acceptodds
Under review as a conference paper at ICLR 2027

Instrumental Variables in Continuous-Time Stochastic Processes

Abstract

Causal questions often arise in systems that evolve in continuous time. While instrumental variable (IV) methods provide a principled approach to identify causal effects in static or discrete-time settings, even in the presence of unobserved confounding, they do not yet extend to continuous-time dynamics. In this work, we develop an IV framework for stationary continuous-time stochastic processes parameterized by stochastic differential equations (SDE), where the instrument itself takes the form of a stochastic process. We encode the structural sparsity and causal effects directly at the infinitesimal level of the drift, and contrast these assumptions with the IV assumptions on the discretized version of the same process. Based on these structural assumptions, we derive moment restrictions using path-valued instruments. We show how these restrictions simplify for Ornstein-Uhlenbeck processes and present an identification condition together with a corresponding consistent estimation method. Finally, we extend our estimation framework to nonlinear causal effects by expressing the path-based moment restrictions leveraging the signature kernel. Empirically, our method recovers causal effects in confounded continuous-time systems where maximum-likelihood-based estimation and discrete-time IV methods are systematically biased.

Primary area:
causal reasoning
Keywords:
stochastic processes, instrumental variables, sde

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_46E2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper develops an instrumental-variable framework for continuous-time stochastic systems with latent confounding. It encodes IV assumptions in the infinitesimal drift and derives moment restrictions relating prediction residuals to the instrument’s history. For Ornstein-Uhlenbeck processes, it establishes identification under a controllability condition and provides a consistent integral-based estimator. It further proposes a signature-kernel estimator for nonlinear drift functions. Experiments on synthetic systems and a diabetes-inspired simulator demonstrate advantages over likelihood estimation that ignores hidden confounding and existing discrete-time IV baselines.

Critical strengths
  1. The paper clearly explains why IV sparsity in continuous-time drift does not generally survive finite-time discretization, identifying a substantive source of model misspecification.

  2. The connection between causal identification and controllability provides a principled explanation for how a single instrument process can identify multiple parameters. The consistency result complements this identification analysis.

  3. Experiments varying confounding and instrument strength support the proposed mechanism and illustrate the expected trade-off between bias reduction and estimation efficiency.

Critical weaknesses
  1. While the paper derives valid conditional moment restrictions, these restrictions do not necessarily identify a unique nonlinear causal drift. Multiple candidate functions may satisfy the same restrictions, and regularization selects among these candidates without necessarily recovering the true mechanism. The paper explicitly acknowledges this limitation, but it leaves a substantial gap between the nonlinear estimation framework and a guarantee of causal recovery.

  2. In the RBF experiments, the estimator uses the same basis-function centers and bandwidth as the ground-truth drift, reducing the task to coefficient estimation within a correctly specified function family. Although this is a useful controlled evaluation, it provides limited evidence that the method can recover nonlinear causal mechanisms when their functional form is unknown.

Questions for authors

N/A

3 comments

W1: Non-identification in the nonparametric setting

Yes, non-identification is an interesting inherent characteristic of the IV setting. More broadly, non-identifiability cannot be addressed on the level of methodology or collecting more (of the same) data. Whether a target parameter is identifiable from observational data solely depends on the modeling assumptions. That leaves to directions: (a) One can add on stronger, more restrictive modeling assumptions until the target parameter is identified. (b) One accepts and grapples with the fact that the target parameter is not identified.

Large parts of statistics intentionally focus on (a) and only deal with models in which the target parameter is identifiable by design. This commonly leads to misspecification. For example, in cause effect estimation, some typical assumptions to render causal effects identified are no unobserved confounding, consistency, and overlap. The IV approach explicitly acknowledges that in many settings the no unobserved confounding assumption is too restrictive. With additional assumptions such as linearity (and in the static case sufficiently rich instruments) full identifiability can be restored. But in the non-linear/non-parametric setting, non-identifiability is an intrinsic property of the IV model (also in the static case).

When multiple causal effects are compatible with the observed data, one can either try to characterize the entire compatible set, or find specific elements. Finding the least norm element is a common approach and has been thoroughly studied in the literature for the static case (Zhang et al., 2023, Bennett et al., 2023, Dikkala et al., 2020). However, one could also find the element closest to the MLE estimator or some other parameter of interest (called "Anchor regression") or aim for bounds of causal effects (Rothenhausler et al., 2021, Kilbertus et al., 2020). These are both active areas of ongoing research.

