Temporal Correlation in Exposure Status

• Temporal correlation in exposure status is a process where exposures persist and cluster over time rather than occurring independently.
• Methodological strategies such as infectious-window models, sequential adjustment, and Markov-chain formulations are used to address this temporal dependence.
• Accounting for exposure clustering is vital for accurate risk estimation and bias correction in vaccine trials, recurrent event studies, and environmental research.

Searching arXiv for the specified papers and closely related work on temporally correlated exposure status. Temporally correlated exposure status denotes a longitudinal exposure process in which exposure at one time point predicts exposure at later time points, so that exposure is not a sequence of independent one-off events but instead occurs in clusters, runs, or persistent states. In the infectious-disease setting, the canonical notation is $E_t$, with direct dependence such as $E_{t-2}\to E_{t-1}$, arising because infectious contacts often remain infectious for several days and because people tend to have repeated contact with the same partners, coworkers, household members, or classmates (Ando et al., 3 Sep 2025). More broadly, the term also covers longitudinal settings in which repeated exposures are highly predictive from one interval to the next, whether handled implicitly through conditioning on past history in causal models (Mork et al., 9 Jun 2025), through structured regularization of adjacent time-indexed predictors (Mork et al., 2021), or through an explicit first-order Markov prior for binary latent status variables (Zhu et al., 2023).

1. Conceptual definition and distinction from related forms of dependence

In the formulation developed for vaccine efficacy trials, exposure at time $t$ is denoted $E_t$, and the defining feature is serial dependence in the latent exposure process itself: exposure at one time makes exposure at the next time more likely (Ando et al., 3 Sep 2025). The paper emphasizes that this is distinct from the more standard problem of general unmeasured heterogeneity in exposure or susceptibility. In that classical setup, a latent factor $U$ makes some individuals systematically more exposed or more susceptible than others. By contrast, temporally correlated exposure status can generate bias even without such a $U$; the key structure is direct persistence in exposure over time.

This distinction matters because it separates two different causal mechanisms. Unmeasured heterogeneity is a between-person problem, whereas temporally correlated exposure status is a within-person, time-varying property of the exposure trajectory. In DAG terms, the vaccine-trial paper highlights the path

$$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$

where $A$ is vaccination and $Y_t$ is infection status (Ando et al., 3 Sep 2025). The role of this path is not merely descriptive: it identifies how treatment-induced survival in the risk set can interact with exposure persistence and thereby alter the latent exposure composition among those still under follow-up.

Related literatures use different language for closely related phenomena. The structural nested recurrent-event paper does not explicitly posit a stochastic model for serial correlation or autocorrelation in the exposure process; instead, it allows temporal dependence implicitly by conditioning the exposure mechanism on past exposure and covariate history (Mork et al., 9 Jun 2025). The tree-based critical-window paper likewise does not model the exposure-generating covariance structure directly; rather, it addresses the consequences of high temporal correlation among adjacent exposures through structured partitions of the ordered time axis (Mork et al., 2021). The massive-access paper is the most explicit stochastic formulation: each binary status variable $\lambda_n^t\in\{0,1\}$ follows an independent across users, first-order, time-homogeneous, steady-state Markov chain with transition probabilities $E_{t-2}\to E_{t-1}$0 and $E_{t-2}\to E_{t-1}$1 (Zhu et al., 2023).

A plausible implication is that “temporally correlated exposure status” is best understood as an umbrella concept spanning at least three inferential regimes: latent exposure persistence that distorts hazard-based estimands, history-dependent exposures handled by sequential exchangeability assumptions, and explicitly parameterized binary state persistence. The common element is serial dependence in exposure trajectories rather than simple cross-sectional exposure variation.

2. Mechanisms that generate temporal dependence

The motivating mechanism in infectious-disease vaccine trials is repeated contact with an infectious source whose infectious period spans several days. If a participant is exposed on one day, the same contact may still be infectious on the next day, and the participant may continue interacting with that contact under ordinary social routines (Ando et al., 3 Sep 2025). Exposure therefore persists across adjacent times, and the latent process contains repeated exposure opportunities within what the paper formalizes as an infectious window.

