Why Use a Negative Control?
An adjusted comparison between a trial population and an external-control cohort can still be biased. Some causes of the outcome may be unmeasured, patients may enter the data sources through different selection processes, or events may be recorded differently. A negative control is an outcome or exposure chosen so that a causal effect is not expected, while relevant sources of bias may still create an observed association. It can therefore act as a diagnostic probe: a discrepancy raises a concern about the comparison, but does not identify its cause by itself.
For example, if trial treatment appears associated with an outcome that the treatment cannot plausibly affect, that association is not evidence that the treatment caused the outcome. It may instead signal residual confounding, selection, or differential measurement. Conversely, a near-null result does not establish that the primary treatment comparison is unbiased. The negative control may simply fail to share the bias, or the data may be too imprecise to reveal it.
These definitions state causal assumptions, not just associations observed in a data table. A candidate outcome is not a valid negative control merely because investigators expect its treatment association to be small. They need a substantive reason the treatment cannot affect it, including through plausible indirect pathways. Likewise, a candidate exposure must not affect the outcome through a direct or indirect causal route. In practice, the assumptions are argued from clinical knowledge and study design; they cannot generally be verified from the observed association alone.
What a Negative-Control Signal Can Diagnose
A useful negative control is chosen because it is susceptible to some of the same distortions as the primary comparison. If unmeasured patient characteristics influence both treatment receipt and a negative-control outcome, residual confounding may produce an association. If trial and external-cohort inclusion depend on patient features related to that outcome, selection may do so as well. Differences in healthcare contact, coding, laboratory testing, or record completeness can create apparent outcome differences even when the underlying condition is unaffected by treatment.
The strength of the probe depends on this shared structure. An unrelated outcome recorded through a different system may have little ability to detect the data-capture artifact of concern. A negative-control exposure that shares no causes with treatment may not expose the relevant confounding. The comparison should therefore be designed so that the negative control is subject to the same eligibility rules, adjustment strategy, data source, follow-up, and outcome ascertainment as the main analysis whenever those features are relevant.
For an external-control analysis, apply the same adjustment or weighting strategy to the negative control that is used for the primary contrast, and define the target population consistently. The earlier material on propensity scores, weighting, covariate balance, and standardization provides the relevant design tools. A negative-control analysis does not replace those methods: it checks whether the resulting comparison still shows an association where the causal effect should be absent.
Bayesian Analysis of a Binary Negative Control
A simple model illustrates how to quantify uncertainty in a diagnostic contrast. Suppose the negative-control event is binary and, after defining the groups and analysis population, the event counts are modeled as independent binomials. Let \(p_T\) and \(p_C\) denote the event risks in the treated and external-control groups. For this illustration, use independent uniform priors, \(p_T\sim\operatorname{Beta}(1,1)\) and \(p_C\sim\operatorname{Beta}(1,1)\). These are modeling choices, not a substitute for checking whether the groups and measurements are comparable.
If a group has \(y\) events among \(n\) people, its likelihood is proportional to \(p^y(1-p)^{n-y}\). Multiplying by a \(\operatorname{Beta}(a,b)\) prior gives a posterior density proportional to \(p^{a+y-1}(1-p)^{b+n-y-1}\). Thus the posterior is \(\operatorname{Beta}(a+y,b+n-y)\). With the stated uniform prior, the posterior mean is \((y+1)/(n+2)\), and the posterior variance is \((y+1)(n-y+1)/[(n+2)^2(n+3)]\). Under independent group priors and likelihoods, the posterior variance of the risk difference is the sum of the two posterior variances.
This model estimates an observed group contrast for the negative-control outcome. It does not turn that contrast into a treatment effect, nor does it adjust automatically for confounding. In a real analysis, covariate adjustment, weighting, or standardization may be needed; the negative-control calculation should then reflect that design rather than silently treating an unadjusted contrast as adequate.