Summary:

Non-identification in our nonlinear setting is an inherent property of the setup. It cannot be fixed by methodology or better estimators. We follow existing literature on the static setting in finding a minimum norm solution in the identified set.

References:

Bennett, A., Kallus, N., Mao, X., Newey, W., Syrgkanis, V., & Uehara, M. (2023). Minimax instrumental variable regression and l2 convergence guarantees without identification or closedness. In G. Neu & L. Rosasco (Eds.), Proceedings of Thirty Sixth Conference on Learning Theory (Vol. 195, pp. 2291–2318). PMLR.

Canay, I. A., Santos, A., and Shaikh, A. M. (2013). “On the Testability of Identification in Some Nonparametric Models With Endogeneity.” Econometrica, 81(6)

Dikkala, N., Lewis, G., Mackey, L., & Syrgkanis, V. (2020). Minimax estimation of conditional moment models. In Proceedings of the 34th International Conference on Neural Information Processing Systems (NeurIPS '20).

Kilbertus, N., Kusner, M. J., & Silva, R. (2020). A class of algorithms for general instrumental variable models. Advances in Neural Information Processing Systems.

Rothenhaeusler, D., Meinshausen, N., Buehlmann, P., & Peters, J. (2021). Anchor regression: Heterogeneous data meet causality. Journal of the Royal Statistical Society.

Zhang, R., Imaizumi, M., Schölkopf, B., & Muandet, K. (2023). Instrumental variable regression via kernel maximum moment loss. Journal of Causal Inference

W2: Performance with a unknown functional form

Thanks for this suggestion. So far, beyond the kernel based experiments, we had one short experiment with a neural network parameterization (l.450 onwards), which is a much more flexible representation of the drift function.

To corroborate our claims in the nonlinear setting, we performed the following additional experiments to show that other flexible parameterizations can also recover the true drift function in the RBF-kernel-based setting (so using the same representation with same centers and width is not necessary). We compare three increasingly flexible parameterizations of the drift function: a full basis-function expansion, a partial neural parameterization fy;θ(Yt,XT)=aYt+bXt+fψ(Xt)f_{y;\theta}(Y_t,X_T) = aY_t + b X_t + f_\psi(X_t), which retains linear terms in YtY_t and XtX_t and uses a neural network to model the nonlinear RBF component, and a fully flexible standalone neural network. For completeness, we also add the naive MLE baseline (which ignores the unobserved confounder) to the comparison.

Table 1 Drift function MSE over the joint distribution of X,YX,Y with standard errors across five seeds.

MethodFunction MSE (SE)
RBF-basis MLE0.1034 (0.0005)
RBF-basis IV0.0134 (0.0018)
Additive neural drift IV0.0354 (0.0041)
Full neural drift IV0.0775 (0.0015)

Our IV method recovers complex drift functions across different types of functional parameterizations even when the full flexibility of a neural network is used. All parameterizations outperform the MLE estimate and we clearly observe that the heavier we restrict the parameterization (while still being able to represent the true drift), the better the MSE. (As parameterizations become more flexible, the identified set grows.)

We added a plot to appendix I.4 (Fig.5-7) showing that the learned networks trace out the complex shape of the RBF-based function with high precision and also present an additional experiment demonstrating that a sine-shaped drift can be recovered as well.

Dear reviewer 46E2, As the discussion window draws to a close, we wanted to gently check in to see if our rebuttal has successfully addressed your concerns. If any questions remain, we are very happy to clarify them before the deadline, and we would appreciate it if you considered re-evaluating your scores. Thank you for your time and effort!

Reviewer_6fTFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript develops an instrumental variable framework for continuous-time stochastic processes with unobserved confounding. It derives moment restrictions using the history of an instrumental process, establishes a controllability-based identification condition and a consistent estimator for Ornstein-Uhlenbeck models, and proposes a signature-kernel approach for nonlinear drift functions. The distinction between infinitesimal structural restrictions and those imposed on discretely sampled processes provides an important motivation for this work. In particular, the connection between controllability and causal parameter identification is a valuable theoretical contribution. The numerical results further illustrate the potential of the proposed methods.

