Causal Sensitivity Testing

• Causal sensitivity testing is a method that quantifies the robustness of causal conclusions by examining how estimated effects change as violations of untestable assumptions are introduced.
• It utilizes formal models like Tukey’s factorization, marginal sensitivity models, and copula factorization to separate observed data components from latent confounding influences.
• The approach yields sharp bounds and interpretable robustness measures that guide researchers in calibrating hidden bias and making informed causal inferences in complex settings.

Causal sensitivity testing is the use of sensitivity analysis to test how robust a causal conclusion is to violations of causal assumptions that are not directly testable from the observed data, especially no unmeasured confounding. In observational studies, causal identification commonly relies on assumptions such as $[Y(0),Y(1)] \perp T \mid X$, but these assumptions are not testable from the observed data; sensitivity analysis therefore asks how conclusions change as one allows departures from ignorability, often by introducing a sensitivity parameter and examining whether estimated effects remain stable across a plausible range of that parameter (Franks et al., 2018, Díaz et al., 2023). Modern formulations place this problem in a partial-identification framework: for a causal query $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$, sensitivity testing seeks lower and upper bounds rather than a single point estimate, and interprets the resulting interval as a robustness statement about hidden bias (Frauen et al., 2023, Frauen et al., 2023).

1. Conceptual role in causal inference

Causal sensitivity testing is distinct from ordinary statistical sensitivity analysis. Statistical sensitivity analysis examines robustness to testable modeling choices, whereas causal sensitivity analysis examines robustness to untestable causal assumptions, such as no unmeasured confounding, correct time ordering, no collider adjustment, assumptions about selection or missingness, and monotonicity or instrumental-variable assumptions (Díaz et al., 2023). A recurrent theme is that a good fit to the observed data does not validate the causal assumptions used to interpret that fit.

A central motivation is the clean separation between what is identified from observed data and what is not. In flexible observational-causal frameworks, the identified part typically includes observed outcome models and observed treatment models, while the unidentified part is the dependence between treatment assignment and unobserved potential outcomes, latent confounders, or missing counterfactuals. The Tukey-factorization approach makes this separation explicit and summarizes the insight as “causal inference is missing data twice” (Franks et al., 2018).

The target of sensitivity testing is not restricted to the average treatment effect. The literature treats causal queries as functionals of potential-outcome distributions and includes ATE, ATT, CATE, CAPO, dose-response functions, quantile treatment effects, mediation and path-specific effects, disparity reduction and disparity remaining, and transported or generalized population effects (Frauen et al., 2023, Frauen et al., 2023, Huang, 2022).

2. Formal models and sensitivity parameterizations

A large share of the literature is organized around sensitivity models: families of full-data distributions that induce the same observed distribution while permitting bounded deviations from ignorability. In generalized formulations, the upper and lower causal bounds are written as

$$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$

so that the interval $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$ is the tightest guarantee under the assumed level of hidden confounding (Frauen et al., 2023).

Several parameterizations recur across the literature. Tukey’s factorization expresses the joint law for each potential-outcome arm as an identified observed-data outcome model, the observed treatment probability, and an unidentified selection function: $$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$ This formulation is attractive because sensitivity parameters affect only the unobserved part of the model and can therefore be applied post hoc on top of flexible observed-data models such as BART, mixtures, or Dirichlet process mixtures (Franks et al., 2018).

Another major family is the marginal sensitivity model and its generalizations. For binary treatment, one representative form bounds the odds ratio between the observed and latent propensities: $$\frac{1}{\Gamma}\le \mathrm{OR}\big(\pi(x),\pi(x,u)\big)\le \Gamma.$$ Generalized treatment sensitivity models replace this with a latent-distribution-shift restriction,

$$\Delta_{x,a}\big(\mathbb{P}(U\mid x),\mathbb{P}(U\mid x,a)\big)\le \Gamma,$$

and subsume the marginal sensitivity model, $f$-sensitivity models, and Rosenbaum’s sensitivity model (Frauen et al., 2023, Javurek et al., 11 May 2026).

