Calibrating Bias with Negative-Control Associations

Why Calibrate With a Negative Control?

A negative-control exposure is a variable that is not expected to cause the outcome, but may share causes with the outcome or treatment. Its adjusted association with the outcome can therefore serve as a diagnostic signal: perhaps some outcome variation attributed to an exposure is actually associated with hidden confounding. Calibration takes that signal one step further by asking how large the bias in the treatment estimate might be if the hidden confounding acts similarly for the negative control and treatment.

That final step is not automatic. A negative-control association is not itself the treatment bias, and the absence of a causal effect of the negative-control exposure does not guarantee that its observed association is entirely due to the same confounding that biases the treatment estimate. A useful calibration states the model and the comparisons it needs, then shows how different plausible comparisons yield different bias magnitudes.

Definition: A negative-control exposure is an exposure that, under the scientific model being considered, has no causal effect on the outcome, but can be measured to assess possible bias. Its association with the outcome is a diagnostic association, not an estimate of a causal effect. Calibration uses that association to quantify a plausible bias in a target treatment estimate by specifying how the relevant confounding relationships compare.

A Linear Model for the Calibration

Consider a continuous outcome $Y$, a treatment or exposure $A$, measured covariates $C$, an unmeasured confounder $U$, and a negative-control exposure $Z$. Suppose the outcome follows an additive linear model with a constant treatment effect:

Y = m(C) + \tau A + \beta U + \varepsilon.

Here $\tau$ is the causal effect of $A$ on $Y$ per one-unit increase in $A$, under this model; $\beta$ is the outcome change per one-unit increase in $U$; and $m(C)$ is the part explained by measured covariates. The model assumes there is no causal effect of $Z$ on $Y$: $Z$ does not appear as a causal term. It also assumes that the stated linear terms adequately represent the relevant relationships. In this section, all regression coefficients are adjusted for $C$.

Write $A_r$, $Z_r$, $U_r$, and $Y_r$ for variables after removing their linear projections on $C$, and center the residuals. Let $\varepsilon_r$ be the corresponding residual error. The formulas below require positive residual variances for $A_r$ and $Z_r$. They also require $\operatorname{Cov}(A_r,\varepsilon_r)=0$; otherwise, the treatment coefficient contains an additional error-association term. These are assumptions, not consequences of having adjusted for $C$.

Key result: Under the linear model and $\operatorname{Cov}(A_r,\varepsilon_r)=0$, the adjusted treatment regression coefficient is $$ b_A = \tau + \beta \frac{\operatorname{Cov}(A_r,U_r)}{\operatorname{Var}(A_r)}. $$ The second term is the omitted-confounding bias in this coefficient relative to $\tau$. A negative-control coefficient $b_Z$ can identify a corresponding component only under additional assumptions: in particular, $\operatorname{Cov}(Z_r,A_r)=0$, $\operatorname{Cov}(Z_r,\varepsilon_r)=0$, and no causal effect of $Z$ on $Y$.

Deriving the Treatment Bias and Negative-Control Association

The treatment formula follows directly from the adjusted regression slope. For a simple regression of $Y_r$ on $A_r$, that slope is $\operatorname{Cov}(A_r,Y_r)/\operatorname{Var}(A_r)$. Substituting the residualized outcome model gives:

\begin{aligned} b_A &= \frac{\operatorname{Cov}(A_r,Y_r)}{\operatorname{Var}(A_r)} \\ &= \frac{\operatorname{Cov}\{A_r,\tau A_r+\beta U_r+\varepsilon_r\}}{\operatorname{Var}(A_r)} \\ &= \tau +\beta\frac{\operatorname{Cov}(A_r,U_r)}{\operatorname{Var}(A_r)} +\frac{\operatorname{Cov}(A_r,\varepsilon_r)}{\operatorname{Var}(A_r)} \\ &= \tau +\beta\frac{\operatorname{Cov}(A_r,U_r)}{\operatorname{Var}(A_r)}. \end{aligned}

The last equality uses the stated zero-covariance assumption for the residual error. Thus, even in this simple model, an adjusted coefficient need not equal the causal effect: it differs when $A_r$ and $U_r$ are associated and $U$ predicts $Y$.

Now regress $Y_r$ on $Z_r$. The negative-control exposure has no causal term in the outcome model, but $Z_r$ can still be associated with $Y_r$. The full slope is:

b_Z = \tau\frac{\operatorname{Cov}(Z_r,A_r)}{\operatorname{Var}(Z_r)} +\beta\frac{\operatorname{Cov}(Z_r,U_r)}{\operatorname{Var}(Z_r)} +\frac{\operatorname{Cov}(Z_r,\varepsilon_r)}{\operatorname{Var}(Z_r)}.

