Negative Control Outcomes

• Negative control outcomes are variables known not to be causally affected by a treatment but that signal underlying bias like unmeasured confounding or selection effects.
• They serve as diagnostic tools in causal inference, helping researchers identify violations in instrumental variable and exchangeability assumptions.
• Advanced methods, such as proximal causal inference, leverage NCOs to construct bridge functions that recover causal effects even in complex, high-dimensional settings.

Searching arXiv for recent and foundational papers on Negative Control Outcomes to ground the article. Negative control outcomes (NCOs) are auxiliary outcomes used to interrogate or exploit causal null relations. Across the modern literature, an NCO is an outcome that is known not to be causally affected by the treatment or exposure of interest, yet remains informative about the latent structure that threatens causal identification—most commonly unmeasured confounding, but in some settings also selection mechanisms, latent heterogeneity, or batch effects (Shi et al., 2020). In instrumental-variable analysis, the term is used more narrowly for an observed proxy of an outcome-side threat to IV validity (Danieli et al., 2023). In proximal and double-negative-control inference, NCOs are outcome confounding proxies used jointly with exposure-side proxies to identify bridge functions and recover causal estimands despite latent confounding (Miao et al., 2018). The resulting literature treats NCOs not as a single heuristic, but as a family of formally defined devices whose role depends on design, estimand, and identification strategy.

1. Conceptual scope and terminology

In the epidemiologic formulation, a negative control outcome is an outcome-like variable that should not respond to the treatment, but should share the same hidden bias mechanism as the primary outcome. One paper states this directly: “A negative control outcome (NCO) is a variable that is not causally affected by the treatment of interest but shares a similar confounding structure with the treatment-outcome relationship” (Portela et al., 30 Oct 2025). The review literature makes the same point in broader terms, defining an NCO $W$ as a variable known not to be causally affected by the treatment $A$, with the additional requirement that it remain associated with the unmeasured confounder $U$ so that it is informative about residual bias rather than merely irrelevant (Shi et al., 2020).

A more formal conceptual refinement appears in the potential-outcomes taxonomy of experimental controls. There, what is commonly called a negative control outcome is best understood as a null contrast-control on a secondary outcome under the primary active-versus-control treatment contrast, not simply as a “null outcome-control” under the active treatment alone. In that formulation, the defining null is

$Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$

which distinguishes relative no-effect claims from absolute statements such as $Y_i^{no}(w_{at})=0$ (Hunter et al., 2021). This distinction matters because a variable can change under both treatment arms and still be a valid NCO if the treatment contrast of interest has no effect on it.

The instrumental-variables literature sharpens the terminology further. In “Negative Control Falsification Tests for Instrumental Variable Designs,” an NCO is an observed proxy for an alternative path outcome (APO) variable, that is, a latent outcome-side threat creating a path from the instrument $Z$ to the outcome $Y$ other than the intended path through treatment $X$. This contrasts with a negative control instrument (NCI), which is an observed proxy for an alternative path instrument (API) variable (Danieli et al., 2023). In that framework, an NCO is not merely “a variable that should not move”; it is a formally constrained proxy for a specific latent threat.

High-dimensional outcome settings introduce a further generalization. In post-integrated inference, a subset $Y_{\mathcal C}$ of outcomes is designated as control outcomes satisfying

$(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$

so that they can be used to estimate latent embeddings $A$0 for downstream direct-effect inference (Du et al., 2024). Here again, the defining feature is not temporal priority but exclusion from the treatment effect together with informativeness about hidden structure.

2. Formal causal conditions

The minimal causal content of an NCO is exclusion from the treatment effect. In the review literature this appears as

$A$1

with $A$2 additionally required to be associated with $A$3 conditional on $A$4 (Shi et al., 2020). In the categorical double-negative-control framework the same exclusion is written as

$A$5

so that $A$6 is causally unaffected by both the primary treatment $A$7 and the auxiliary exposure $A$8 (Shi et al., 2018). The test-negative-design vaccine-effectiveness literature uses analogous exclusion restrictions, requiring an NCO $A$9 to be known a priori not to be a causal effect of either vaccination $U$0 or the negative control exposure $U$1, and imposing

$U$2

together with an additional sampling condition in the selected sample (Li et al., 2022).

In IV designs, the formal structure is different because the treatment threat is mediated by the instrument. The core IV validity condition under the paper’s notation is

$U$3

An APO variable $U$4 satisfies

$U$5

together with a path-indication condition, and an NCO $U$6 is defined by the existence of such a $U$7 for which

$U$8

The first clause is the NCO exclusion restriction, while the second is the proxy-relevance requirement (Danieli et al., 2023).