Critical strengths

None

Critical weaknesses
  1. The novelty claims should be positioned more carefully relative to existing continuous-time IV methods for system identification, including irregular sampling and closed-loop systems. Distinguishing causal parameter estimation from dynamical model identification does not, by itself, establish methodological novelty. The authors should clarify which aspects of their latent-confounding setup and theoretical results extend beyond existing approaches.
  2. The unpenalized V-statistic criterion in Equation (19) can be interpreted as a GMM objective with identity weighting in the instrument RKHS. Could an optimal weighting operator improve efficiency, particularly with many correlated instruments? For dependent path segments, this would involve the long-run covariance operator of the RKHS-valued moments, whose inversion may require regularization. Although unnecessary for consistency, discussing optimal weighting would clarify the efficiency implications of the proposed criterion.
  3. The authors should clarify whether the instruments used in estimation preserve sufficient identifying information. Equation (9) is equivalent to Equation (5) when imposed for all bounded measurable path functionals g, despite each g being scalar-valued. Does this equivalence remain valid after restriction to the proposed RKHS and finite lookback windows? Although Equation (16) combines multiple lag-based moment conditions through least squares, further guidance is needed on selecting lags and instrument functions to address weak or redundant moments and numerical instability. Instrument validity should also be distinguished from identification, which requires additional conditions, such as a rank condition or injectivity of the relevant conditional moment operator.
  4. The nonlinear extension, presented as a main contribution, requires a more comprehensive theoretical treatment. The relationship between the minimizer of Equation (17), the true structural drift, and a minimum-norm element of the identified set remains unclear. The authors should distinguish ill-posedness from nonidentification, establish identification conditions, and characterize the selected solution when point identification fails. Consistency and convergence rates should also be investigated, accounting for kernel and regularization choices, temporal dependence, finite lookbackwindows, and numerical integration. These developments are essential to clarify what the nonlinear procedure estimates and when reliable recovery can be expected.
  5. The empirical evaluation would benefit from a more systematic assessment of finite-sample performance and the value of path-based instruments. The principal OU experiment uses 8000 trajectories per system, while the nonlinear RBF experiment uses the true basis centers and bandwidths. Performance with fewer trajectories, a single long trajectory, and sparser observations remains unclear. Comparisons with instruments constructed from the current state, a finite collection of lagged states, or simple path summaries would also help establish whether the signature kernel provides a substantive advantage.
Questions for authors

How to select instrument variables in practical applications? How is the effect of the selection of instrument variables on estimation?

7 comments

We thank the reviewer for the feedback and address each question within each mentioned weakness one by one. Given the breadth of the raised questions, we provide a response organized into separate comments to fully address each point. If any points remain unclear, we would appreciate additional clarifications.

W1: connections to system identification.

Our key novelty and differentiation is in the results provided for the IV-structured SDE model. We will sharpen the wording for the distinction from existing approaches. Importantly, in our setting,

  • in eq. (1), both XX and YY evolve stochastically,
  • an unobserved process continuously influences their dynamics
  • and feedback between XX and YY is allowed.
  • We estimate the local drift fy(Yt,Xt)f_y(Y_t,X_t) by using ZZ to eliminate the confounder from the moment equations.
  • All this requires no fitted model of the confounder trajectory or dynamics. In the OU estimator, even the confounder dimension and its drift and coupling parameters need not be estimated.

The settings motivating instrumental variable methods in system identification differ substantially from our causal setting, as do the resulting estimation methods. We do not simply label the target parameter as "causal".

The closest work in system identification is the Refined Instrumental Variable Method presented in Young et al. (2008), which uses instruments to estimate the system parameters in discrete and continuous-time transfer function models. In contrast to our submission, the method assumes additive observation noise on the output process instead of the existence of a true continuous-time confounding mechanism. This allows the method to parametrically write down the transfer function between input and output. Our proposed setting is considerably harder, as the confounding mechanism acts continuously throughout time, which prevents us from even parametrically writing down the transition densities of the process (in the linear case we do not know the dimension of the unobserved confounder and thus cannot even write down the true matrix exponential parametrically).

Other methods, such as Escobar et al. (2017) allow for confounding via a common Brownian motion together with a steerable control input, whereas we are explicitly interested in the purely observational setting.

A different branch in continuous-time system identification instead relies on Kalman filters (Halmos et al. 2023). These models make structural assumptions about the unobserved confounding process, which we explicitly avoid in our paper. (We nevertheless include a comparison with Kalman filters in the revised version demonstrating that explicitly modeling the latent process yields subpar performance.)

