Proximal Causal Inference

Proximal causal inference is a non-parametric and semiparametric methodology in epidemiology, biostatistics, and public health introduced by statisticians Wang Miao, Zhi Geng, and Eric J. Tchetgen Tchetgen in 2018 and developed into a general longitudinal and semiparametric theory by Tchetgen Tchetgen, Andrew Ying, Yifan Cui, Xu Shi, and Wang Miao in 2020–2024. The framework provides formal point identification and doubly robust estimation of causal treatment effects in observational cohort studies and electronic health record databases when unmeasured confounding invalidates standard adjustment methods—without requiring a valid instrumental variable or knowing the functional form of the unmeasured confounder distribution.

Classical observational causal inference methodologies—including the parametric G-formula, marginal structural models, propensity score matching, targeted maximum likelihood estimation, and G-estimation of structural nested models—rely on the assumption of conditional exchangeability (no unmeasured confounding given measured covariates X). In practice, investigators rarely observe the true underlying biological or behavioral confounders U (such as baseline frailty, subclinical disease severity, or health-seeking behavior) directly; instead, clinical registries record noisy proxies of U. Proximal causal inference partitions available proxy variables into two complementary sets—treatment-inducing confounding proxies Z (generalizing negative-control exposures) and outcome-inducing confounding proxies W (generalizing negative-control outcomes)—and proves that if (Z, W) satisfy structural exclusion and completeness conditions, the causal effect of treatment A on outcome Y is non-parametrically identified via an outcome confounding bridge function h(W, A, X) that solves a Fredholm integral equation of the first kind.

Why Conditioning on Noisy Confounder Proxies Fails

Let A denote a binary or continuous treatment exposure, Y denote a clinical outcome, Y(a) denote the potential outcome under intervention A = a, X denote accurately measured baseline covariates, and U denote latent (unobserved) common causes of both A and Y. If U were observed alongside X, latent conditional exchangeability Y(a) ⊥ A | (U, X) would hold, and the marginal counterfactual expectation E[Y(a)] would be identified by Robins's G-computation formula:

When U is unmeasured, conditioning only on X fails (Y(a) ̸⊥ A | X), a ubiquitous bias in pharmacoepidemiology known as confounding by indication. Even if the investigator records noisy proxy measurements W of the latent confounder U (such as an administrative billing code or single baseline lab value W as a surrogate for physiological frailty U), substituting W for U in standard regression, propensity score matching, or inverse probability weighting leaves residual confounding bias whenever W is measured with error (Y(a) ̸⊥ A | (W, X)).

From Qualitative Negative Controls to Quantitative Confounding Elimination

Historically, epidemiologists used negative controls (Lipsitch, Tchetgen Tchetgen, and Cohen 2010) as a diagnostic check to detect unmeasured confounding rather than to correct for it:

1. Negative-Control Exposure (NCE) Z: An auxiliary exposure variable Z associated with the latent confounder U (and possibly correlated with primary treatment A) that is known on mechanistic grounds to have zero direct causal effect on the primary outcome Y: Y(a, z) = Y(a). For example, in a study of whether annual influenza vaccination A reduces winter mortality Y among older adults (confounded by unmeasured healthy-user frailty U), a prescription refill indicator for statins or ocular medications Z has no direct mechanism to prevent influenza mortality Y, yet reflects the same healthy-user adherence U.
2. Negative-Control Outcome (NCO) W: An auxiliary outcome variable W affected by the latent confounder U (such as pre-season hospitalization before influenza circulates, or hospitalization for accidental injury) that is causally unaffected by both the primary treatment A and the negative-control exposure Z: W(a, z) = W.

While observing a non-null association between Z and Y or between A and W alerts the investigator that unmeasured confounding U is present, Miao, Geng, and Tchetgen Tchetgen (2018) proved that measuring both a proxy Z and a proxy W in the same study enables non-parametric subtraction of the unmeasured confounding bias without assuming linear additive structural equations.

Three Canonical Proxy Roles (X, Z, W)

In the proximal causal inference framework, observed covariates are partitioned into three structural types (X, Z, W) relative to treatment A, outcome Y, and latent confounder U:

• Type 1 — Common Measured Confounders (X): Measured baseline covariates that may directly affect A, Y, Z, and W, and may correlate with U.
• Type 2 — Treatment-Inducing Confounding Proxies (Z): Proxies of U that may be causally linked to A (Z → A or A → Z alongside U → Z) but do not directly affect Y or W once (A, U, X) are held fixed.
• Type 3 — Outcome-Inducing Confounding Proxies (W): Proxies of U that may directly affect Y (W → Y alongside U → W), but are conditionally independent of (A, Z) given (U, X).