Worked Example: A Negative-Control Outcome Differs Between Groups
Hypothetical data. Suppose a binary outcome known, on clinical grounds, not to be affected by the study treatment occurs in 12 of 100 treated participants and 5 of 100 external controls. For a simple illustration, assume independent binomial sampling within groups and independent \(\operatorname{Beta}(1,1)\) priors on the two risks. This calculation does not adjust for covariates; its interpretation therefore depends on whether that simple comparison is appropriate.
The posterior risks are \(p_T\mid\text{data}\sim\operatorname{Beta}(13,89)\) and \(p_C\mid\text{data}\sim\operatorname{Beta}(6,96)\). Their posterior means are \(13/102=0.1275\) and \(6/102=0.05882\), so the posterior mean difference, treated minus external control, is \(7/102=0.06863\), or about 6.86 percentage points.
The posterior variance of the difference is \(13(89)/(102^2\cdot103)+6(96)/(102^2\cdot103)=1733/1{,}071{,}612\approx0.001617\). The posterior standard deviation is approximately \(0.04021\). A normal approximation to the posterior gives a 95% credible interval of \(0.06863\pm1.96(0.04021)\), or approximately \(-0.0102\) to \(0.1474\). Under this approximation, the posterior probability that the difference is positive is \(\Phi(0.06863/0.04021)\approx0.956\).
The data show a potentially concerning positive contrast in an outcome assumed to be unaffected by treatment, but the interval is wide and includes zero. This is a diagnostic signal to investigate comparability, selection, and ascertainment; it is not evidence that the treatment caused the negative-control outcome. The approximation summarizes posterior uncertainty under the stated model and does not account for model misspecification.
Worked Example: A Negative-Control Exposure Shows an Outcome Association
Hypothetical data. In an external-control cohort, a candidate exposure that is assumed not to cause a binary clinical outcome is present for 80 people, of whom 8 have the outcome, and absent for 80 people, of whom 4 have the outcome. Thus, there are 8 events among 80 exposed people and 4 events among 80 unexposed people. Assume independent binomial sampling and independent \(\operatorname{Beta}(1,1)\) priors on the two risks.
The posterior risk among exposed people is \(\operatorname{Beta}(9,73)\), with mean \(9/82\); among unexposed people it is \(\operatorname{Beta}(5,77)\), with mean \(5/82\). The posterior mean risk difference is \(9/82-5/82=4/82=0.04878\), or 4.88 percentage points.
The posterior variance of the difference is \([9(73)+5(77)]/(82^2\cdot83)=1042/558092\approx0.0018671\), so the posterior standard deviation is approximately \(0.04321\). A normal approximation gives a 95% credible interval of \(0.04878\pm1.96(0.04321)\), or approximately \(-0.0359\) to \(0.1335\). The approximate posterior probability of a positive difference is \(\Phi(0.04878/0.04321)\approx0.8705\).
Because the exposure is assumed not to cause the outcome, the observed contrast could reflect shared causes of exposure and outcome, selection, or measurement differences. The wide interval also shows that the data do not pin down the size or direction of the association precisely. Neither the posterior probability nor the interval is a probability that the causal assumption is true; that assumption must be justified separately.
Choosing and Interpreting Controls
A candidate negative control is most informative when its causal null is credible and it shares plausible bias pathways with the primary comparison. Selection should ideally be prespecified using clinical and data-generation knowledge, rather than chosen after inspecting which outcomes produce striking results. If several controls are examined, report the set and the reasoning for selecting them. A pattern across multiple controls can be more informative than a single result, but it does not remove the need to consider each control's validity and precision.
For a negative-control outcome, ask whether treatment could affect the outcome through any plausible pathway and whether the outcome is captured similarly in the trial and external data. For a negative-control exposure, ask whether it can affect the outcome and whether it shares relevant causes or selection mechanisms with treatment. Also check whether the negative control has enough events and whether the comparison has adequate overlap. Poor overlap or sparse data can make a diagnostic contrast uninformative even if its causal assumptions are sound.
An apparent association should prompt investigation, not automatic correction. It does not reveal which bias generated the signal, how large the bias in the primary treatment effect is, or whether the bias acts in the same direction for the primary outcome. Subtracting the negative-control contrast from the treatment effect, or treating it as a direct estimate of bias, requires additional assumptions that are not supplied by the negative-control design alone.