Proximal causal inference recasts the same ideas in proxy language. With observed data $U$9, the outcome proxy $Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$0 is required to satisfy

$Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$1

while the treatment proxy $Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$2 satisfies

$Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$3

and latent exchangeability is written

$Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$4

Under this formulation, $Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$5 is not simply an unaffected auxiliary endpoint; it is an outcome-side confounding proxy whose observed conditional distribution contains information about $Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$6 (Liu et al., 2024).

Single-proxy control introduces an even more specific condition targeted to the untreated potential outcome. There the NCO $Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$7 satisfies

$Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$8

so that $Y_i^{nc}(w_{at}) - Y_i^{nc}(w_{ct}) = 0,$9 functions as a proxy for the treatment-free counterfactual outcome $Y_i^{no}(w_{at})=0$0 rather than for an abstract latent confounder (Park et al., 2023).

Randomized-trial adjustment with an NCO uses the sharpest exclusion statement in this literature. A valid post-randomization NCO $Y_i^{no}(w_{at})=0$1 must satisfy

$Y_i^{no}(w_{at})=0$2

which permits adjustment for $Y_i^{no}(w_{at})=0$3 without inducing post-randomization bias because the auxiliary outcome is completely unaffected by treatment (Ashby et al., 2024).

3. Falsification and diagnostic use

The oldest and most widely used role of an NCO is diagnostic. The basic logic is asymmetrical: if treatment appears to affect an outcome it cannot causally affect, the causal design is suspect; if no such association is found, the result is at most indirect support. The causal machine-learning simulation study states this directly: because the NCO is known a priori to have no true causal association with treatment, any observed association suggests residual confounding, whereas the absence of such association may provide only indirect evidence supporting conditional exchangeability (Portela et al., 30 Oct 2025). The review literature makes the same point more generally for observational epidemiology: a non-null association between $Y_i^{no}(w_{at})=0$4 and $Y_i^{no}(w_{at})=0$5 after adjustment for measured $Y_i^{no}(w_{at})=0$6 indicates some failure of the assumed causal structure, often residual confounding, while a null finding is not proof of validity (Shi et al., 2020).

IV designs provide the cleanest formal falsification theorems. For an NCO test, the null is

$Y_i^{no}(w_{at})=0$7

or, with controls,

$Y_i^{no}(w_{at})=0$8

If $Y_i^{no}(w_{at})=0$9 is a valid NCO, then $Z$0 implies that either outcome independence or the exclusion restriction is violated, hence the IV design is invalid. The paper is equally explicit that non-rejection does not validate the IV: the NCO may be weak, the sample may be small, or the true threat may not be captured by the chosen proxy (Danieli et al., 2023).

This diagnostic role extends naturally to subgrouped causal machine learning. In the simulation study of individualized treatment effects, causal forest and X-learner models were trained on the primary outcome, individuals were ranked by predicted benefit, and quartiles $Z$1–$Z$2 were formed. Within each quartile, the apparent treatment effect on the NCO was then estimated by logistic regression. Under no unmeasured confounding, subgroup NCO effects were near zero; when conditional exchangeability was violated, non-zero subgroup effects emerged, sometimes most strongly in the subgroup where the primary-outcome treatment effect was most distorted (Portela et al., 30 Oct 2025). In that usage, the NCO is a post-estimation falsification test and model-criticism device for local rather than global credibility.

The same literature supplies concrete examples. Martin and Yurukoglu’s Fox News design uses lagged Republican vote share as an NCO for local conservatism; Angrist and Evans, with Rosenzweig and Wolpin’s expenditure channel, motivate clothing expenditure as an NCO for an exclusion-restriction threat; Deming’s school-lottery application shows that multiple NCOs and joint tests can reveal an omitted control in IV construction; and shift-share designs gain power when many predetermined labor-market variables are used as NCOs jointly (Danieli et al., 2023). The general lesson is that NCOs diagnose specific latent threats rather than serving as generic placebo outcomes.

4. Identification with double negative controls

Under stronger assumptions, NCOs do more than falsify. In double-negative-control inference they participate directly in nonparametric identification. The central object is the outcome confounding bridge function $Z$3, defined so that

$Z$4

Under latent ignorability, NCO, bridge, NCE, and completeness assumptions, this yields

$Z$5

while the bridge itself is identified from the observed-data equation

$Z$6

This framework, introduced for double negative control inference on causal effects, turns the NCO from a detector of bias into an observable vehicle for representing the confounding component of the primary outcome (Miao et al., 2018).