This expression matters because “$Z$ has no causal effect on $Y$” alone does not remove the treatment term: if $Z$ and $A$ are associated, the causal effect of $A$ can contribute to the $Z$-outcome regression slope. For the simple calibration below, impose both $\operatorname{Cov}(Z_r,A_r)=0$ and $\operatorname{Cov}(Z_r,\varepsilon_r)=0$. Then $b_Z=\beta\operatorname{Cov}(Z_r,U_r)/\operatorname{Var}(Z_r)$: the association is attributable to the shared $U$ component in this model.

Define the confounder association per exposure unit as $\kappa_A=\operatorname{Cov}(A_r,U_r)/\operatorname{Var}(A_r)$ for treatment and $\kappa_Z=\operatorname{Cov}(Z_r,U_r)/\operatorname{Var}(Z_r)$ for the negative control. If $\kappa_Z\ne 0$, define the comparability factor $\lambda=\kappa_A/\kappa_Z$. The treatment bias is then $\beta\kappa_A=\lambda b_Z$. This is the calibration: the observed negative-control coefficient is multiplied by a stated assumption about the relative confounder associations.

Calibration formula: Under the assumptions just stated, $$ \text{bias in } b_A = \lambda b_Z, \qquad \tau_{\text{calibrated}} = b_A-\lambda b_Z, \qquad \lambda = \frac{\operatorname{Cov}(A_r,U_r)/\operatorname{Var}(A_r)} {\operatorname{Cov}(Z_r,U_r)/\operatorname{Var}(Z_r)}. $$ The data may estimate $b_A$ and $b_Z$, but they do not, by themselves, establish the value of $\lambda$.

Worked Calibrations

Worked Example: A Specified Comparability Factor

Suppose a hypothetical adjusted analysis estimates a treatment coefficient of $b_A=2.40$ outcome units per treatment unit. A negative-control exposure has an adjusted outcome coefficient $b_Z=0.30$ outcome units per negative-control unit. Investigators judge, for illustration, that the treatment’s confounder association per unit is 1.5 times the negative control’s, so $\lambda=1.5$. Assume the linear model and the required no-causal-effect, zero-covariance, and residual-error conditions are reasonable.

The calibrated bias is:

\text{bias}= \lambda b_Z =1.5(0.30) =0.45\text{ outcome units per treatment unit}.

Subtracting this assumed bias gives:

\tau_{\text{calibrated}} =b_A-\lambda b_Z =2.40-0.45 =1.95\text{ outcome units per treatment unit}.

Under this model and comparability judgment, the estimated treatment association would be reduced by 0.45 outcome units per treatment unit, from 2.40 to 1.95. The calculation does not prove that the causal effect is 1.95; it shows what follows from the chosen calibration assumptions.

Worked Example: Checking Comparability From Covariances

Consider hypothetical residualized variables with $\operatorname{Var}(A_r)=4$, $\operatorname{Cov}(A_r,U_r)=1.2$, $\operatorname{Var}(Z_r)=1$, and $\operatorname{Cov}(Z_r,U_r)=0.20$. Suppose $\beta=3$ outcome units per unit of $U$, and assume $\operatorname{Cov}(A_r,\varepsilon_r)=\operatorname{Cov}(Z_r,\varepsilon_r)=\operatorname{Cov}(Z_r,A_r)=0$. These covariance and model assumptions are stipulated for the example.

The confounder associations per exposure unit are:

\kappa_A=\frac{1.2}{4}=0.30, \qquad \kappa_Z=\frac{0.20}{1}=0.20.

Therefore $\lambda=\kappa_A/\kappa_Z=0.30/0.20=1.5$. The bias in the treatment coefficient is $\beta\kappa_A=3(0.30)=0.90$ outcome units per treatment unit. The negative-control coefficient is $b_Z=\beta\kappa_Z=3(0.20)=0.60$ outcome units per negative-control unit. Calibration agrees:

\lambda b_Z=1.5(0.60)=0.90\text{ outcome units per treatment unit}.

The matching values are not an empirical validation of $\lambda$; they follow from the covariances and common-$\beta$ model supplied in this example. In an application, the relative confounder associations are usually not directly observed.

Worked Example: A Range of Plausible Biases

Suppose, hypothetically, that $b_A=1.80$ outcome units per treatment unit and $b_Z=0.24$ outcome units per negative-control unit. Analysts consider $\lambda$ values from 0.5 to 2.0 because they cannot defend a single relative confounder association. Assume the simple calibration conditions hold and that these values are treated as sensitivity assumptions.

