Why Use a Negative-Control Exposure?
Instrumental-variable analysis requires more than an association between an instrument and treatment. In particular, independence requires the instrument to be unrelated, conditional on measured covariates, to the potential outcomes and potential treatment values that would arise under different instrument assignments. That condition cannot usually be verified from observed data: for each person, only one instrument value and one outcome are observed.
A negative-control exposure offers a way to look for warning signs. It is a placebo exposure that is not itself a cause of the outcome, but is chosen because it may reflect some of the same assignment processes or unmeasured factors as the instrument. If the placebo exposure predicts the outcome after appropriate adjustment, that association can suggest that outcome-related factors are connected to the processes under study. It does not, by itself, show that the instrument violates independence.
For a binary instrument and treatment, the usual conditional independence condition is
Here \(Y(d)\) is the outcome that would be observed under treatment level \(d\), and \(D(z)\) is the treatment that would be received under instrument value \(z\). The condition says that, within levels of \(X\), the instrument is independent of these potential outcomes and treatment values. It is a distinct requirement from instrument relevance and the exclusion restriction. A negative-control exposure is a diagnostic for possible threats to independence; it does not test all instrumental-variable assumptions.
Designing and Analyzing the Diagnostic
The negative-control exposure must be selected using substantive knowledge, not merely because it is available in a dataset. Ask whether it could affect the outcome through any route, including through treatment, healthcare-seeking, or another post-exposure variable. Also ask whether it plausibly reflects the same assignment process or unmeasured causes that could connect the proposed instrument to potential outcomes. If it reflects unrelated factors, its association—or lack of association—with \(Y\) may say little about the instrument.
A basic diagnostic compares outcomes across values of \(Z^{NC}\) after adjustment for measured pre-instrument covariates. If the actual instrument can affect \(Y\) through \(D\), and \(Z^{NC}\) is associated with \(Z\), an unadjusted association between \(Z^{NC}\) and \(Y\) could simply reflect the instrument’s genuine effect. One option is therefore to compare the negative-control groups within strata of both \(X\) and \(Z\), provided the design and analysis make that conditioning appropriate. The diagnostic should be specified with the assignment structure in mind; conditioning on variables affected by the instrument or by the negative-control exposure can introduce other biases.
For a binary outcome, a stratum-specific contrast is a risk difference. With stratum \(s\), define \(\hat p_{1s}\) as the outcome proportion for \(Z^{NC}=1\) and \(\hat p_{0s}\) as the proportion for \(Z^{NC}=0\). A standardized contrast can average these differences using prespecified stratum weights \(w_s\), where the weights sum to one:
A nonzero contrast is evidence of an adjusted association between the placebo exposure and the observed outcome in the analyzed population. Whether that association is a warning about instrument independence depends on the negative-control assumptions and the relationship between the placebo and instrument assignment processes. Report its magnitude and uncertainty, not only a significance test.
Worked Example: A Conditional Association With a Binary Outcome
Hypothetical data. Suppose a study records a binary outcome, a binary negative-control exposure, and a binary instrument. Investigators compare negative-control groups within four strata defined by covariate \(X\) and instrument \(Z\). There are 50 people in each negative-control group in every stratum. The outcome counts are:
| Stratum | Outcome events, \(Z^{NC}=0\) | Outcome events, \(Z^{NC}=1\) |
|---|---|---|
| 1 | 10 of 50 | 16 of 50 |
| 2 | 15 of 50 | 21 of 50 |
| 3 | 20 of 50 | 26 of 50 |
| 4 | 25 of 50 | 31 of 50 |
The four stratum-specific risk differences are \(16/50-10/50=0.12\), \(21/50-15/50=0.12\), \(26/50-20/50=0.12\), and \(31/50-25/50=0.12\). With equal weights of \(0.25\), the standardized estimate is
Thus, within these strata, the outcome risk is estimated to be 12 percentage points higher for \(Z^{NC}=1\). For an approximate standard error, use the independent-binomial variance within each stratum and the variance of a weighted sum:
For example, the first stratum contributes \(0.25^2[(0.32)(0.68)/50+(0.20)(0.80)/50]\approx0.000472\) to the estimated variance; summing the four contributions gives approximately \(0.002258\), whose square root is \(0.04752\). The approximate test statistic is \(0.12/0.04752=2.525\), corresponding to a two-sided normal-approximation \(p\)-value of about \(0.0116\). Under the assumed sampling model, the data provide evidence of a conditional association. They do not prove that the proposed instrument is dependent on potential outcomes: that conclusion additionally requires the negative control to reflect relevant sources of instrument confounding and to have no causal pathway to the outcome.
Why the Signal Is Suggestive, Not Conclusive
The diagnostic’s logic can be made explicit. Suppose an unmeasured factor \(U\) affects the outcome and also influences the negative-control exposure. Even if \(Z^{NC}\) has no causal effect on \(Y\), the observed outcome can differ across negative-control groups because those groups have different distributions of \(U\). In a simplified setting with no other relevant causes, this can be represented as
If outcome risk varies with \(U\), and the distribution of \(U\) varies with \(Z^{NC}\), the two sums can differ even though changing \(Z^{NC}\) itself would not change anyone’s outcome. The diagnostic therefore detects a pattern compatible with shared confounding; it does not identify the cause of that pattern. To implicate instrument independence, there must also be a credible reason to believe that the factors reflected by \(Z^{NC}\) are connected to the assignment of \(Z\).