The categorical extension weakens the original invertibility requirements. With categorical $Z$7, $Z$8, and $Z$9, the paper defines matrices such as $Y$0, $Y$1, and $Y$2, and shows that identification of the ATE can proceed under

$Y$3

provided the latent proxy matrices have rank $Y$4 (Shi et al., 2018). This distinguishes the estimand from the bridge: even when the bridge vector $Y$5 is not unique, the causal estimand remains uniquely identified.

Test-negative design studies of vaccine effectiveness adapt the same logic to outcome-dependent sampling. There the NCO $Y$6 is not merely a bias check; it identifies the treatment confounding bridge $Y$7 through observed-data equations among test-negative controls. Under treatment-independent sampling, negative-control assumptions, and a rare infection approximation, the bridge satisfies an observed-data relation in the $Y$8 subset, which then yields a de-biased estimator of the causal log-risk-ratio and vaccine effectiveness (Li et al., 2022). The NCO is indispensable because it makes the bridge calibratable from selected controls.

5. Single-proxy, regression, kernel, and automated discovery methods

A large methodological literature develops estimation strategies once NCO assumptions are granted. Regression-based proximal causal inference replaces abstract integral equations with two-stage generalized linear models. In this approach, $Y$9 is modeled in a first-stage regression such as $X$0, and the fitted linear predictor

$X$1

is carried into a second-stage model for $X$2, for example

$X$3

For continuous, count, binary, and polytomous outcomes, the treatment coefficient satisfies $X$4 under the proxy assumptions, so the NCO-derived regressor $X$5 functions as a proximal control for latent confounding (Liu et al., 2024).

Kernel and RKHS methods provide a nonparametric analogue. In “Kernel Methods for Unobserved Confounding,” the NCO $X$6 enters a confounding bridge equation

$X$7

where the first stage estimates the conditional mean embedding of $X$8, and the second stage estimates the bridge $X$9. This yields nonparametric estimators of dose-response curves, ATT-type functionals, CATEs, and distribution-shifted effects under negative-control and completeness assumptions (Singh, 2020).

Single-proxy control shows that one NCO can suffice for the effect of treatment on the treated. With

$Y_{\mathcal C}$0

the paper develops two nonparametric identification routes. The first uses an extended propensity score

$Y_{\mathcal C}$1

linked to the observed data through an integral equation involving $Y_{\mathcal C}$2. The second posits a COCA confounding bridge $Y_{\mathcal C}$3 satisfying

$Y_{\mathcal C}$4

which yields

$Y_{\mathcal C}$5

The paper further derives a doubly robust influence-function-based estimator (Park et al., 2023).

Negative-control outcomes also appear in nonparametric causal hypothesis testing. For a single NCO $Y_{\mathcal C}$6 satisfying

$Y_{\mathcal C}$7

the null hypothesis

$Y_{\mathcal C}$8

implies the bridge equation

$Y_{\mathcal C}$9

This motivates the Proxy Maximum Characteristic Restriction and a kernel-based test that exploits characteristic-function restrictions rather than only first moments. The same paper also shows that a single NCO can fail to identify the null under alternatives, and introduces an NCE $(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$0 to restore identifiability through a stronger bridge condition involving $(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$1 and $(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$2 (Wu et al., 20 Oct 2025).

Finally, DANCE adds a data-driven discovery layer for a special subclass of negative controls: disconnected negative controls. Under a simple NC model with a single latent $(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$3, DANCE searches over candidate triplets and validates them using six vanishing tetrad tests. Validated variables can then serve as ordered $(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$4 pairs in double-NC estimation, and estimates are aggregated across pairs (Kummerfeld et al., 2022). This is not a general NCO validator, but it is a formal search procedure for NCO-compatible proxies under a restricted graph class.

6. Specialized domains and applications

Several papers adapt NCOs to specialized inferential settings. In right-censored survival analysis, regression-based proximal causal inference for additive hazards models allows three NCO types: continuous, count, and right-censored time-to-event variables. For a survival NCO $(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$5, the first stage fits an additive hazards model

$(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$6

and carries the fitted linear predictor $(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$7 into the second-stage additive hazards model for the primary event time $(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$8 (Li et al., 2024). Here the NCO is an outcome confounding proxy within a fully survival-analytic proximal framework.