At the low end, the bias is $0.5(0.24)=0.12$ outcome units per treatment unit, giving $1.80-0.12=1.68$. At the high end, the bias is $2.0(0.24)=0.48$, giving $1.80-0.48=1.32$. Thus the calibrated point estimates range from 1.32 to 1.68 outcome units per treatment unit over this $\lambda$ range.

This range reflects uncertainty about confounding comparability, not a confidence interval. Sampling uncertainty in $b_A$ and $b_Z$ would require additional uncertainty calculations; a range of assumed $\lambda$ values does not include that uncertainty automatically.

What Makes the Calibration Credible?

The key scientific judgment is whether the negative control and treatment have sufficiently comparable relationships with the hidden confounding that matters for the outcome. The formula uses $\lambda$ to make that judgment visible. Setting $\lambda=1$ asserts equality of the confounder association per exposure unit; it is not a default truth. Different units or distributions for $A$ and $Z$ can make equal raw-unit associations implausible. Standardizing exposures can help describe comparisons, but does not establish that the underlying confounding relationships are equal.

The simple derivation also assumes a shared outcome relationship $\beta$ with the relevant unmeasured confounder, no additional $Z$-outcome pathway, and no residual error association with either exposure. If several unmeasured causes exist, the negative-control association may reflect a different mixture of causes than the treatment bias. A single $\lambda$ can then be a rough summary rather than a literal ratio. The calibration should be presented as a model-based sensitivity analysis, with a range of values when the relative confounding strength is uncertain.

A nonzero negative-control association can be informative, but a near-zero estimate does not establish the absence of treatment confounding. The negative control may be weakly associated with the relevant hidden cause, or its association may be estimated imprecisely. Conversely, a nonzero association can arise from violations of the model, including an unrecognized causal pathway or residual association not represented by $U$.

Worked Example: Treatment and Negative Control Are Associated

Suppose the stipulated hypothetical model has $\tau=1$, $\beta=1$, $\operatorname{Var}(Z_r)=1$, $\operatorname{Cov}(Z_r,A_r)=0.50$, $\operatorname{Cov}(Z_r,U_r)=0.20$, and $\operatorname{Cov}(Z_r,\varepsilon_r)=0$. The negative control has no causal effect on $Y$, but it is associated with the treatment. Its adjusted outcome slope is:

b_Z =1\left(\frac{0.50}{1}\right) +1\left(\frac{0.20}{1}\right) +0 =0.70.

Only $0.20$ of this slope is the shared-$U$ component; $0.50$ arises because $Z_r$ is associated with the causally active treatment $A_r$. Treating all $0.70$ as a confounding signal and multiplying it by $\lambda$ would therefore miscalibrate the treatment bias. The simple formula requires $\operatorname{Cov}(Z_r,A_r)=0$, or a different justified method to separate the treatment-associated component. Merely knowing that $Z$ has no causal effect is insufficient.

Common Mistakes and Practical Reporting

Wrong: “The negative control has no causal effect, so its outcome coefficient equals the confounding bias.” Correction: The coefficient can include treatment-associated and residual-error components. Derive its terms and state conditions that remove them.
Wrong: “The negative-control coefficient is the bias.” Correction: It is measured per negative-control unit. The bias is $\lambda b_Z$, where $\lambda$ compares confounder associations on the relevant scales.
Wrong: “A calibration with $\lambda=1$ is objective because both variables are exposures.” Correction: Equal confounder association per unit is a substantive assumption; justify it or show results over a range.
Wrong: “The calibrated estimate is the corrected causal effect.” Correction: It is the estimate implied by the model and chosen calibration, and it remains vulnerable to violations of those assumptions.

A clear report gives the adjusted treatment coefficient and negative-control coefficient with units; states the no-causal-effect and residual-association assumptions; explains why the chosen $\lambda$ values are plausible; and reports the resulting calibrated bias or range. The negative control then contributes a quantitative benchmark without being mistaken for proof that confounding has been removed.

Check Your Understanding

Use the model and assumptions in this tutorial to answer the following.

In the linear model, which covariance term represents the omitted-confounding bias in the adjusted treatment coefficient?
Why does the condition that $Z$ has no causal effect on $Y$ fail, by itself, to identify its outcome association with the shared confounder?
If $b_Z=0.20$ outcome units per negative-control unit and $\lambda=1.25$, what is the calibrated bias in outcome units per treatment unit?
What does setting $\lambda=1$ assert, and why is it not guaranteed by the data?
Why is a range of calibrated point estimates over several $\lambda$ values not automatically a confidence interval?