.DOI
  29. [31]Rachel C. Nethery, Fabrizia Mealli, and Francesca Dominici. Estimating population average causal effects in the presence of non-overlap: The effect of natural gas compressor station exposure on cancer mortality. The Annals of Applied Statistics, 13(2):1242–1267, 2019. doi: 10.1214/18-A OAS1231.DOI
  30. [32]Maziar Raissi, Paris Perdikaris, and George E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019.DOI
  31. [33]James M. Robins. A new approach to causal inference in mortality studies with a sustained exposure period: Application to control of the healthy worker survivor effect. Mathematical Modelling, 7(9–12):1393–1512, 1986.DOI
  32. [34]Domenica Rubino, Niclas Abrahamsson, Melanie Davies, et al. Effect of continued weekly subcutaneous semaglutide vs placebo on weight loss maintenance in adults with overweight or obesity: The STEP 4 randomized clinical trial. JAMA, 325(14):1414–1425, 2021.DOI
  33. [35]Nabeel Seedat, Fergus Imrie, Alexis Bellot, Zhaozhi Qian, and Mihaela van der Schaar. Continuous-time modeling of counterfactual outcomes using neural controlled differential equations. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 19497–19521, 2022.arxiv.org/abs/2206.08311
  34. [36]Mark J. van der Laan and Susan Gruber. Targeted minimum loss based estimation of causal effects of multiple time point interventions. The International Journal of Biostatistics, 8(1), 2012.DOI
  35. [37]Sam West, Jadine Scragg, Paul Aveyard, Jason L. Oke, et al. Weight regain after cessation of medication for weight management: Systematic review and meta-analysis. BMJ, 392:e085304, 2026. doi: 10.1136/bmj-2025-085304.DOI
  36. [38]John P. H. Wilding, Rachel L. Batterham, Salvatore Calanna, et al. Once-weekly semaglutide in adults with overweight or obesity. New England Journal of Medicine, 384(11):989–1002, 2021.DOI
  37. [39]John P. H. Wilding, Rachel L. Batterham, Melanie Davies, et al. Weight regain and cardiometabolic effects after withdrawal of semaglutide: The STEP 1 trial extension. Diabetes, Obesity and Metabolism, 24(8):1553–1564, 2022. doi: 10.1111/dom.14725.DOI
  38. [40]Dingling Yao, Kahaan Gandhi, Valentin Duruisseaux, Boris Bonev, Francesco Locatello, and Anima Anandkumar. Counterfactual predictions in scientific emulators without controlled experiments, 2026. arXiv:2610.02252.arxiv.org/abs/2610.02252
  39. [41]Yuan Yin, Vincent Le Guen, Jérémie Dona, Emmanuel de Bézenac, Ibrahim Ayed, Nicolas Thome, and Patrick Gallinari. Augmenting physical models with deep networks for complex dynamics forecasting. In International Conference on Learning Representations, 2021.
  40. [42]Bob Junyi Zou, Matthew E. Levine, Dessi P. Zaharieva, Ramesh Johari, and Emily B. Fox. Hybrid² neural ODE causal modeling and an application to glycemic response. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 62934–62963, 2024.arxiv.org/abs/2402.17233

Paper details

vlnc:W_3q4re5jz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents