Why Look at an Exposure That Happens Later?
A useful falsification test asks whether an analysis detects an association where the proposed causal explanation says there should be none. A future exposure can provide such a check: if it truly begins after the outcome has occurred, it cannot ordinarily cause that earlier outcome. Yet it may still be related to the same unmeasured circumstances that influence both the exposure of interest and the outcome. An association between the future exposure and the earlier outcome can therefore flag residual bias.
This logic depends on more than the calendar order of two recorded dates. A planned future exposure may change behavior before it begins; an ongoing exposure may have started earlier than its recorded date; and the earlier outcome may influence later exposure decisions. In each case, an apparent “future exposure” association may have explanations other than shared unmeasured confounding.
Conditions for a Meaningful Check
Suppose covariates \(L\) are measured before the exposure of interest. Let \(H\) represent relevant exposure history before \(Y\), including \(A\) when needed. The diagnostic compares outcomes across levels of \(A^+\) among people with comparable \(L\) and \(H\). For a binary outcome, one possible adjusted association is the standardized risk difference
where \(w(l,h)\) are prespecified weights over covariate and exposure-history strata with observations at both levels of \(A^+\). This is a standardized association. It is not the causal effect of \(A^+\), because the control is selected precisely on the premise that \(A^+\) cannot cause the earlier outcome. For continuous outcomes, the same idea can be expressed as a conditional mean difference or a regression coefficient, with the model and scale stated.
The check is informative only under several substantive conditions. First, the actual causal exposure must not begin before the outcome merely because its recorded date is later. Second, the future exposure must not affect \(Y\) through anticipation: people must not change relevant behavior in response to knowing or expecting that exposure. Third, \(Y\) must not itself cause, or substantially determine, the later exposure. Fourth, the adjustment must handle measured common causes and relevant prior exposure history without introducing inappropriate conditioning. Finally, the candidate control should share at least some causes or assignment mechanisms relevant to the suspected residual confounding. If it shares no such sources, a null association may tell little about bias in the main analysis.
These conditions are design claims, not facts established by a nonsignificant test. In particular, no association is expected only if the control is valid and the analysis can detect the association of interest. The earlier material on negative control exposures and falsification tests supplies the broader framework; the distinctive issue here is whether “future” really rules out every pathway from the candidate exposure to the earlier outcome.
Why a Nonnull Association Can Be a Warning
A simplified causal structure explains the diagnostic. Suppose an unmeasured factor \(U\), such as underlying disease severity or access to care, influences both \(Y\) and the later exposure \(A^+\). If \(A^+\) cannot affect \(Y\), an adjusted association between them may remain because \(U\) is not controlled. If that same \(U\), or a related process, also influences \(A\), the finding raises concern that the estimated association between \(A\) and \(Y\) may retain confounding.
The reasoning is conditional: if there is no causal path from \(A^+\) to \(Y\), no reverse path from \(Y\) to \(A^+\), and no other unaccounted pathway, then a remaining association is evidence that some part of the presumed separation has failed. Shared unmeasured causes are one important explanation. The result does not identify which cause is responsible, quantify the bias in the main effect, or prove that the main estimate is wrong by a particular amount.
The diagnostic also depends on scale and adjustment. A conditional odds ratio, a risk difference, and an unadjusted difference can behave differently, especially when covariates are imbalanced. State the chosen measure and adjustment set, and inspect whether there is adequate overlap: if a stratum contains only one level of \(A^+\), its within-stratum contrast cannot be estimated from the data.
Worked Example: A Residual Association After Standardizing
In a hypothetical study, researchers examine whether an earlier treatment \(A\) is associated with a binary outcome \(Y\). They also record a candidate future exposure \(A^+\). To keep this calculation focused on the control check, suppose comparisons are made within the same prior-treatment history, and the only adjustment variable shown is baseline severity \(L\). The following counts are hypothetical.
| Severity \(L\) | \(A^+\) | People | People with \(Y=1\) |
|---|---|---|---|
| Lower | Yes | 30 | 6 |
| Lower | No | 70 | 7 |
| Higher | Yes | 70 | 28 |
| Higher | No | 30 | 9 |
The lower-severity outcome risks are \(6/30=0.20\) for \(A^+=1\) and \(7/70=0.10\) for \(A^+=0\). The higher-severity risks are \(28/70=0.40\) and \(9/30=0.30\), respectively. Giving the two severity strata equal standardization weights,
The adjusted association is a 10-percentage-point higher outcome risk among those with the future exposure, under this standardization. The crude risks are \(34/100=0.34\) and \(16/100=0.16\), a difference of 18 percentage points; the change illustrates why the adjustment and target weights matter. If timing records and subject-matter knowledge support the negative-control assumptions, the adjusted association is a warning of residual structure not accounted for by the stated adjustment. It is not evidence that the future exposure caused the earlier outcome, and it does not show that the estimated effect of \(A\) is biased upward by 10 percentage points.
Timing Problems That Can Break the Interpretation
A recorded date is not always the date an exposure begins to affect people. A prescription may be filled after an outcome even though the treatment was planned earlier. A policy may formally take effect later while people respond to its announcement beforehand. These are anticipation pathways: the future exposure, or information about it, can influence behavior before the recorded start. In that situation, temporal ordering alone does not establish a valid negative control.
Persistence creates a different problem. If the candidate exposure is part of a continuing process, the future measurement may be a proxy for exposure that was already present before \(Y\). For example, a condition recorded after an event may reflect a long-running state that began earlier. That prior state could affect the outcome, so the association cannot be attributed solely to unmeasured confounding shared with the exposure of interest.
