Source-linked AI summary
On Measurement Bias in Causal Inference
Judea Pearl
TL;DR
The paper addresses estimating the causal effect of X on Y when a sufficient confounder Z is observed only through a noisy proxy W, which ordinary adjustment cannot adequately control. It develops effect-restoration methods using the measurement-error mechanism and analyzes their causal conditions. With the mechanism specified, modified adjustment can yield an asymptotically unbiased effect estimate, subject to limitations from misspecification and incompatibility.
Problem
When sufficient confounders Z are unobserved and only noisy proxies W are available, ordinary adjustment leaves the back-door path from X to Y open.
Source: §1 · p. 1Method
Results
Given P(w|z), modified adjustment can produce an asymptotically unbiased point estimate of P(y|do(x)).
Source: §1 · p. 1Takeaways & Limitations
The paper describes synthetic-sample construction and modified inverse-probability weighting as ways to address measurement bias, alongside conditions for removing it in linear systems.
Source: §6 · p. 7
Abstract
from arXiv · showhide
This paper addresses the problem of measurement errors in causal inference and highlights several algebraic and graphical methods for eliminating systematic bias induced by such errors. In particulars, the paper discusses the control of partially observable confounders in parametric and non parametric models and the computational problem of obtaining bias-free effect estimates in such models.
1 INTRODUCTION
The introduction frames causal effect estimation with an unobserved sufficient confounder and a noisy proxy as a problem that ordinary adjustment cannot solve. It motivates effect restoration using a specified measurement-error mechanism, while emphasizing the causal perspective and the method’s assumptions and aims.
1 INTRODUCTION
- When sufficient confounders Z are unobserved and only noisy proxies W are available, adjusting for W leaves the back-door path from X to Y open.The causal effect is identified when Z is observed, but not by ordinary adjustment for W alone.Source: §1 · p. 1
- Given the error mechanism P(w|z), modified adjustment can recover an asymptotically unbiased point estimate of P(y|do(x)) despite uncertainty about each latent Z.The correction is presented as asymptotically equivalent to observing and adjusting for Z.Source: §1 · p. 1
- When P(w|z) is unavailable, the paper describes using priors or bounds for that mechanism, or estimating it by retesting Z in a sampled subpopulation.These options yield distributions or bounds for the causal effect, or an estimate of the measurement mechanism.Source: §1 · p. 1
- Effect restoration increases sampling variability as measurement error grows, and a misspecified P(w|z) can leave estimates biased even with unbounded sample size.In extreme cases, the postulated mechanism can conflict with observed data so that no estimate is obtained.Source: §1 · p. 2
- The paper distinguishes causal analysis, which targets P(y|do(x)), from regression analysis, which expresses that target using regression quantities only under plausible causal assumptions.It argues that specifying P(w|z) and judging its transportability also require causal considerations.Source: §1 · p. 2
- The paper identifies a mutual gap: partially specified causal parameters are rarely treated in causal inference, while graphical models are little used in measurement-error research.It presents its aims as analyzing feasibility, computational problems, and features that contribute to bias or robustness, with an asymptotic focus.Source: §1 · p. 2
2 EFFECT RESTORATION BY MATRIX ADJUSTMENT
Inverting the measurement mechanism P(w|z) reconstructs the joint distribution involving the latent confounder Z, enabling causal-effect estimation when Z is sufficient. Factorization, sparse inversion, and propensity-score methods address computational and high-dimensional data challenges.
2 EFFECT RESTORATION BY MATRIX ADJUSTMENT
- Inverting M(w, z) = P(w|z) reconstructs P(x, y, z), from which the causal effect is estimable when Z is sufficient and the error mechanism is known.The assumption that W depends only on Z is termed nondifferential error.Source: §2 · pp. 2–3
- Measurement errors in X and Y can be absorbed into the proxy vector W without creating a conceptual problem.Source: §2 · p. 3
- Independent local noise mechanisms factor P(w|z) and make its inverse decomposable, while sparse couplings permit parsimonious parameterization.The paper suggests Bayesian networks for representing sparse couplings.Source: §2 · p. 3
- Although Eq. (4) entails exponentially long summations over w, sparse nonnegligible entries of I(z, w) allow efficient computation using Bayesian-network inference.The computational shortcut relies on only a small number of w values having nonnegligible I(z, w) for each z.Source: §2 · p. 3
- High-dimensional Z and W can create empty cells that prevent reliable estimation of P(x, y, w), motivating propensity-score methods that map Z cells to a scalar.The error-free score L(z) can be computed from I(z, w), L(w), and P(w), then used to estimate the causal effect.Source: §2 · pp. 3–4
- Synthetic samples of P(x, y, z) that mirror empirical P(x, y, w) samples offer an approach to constructing P(l) and P(y|x, l).A preceding proposal does not fully use the inversion or graphical methods.Source: §2 · p. 4
3 EFFECT RESTORATION IN BINARY MODELS
For binary variables, the paper expresses effect restoration as an inverse-matrix transformation of observed cells into a synthetic distribution over latent cells. The transformation has compatibility and interpretability limits, while modified weighting factors can also assess bias from ignoring measurement error.