Summary:

The continuous-time IV settings considered in the field of system identification differ significantly in terms of assumed data generating mechanism, confounding mechanism, and estimation methods from the setting presented in this paper.

Escobar, J., & Enqvist, M. (2017). Instrumental variables and LSM in continuous-time parameter estimation. ESAIM: Control, Optimisation and Calculus of Variations

Halmos, P., Pillow, J., & Knowles, D. A. (2023). System identification for continuous-time linear dynamical systems.

Young, P. C. (2008). The refined instrumental variable method: Unified estimation of discrete and continuous-time transfer function models. Journal Européen des Systèmes Automatisés

W2: weighted GMM.

Thanks for the suggestion. Yes, appropriate weighting can improve efficiency. In general, our submission establishes the novel continuous-time IV setting, the relevant moment restrictions, identification results (for OU), and (consistent) estimation techniques for (non)linear settings. It thereby opens the door for ongoing developments of more efficient estimators tailored towards specific circumstances. Optimal weighting can be considered one such improvement, especially for correlated moments. Let us lean into that possible extension.

Theory.

From a theoretical perspective, additional analysis is required to maintain consistency. Concretely, for the weighted estimator

K^n,W:=B^nWnG^n⊤(G^nWnG^n⊤)−1→pBWG⊤(GWG⊤)−1=K\widehat K_{n,W} := \widehat B_n W_n \widehat G_n^\top \bigl(\widehat G_n W_n \widehat G_n^\top\bigr)^{-1} \xrightarrow{p} B W G^\top (G W G^\top)^{-1} = K

a number of additional requirements must be satisfied by WW: It needs to be symmetric, positive semidefinite, converge in probability to a fixed matrix, and preserve sufficient moment information. Under those assumptions, the consistency proof goes through exactly as stated in the submission and we will add this extension to the revised version. In practice, one could estimate WW either from a single long path or from multiple paths. We will include this discussion in the revised version.

Experiments.

We empirically test GMM weighting by estimating the weight matrix across paths and using the estimator described above. We sample five systems and evaluate the integral version of the estimator with and without GMM weighting. We notice that the moments used by the integral estimator are already highly correlated as they use the same estimation of the initial covariance and overlapping observations from the same trajectories. The results show slight improvements in estimation error. The performance gains are (perhaps expected) most visible in the small sample setting where increased efficiency manifests the most.

Table 1: RMSE of the linear estimator and standard error across five systems

PathsIV-SDE no GMMIV-SDE GMMMLE
2000.1019 (0.0608)0.0802 (0.0225)0.1498 (0.0326)
5000.0655 (0.0195)0.0336 (0.0084)0.1504 (0.0332)
10000.0325 (0.0103)0.0242 (0.0088)0.1441 (0.0347)
50000.0144 (0.0047)0.0156 (0.0040)0.1446 (0.0343)

W3: Which identifying information is retained?

Q1: "whether the instruments preserve sufficient identifying information"

Identification depends on the combination of considered drift function class and the moment conditions. In the linear case, we answer this question in Proposition 2 as a key contribution. In the nonlinear case, Section 6 makes precise that the defining IV conditions do not generally allow for identification when allowing for general nonparametric function classes. Without additional assumptions, the standard approach is to single out a specific member of the identified set. Following work in the static setting, we decide to retrieve the minimum L2 norm solution (see following replies).

Q2: "equivalence between Eq (9) and Eq(5)"

Technically, the scalar output of each gg causes no information loss when all bounded measurable functionals are included. For an integrable residual

E[R∣Ftz]=0⟺E[Rg(Z[0,t])]=0\mathbb{E}[R\mid\mathcal{F}_t^z]=0\quad\Longleftrightarrow\quad\mathbb{E}[Rg(Z_{[0,t]})]=0 for all bounded measurable gg.

This equivalence is used in eqs. (5) and (9). The questions thus becomes what happens when family of functionals, the history they observe, and the residual horizons are restricted.

Restriction to RKHS.

When considering an RKHS, a sufficient criterion is that it can distinguish between all residual-weighted signed measures A↦E[Rf1P∈A]A \mapsto \mathbb{E}[R_f \mathbf{1}\\{P \in A\\}] for a chosen path variable PP. Therefore, characteristicness of the kernel underlying the RKHS is a sufficient condition for eq. (5) and (9) being equivalent. This motivates the choice of the signature kernel in the nonlinear estimator. In the case of an OU process, linear kernels are sufficient.

Finite lookback and fixed horizon.