Alongside consistency (Y = Y(A) and W = W(A, Z)) and positivity (0 < P(A = a | Z, U, X) < 1), proximal causal identification rests on two structural conditional independence conditions involving U:

1. Proximal Conditional Independence for Y (No Direct Effect of Z on Y):

1. Proximal Conditional Independence for W (No Direct Effect of (A, Z) on W):

These assumptions allow A to be confounded with Y by U (U → A and U → Y), allow Z to be associated with A, and allow W to directly cause Y. Thus Z is not required to be a valid instrumental variable (unlike Mendelian randomization, where the instrument must be marginally independent of U and must affect Y through A). In proximal causal inference, Z is confounded by U (U → Z) and does not need to cause A.

Fredholm Integral Equation of the First Kind

Because W ⊥ (A, Z) | (U, X) and Y ⊥ Z | (A, U, X), any association observed between the proxy Z and either Y or W conditional on (A, X) flows through the latent confounder U. Miao, Geng, and Tchetgen Tchetgen (2018) defined an outcome confounding bridge function h(w, a, x)—a function of the observable variables (W, A, X)—whose conditional expectation given (U, A, X) reproduces the latent regression surface E[Y | A = a, U = u, X = x]:

Integrating both sides with respect to the conditional distribution dF(u | A = a, Z = z, X = x) and invoking the two proximal conditional independencies eliminates the unobserved U, yielding an observable Fredholm integral equation of the first kind:

Every quantity in this equation is a function of the observed data distribution P(Y, W, A, Z, X). As Z varies while (A = a, X = x) are held fixed, Z perturbs the conditional distribution of U; the bridge function h(W, a, x) is the transformation of the outcome proxy W whose conditional mean shifts across levels of Z in lockstep with E[Y | A = a, Z = z, X = x].

Completeness Conditions and the Proximal G-Formula

To ensure that any solution h(W, a, x) to the observable Fredholm equation also satisfies the latent bridge equation with respect to U, proximal causal inference replaces untestable conditional exchangeability with a completeness condition:

• Completeness of U given (Z, A, X): For any square-integrable function g(U), if E[g(U) | Z = z, A = a, X = x] = 0 almost surely for all z, then g(U) = 0 almost surely. This requires that variation in the negative-control exposure Z is sufficiently informative to detect any non-zero variation in the latent confounder U (for categorical U with k categories, Z and W must each have at least k categories; for continuous U, completeness holds across canonical exp(η · T(U)) exponential-family distributions and location-scale models).

Under proximal conditional independence and completeness, Miao, Geng, and Tchetgen Tchetgen (2018) proved that the counterfactual mean E[Y(a)] is non-parametrically identified by the Proximal G-Formula:

Unlike Robins's classical G-computation formula (which evaluates E[Y | A = a, W, X] and suffers from residual confounding when W is a noisy proxy of U), the Proximal G-Formula averages the de-confounded bridge function h(W, a, X) over the marginal distribution of (W, X).

Closed-Form Linear Illustration (Structural Error-in-Variables)

Suppose the true data-generating process (suppressing X) is linear: E[Y | A, Z, U] = β₀ + β_a A + β_u U, and E[W | A, Z, U] = α₀ + α_u U (with α_u ≠ 0). Substituting U = (E[W | U] − α₀) / α_u into E[Y | A, U] shows that the outcome confounding bridge function is linear:

Ordinary least-squares regression of Y on (A, W) fails to recover (b_a, b_w) because measurement noise in W attenuates the slope on W and biases the coefficient on A. However, the observable bridge equation E[Y − b₀ − b_a A − b_w W | A, Z] = 0 uses Z as an instrument for the noisy proxy W while holding A fixed, recovering the attenuation-free bridge slope b_w = Cov(Y, Z | A) / Cov(W, Z | A) and the true causal effect b_a = β_a.

The Treatment Confounding Bridge q(Z, a, X) (Proximal IPW)

Just as classical causal inference admits both outcome regression (G-computation) and inverse probability weighting (marginal structural models), Yifan Cui, Hongming Pu, Xu Shi, Wang Miao, and Eric J. Tchetgen Tchetgen (2024) established a dual identification route via a treatment confounding bridge function q(Z, a, X) whose conditional expectation given (U, A = a, X) equals the latent inverse propensity weight 1 / f(A = a | U, X). Integrating out U against dF(u | A = a, W = w, X = x) yields the dual observable Fredholm equation:

Under completeness of W for U given (A, X), the counterfactual mean is identified by Proximal Inverse Probability Weighting (Proximal IPW):

The Doubly Robust Proximal Efficient Influence Function

Combining both bridge functions h(W, a, X) and q(Z, a, X), Cui et al. (2024) derived the semiparametric efficient influence function (EIF) for Ψ(a) = E[Y(a)] in the non-parametric proximal model:

The resulting estimator Ψ̂_DR(a) is doubly robust (analogous to targeted maximum likelihood estimation): it is consistent and asymptotically normal if either the outcome bridge h(W, a, X) or the treatment bridge q(Z, a, X) is correctly specified, and achieves the semiparametric efficiency bound when both are estimated via cross-fitted machine learning at o(n^−1/4) convergence rates.

