Future Outcomes as Negative Controls

Why Examine an Outcome Measured Before Treatment?

A treatment cannot ordinarily change an outcome that was already determined and measured before treatment began. Yet the treatment group and comparison group may differ on that earlier outcome. Such a difference cannot be a causal effect of the later treatment; it can instead flag differences between the groups that may also threaten a causal analysis of the primary outcome. This makes a pre-treatment outcome a possible negative control outcome.

The test is a diagnostic, not a method for estimating the treatment effect or a direct measure of its bias. A nonzero association can reveal a problem with comparability, but a null association does not establish that confounding is absent. Whether the test is informative depends on the outcome’s timing, how it was measured, and whether it shares relevant causes with the primary outcome.

Chronology and the Causal Null

The key temporal claim is stronger than saying that a database date precedes a treatment date. The outcome itself must have been determined before treatment could affect it. Its measurement must also represent that earlier state rather than a later event, later diagnosis, or post-treatment reinterpretation of an earlier record.

In potential-outcome notation, write \(Y_{\mathrm{pre}}(a)\) for the pre-treatment outcome that would be observed under treatment value \(a\). If the outcome is genuinely fixed before treatment and there is no anticipation or other path by which treatment can affect it, then for each individual:

This individual-level causal null does not imply that observed outcomes are distributed identically in the two treatment groups. Treatment may be more common among people who already differ in health, access to care, or other characteristics. For example, a prior symptom could influence treatment initiation and also predict the primary outcome. The treatment and control groups can then differ in the symptom even though treatment did not cause it.

For a negative-control test to be diagnostically useful, two distinct conditions matter. First, the outcome must be unaffected by treatment. Second, it must be relevant to the bias question: it should share causes of treatment and the primary outcome, or otherwise be informative about a source of bias that could affect the primary analysis. A pre-treatment outcome that has no meaningful connection to treatment selection or the primary outcome can pass the test while the primary analysis remains confounded.

Specify the Association Before Testing

The test should use an association scale that fits the outcome and matches the analysis question. For a continuous outcome, one simple marginal contrast is the difference in group means, \(\bar{Y}_{\mathrm{pre},1}-\bar{Y}_{\mathrm{pre},0}\). For a binary outcome, a risk difference is \(\hat{p}_1-\hat{p}_0\), where \(\hat{p}_a\) is the observed proportion with the outcome in treatment group \(a\). A conditional or adjusted contrast can be used when appropriate, but its covariates and target contrast should be specified in advance.

If the primary analysis uses matching, weighting, or covariate adjustment, the negative-control analysis should generally use a comparable design and clearly state its target population. Covariates used for adjustment should be measured before treatment. A test based on an unadjusted comparison answers a different question from one based on a matched or adjusted comparison; neither should be described vaguely as “the” treatment association.

The null hypothesis is typically a zero association on the chosen scale, such as \(H_0:\mu_1-\mu_0=0\) for a continuous outcome or \(H_0:p_1-p_0=0\) for a binary outcome. A confidence interval communicates the range of association sizes reasonably compatible with the data under the model. A p-value summarizes how surprising the test statistic would be under the specified null and test assumptions; it is not the probability that the null is true.

Worked Example: A Binary Pre-Treatment Outcome

When the Dates Do Not Validate the Outcome

Chronology must be established for the underlying event, not inferred from the order in which a database displays records. A diagnosis entered after treatment may describe disease that existed beforehand; conversely, a diagnosis recorded before treatment may reflect a process already influenced by anticipation of treatment or by a treatment decision made earlier than the recorded start date. If treatment is ongoing rather than newly initiated, a nominally “pre-index” measure may actually follow earlier exposure.

Measurement also matters. If treated people have more clinical contact before treatment, their earlier conditions may be detected and recorded more often than those of untreated people. The observed association can then reflect differential ascertainment rather than a difference in the underlying condition. Such a result still flags a comparability or measurement concern, but it is not clean evidence about a shared biological confounder.

Deriving the Diagnostic Logic

The logic can be stated in two steps. First, if treatment occurs after the outcome is fixed, the causal contrast for that outcome is zero:

Second, the observed difference compares people selected into different treatment groups, rather than assigning the same people both treatment values. In general:

The left side can be nonzero because the groups differed before treatment. This explains both the value and the limit of the test: a nonzero association is inconsistent with treating the observed groups as alike on that outcome, while a zero association only says that this particular contrast did not detect a difference. It does not show that all confounding pathways are absent.

Diagnostics, Alternatives, and Limitations

Before interpreting a result, check the treatment start date, outcome event date, and measurement window; verify that no prior treatment episode or anticipatory pathway invalidates the temporal argument; and compare outcome definitions and ascertainment across groups. Inspect group sizes and outcome variability, and report an interval as well as a p-value. A nonsignificant result with a wide interval may be uninformative. If the analysis adjusts for covariates or uses a matched sample, confirm that the same design is appropriate for the control outcome and that adjustment has not introduced variables measured after treatment.

A pre-treatment outcome test complements, rather than replaces, other approaches. The earlier material on future exposures as negative controls concerns a different falsification strategy: it asks whether a later exposure is associated with an earlier outcome. Placebo treatment tests instead assign a fictitious treatment time or exposure definition. Covariate adjustment, matching, and weighting aim to improve comparability under their own assumptions; a negative-control result can help diagnose remaining problems but does not guarantee those assumptions.

The most important limitations are that the control outcome may not share the relevant confounding structure, that multiple sources of bias can cancel on the chosen scale, and that measurement error or limited sample size can hide an association. A nonnull result does not identify a unique bias mechanism, while a null result is not proof of no bias. Unlike a calibrated bias analysis, this diagnostic alone does not provide a numerical correction to the primary effect estimate.