Let Pℓ=Z[t−ℓ,t]P_\ell=Z_{[t-\ell,t]} and Ft,ℓz=σ(Pℓ)\mathcal F_{t,\ell}^z=\sigma(P_\ell). The tower property gives

E[R∣Ftz]=0⟹E[R∣Ft,ℓz]=0.\mathbb E[R\mid\mathcal F_t^z]=0 \quad\Longrightarrow\quad \mathbb E[R\mid\mathcal F_{t,\ell}^z]=0.

Thus, finite lookback preserves validity, but may discard identifying information. Since autocovariances typically decay very quickly in most processes, a fixed lookback window is a practical choice and has not compromised identification power in our experiments.

Finite OU construction is sufficient.

For the OU setting, we present a positive identification result with finite lags: As described, joint Gaussianity makes linear functionals sufficient, and stationarity yields the covariance formulation in Section 5 and Appendix D. Proposition 2 then proves that, with a=dx+dya=d_x+d_y,

rank⁡[b,Acb,…,Aca−1b]=a\operatorname{rank}[b,A_cb,\ldots,A_c^{a-1}b]=a

together with the stated stability and nondegenerate instrument-diffusion assumptions ensures that aa lags drawn from an absolutely continuous distribution give a full-row-rank integrated moment matrix almost surely. Thus the finite moment system identifies (βyy,βyx)(\beta_{yy},\beta_{yx}). This directly answers whether the moments in Eq. (16) retain enough identifying information.

Q3: "Lag selection and instrument functions"

The estimator presented in eq. 16 does not rely on instrument functions. The estimator only considers the linear case, in which linear combinations of instruments are sufficient for identification. Practical guidance is to use all available moments where the covariances are above a certain threshold, and lags that cover the relevant dynamical time scales. Choosing closely spaced lags may generate nearly collinear moments, which can be countered by the typical regularization measures (ridge/spectral). The case of weak instruments leaves the main contributions of our submission (setting, moment restrictions, identification results, estimator consistency) unaffected. They primarily lead to subtleties and challenges in the finite sample estimation and are not specific to our setting. We leave such extensions for future work and will acknowledge the implied challenges more explicitly in the revised pdf. Introducing a weighing matrix (as suggested) might be one fruitful approach.

Q4: "Instrument validity should be distinguished from identification, which requires additional conditions"

We clearly distinguish these two aspects in the submission: The structural assumptions in Section 3 and Proposition 1 explicitly imply validity. Identification then asks whether the resulting restrictions can separate the considered drift functions, which depend on the model. We derive precise identification conditions in the linear case in Proposition 2, which contains the requested rank condition in form of the controllability matrix. In the nonlinear case, unrestricted identification cannot be established as described in detail in Section 6. In the RKHS setting, kernel richness avoids losing conditional-moment information, but injectivity of the drift-to-moment map is a separate question, which we answer next.

Reviewer_FmNFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper presents a continuous generalization of the VAR(1) case of prior work on the instrumental variables method and further introduces a non-linear drift function c

Critical strengths
  1. An interesting and meaningful direction for providing a continuous-time generalization of prior discrete-time and static work.
  2. The analysis seems to be sound in providing the moment conditions and then translating them to the OU process.
  3. The non-linear extension scheme also seems to be sound
Critical weaknesses
  1. While the work seems theoretically sound, it is quite a narrow piece of development. As a continuous-time generalization of the VAR(1) in the 2024 JMLR paper, there is a natural and broad family of CAR(p), while the authors have only explored the most basic CAR(1) case.
  2. The experimental scope is also narrow. In terms of baselines, it would be most interesting to see comparisons to the other VAR(p) families, while the authors only compare their methods to VAR(1) for one experiment, and all the other experiments compare to the naive baseline, which seems limited. There is also no direct ablation of the non-linear framework by comparing the linear to the non-linear case. The number of datasets also seems to be too few to claim superiority of the methods.
Questions for authors

1, Can you include more comparisons to the VAR(p) family in the 2024 JMLR paper or potentially other, more capable baselines in addition to the naive one? 2. Can you make more comparisons between linear and non-linear models? 3. Can you explain why the current exploration is sufficient in not including the CAR(p) equivalence of the VAR(p) methods?

4 comments

We thank the reviewer for their feedback and address comments and questions one by one.