In multi-treatment settings, copula factorization supplies a different decomposition. With treatment vector $T$, scalar outcome $Y$, observed covariates $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$0, and latent confounders $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$1, the conditional outcome model is written as

$$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$2

where the copula density $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$3 controls the hidden association between $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$4 and $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$5 given $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$6. This preserves the observed conditional marginal $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$7 and the latent treatment model $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$8, while varying only the outcome–confounder dependence (Zheng et al., 2021).

Framework	Key restriction	Representative setting
Tukey’s factorization	unidentified selection function $$Q(x,a,\mathbb{P})=\mathcal{F}(\mathbb{P}(Y(a)\mid x))$$9	observational causal inference (Franks et al., 2018)
Marginal sensitivity model	$$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$0	binary treatment (Frauen et al., 2023)
Generalized treatment sensitivity model	$$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$1	binary and continuous treatments (Javurek et al., 11 May 2026)
Copula factorization	$$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$2	multiple simultaneous treatments (Zheng et al., 2021)

3. Partial identification, sharp bounds, and ignorance regions

The contemporary literature treats sensitivity testing as a bounding problem. Under generalized marginal sensitivity models, sharp bounds can be derived for a large class of causal effects, including expectation-based effects, distributional effects, mediation effects, and path-specific effects, and the partial-identification problem is interpreted as a distribution shift in latent confounders while evaluating the causal effect of interest (Frauen et al., 2023). In the binary-treatment special case, these bounds coincide with recent optimality results for causal sensitivity analysis (Frauen et al., 2023).

In multi-treatment observational studies, a key negative result is that even if one somehow knew the conditional distribution of the confounders given treatment, $$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$3, point identification still generally fails, because the remaining unknown dependence between $$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$4 and $$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$5 leaves a range of compatible causal effects (Zheng et al., 2021). The copula construction addresses this by producing a family of data-compatible causal models and, in the linear Gaussian special case, an omitted-variable bias formula and a corresponding ignorance region for the PATE (Zheng et al., 2021).

Weighting-based sensitivity analyses often translate bound computation into tractable optimization. For weighted causal decomposition estimators, the marginal sensitivity model constrains the ratio of ideal and observed RMPW weights,

$$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$6

and the extrema over the hidden-bias functions $$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$7 are computed via fractional linear programming, yielding sensitivity bounds and percentile-bootstrap confidence intervals for the counterfactual mean (Shen et al., 2024). For risk-ratio-based or odds-ratio-based marginal sensitivity models with binary or multivalued treatments, the partially identified point-estimate interval is similarly obtained by linear fractional programming or linear programming after Charnes–Cooper transformation, followed by percentile bootstrap for uncertainty quantification (Basit et al., 2023, Basit et al., 2023).

Sensitivity testing has also been extended to limited overlap. There the problem is not hidden confounding alone, but the bias introduced by trimming or truncating weights when overlap is weak. The proposed framework places the outcome function in a Lipschitz class

$$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$8

treats $$Q^+_{\mathcal{M}}(x,a)=\sup_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}), \qquad Q^-_{\mathcal{M}}(x,a)=\inf_{\mathbb{P}\in\mathcal{M}} Q(x,a,\mathbb{P}),$$9 as the sensitivity parameter, and derives minimax confidence intervals with finite-sample coverage guarantees for the non-overlap region (Ma et al., 27 Nov 2025).

4. Calibration, robustness measures, and interpretability

A central practical question is how to choose or interpret the sensitivity parameter. Several frameworks recast hidden bias in terms of quantities that can be benchmarked. In copula-based multi-treatment sensitivity analysis, the Gaussian-copula sensitivity vector is calibrated by the partial $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$0

$[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$1

which is the main calibration scale in that paper. The sensitivity vector is further parameterized by a scalar magnitude $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$2 and a direction $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$3, so that confounding strength can be discussed separately from the direction of worst-case bias (Zheng et al., 2021).