Diagnostics, Alternatives, and Limitations
Several patterns can indicate that a negative-control strategy is failing. A control outcome may be affected by treatment after all; a candidate exposure may have an unrecognized causal pathway; measurement definitions may differ between data sources; or the control may not share the bias mechanisms under study. Unexpectedly large differences can also arise from random variation, especially when many controls are tested. A result near zero can be falsely reassuring when the control is rare, the sample is small, or the relevant bias does not influence that control.
Negative controls complement, but do not replace, covariate adjustment and the earlier material on sensitivity analysis for residual confounding and quantitative bias analysis. Adjustment addresses measured covariates under its modeling and design assumptions. Quantitative bias analysis evaluates how specified unmeasured-confounding assumptions could alter an estimate. Negative controls instead probe whether an association appears in a setting where the causal effect is assumed absent. A concurrent randomized comparison can provide a stronger basis for estimating a treatment effect because randomization supports exchangeability on average; a negative control in an external comparison is a diagnostic, not a substitute for randomization.
A negative-control analysis is limited by the fact that the key assumptions are partly untestable. It may detect some residual problems but cannot certify the absence of all bias. Its value is greatest when interpreted alongside the clinical rationale, data provenance, covariate diagnostics, and sensitivity analyses, with uncertainty reported rather than reduced to a binary pass-or-fail label.
Worked Example: A Data-Capture Negative Control
Hypothetical data. Suppose an unrelated historical condition, believed not to be affected by the current treatment, is recorded in 18 of 120 treated participants and 7 of 120 external controls. The condition is included because its recording may be sensitive to healthcare contact and documentation practices. Assume independent binomial sampling, independent \(\operatorname{Beta}(1,1)\) priors, and comparable ascertainment definitions for this illustration.
The posterior risks are \(\operatorname{Beta}(19,103)\) and \(\operatorname{Beta}(8,114)\), with means \(19/122=0.1557\) and \(8/122=0.06557\). The posterior mean difference is \(11/122=0.09016\), or about 9.02 percentage points. Its posterior variance is \([19(103)+8(114)]/(122^2\cdot123)=2869/1830732\approx0.001567\), and its posterior standard deviation is approximately \(0.03959\). The normal-approximation 95% credible interval is \(0.09016\pm1.96(0.03959)\), or approximately \(0.0126\) to \(0.1678\).
The positive contrast is compatible with a difference in recording or healthcare contact, among other explanations. It cannot establish that data capture caused the contrast, and it does not quantify the bias in the primary treatment effect. The appropriate response is to inspect the data-collection processes and assess whether the primary outcome is vulnerable to the same differences.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A candidate negative-control outcome has a small observed difference between treated and external-control groups. Does that establish that the primary treatment estimate is unbiased? Solution: No. The control may lack power or fail to share the relevant confounding, selection, or capture mechanisms. A small contrast is not proof that bias is absent.
- Under a \(\operatorname{Beta}(1,1)\) prior, 10 events among 98 people give what posterior distribution and mean risk? Solution: The posterior is \(\operatorname{Beta}(11,89)\), since there are 88 non-events and the prior adds one to each count. Its mean is \(11/(11+89)=0.11\).
- Why is a treatment-affected outcome not a valid negative-control outcome, even if its expected effect is small? Solution: The defining assumption requires the causal effect to be absent, not merely small or believed to be negligible. A true treatment effect could produce an observed contrast that would be misread as bias.
- A negative-control exposure is associated with the outcome. Name two possible explanations other than a causal effect of that exposure. Solution: Shared causes of exposure and outcome may leave residual confounding; selection into the analyzed cohort or differential measurement may also create an association. The signal alone does not distinguish among these explanations.
- Should an analyst automatically subtract the negative-control risk difference from the primary treatment effect? Solution: No. That correction requires additional assumptions that the negative-control contrast measures the same bias, on a compatible scale and in the same direction, as the bias in the primary estimate. The negative-control result is ordinarily diagnostic rather than a direct bias estimate.