Q1 ("Can you include var(p) or other baselines?") + first sentence of W2 ("var(p) family comparisons missing; few datasets"):

TL;DR: We add further baselines, present additional results in a new setting where we estimate linear effects under nonlinear confounding and demonstrate functional recovery of complex drift shapes with neural parametrizations.

Beyond the previous estimates obtained from VAR(1)VAR(1) and MLE as well as the ratio estimates via integrated covariances, we added further VAR(p)VAR(p)-based estimates (suggested) and also a Kalman filter based approach. Importantly, we are pursuing an estimation task, not a prediction or raw modeling task. While higher order VAR(p)VAR(p) processes and Kalman filters provide a modeling/prediction framework, we must still define our estimation target in these models. For VAR(p)VAR(p) processes, the natural choice are the parameters describing the one timestep influence of the covariates onto the outcome. Under the experiment settings in Figure 2, we obtain:

Table 1: Parameter RMSE with standard errors across five systems.

MethodMean RMSE (SE)
IV-diff0.0160 (0.0041)
VAR(1)0.0626 (0.0147)
VAR(2)0.2019 (0.0330)
VAR(3)0.1969 (0.0306)
VAR(5)0.1925 (0.0286)

Our principled IV-SDE method strongly outperforms all higher order VAR(p)VAR(p)-based approaches for cause-effect estimation. The hypotheses for why these methods are not competitive, presented in appendix B, hold generally for VAR(p)VAR(p) processes.

An alternative prominent method for estimating parameters under unobserved confounding is the Kalman filter. A major difference from our approach is that the Kalman filter explicitly models the latent processes as well. Hence, in contrast to our method, Kalman filters require prior knowledge and explicit specification of these latent processes. This typically leads to high sensitivity to misspecifications. Experimentally, we compare our continuous-time IV estimator with the Kalman filter in two settings: (a) An OU setting, where the Kalman filter assumes a scalar latent process for two unobserved covariates; (b) linear observed dynamics, but a nonlinear latent process. Our IV method achieves robust and accurate estimates in both settings, outperforming the Kalman filter by a substantial margin.

Table 2: Mean parameter RMSE and standard error across five systems.

ConfounderMethodRMSE (SE)
OUIV-diff0.0320 (0.0031)
OUKalman0.2284 (0.0350)
PhaseIV-SDE0.0245 (0.0052)
PhaseKalman0.1546 (0.0191)

Additionally, we perform new experiments with the nonlinear estimator in the setup described as kernel-based estimation of nonlinear causal effects. Instead of parameterizing the drift function with a basis function expansion, we use neural network-based function classes of varying flexibility. The results (as shown in the updated PDF in appendix I.4.) clearly demonstrate that we can recover complex functional drift shapes with neural drift functions.

All details are in the appendix of the revised submission.

Summary:

Our principled estimator is tailored for IV estimation exploiting the structural information while being agnostic to how precisely the latents enter. It therefore robustly outperforms all different baselines in a variety of settings which either underperform or exhibit substantial estimation bias.

We emphasize that our experimental results comprise a large number of datasets from different systems. Figure 2 aggregates performance across 310 linear systems and we assess three major nonlinear settings across nine different configurations.

Q2 ("comparisons between linear/non-linear?") + second part of W2 ("ablation of non-linear by comparing to linear"):

Since we are not sure we fully understand the question, let us start with our understanding (please correct us if we misunderstood).

We are estimating causal effects in IV models. If we are willing to assume a linear parametric model, the estimation targets are simply (vectors of) numbers (βyy,βyx\beta_{yy},\beta_{yx} in eq. 10). A linear method (exploiting this linearity assumption) will thus just output these numbers. In this case, RMSE directly of the linear parameters are a natural performance metric.

If we instead work with a non-linear model, the estimation target generally is a function (fY{f}_{Y} in eq. 1). Depending on the chosen parameterization, this function estimate can be represented, for example, via kernel based techniques (l.436) or neural networks (l.450). In this case, the relevant performance metric is function-norm, e.g., L2 norm of functions.

The two possible comparisons we can envision are the following: (a) use a non-linear method requiring fewer assumptions in a setting that happens to be linear (well-specified); (b) use a linear method even though the true processes are non-linear (misspecified).

(a) Non-linear method for linear setting.