Re-Evaluating Right Heart Catheterization in the SUPPORT Cohort

An empirical benchmark of proximal causal inference is the re-analysis of the SUPPORT critical care cohort evaluating the causal effect of invasive right heart catheterization (RHC) (A) during the first 24 hours in the intensive care unit on 30-day survival (Y). In Connors et al.'s (1996) propensity-score analysis, RHC appeared to increase 30-day mortality, raising concern that unmeasured acute physiological instability (U) drove both catheter insertion and patient death via confounding by indication. Tchetgen Tchetgen et al. (2024) and Cui et al. (2024) classified serum pH (*ph*1) and hematocrit (*hema*1) as outcome-inducing confounding proxies W, paired with arterial oxygenation and serum creatinine markers (*pafi*1, *crea*1) as treatment-inducing confounding proxies Z, revealing stronger negative confounding by latent physiological severity U than captured by standard propensity-score or targeted maximum likelihood estimation.

Longitudinal Proximal G-Estimation

In longitudinal cohort studies with time-varying treatments Ā_K and time-varying latent confounders Ū_K, Tchetgen Tchetgen et al. (2024) generalized Robins's G-methods (parametric G-formula, marginal structural models, and G-estimation of structural nested models) to the Proximal G-Computation Algorithm, solving recursive outcome and treatment bridge functions backward across visits m = K, …, 0 using longitudinal proxy pairs (Z_m, W_m).

How does proximal causal inference differ from instrumental variable (IV) analysis?

An instrumental variable Z (such as a genetic variant in Mendelian randomization) must be unconfounded with the outcome (Z ⊥ U) and must cause the exposure (Z → A). Proximal causal inference drops both requirements: the exposure proxy Z is allowed to be confounded by U (U → Z) and does not need to cause A, provided it is paired with an outcome proxy W (U → W) that is conditionally independent of (A, Z) given (U, X).

What happens if the Fredholm integral equation of the first kind is ill-posed?

When W and Z are continuous, solving E[Y | A, Z, X] = ∫ h(w, A, X) dF(w | A, Z, X) for h is a linear inverse problem stabilized via sieve generalized method of moments with Tikhonov (L₂ ridge) regularization or reproducing kernel Hilbert space (RKHS) minimax estimation. Because the target causal parameter Ψ(a) = E[h(W, a, X)] is a smooth linear functional of h, it remains √n-estimable when the dual bridge q(Z, a, X) is bounded.

Can two noisy measurements of the same confounder U serve as Z and W?

Yes. If a latent baseline confounder U is measured twice via independent assays Z = U + ε₁ and W = U + ε₂ (where ε₁ ⊥ ε₂ | U), and neither measurement noise ε₁ nor ε₂ directly drives treatment assignment or the clinical outcome beyond U, then Z and W satisfy both proximal conditional independence conditions.

How many categories or dimensions must Z and W have relative to U?

Under the completeness condition, (Z, W) must have at least as many effective degrees of variation as U: if U has k latent classes, Z and W must each have at least k categories with a full-rank conditional probability matrix P(W | Z, A, X); if U is d-dimensional continuous, Z and W should each comprise at least d continuous proxy measurements.

1. Miao, W., Geng, Z., & Tchetgen Tchetgen, E. J. (2018). "Identifying causal effects with proxy variables of an unmeasured confounder." Biometrika, 105(4), 987–993.
2. Tchetgen Tchetgen, E. J., Ying, A., Cui, Y., Shi, X., & Miao, W. (2024). "An introduction to proximal causal inference." Statistical Science, 39(3), 375–390.
3. Cui, Y., Pu, H., Shi, X., Miao, W., & Tchetgen Tchetgen, E. J. (2024). "Semiparametric proximal causal inference." Journal of the American Statistical Association, 119(546), 1348–1359.
4. Lipsitch, M., Tchetgen Tchetgen, E., & Cohen, T. (2010). "Negative controls: A tool for detecting confounding and bias in observational studies." Epidemiology, 21(3), 383–388.
5. Shi, X., Miao, W., Nelson, J. C., & Tchetgen Tchetgen, E. J. (2020). "Multiply robust causal inference with double-negative control adjustment for categorical unmeasured confounding." Journal of the Royal Statistical Society: Series B (Statistical Methodology), 82(2), 521–540.