The infectious-window model makes this persistence explicit. Individuals may enter a temporary state during which they remain exposed for several time points, with duration

$E_{t-2}\to E_{t-1}$2

Entry into an infectious window occurs with a small probability $E_{t-2}\to E_{t-1}$3 on any day, assumed i.i.d. across people and time, and during the window each time point carries infection probability $E_{t-2}\to E_{t-1}$4 (Ando et al., 3 Sep 2025). Under one-time infection, once infected a participant exits the risk pool permanently and cannot re-enter an infectious window. Temporal dependence therefore arises from the persistence of the exposure episode itself rather than from a permanent person-level risk type.

Other fields instantiate the same phenomenon differently. In longitudinal causal inference for recurrent events, exposure histories are indexed at discrete times $E_{t-2}\to E_{t-1}$5, with average exposure $E_{t-2}\to E_{t-1}$6 over $E_{t-2}\to E_{t-1}$7, and temporal dependence is allowed through the conditional exposure mechanism

$E_{t-2}\to E_{t-1}$8

This formulation explicitly permits $E_{t-2}\to E_{t-1}$9 to depend on earlier exposure $t$0, contemporaneous covariates, and survival status under the relevant intervention, without making autocorrelation the scientific target (Mork et al., 9 Jun 2025).

In high-resolution environmental epidemiology, temporally correlated exposure status appears as adjacent repeated measurements that behave like highly collinear predictors. For subject $t$1, exposure history is observed at equally spaced times $t$2, either as a single vector $t$3 or as mixture-specific histories $t$4 (Mork et al., 2021). The central challenge is not that temporal dependence is absent from the model, but that naively regressing on each week-specific exposure yields unstable lag-effect estimates because neighboring weeks compete to explain the same signal.

The explicit Markov-chain formulation in the massive-access literature provides a compact state-based representation of persistence. For each user $t$5,

$t$6

Here $t$7 is the switch-on probability and $t$8 the stay-on probability (Zhu et al., 2023). Under stationarity,

$t$9

which separates marginal prevalence from temporal persistence. This suggests a general principle: two exposure processes can have the same marginal exposure frequency yet differ materially in inferential behavior because their persistence structures differ.

3. Selection-through-risk-set bias in vaccine efficacy trials

The vaccine-trial literature represented by "Temporal Exposure Dependence Bias in Vaccine Efficacy Trials" (Ando et al., 3 Sep 2025) identifies a specific consequence of temporally correlated exposure status: downward bias in time-to-event estimates of per-contact vaccine efficacy when standard Cox proportional hazards models do not condition on exposure. The core mechanism is selection through the evolving risk set. If vaccination lowers the probability that a given exposure causes infection, vaccinated individuals remain uninfected longer and therefore continue to be available for later exposures arising from the same infectious episode. Unvaccinated individuals are more likely to be infected earlier and to leave follow-up sooner.

This produces arm-specific exposure imbalance among those still at risk. Vaccinated participants tend to accumulate more exposure opportunities conditional on still being uninfected, not because randomization failed, but because treatment changes survival through prior exposures (Ando et al., 3 Sep 2025). Standard hazard-based analyses compare risks without observing the latent exposure process $E_t$0, so the estimated hazard ratio combines biological protection per exposure with differential exposure accumulation over time.

The paper formalizes the distinction between marginal hazard and per-contact hazard. The discrete-time marginal hazard is

$E_t$1

while the per-contact hazard is

$E_t$2

Using the law of total probability and the rare-entry approximation, the marginal hazard is decomposed as

$E_t$3

Cox-based vaccine efficacy is then defined by

$E_t$4

whereas per-contact vaccine efficacy is

$E_t$5

Substitution yields the key bias equation

$E_t$6

If the relative exposure probability exceeds $E_t$7, then $E_t$8, implying downward bias in the Cox-based estimate (Ando et al., 3 Sep 2025).

Under the infectious-window model, the relative exposure probability admits the closed-form approximation

$E_t$9

For practical use, the paper also proposes a truncated version with $U$0,

$U$1

leading to

$U$2

These expressions make the determinants of bias explicit: per-contact transmissibility $U$3, per-contact vaccine efficacy $U$4, and the persistence distribution $U$5 (Ando et al., 3 Sep 2025).

The empirical importance is not presented as universal. The paper states that the bias is “typically small” in many scenarios but can still be nontrivial. Under “realistic epidemiological conditions” such as $U$6, $U$7, and per-contact VE $U$8, the gap between per-contact VE and Cox-based VE exceeds $U$9 (Ando et al., 3 Sep 2025). This suggests that temporally correlated exposure status is not merely a theoretical nuisance; in some trial settings it can affect interpretation and threshold-based decision rules.

4. Modeling strategies across research areas

The available arXiv literature presents three distinct methodological responses to temporally correlated exposure status.

Paper	Exposure representation	How temporal correlation is handled
(Ando et al., 3 Sep 2025)	Latent $U$0 within an infectious-window model	Explicit persistence through $U$1; analytic bias approximation
(Mork et al., 9 Jun 2025)	Time-varying $U$2 with history $U$3	Conditioning on past exposure/covariate history; sequential ignorability
(Mork et al., 2021)	Time-resolved predictors $U$4	Contiguous time partitions and shrinkage over adjacent weeks
(Zhu et al., 2023)	Binary latent status $U$5	First-order Markov prior with $U$6

In the vaccine-trial framework, temporal dependence is an object of substantive concern because it induces a discrepancy between marginal hazard ratios and per-contact causal protection (Ando et al., 3 Sep 2025). The model is therefore mechanistic: it specifies entry into infectious windows, duration $U$7, and per-contact infection risk $U$8, and derives analytic expressions for resulting bias.

The SNCURE framework of "A structural nested rate model for estimating the effects of time-varying exposure on recurrent event outcomes in the presence of death" (Mork et al., 9 Jun 2025) takes a different stance. It is designed to estimate short-term and delayed marginal causal effects of exposures on recurrent event rates while accounting for death as a terminating event. Exposure history enters both as the object of intervention and as the conditioning set used for confounding adjustment. The lag-$U$9 structural nested rate model is

$$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$0

Temporal correlation is thus accommodated by history dependence in the exposure mechanism rather than by explicit autocorrelation parameters.

The structured Bayesian regression tree-pair framework of "Estimating Perinatal Critical Windows of Susceptibility to Environmental Mixtures via Structured Bayesian Regression Tree Pairs" (Mork et al., 2021) addresses the instability caused by high temporal correlation among repeated exposures. For a single exposure, it replaces unconstrained week-specific coefficients $$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$1 with a piecewise constant lag function induced by binary trees over time. For multiple exposures, tree pairs yield blockwise constant interaction surfaces over $$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$2. The paper states directly that by keeping the number of terminal nodes small, the treed distributed lag model introduces a necessary constraint on the distributed lag function to account for temporal correlation by assuming that exposures within the same terminal node share an equal effect. Here temporal correlation is not modeled as a stochastic exposure process; instead, inference is regularized to respect temporal ordering and adjacency.

The massive-access formulation in "Cooperative Multi-Cell Massive Access with Temporally Correlated Activity" (Zhu et al., 2023) treats temporal correlation as the central prior structure. Binary status variables $$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$3 are linked across time by Markov transitions and, conditional on status, gate continuous latent coefficients through a Bernoulli-Gaussian prior

$$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$4

This architecture is not an exposure-effect model in the epidemiologic sense, but it is directly relevant to generic temporally correlated exposure-status inference because it separates state persistence from measurement generation.

A plausible synthesis is that these approaches correspond to three inferential priorities: bias correction for latent persistent exposure, causal effect identification under history-dependent treatment, and stable estimation under temporally ordered collinearity. They are complementary rather than competing descriptions of the same phenomenon.

5. Estimation, identification, and computational inference

In the vaccine-trial setting, standard Cox analysis estimates

$$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$5

with $$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$6 obtained by maximizing the partial likelihood

$$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$7

The critique is not directed at Cox regression as a generic tool but at what the hazard ratio identifies when exposure is latent and temporally persistent (Ando et al., 3 Sep 2025). Because the Cox risk set mixes exposed and unexposed individuals and does not condition on current exposure, treatment-induced changes in survival through repeated exposure windows alter the exposure composition of the risk set.

The paper supplements this critique with a practical adjustment strategy. In simulations with $$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$8 individuals, $$A \to Y_{t-1} \leftarrow E_{t-2} \to E_{t-1} \to Y_t,$$9 days of follow-up, $A$0, $A$1, $A$2, and $A$3, it compares the Cox-based estimator $A$4 with an adjusted estimator $A$5 obtained by solving

$A$6

using the truncated approximation with $A$7 (Ando et al., 3 Sep 2025). Across simulations, $A$8 is biased downward relative to the true per-contact VE, whereas the adjusted estimator substantially reduces the bias.

The structural nested recurrent-event framework uses estimating equations based on residualized exposures

$A$9

where $Y_t$0 is the conditional mean of exposure given past exposure and covariates (Mork et al., 9 Jun 2025). For $Y_t$1, $Y_t$2 solves

$Y_t$3

For delayed effects $Y_t$4, the estimating equations incorporate survival weights

$Y_t$5

which map observed survivor populations to counterfactual survivor populations under blipped exposure histories. Identification relies on consistency, sequential no-unmeasured-confounding assumptions, and positivity (Mork et al., 9 Jun 2025). The paper explicitly notes that if exposure is highly persistent, positivity may be practically weak because the residualized exposure can have little variation.

The structured tree-pair approach uses MCMC over trees, exposure assignments, node effects, and shrinkage parameters (Mork et al., 2021). Its inferential logic is not causal identification through sequential exchangeability or hidden-state smoothing through Markov recursion. Instead, it searches over structured low-dimensional representations of lag functions and interaction surfaces. A tree partitions the ordered time axis into contiguous intervals, and the additive ensemble allows the final lag function to be approximately smooth or piecewise constant. Variable selection operates both at the exposure level and, indirectly, at the time-window level through the learned tree structure.

The DCS-MMV-GAMP algorithm combines discrete message passing over temporal status variables with continuous latent-variable estimation via GAMP (Zhu et al., 2023). The posterior over a sliding window $Y_t$6 factors as

$Y_t$7

Forward-backward temporal message passing refines the activity likelihood using both previous and future frames inside the window. This is analogous to hidden Markov smoothing, but embedded in a larger factor graph with multiple measurement vectors. The generalized sliding-window strategy balances temporal context against complexity and latency (Zhu et al., 2023).

6. Applications, limitations, and interpretive issues

The direct empirical illustration in the vaccine-trial paper uses the Hispanic/Latinx subgroup from the ChAdOx1 nCoV-2019 trial reported by Falsey et al. (2021) (Ando et al., 3 Sep 2025). The reported Cox-based VE is $Y_t$8 with 95% CI $Y_t$9. Under $\lambda_n^t\in\{0,1\}$0, $\lambda_n^t\in\{0,1\}$1, and $\lambda_n^t\in\{0,1\}$2, the estimated per-contact VE values are $\lambda_n^t\in\{0,1\}$3, $\lambda_n^t\in\{0,1\}$4, and $\lambda_n^t\in\{0,1\}$5, with corresponding confidence intervals $\lambda_n^t\in\{0,1\}$6, $\lambda_n^t\in\{0,1\}$7, and $\lambda_n^t\in\{0,1\}$8. The adjusted estimates are therefore modestly higher than the original Cox-based estimate, and the lower confidence bounds exceed the FDA’s $\lambda_n^t\in\{0,1\}$9 EUA threshold, unlike the original lower bound of $E_{t-2}\to E_{t-1}$00.

The recurrent-event literature provides a different application profile. SNCURE is applied to $E_{t-2}\to E_{t-1}$01 Medicare beneficiaries to estimate the effects of fine particulate matter air pollution on recurrent cardiovascular hospitalizations in the presence of death (Mork et al., 9 Jun 2025). The main relevance to temporally correlated exposure status is methodological: exposures may be substantially history-dependent, but the framework treats that dependence as allowable nuisance structure so long as the relevant history is observed and the sequential ignorability assumptions remain credible.

The perinatal environmental-mixtures application uses weekly gestational exposures over $E_{t-2}\to E_{t-1}$02 weeks in a Denver birth cohort of $E_{t-2}\to E_{t-1}$03 singleton full-term births (Mork et al., 2021). Same-week pairwise exposure correlations range from $E_{t-2}\to E_{t-1}$04 to $E_{t-2}\to E_{t-1}$05, and the method identifies broad contiguous critical windows rather than unstable week-level spikes. This is important because a common misconception is that temporally correlated exposures can always be localized precisely to individual time points. The paper’s results instead show that under strong temporal correlation, broad windows may be more stable and interpretable than isolated weeks.

