Use Negative Controls to Probe Residual Bias

Why Use a Negative Control?

An adjusted regression can account for measured covariates, but an association that remains after adjustment is not automatically causal. Some common causes may be unmeasured, measured poorly, or represented inadequately in the model. A negative control is a carefully chosen exposure or outcome whose causal relationship to the main variables is known to be absent in a specified direction. Because that causal effect should be zero, an unexpected adjusted association can act as a warning that the analysis may still contain bias.

The control is informative only if it is chosen for a reason. In particular, it should be subject to some of the same sources of bias as the main analysis. A variable that is simply unrelated to everything in the study is not necessarily a useful negative control: it may have no reason to reveal the residual confounding that could affect the exposure–outcome association of interest.

The word “negative” refers to the causal relationship that is expected to be absent, not to a negative coefficient. For a negative-control exposure, the causal effect of \(A^{-}\) on \(Y\) is absent. For a negative-control outcome, the causal effect of \(A\) on \(Y^{-}\) is absent. Either control can nevertheless have a nonzero adjusted association in observed data when confounding or another source of bias remains.

Regression Checks and Their Interpretation

Suppose the main analysis uses a regression model relating \(Y\) to \(A\) and \(C\). To examine a negative-control exposure, fit a corresponding model relating \(Y\) to \(A^{-}\) and \(C\). To examine a negative-control outcome, relate \(Y^{-}\) to \(A\) and \(C\). In each case, the coefficient of interest is an adjusted association between the control variable and the other variable, conditional on the covariates included in the model.

The first equation is the main outcome regression. The second is the negative-control exposure check; if the control is valid, its causal effect is zero, so a nonzero \(\eta_{-}\) is unexpected. The third is the negative-control outcome check; if \(A\) does not cause \(Y^{-}\), a nonzero \(\lambda_A\) is likewise unexpected. Regression coefficients describe associations in these models. They do not, by themselves, establish the causal conditions needed to interpret \(\beta_A\) as an effect.

For an ordinary coefficient test, the null hypothesis for a negative-control check is that its adjusted association is zero. A confidence interval and p-value can summarize the estimate’s uncertainty under the fitted model. Their usual interpretation depends on the model and standard-error assumptions being reasonable, including an appropriate mean structure and standard errors for the data’s dependence pattern. With a sufficiently large sample, a coefficient divided by its standard error can often be compared with the standard normal distribution. A small p-value is evidence against a zero adjusted association under that model; it is not the probability that the control is valid or that the null hypothesis is true.

Why Residual Confounding Can Show Up

The basic logic can be seen with a short derivation. Suppose \(U\) is an unmeasured variable that affects \(Y\), and that, after adjustment for \(C\), it is still associated with a negative-control exposure \(A^{-}\). Consider the linear relation \(Y=\eta_0+\eta_{-}A^{-}+\boldsymbol{\eta}_C^{T}C+\varepsilon\). For this explanation, assume the true causal coefficient of \(A^{-}\) is zero, \(U\) has an additive effect \(\gamma\) on \(Y\), and the remaining error has mean zero and is uncorrelated with \(A^{-}\) after conditioning on \(C\).

Remove the linear contributions of \(C\) from \(A^{-}\), \(U\), and \(Y\), and call the resulting residuals \(A^{-}_r\), \(U_r\), and \(Y_r\). The residualized outcome relation is \(Y_r=\gamma U_r+\varepsilon_r\). The coefficient from regressing \(Y\) on \(A^{-}\) and \(C\) is the slope of \(Y_r\) on \(A^{-}_r\). Therefore,

The final equality uses the assumption \(\operatorname{Cov}(A^{-}_r,\varepsilon_r)=0\). If \(U\) affects \(Y\) (\(\gamma\ne 0\)) and remains associated with \(A^{-}\) after adjustment (\(\operatorname{Cov}(A^{-}_r,U_r)\ne 0\)), then \(\eta_{-}\) can be nonzero even though \(A^{-}\) has no causal effect on \(Y\). This explains how a negative-control association can flag residual confounding. It does not show that the bias in the main \(A\)-\(Y\) analysis has the same size or direction: that would require additional assumptions about how the control and main variables share confounding.

Choosing and Combining Controls

A negative control is most useful when its causal status and its relationship to possible bias can be defended before examining its result. A sensible choice requires two separate arguments:

• Why the causal effect should be absent: explain why the proposed negative-control exposure cannot affect the outcome, or why the exposure of interest cannot affect the proposed negative-control outcome.
• Why it can reveal bias: explain which causes, measurement processes, or features of treatment assignment it shares with the main analysis.

Then use a regression adjustment that is justified for that particular control, and report the adjusted estimate, its uncertainty, and the assumptions behind the inference. The same covariates may be a reasonable starting point, but a control model still needs a defensible specification for its own outcome and exposure. A control should not be selected merely because it gives a convenient or statistically significant result.

Several controls can provide complementary checks. If multiple plausible controls show unexpected associations in a similar direction, that pattern may strengthen concern that the analysis has residual bias. If one control shows an association and another does not, the discrepancy calls for scrutiny of how well each control meets its causal and bias-sharing criteria. A null result for one control is not a clean bill of health: the control may be weakly related to the relevant unmeasured causes, or its estimate may be imprecise.

Common Mistakes and a Practical Reading

• “The negative control is significant, so it proves confounding.” A nonzero association is a warning under the control’s assumptions. Model misspecification, chance, or a flawed control definition may also explain it.
• “The negative control is not significant, so there is no residual bias.” A nonsignificant result may reflect low precision or a control that does not share the relevant bias. Inspect the estimate and interval, not only the p-value.
• “The control coefficient estimates the bias in the main effect.” It generally does not. A control association reveals a possible problem, but translating it into a correction requires additional assumptions not established by the negative-control check alone.
• “Any variable unrelated to the exposure is a negative control.” A valid negative control needs a specified absent causal relationship and a plausible connection to sources of bias in the main analysis.

The practical conclusion should therefore be calibrated: state which causal relationship makes the variable a proposed control, what adjusted association was estimated, and why the result raises or does not raise concern. Negative controls are diagnostic tools. They complement adjustment and sensitivity analyses; they do not replace the identification assumptions needed for causal interpretation.