Here, our linear estimator can actually be retrieved analytically from the nonlinear estimator. Using a linear kernel we obtain klin,δm(Z[0,t]i,Z[0,t]j)=(Zt−δmi)TZt−δmj,∀δm∈[0,t].k_{\mathrm{lin},\delta_m}(Z^i_{[0,t]},Z^j_{[0,t]})=(Z^i_{t-\delta_m})^TZ^j_{t-\delta_m},\forall\delta_m\in[0,t].

Hence, the minimum of the moment condition in equation 17 is obtained for

E[Rt,t+h;θZt−δmT]=0,∀δm∈[0,t],∀h>0.\mathbb E[R_{t,t+h;\theta}Z_{t-\delta_m}^T]=0,\forall\delta_m\in[0,t],\forall h>0.

Using a linear drift parametrization (following the assumption) and the moment transformation based on the process stationarity described in appendix D, we obtain the following moment conditions

0=E[Rt,t+τℓ;θZtT]=CYZ(τℓ)−CYZ(0)−βyy∫0τℓCYZ(s)ds−βyx∫0τℓCXZ(s)ds,ℓ=1,…,L.0=\mathbb E[R_{t,t+\tau_\ell;\theta}Z_t^T]=C_{YZ}(\tau_\ell)-C_{YZ}(0)-\beta_{yy}\int_0^{\tau_\ell}C_{YZ}(s)ds-\beta_{yx}\int_0^{\tau_\ell}C_{XZ}(s)ds,\ell=1,\dots,L.

for LL integration horizons. This exactly recovers the linear moment conditions described in Proposition 2 and can be used for parameter estimations with eq. 16. We also experimentally verify that the nonlinear signature-kernel-based method can estimate linear parameters in the settings described in Appendix I. Here, the non-linear method achieves similar estimation performance to the linear setting.

Table 3: Parameter RMSE for linear IV and signature IV with standard errors across five systems.

MethodParameter RMSE (SE)
Linear IV0.0212 (0.0030)
Signature IV0.0236 (0.0073)

(b) Linear method for non-linear setting.

We also conduct an additional ablation to test if a fixed number of moments defined by past instrument states, as used in the linear estimator, might be similarly effective as using the full signature-kernel in the nonlinear estimator. We experiment using the nonlinear RBF-kernel-based setup described in Section 7. We run both the estimator based on the full signature kernel as well as the adapted estimator using 10 past instrument states. Here, the estimation target is full functions, so we report the function output MSE over the distribution of XX (for linear methods, the estimated parameters represent a linear function).

Table 4: Function MSE for linear IV and signature IV with standard errors across five seeds.

MethodFunction MSE (SE)
Linear IV (10 past Z)0.1273 (0.0380)
Signature IV0.0125 (0.0027)

This demonstrates that the nonlinear moment conditions leveraging the signature-kernel retrieve more information compared to the reduced linear comparison. Due to the misspecification of the linear method in this setting, this is expected.

Summary: The more general non-linear method directly reduces to the linear one analytically and also experimentally recovers the true parameters in linear settings. When the linear model is used in nonlinear settings (misspecified), it generally performs worse than our nonlinear estimator.

Q3 (why not CAR(p)CAR(p)?) + W1 (only CAR(p)CAR(p)):

The main reason for not including CAR(p)CAR(p) processes in the main setting is that they do not admit the defining graphical instrumental variable conditions and therefore do not allow for the formulation of a cause-effect estimation task. Crucially, formulating the IV conditions in continuous time, relies on infinitesimal criteria of the process. For SDEs (first-order continuous processes), this was already a substantive undertaking (a key contribution of this work). In higher-order CAR(p)CAR(p) processes, the underlying task of cause-effect estimation becomes ambiguous in the following sense.

The model parameters describing the causal effect (the target parameter) are specific to the chosen model. Neither our SDE construction nor the estimand in Thams et. al, (2024) transfer to higher-order CAR(p)CAR(p) models. One possible definition of the target parameter in such settings could be the interventional distribution of the outcome under path-valued interventions on the covariate process. These cannot easily be summarized as a single parameter, but form an entire family of interventional, path-valued distributions. In the following, we explore the much narrower example of point interventions. This setting allows for the definition of a proposed estimand, which arguably captures a meaningful notion of a total effect, analogous to the one mentioned in Sec. 2.1, Thams et. al, (2024) .

