Why Candidate Selection Requires Two Questions
A negative control exposure is useful only when it meets two distinct requirements. First, changing it must not plausibly change the study outcome. Second, it should share relevant sources of confounding with the treatment or exposure of interest. The first condition protects against mistaking a genuine causal effect for evidence of bias. The second gives the control a reason to reveal residual bias. An exposure satisfying only one condition is not a strong candidate.
Suppose researchers estimate the effect of a treatment \(A\) on an outcome \(Y\) using observational data. A candidate negative control exposure, written \(A^{NC}\), is an exposure selected to help probe whether the observed treatment–outcome association may reflect residual confounding. Its suitability is a causal and subject-matter judgment, not a conclusion reached simply by finding an exposure that is convenient to measure or appears unrelated to the outcome in one data set.
The causal null for the candidate can be expressed using potential outcomes. If \(Y(a,n)\) denotes the outcome that would occur under treatment level \(a\) and candidate exposure level \(n\), then a necessary no-effect condition is that changing \(n\) leaves the outcome unchanged at each fixed \(a\):
This condition is not established by a nonsignificant association in the observed data. It is a statement about what would happen under intervention, supported by the causal structure and subject-matter knowledge. Also check that the candidate does not cause the actual treatment: if \(A^{NC}\) changes \(A\), and \(A\) affects \(Y\), then the candidate can affect \(Y\) indirectly.
A Practical Selection Framework
Start with the target treatment, outcome, population, and time window. Then state, in words, why the candidate could not affect the outcome and which unmeasured factors could make it related to both treatment assignment and outcome. The rationale should be specific enough that another researcher could challenge it. “This exposure is unrelated to the outcome” is not an adequate justification; the relevant question is whether it could causally influence the outcome in the study’s setting and follow-up period.
Specify when \(A\) is assigned, when \(A^{NC}\) occurs, and when \(Y\) is measured. Prefer a candidate measured before the treatment and outcome period, and check that it is not a consequence of either.
Consider biological, behavioral, clinical, and health-care pathways, including effects mediated by treatment, co-interventions, or changes in outcome measurement. Reject a candidate when an important route is plausible and cannot be excluded.
For example, access to care, health-seeking behavior, disease severity, or a risk factor might influence treatment assignment and the candidate exposure, while also relating to the outcome. A candidate with no reason to share such causes may be uninformative.
Ask whether the candidate is recorded comparably across treatment groups, whether it is common enough to study, and whether its meaning is stable in the target population. Differential recording or a highly selected candidate group can undermine its usefulness.
State the assumptions that support the choice, the main plausible violations, and why a close alternative was rejected. Treat the result as a probe of residual bias, not as a direct measurement of the treatment effect’s bias.
A causal diagram can help organize these judgments. Let \(U\) denote an unmeasured factor. A candidate is more promising when there is a defensible reason to believe that factors such as \(U\) contribute to both treatment assignment and candidate exposure, and can also relate to the outcome. The exact shared pathways require careful causal reasoning; a superficial resemblance between two exposures is not enough.
Worked Examples
Worked Example: A Candidate for an Asthma Treatment Study
Suppose a hypothetical observational study estimates the effect of starting a biologic medication \(A\) on asthma exacerbations \(Y\) over the next six months. Researchers consider a dermatologist visit in the six months before treatment assignment as \(A^{NC}\). They propose that atopic predisposition and access to specialist care may relate to both a dermatology visit and biologic treatment; some of these factors may also relate to later exacerbations.
The researchers review records for 500 patients. Of 200 with a prior dermatology visit, 50 later start the biologic; of 300 without a visit, 45 start it. The observed treatment-start proportions are:
The descriptive difference is \(0.250-0.150=0.100\), or 10.0 percentage points. This difference does not prove the candidate shares confounding with treatment: age, disease severity, and health-care access could explain it, and the figures are not an inferential test. It does show that the candidate is not simply absent from the treatment-assignment context.
The no-effect screen is plausible but not automatic. A routine dermatology visit itself is unlikely to change asthma exacerbations, but systemic medications prescribed at the visit could affect them. The researchers therefore define the candidate narrowly as a visit without systemic treatment relevant to asthma and check its timing. Under that restriction, the visit is a defensible candidate if the assumed shared causes are credible. The choice should still be reported as provisional, since unmeasured factors may not relate to the candidate and treatment in comparable ways.
Worked Example: Rejecting a Plausible-Looking Candidate
In a hypothetical study of statin initiation \(A\) and heart attack within one year \(Y\), researchers consider physical therapy after a knee injury as a negative control exposure. The proposed shared causes include health-care access and general health. However, physical therapy can change mobility and activity, which may affect cardiovascular risk. It may also be associated with pain treatment or other care that changes risk.
In a sample of 400 patients, 80 of the 160 who received physical therapy start a statin, while 72 of the 240 who did not receive therapy start one:
The difference is \(0.500-0.300=0.200\), or 20.0 percentage points. This association may be consistent with shared health-care access or health status, but it cannot rescue the candidate from the causal-effect concern. Because physical therapy may influence the outcome through mobility or related care, the no-effect requirement is not convincingly satisfied. The candidate should be rejected or replaced, even if it appears to track treatment assignment well.