The Tukey-factorization literature proposes an “implicit $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$4” heuristic for calibrating selection parameters against the amount of variation in treatment assignment explained by observed covariates. The practical suggestion is to choose a target $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$5, set it equal to the partial $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$6 of a familiar covariate or covariate block, and back out the corresponding $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$7 (Franks et al., 2018). Related $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$8-based parameterizations appear in decomposition analysis, where partial $[Q^-_{\mathcal{M}}(x,a),Q^+_{\mathcal{M}}(x,a)]$9 values for the mediator–confounder and outcome–confounder relationships yield robustness values for disparity reduction and disparity remaining (Park et al., 2022).

Generalization sensitivity analysis introduces a three-parameter decomposition of bias into $$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$0, $$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$1, and $$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$2: residual imbalance in the omitted moderator, alignment with treatment-effect heterogeneity, and the variance of individual treatment effects. This framework emphasizes bounded sensitivity parameters, bias contour plots, killer-confounder regions, robustness values, extreme-scenario analyses, and benchmarking against observed covariates (Huang, 2022). In multi-study generalization, the same covariance logic reappears with

$$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$3

together with robustness values and minimum relative effect-modifying strength benchmarks (Liu et al., 24 Oct 2025).

Other frameworks target interpretability more directly. In weighted causal decompositions, the one-parameter $$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$4 model is amplified to a two-parameter bias decomposition,

$$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$5

where $$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$6 is the outcome impact of $$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$7 and $$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$8 is the imbalance in $$f_\psi(Y(t), T \mid X) = f^{\rm obs}(Y(t)\mid T=t, X)\, f(T=t \mid X)\cdot \frac{f_\psi(T\mid Y(t), X)}{f_\psi(T=t\mid Y(t), X)}.$$9 induced by weighting. The same work defines critical values $$\frac{1}{\Gamma}\le \mathrm{OR}\big(\pi(x),\pi(x,u)\big)\le \Gamma.$$0 and $$\frac{1}{\Gamma}\le \mathrm{OR}\big(\pi(x),\pi(x,u)\big)\le \Gamma.$$1 to summarize how much hidden bias is needed to move the point estimate or confidence interval across the null (Shen et al., 2024). More generally, several papers recommend calibration by leave-one-covariate-out comparisons, or by asking whether an omitted confounder would need to be as prognostic or as imbalanced as a named observed covariate (Lu et al., 2023).

5. Extensions to structured causal settings

Causal sensitivity testing is now formulated for a wide range of causal structures. In multivalued-treatment settings, generalized propensity score models support additive causal estimands $$\frac{1}{\Gamma}\le \mathrm{OR}\big(\pi(x),\pi(x,u)\big)\le \Gamma.$$2, with sensitivity analyses based on modified marginal sensitivity models and efficient linear-programming implementations (Basit et al., 2023, Basit et al., 2023). In multi-treatment observational studies, copula-based approaches exploit latent treatment factor models and can identify the treatment-confounding side even when outcome-confounder dependence remains unknown (Zheng et al., 2021).

Decomposition and mediation analyses introduce another layer of structure because the key assumption is no omitted mediator–outcome confounding rather than no treatment–outcome confounding alone. Interventional disparity reduction and disparity remaining can be given regression-coefficient and $$\frac{1}{\Gamma}\le \mathrm{OR}\big(\pi(x),\pi(x,u)\big)\le \Gamma.$$3-based sensitivity analyses, including settings with pre-exposure and intermediate unobserved confounding (Park et al., 2022). Weighted decomposition estimators based on ratio-of-mediator-probability weighting similarly admit marginal-sensitivity-model bounds, asymptotically valid percentile-bootstrap intervals, and two-parameter amplification (Shen et al., 2024).