Reverse causation also deserves explicit attention. An earlier outcome may trigger follow-up, change eligibility, or alter a later treatment decision. Then \(Y\) can cause \(A^+\), producing an association even if \(A^+\) did not cause \(Y\). Examine the clinical or administrative decision process, not merely the timestamps. Excluding exposures decided in response to the outcome may help in some settings, but exclusions can change the study population and do not automatically restore validity.
Worked Example: An Anticipated Future Treatment
In a hypothetical clinic, a treatment is scheduled to begin in the next month. Patients are told about the plan before the outcome window ends and may change their diet immediately. Suppose the earlier outcome is a symptom score, where higher values are worse. Among 100 patients scheduled for the future treatment, 30 have a high symptom score; among 100 not scheduled, 20 have a high score.
The crude risk difference is
Thus, the observed high-symptom risk is 10 percentage points higher in the scheduled group. But the schedule was known before symptoms were measured, and patients changed behavior in response. A pathway from the future treatment plan to behavior before the outcome is therefore plausible. The association cannot be interpreted as a clean negative-control signal for residual confounding: anticipation may have affected \(Y\). The appropriate conclusion is that this candidate fails, or at least requires further justification, as a future-exposure control in this design.
Calendar time can cause a subtler failure. If the frequency of a future exposure changes over time and the outcome also has a time trend, a crude association may appear even when there is no within-period association. Conversely, time trends can obscure an association. Calendar period, season, or changes in measurement and eligibility should be considered as possible common causes, and adjusted estimates should be compared with crude ones.
Worked Example: A Crude Signal Explained by Calendar Time
Consider a hypothetical study with two calendar periods. The future exposure has no association with the outcome within either period. The counts are:
| Period | \(A^+\) | People | People with \(Y=1\) |
|---|---|---|---|
| Earlier | Yes | 80 | 16 |
| Earlier | No | 20 | 4 |
| Later | Yes | 20 | 8 |
| Later | No | 80 | 32 |
Within the earlier period, both risks are \(16/80=0.20\) and \(4/20=0.20\). Within the later period, both are \(8/20=0.40\) and \(32/80=0.40\). The within-period differences are both zero. Yet the crude risks are \(24/100=0.24\) for \(A^+=1\) and \(36/100=0.36\) for \(A^+=0\), giving a crude difference of \(-0.12\), or 12 percentage points lower in the future-exposure group. Standardizing to equal weights for the two periods gives
Here the crude association reflects different distributions across calendar periods, not a within-period signal. This example does not prove that time adjustment is sufficient in any real study; it shows why temporal trends and the distribution of \(A^+\) across time must be examined.
Diagnostics, Alternatives, and Limits
Before interpreting a future-exposure test, check whether the candidate exposure truly begins after the outcome, whether its timing is measured reliably, and whether the outcome can influence its occurrence or recording. Check overlap in the adjustment strata and compare crude and adjusted associations. Consider whether the future exposure plausibly shares causes of the main exposure and outcome. Where possible, repeat the check using reasonable alternative definitions of the future-exposure window and covariate adjustment, while explaining why those choices are appropriate.
A nonnull association is a falsification signal, not a diagnosis of one specific bias. It may reflect shared unmeasured causes, anticipation, persistence, reverse causation, residual calendar-time confounding, selection, or measurement processes. A null finding is also limited: the test may have low precision, the candidate exposure may not share the relevant confounding pathway, or opposing mechanisms may cancel. A confidence interval and an assessment of whether it excludes associations considered meaningful are more informative than a nonsignificant test alone.
A future-exposure check differs from a placebo treatment, which uses a deliberately assigned or defined exposure that should not produce the effect under examination. The future-exposure approach gets its negative-control logic from timing, but that timing must be credible. Other negative-control designs may be preferable when future exposure is affected by the earlier outcome or when anticipatory effects are plausible. None of these checks replaces design-based control of confounding, and a negative-control result by itself does not establish the causal effect of \(A\).
Define when the outcome occurs and when the candidate exposure begins, including when it is planned or communicated.
Specify measured covariates, relevant prior exposure history, the association scale, and the standardization or model used.
Assess anticipation, persistence, reverse causation, time trends, overlap, and whether the candidate shares relevant confounding pathways.
Report the estimate and its uncertainty as evidence about the assumptions, not as a causal effect of the future exposure or a direct correction to the main estimate.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- Why can a future exposure be associated with an earlier outcome even when it has no causal effect on that outcome?
- A future treatment is announced before the outcome is measured, and patients change their behavior after the announcement. Which negative-control condition is threatened?
- In the calendar-time example, what is the crude risk difference, and what is the standardized difference using equal period weights?
- Does a null future-exposure association establish that the main exposure-outcome estimate is unconfounded? Give two reasons.
- Why should prior exposure history sometimes be included when estimating the future-exposure association?
Solutions. (1) Shared unmeasured causes can influence both variables; the outcome may also influence later exposure, or time-related factors may affect both. (2) The no-anticipation condition is threatened because information about the planned exposure may affect behavior before the recorded exposure begins. (3) The crude difference is \(24/100-36/100=-0.12\); equal-weight standardization gives \(0\), since both period-specific differences are zero. (4) No. The test may be imprecise, or the candidate may not share the confounding pathways relevant to the main exposure; a null result can also conceal offsetting mechanisms. (5) Earlier exposure may affect both the outcome and the later exposure measurement, so comparisons that ignore exposure history may mix that structure with the association being used as a diagnostic.