3 EFFECT RESTORATION IN BINARY MODELS
- The inverse-matrix transformation maps observed P(x, y, w) cells to a synthetic P(x, y, z) distribution, which can be substituted into the causal-effect estimand.For K binary confounders with K local indicators, the paper also describes generating synthetic samples to estimate the true propensity score and causal effect.Source: §3 · pp. 4–5
- Each observed cell-pair's combined weight P(x, y) is reassigned between latent cells according to P(w1|x, y) and the error rates.As P(w1|x, y) approaches 1−ϵ, most weight goes to z1; as it approaches δ, most goes to z0.Source: §3 · p. 4, Figure 2 · p. 4
- When ϵ + δ = 1, W carries no information about Z and the inverse does not exist; synthetic probabilities outside (0, 1) indicate incompatibility with the modeled structure.The incompatibility is between the observed distribution and the postulated error mechanism under Fig. 1's structure.Source: §3 · p. 4
- The extra factors in Eq. (8) modify inverse-probability weights for bias-free estimation and can assess bias from treating W as a faithful measure of Z.Source: §3 · pp. 4–5
- With two error parameters and eight cells, the causal-effect expression does not yield an intuitive account of how ϵ and δ affect the estimand.The paper says linear models facilitate this evaluation.Source: §3 · p. 5
4 EFFECT RESTORATION IN LINEAR MODELS
Linear models clarify which measurement-error parameters are needed to identify the causal effect coefficient c0. With multiple proxies, the pivotal product can be estimated from data; with one proxy, it must be assessed externally.
4 EFFECT RESTORATION IN LINEAR MODELS
- With two independent indicators of Z, the product c3^2 var(Z) is identifiable, even though factor loadings may remain only partially identifiable.Partial identifiability of the factor loadings does not prevent identifying c0.Source: §4 · p. 5
- The product c3^2 var(Z) is sufficient to identify c0 and can be estimated from data when several proxies for Z are available.With only one proxy, the product must be estimated externally using a pilot study or judgmental assessment.Source: §4 · pp. 5–6
- Under a Gaussian error model, c3 and var(ϵW) characterize f(w|z), while estimating c3 alone is insufficient to neutralize confounding.The paper also notes that latent-factor analysis can identify causal effects without separately identifying c3 from var(Z).Source: §4 · p. 6
- In the noiseless case, adjusting for W gives the same bias-free coefficient c0 as adjusting for Z, and c3 plays no role.The section contrasts this with error-prone measurement, where c0 is expressed using a different formula.Source: §4 · p. 6
5 MODEL TESTING WITH MEASUREMENT ERROR
The section develops observable covariance constraints for testing graphical claims when confounders are measured through proxies. It also describes a sample-based surrogate test and extends the diagnostic approach to missing edges involving latent variables.
5 MODEL TESTING WITH MEASUREMENT ERROR
- Measurement error makes the usual conditional-independence tests unavailable because proxies of Z no longer d-separate X and Y.Surrogate tests are possible when the error-process parameters are known or can be estimated.Source: §5 · p. 6
- If latent Z d-separates X and Y and has proxy W, their covariance must satisfy a constraint that can serve as a surrogate test.The usual vanishing partial-regression condition is modified to require separate estimation of c^2 var(Z); with other proxies, the condition takes TETRAD form.Source: §5 · p. 7
- The sample procedure transforms observations into fictitious Yi, Vi pairs, regresses Yi on Vi, and treats a sufficiently confident zero slope as compatible with X ⊥⊥ Y | Z.The listed steps collect Xi, Yi, Wi, form Vi = Xi − c1/αWi, and test whether the fitted coefficient a vanishes.Source: §5 · p. 7
- The diagnostic generalizes to missing edges between latent variables and between latent and observed variables, using the implied covariance constraint before or instead of a global fit test.The condition is that filling in the missing edge permits identification of its coefficient.Source: §5 · p. 7
6 CONCLUSIONS
The paper addresses effect restoration when confounders are mismeasured or misclassified. It describes computational and graphical routes for removing measurement bias.
6 CONCLUSIONS
- The paper explains how to create synthetic samples from empirical samples to remove measurement bias.Source: §6 · p. 7
- It shows how inverse-probability weighting can be modified to account for measurement error.Source: §6 · p. 7
- For linear systems, the paper analyzes measurement bias and gives graphical conditions under which it can be removed.Source: §6 · p. 7