Worked Example: Comparing Preventive-Care Candidates
Consider a hypothetical study of influenza vaccination \(A\) and laboratory-confirmed influenza hospitalization \(Y\) during the following winter. Researchers compare two candidates measured before vaccination: a routine eye examination and a prescription for an antiviral medication. They hypothesize that access to care and preventive-care engagement may influence both vaccination and having an eye examination, and may also relate to the chance of hospitalization.
Among 600 patients, 210 of 300 with a routine eye examination were vaccinated, compared with 150 of 300 without one:
The vaccination difference is \(0.700-0.500=0.200\), or 20.0 percentage points. The eye examination is a plausible candidate if it occurred before the vaccination period and has no plausible effect on influenza hospitalization during follow-up. Its association with preventive-care access gives a reason to investigate whether it shares relevant causes with vaccination, but does not establish that it does.
An antiviral prescription is a poor candidate: it may directly reduce influenza severity or hospitalization, violating the no-effect condition. On the information given, the eye examination is the stronger candidate. The conclusion depends on the clinical setting and timing, not on which exposure has the larger observed association with vaccination.
What the Candidate Can—and Cannot—Tell You
A negative control exposure is not a substitute for adjusting for measured confounders. Adjustment uses measured variables to address confounding under the relevant assumptions; a negative control exposure is selected as a probe for residual problems that may remain. Nor is it an instrumental variable: an instrumental variable is used to identify a causal effect under its own assumptions, whereas a negative control exposure is chosen not to cause the outcome.
A negative control outcome instead uses an outcome that the treatment cannot plausibly cause. The two approaches can complement one another, but they require different causal justifications. A candidate exposure should not be called a valid control merely because it is medically unrelated to the outcome: it must also be relevant to the confounding concern.
Selection does not establish that all important unmeasured confounding is shared with the candidate. A candidate can fail to reveal bias because it shares only some causes, because its measurement is weak, or because the sample provides little information. Conversely, an association involving the candidate could arise from a mistaken no-effect assumption, a shared cause, selection, or other design features. The candidate’s role is diagnostic, not a guarantee that the main estimate is unbiased or biased by a known amount.
Common Mistakes and Diagnostics
- Choosing an exposure solely because it is unrelated by definition. A candidate that cannot plausibly affect \(Y\) may still be unhelpful if it shares no relevant sources of confounding with \(A\). Explain both parts of the rationale.
- Treating a small observed association as proof of no causal effect. Observational data cannot establish the intervention-based no-effect condition. Use timing, clinical knowledge, and plausible causal pathways to justify it.
- Ignoring indirect effects. Check whether the candidate could affect the treatment, a co-intervention, follow-up, or outcome measurement. A route through another variable still counts as a causal route if it can change \(Y\).
- Assuming similar exposures have similar confounding. Two health-care encounters may reflect different patient groups, access barriers, or clinical processes. Name the hypothesized shared causes rather than relying on surface similarity.
- Selecting after looking for a favorable result. Choosing a control because its observed association supports a preferred conclusion risks selective interpretation. Record the causal rationale before interpreting its empirical association.
Useful diagnostics at the selection stage are therefore substantive: check dates, eligibility, coding definitions, treatment pathways, co-interventions, and the likely causes of both exposures. If the no-effect claim depends on excluding a common clinical pathway, or if the proposed shared causes are vague, report the candidate as weak or choose another. A candidate’s observed data patterns can inform later assessment, but they do not repair an implausible causal rationale.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- In a study of a medication’s effect on hospitalization, why is an exposure that strongly predicts medication use not automatically a suitable negative control exposure?
- A candidate occurs after treatment initiation. What timing concern should be raised?
- In the asthma example, calculate the treatment-start proportion difference between patients with and without a dermatology visit. What does that calculation establish—and what does it not establish?
- Why is physical therapy after knee injury a weak candidate in the statin and heart-attack example, even though it may share health-care access as a cause with statin use?
- How does a negative control exposure differ in purpose from an instrumental variable?
Solutions. (1) Prediction alone does not show that the candidate shares relevant confounding, and the candidate may itself cause the outcome. Both conditions require a causal rationale. (2) It may be a consequence of treatment, or treatment may affect its measurement; either possibility undermines its role as a pre-treatment control. (3) The difference is \(50/200-45/300=0.250-0.150=0.100\), or 10.0 percentage points. It describes the sample’s treatment-start proportions; it does not prove shared confounding or validate the no-effect condition. (4) Physical therapy may affect mobility, activity, or related care and thereby plausibly affect heart-attack risk. A shared cause cannot compensate for a plausible causal effect. (5) A negative control exposure is selected not to cause the outcome and is used to probe residual bias; an instrumental variable is used to identify a causal effect under separate assumptions.