The massive-access paper shows, in a different domain, that exploiting temporal correlation in binary status detection improves performance relative to methods that ignore it, and that gains increase with persistence parameter $E_{t-2}\to E_{t-1}$06 (Zhu et al., 2023). A plausible implication is that whenever exposure status is latent and repeatedly measured through noisy proxies, forward-backward smoothing with an explicit persistence prior can materially improve state inference.

Across these literatures, the main limitations are consistent. In the vaccine-trial setting, exposure is latent, and the correction does not identify per-contact VE from observed data alone; it requires specification of $E_{t-2}\to E_{t-1}$07 and the infectious-window distribution $E_{t-2}\to E_{t-1}$08, relies on the rare-entry approximation $E_{t-2}\to E_{t-1}$09, and treats multiple simultaneous exposures as negligible (Ando et al., 3 Sep 2025). In SNCURE, strong persistence can create weak positivity and low residual variation in lagged exposures, making delayed effects unstable (Mork et al., 9 Jun 2025). In structured tree-pair models, temporal correlation is regularized rather than explicitly modeled, so exact week-level localization can remain uncertain under extreme collinearity (Mork et al., 2021). In Markov state-space formulations, independence across units and known transition parameters may be unrealistic in many exposure settings (Zhu et al., 2023).

A recurring interpretive issue is whether temporally correlated exposure status should be viewed as a confounding problem, a measurement problem, or a dynamic selection problem. The available evidence suggests that the answer depends on the estimand. In vaccine efficacy trials it is primarily a dynamic selection-through-risk-set problem; in structural nested models it is a history-dependent treatment process requiring sequential adjustment; in distributed lag regression it is a collinearity and regularization problem; and in latent-state inference it is a persistence prior over binary statuses. The term therefore names a property of the exposure trajectory, but its inferential consequences are model-specific.

7. Broader significance and future methodological directions

The clearest general lesson is that time-indexed exposure should not be treated as i.i.d. by default. In infectious-disease trials, failure to account for persistent latent exposure can bias marginal hazard-based estimates downward relative to per-contact biological protection (Ando et al., 3 Sep 2025). In recurrent-event causal inference, exposure histories must be modeled conditionally on observed past and on survival processes if short-term and delayed causal effects are to be estimated coherently (Mork et al., 9 Jun 2025). In high-resolution exposure-response modeling, temporally ordered collinearity motivates architectures that learn effects over contiguous windows rather than over isolated time points (Mork et al., 2021). In latent binary-state problems, explicit Markov persistence can be exploited through forward-backward smoothing within a sliding window (Zhu et al., 2023).

The vaccine-trial paper recommends several practical responses: recognize temporally correlated exposure explicitly in infectious-disease VE trials; quantitatively assess its possible magnitude using plausible values of $E_{t-2}\to E_{t-1}$10 and $E_{t-2}\to E_{t-1}$11 even when exposure is unmeasured; consider the analytic correction as a supplementary sensitivity analysis; collect data that inform exposure timing, contact persistence, or infectious periods when possible; and explore design or modeling adaptations that better account for latent exposure processes (Ando et al., 3 Sep 2025). These proposals remain supplementary rather than definitive because the underlying exposure process is usually not observed.

The causal recurrent-event paper points toward another direction: flexible nuisance estimation of the exposure mechanism $E_{t-2}\to E_{t-1}$12 and the blipped outcome regression $E_{t-2}\to E_{t-1}$13 can absorb complex serial dependence without requiring a parametric time-series law (Mork et al., 9 Jun 2025). The tree-pair paper suggests a complementary route: when the scientific goal is window identification rather than exposure-process modeling, temporally correlated exposure status can be handled by structured time partitions, sparse exposure selection, and shrinkage priors that stabilize neighboring coefficients (Mork et al., 2021). The massive-access paper suggests yet another route: if status persistence itself is substantively meaningful and sufficiently simple, a Markov prior plus message passing may offer an efficient way to combine temporal and multi-source information (Zhu et al., 2023).

This suggests that temporally correlated exposure status is not a niche complication confined to one application area. It is a general feature of longitudinal exposure processes whose consequences depend on whether the scientific target is a per-contact causal effect, a lagged marginal causal effect, a critical exposure window, or a latent exposure-state trajectory. The arXiv literature indicates that the methodological challenge is therefore not merely to acknowledge temporal dependence, but to align the representation of that dependence with the estimand, the observation process, and the source of inferential instability.