We start with a CAR(p)CAR(p) process for the outcome (Brockwell et al., 2007):

dYt(p−1)=(∑k=0p−1akYt(k)+∑k=0p−1bkXt(k)+Ut)dt+σYdWtY,dY_t^{(p-1)}=(\sum_{k=0}^{p-1}a_kY_t^{(k)}+\sum_{k=0}^{p-1}b_kX_t^{(k)}+U_t)dt+\sigma_YdW_t^Y,

which we rewrite equivalently as a CAR(1)CAR(1) process via the standard augmentation

$ d\begin{pmatrix} Y_t\\ Y_t^{(1)}\\ \vdots\\ Y_t^{(p-1)} \end{pmatrix} =\begin{pmatrix} Y_t^{(1)}\\ Y_t^{(2)}\\ \vdots\\ \displaystyle\sum_{k=0}^{p-1}a_kY_t^{(k)}+\sum_{k=0}^{p-1}b_kX_t^{(k)}+U_t \end{pmatrix}dt +\sigma_Y\begin{pmatrix} 0\\ 0\\ \vdots\\ 1 \end{pmatrix}dW_t^Y. $

In this model, we can define the Continous-Time Total Causal Effect θ(h)\theta (h) of an intervention on XtX_{t} on the outcome variable Yt+hY_{t+h} via

θ(h):=∂Yt+h∂Xt=cYTeAheX.\theta(h):=\frac{\partial Y_{t+h}}{\partial X_t}=c_Y^Te^{\mathcal{A}h}e_X.

where A\mathcal{A} describes the drift matrix of the full augmented CAR(1)CAR(1) process, cYc_{Y} indexes the output dimension of YY and eXe_{X} the selected input covariates.

This total causal effect satisfies the moment condition E[(Yt+h−θ(h)Xt−α(h)TRt)ZtT]=0.\mathbb E[(Y_{t+h}-\theta(h)X_t-\alpha(h)^TR_t)Z_t^T]=0.

where RtR_{t} now includes the remaining augmented-state components, such as YtY_{t} and, importantly, all p−1p-1 derivatives of YtY_{t}. While this can be argued to be a meaningful definition of the causal effect, turning this moment condition into a practical estimator poses severe challenges. It would require direct observation or measurement of higher-order derivatives of the outcome process, which are typically not available. For stochastic processes, potentially observed at irregular timesteps, these are also notoriously hard to estimate.

Summary: A natural extension to CAR(p)CAR(p) processes from existing work fails, because one must first (a) define the estimation target (what is a causal effect in such models), (b) formulate meaningful structural IV assumption, and, even when specific formulations for (a) and (b) are chosen as we did, (c) faces fundamentally novel challenges in estimation. Such novel settings are interesting directions for future work. We will add the above example and general discussion to the revised manuscript.

For completeness, we highlight that Thams et al, (2024) develop identification arguments and an estimation algorithm only for the VAR(1)VAR(1) case and not for higher-order VAR(p)VAR(p) models.

References:

Brockwell, P. J., Davis, R. A., & Yang, Y. (2007). Continuous-time Gaussian autoregression. Statistica Sinica

Thams, N., Søndergaard, R., Weichwald, S., & Peters, J. (2024). Identifying causal effects using instrumental time series: Nuisance IV and correcting for the past. Journal of Machine Learning Research

Dear reviewer FmNF,

As the discussion window draws to a close, we wanted to gently check in to see if our rebuttal has successfully addressed your concerns. If any questions remain, we are very happy to clarify them before the deadline, and we would appreciate it if you considered re-evaluating your scores.

Thank you for your time and effort!

Comments on the paper

1 comment

Dear AC, dear reviewers,

We sincerely thank all of you for the constructive feedback, which has helped us improve our work. We have addressed all reviewer concerns through additional experiments, thorough theoretical discussion of alternative models, more clearly stating the novelty of our approach, and many smaller revisions (all highlighted in blue in the revised version). The improvements are summarized as:

  • We discuss effect estimation in CAR(p) processes
  • We more clearly contrast our submission with related work in the field of system identification
  • We explore moment weighting and empirically show its benefits in the small-sample setting
  • Experimental extension:
    • We introduce further baselines (VAR(p), Kalman) for the linear estimator
    • We demonstrate that the nonlinear estimation method can retrieve drift functions with an unknown functional form


We believe the revised submission makes a strong and novel contribution at the intersection of IV cause-effect estimation and continuous-time dynamical systems.

Even though not much time remains, we would still highly appreciate any comments or questions on our rebuttals from the reviewers who have not yet had the chance to reply.

We highly appreciate your and the reviewers’ time and feedback!

Kind regards,

the authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.