Generalization and transportability add sensitivity to omitted effect modifiers. For a single randomized study, weighted estimators can be analyzed through the covariance between weight error and treatment-effect heterogeneity, without assuming a parametric data-generating process (Huang, 2022). In multiple studies, the same bias structure extends to combinations of randomized trials and observational studies, and the multi-study setting also permits an exploratory Wald-type diagnostic for unobserved effect modification across studies (Liu et al., 24 Oct 2025). Regulatory-science guidance further emphasizes that such sensitivity analyses should be pre-specified, use sensitivity parameters with scientific meaning, and should not be treated as a substitute for good design (Díaz et al., 2023).

Longitudinal, clustered, and interference settings require additional structure. For clustered observational data, the true and confounded treatment effects are linked by a bias factor $$\frac{1}{\Gamma}\le \mathrm{OR}\big(\pi(x),\pi(x,u)\big)\le \Gamma.$$4, and robustness can be summarized by the minimal bias factor needed to explain away the observed effect in single studies or meta-analysis (Ou et al., 2023). For time-varying treatment and time-varying unmeasured confounding, Bayesian latent-variable sensitivity analysis and Bayesian sensitivity-function methods extend g-computation and marginal structural models to longitudinal settings (Zou et al., 12 Jun 2025). Under interference, weighting-based sensitivity analysis decomposes bias into baseline-outcome, spillover, and transportability components, allowing practitioners to assess unmeasured confounding, ignored interference, and lack of transportability simultaneously (Ortyashov et al., 26 Nov 2025).

6. Computation, neural methods, and newer applications

Recent work has shifted from hand-derived closed forms toward flexible computational engines. NeuralCSA formulates generalized causal sensitivity analysis with two conditional normalizing flows: one to learn the observational outcome mechanism and one to learn the confounding-induced latent shift. This makes the method compatible with the marginal sensitivity model, $$\frac{1}{\Gamma}\le \mathrm{OR}\big(\pi(x),\pi(x,u)\big)\le \Gamma.$$5-sensitivity models, and Rosenbaum’s sensitivity model, and with binary or continuous treatments, multiple outcomes, and a broad class of causal queries (Frauen et al., 2023).

A second computational direction is amortization. Prior-data-fitted networks have been used to turn causal sensitivity analysis from a per-instance optimization into an in-context-learning problem: a model is trained on tuples $$\frac{1}{\Gamma}\le \mathrm{OR}\big(\pi(x),\pi(x,u)\big)\le \Gamma.$$6, and labels are generated by Lagrangian scalarization over a generalized treatment sensitivity model. Under convexity and linearity conditions, the Lagrangian sweep recovers the full Pareto frontier, and the resulting test-time computation is reported to be orders of magnitude faster than per-instance methods (Javurek et al., 11 May 2026).

The notion of sensitivity testing has also broadened in application. In Bayesian Causal Impact models for marketing interventions, a simulation-based framework perturbs observed data, injects controlled degradations, and estimates alarm activation probabilities across effect magnitudes, confidence levels, and monitoring horizons. Two alarm criteria are studied: a proportion-based rule using predictive lower bounds and a persistence-based rule based on three consecutive days of negative cumulative impact; the latter is reported to provide more stable and operationally meaningful detection behavior (Pellegrini, 6 Jul 2026). In causal fairness, distribution-based potential-outcomes formulations test whether factual and counterfactual outcome distributions are sufficiently close under intervention on a sensitive attribute, with the closeness parameter $$\frac{1}{\Gamma}\le \mathrm{OR}\big(\pi(x),\pi(x,u)\big)\le \Gamma.$$7 functioning as an explicit sensitivity knob (Fu et al., 18 Feb 2025).

A persistent misconception in this literature is that identifying more of the observed-data or treatment-assignment structure automatically identifies the causal effect. The modern results are more restrictive: learning the latent structure in treatment alone is generally insufficient, the observed-data model can be fit extremely well while causal assumptions still fail, and sensitivity analysis does not prove causality. Its role is to quantify how large a violation would need to be to overturn a substantive conclusion, and to make that quantification interpretable through bounds, calibration, and explicitly chosen sensitivity parameters (Zheng et al., 2021, Franks et al., 2018, Díaz et al., 2023).