Randomized trials invert the usual observational emphasis. Because treatment is randomized, the ATE is already identified, and the NCO becomes a precision device rather than an identification device. In early-phase vaccine trials, a valid post-randomization NCO $(Y_{\mathcal C^c},X)\perp Y_{\mathcal C}\mid U,$9 satisfies

$A$00

so it can be used in augmented estimators analogously to a prognostic baseline covariate. The paper shows that adjustment for $A$01 is asymptotically at least as efficient as adjustment for $A$02 alone, and recommends parsimonious working models, HC3 variance corrections, and quantile transforms when NCOs are skewed or subject to detection limits (Ashby et al., 2024).

High-dimensional outcome integration provides another extension. In post-integrated inference, control outcomes $A$03 are used to estimate latent embeddings $A$04, and inference then targets projected direct effects such as

$A$05

The paper emphasizes that these estimands remain statistically meaningful under model misspecification and with error-prone embeddings, provided the NCO subset satisfies the required exclusion and completeness conditions (Du et al., 2024).

Causal machine learning uses NCOs yet differently. The cited simulation study does not use the NCO to debias the learner directly. Instead, NCOs are incorporated after ITE estimation as subgroup-level diagnostics: predicted-benefit quartiles $A$06–$A$07 are formed from the primary-outcome model, and apparent treatment effects on the NCO are estimated within each subgroup. Under ideal NCO assumptions, the largest NCO deviation often appears in the subgroup where the primary-outcome estimates are most distorted by unmeasured confounding (Portela et al., 30 Oct 2025).

7. Limitations, failure modes, and interpretation

The NCO literature is unusually explicit about failure modes. First, NCOs are indirect diagnostics. A null result does not prove exchangeability, IV validity, or absence of hidden bias; it only provides supporting evidence under the maintained NCO assumptions (Portela et al., 30 Oct 2025). The broader experimental-controls framework makes the same asymmetry explicit: a violated control can indicate either that prior scientific knowledge was wrong or that the study has a flaw, while a satisfied control cannot prove that the study is sound (Hunter et al., 2021).

Second, not every placebo outcome is a valid NCO. The IV falsification paper stresses that an NCO must be an observed proxy for a specific latent threat and must satisfy both an exclusion-type restriction and a relevance condition. A variable may look pre-treatment or irrelevant yet fail because it has another path to the instrument. The pollution example makes this concrete: non-respiratory hospital admissions were used as placebo outcomes, but Guidetti et al. showed that pollution affected them through hospital congestion, violating the NCO assumption (Danieli et al., 2023).

Third, regression implementations may test more than the causal assumption. In IV settings, common linear NCO regressions can reject not only because outcome independence or exclusion fails, but also because the linear specification for controls is wrong. Under the paper’s “rich covariates” assumption,

$A$08

conditional independence implies

$A$09

so a nonzero OLS coefficient on the NCO can signal either causal invalidity or failure of the functional-form restriction (Danieli et al., 2023).

Fourth, identification strategies using NCOs depend on strong completeness, rank, or relevance conditions. Double-negative-control bridge methods require proxy richness; kernel and proximal methods become unstable when the NCO is only weakly related to the hidden confounder; single-NCO causal testing can fail when the bridge equation is solvable under alternatives; and test-negative-design bridge identification depends on a rare-disease approximation outside the null-preserving case (Miao et al., 2018). The causal-machine-learning study adds a practical version of the same point: NCOs are most effective when the event is sufficiently common, and imperfect NCOs may still flag bias while misidentifying the most biased subgroup (Portela et al., 30 Oct 2025).

Finally, some methods are deliberately narrow. DANCE validates only disconnected negative controls under a simple NC model with one latent confounder and linear acyclic measured-variable structure; it does not provide a general test of NCO validity (Kummerfeld et al., 2022). Randomized-trial NCO adjustment requires the unusually strong unit-level null $A$10 for every participant (Ashby et al., 2024). These limitations are not peripheral; they define the scope of what each NCO method can legitimately claim.

Taken together, the literature presents NCOs as formally structured auxiliary outcomes whose usefulness depends on two features held in tension: exclusion from the treatment effect and informativeness about the hidden structure that threatens causal inference. When used only for falsification, they provide asymmetrical evidence against a design. When combined with additional proxy variables, bridge functions, completeness conditions, or semiparametric estimating equations, they can support deconfounding, efficiency gains, and even causal hypothesis testing. The conceptual unity of the field lies precisely in that dual role: NCOs are outcomes that should not move causally, but should still move with the bias one seeks to detect or remove.