The interpretation is asymmetric. A clear association can raise concern, but it may have explanations unrelated to instrument independence, such as a missed causal pathway, selection into the analytic sample, measurement artifacts, or chance. A null result is not reassurance that independence holds: the placebo may be a poor proxy for the relevant factors, the association may be too small to detect, or opposing sources of bias may cancel. Independence concerns potential outcomes, which are not all observed, so no single observed-data test can verify the condition.
Worked Example: A Null Result Does Not Establish Validity
Hypothetical data. In a simpler study, 80 people with \(Z^{NC}=0\) have 40 outcome events, and 80 people with \(Z^{NC}=1\) also have 40 events. The estimated risks are both \(40/80=0.50\), so the estimated risk difference is \(0.50-0.50=0\).
Using the usual independent-binomial approximation, the standard error is
An approximate 95% confidence interval for the risk difference is \(0\pm1.96(0.07906)\), or approximately \((-0.155,\,0.155)\). These data do not show an association, but they are compatible with risk differences as large as about 15.5 percentage points in either direction under this approximation. More importantly, a null estimate would not rule out a compromised instrument if the negative-control exposure fails to capture the source of dependence that matters for \(Z\). The result is limited evidence about this particular diagnostic contrast, not proof of instrument validity.
Assumptions and Practical Checks
Interpret a negative-control diagnostic only after checking the assumptions that connect it to the question about \(Z\). The checks below do not turn an untestable independence assumption into a testable one, but they make the diagnostic more informative and less vulnerable to avoidable misinterpretation.
State why the negative-control exposure cannot affect \(Y\), directly or through \(D\). If that claim is not credible, an association with \(Y\) may be causal rather than diagnostic.
Describe which assignment processes or unmeasured factors the negative-control exposure is intended to reflect, and why those factors could also threaten the independence of \(Z\).
Adjust for relevant measured pre-instrument covariates. Consider the role of the observed \(Z\) if it can create an outcome pathway, but avoid conditioning mechanically on variables affected by \(Z\) or \(Z^{NC}\).
Give the adjusted contrast, its units, and an interval or standard error. A \(p\)-value alone does not describe the magnitude or precision of the diagnostic association.
Describe a positive result as a warning signal under stated assumptions, and a null result as a lack of detected association. Neither result proves or disproves instrument independence.
Worked Example: A Placebo That Tracks the Actual Instrument
Hypothetical data. Suppose \(Z\) is a randomized encouragement, treatment uptake is 20% when \(Z=0\) and 60% when \(Z=1\), and \(Z^{NC}\) is recorded as a marker that exactly equals \(Z\). Assume, for illustration, outcome risk is 10% without treatment and 50% with treatment, and the encouragement affects the outcome only by changing treatment. The outcome risks by encouragement group are then
Because \(Z^{NC}=Z\), the unadjusted risk difference across the negative-control groups is \(0.34-0.18=0.16\), or 16 percentage points. This association arises from the real encouragement-to-treatment-to-outcome pathway, not from confounding of randomized \(Z\). In this construction, conditioning on \(Z\) leaves no variation in \(Z^{NC}\), so the within-\(Z\) diagnostic cannot be estimated. The example shows why a placebo marker must be chosen and analyzed with the actual instrument’s causal pathways in mind; a raw association is not automatically evidence against independence.
Common Mistakes and Why the Diagnostic Matters
- Calling a positive association proof of invalidity. The association may reflect a causal pathway, selection, measurement problems, or chance. First assess whether the negative-control assumptions and shared-source rationale are credible.
- Calling a null association proof of validity. A weak or poorly chosen placebo may miss the relevant dependence, and finite samples may be imprecise. Report the uncertainty and the limits of the proxy.
- Ignoring the actual instrument’s causal effect. If the placebo is associated with \(Z\), the path from \(Z\) through treatment to outcome can create an association. Consider the assignment design and the appropriate role of \(Z\) in the diagnostic analysis.
- Adjusting for post-assignment variables without justification. Conditioning on treatment or another variable affected by assignment can change the comparison and may introduce bias. Choose the adjustment set from the causal structure, not from a desire to make groups look similar.
- Confusing this diagnostic with the full independence condition. An observed association involving \(Z^{NC}\) is not a direct test of \(Z \perp \{Y(0),Y(1),D(0),D(1)\}\mid X\). It is one empirical probe whose meaning depends on design assumptions.
Check Your Understanding
Answer each question by referring to the assumptions and limits of the diagnostic.
- State the conditional independence condition for a binary instrument, treatment, and outcome using potential-outcome notation.
- Why must a negative-control exposure be unable to affect the outcome through treatment as well as directly?
- In a study where the actual instrument can affect the outcome, why might an unadjusted association between a placebo exposure and the outcome be misleading?
- A negative-control exposure has no detectable association with the outcome, but it is a weak proxy for the instrument’s assignment process. What can and cannot be concluded?
- Give two explanations, other than instrument confounding, for an observed association between a negative-control exposure and the outcome.