The study, conducted by Jonathon Mellor, Maria Tang, Robert S. Paton, Thomas Ward and colleagues, takes the form of a narrative review of the leading indicator literature. Rather than introducing new data, the authors systematically examined how previous studies have related one epidemic time series to another, and where the analytical machinery of those studies tends to break down. Their central contention is that many published analyses of leading indicators suffer from methodological limitations in both reporting and analysis, limitations that are often invisible to readers but that can dramatically overstate how useful a signal would be during an actual outbreak. From this review, the team distilled a checklist and workflow aimed specifically at the constraints of real-time epidemic analysis, not the comfortable conditions of retrospective study.

The distinction between retrospective and real-time analysis is the conceptual heart of the paper. In a retrospective study, an analyst works with a finished dataset: every case has been counted, every delay has resolved, every revision has been made. In a real-time setting, the most recent data are always the least reliable. Test results arrive days or weeks after the infections they represent, death registrations trail hospitalizations, and administrative datasets are silently revised as late reports trickle in. A leading indicator that appears to predict hospital admissions beautifully when evaluated on finalized data may lose almost all of its predictive advantage once reporting delays and data revisions are taken into account. The authors therefore place reporting delays and data revisions at the center of their recommendations, arguing that any claim about a leading indicator’s value must be evaluated on data as they would have actually been available to decision makers at the moment a forecast or warning would have been issued.

Among the most important of the paper’s recommendations is the call for explicit reporting of causal mechanisms. A correlation between an early signal and a later outcome, however strong, is fragile evidence on its own. If analysts can articulate why a signal should lead, for example, because viral shedding in feces begins before symptom onset, or because test-seeking behavior changes before care-seeking does, then the relationship has a plausible biological or behavioral basis that makes it more likely to hold under changing conditions. Without such a mechanism, a leading relationship observed in one outbreak may simply be an artifact of shared trends, seasonality, or coincidental timing, and may evaporate precisely when it is needed. The review urges authors to state the presumed causal pathway connecting indicator to outcome, and to consider how interventions, behavioral shifts, or pathogen evolution might sever that pathway mid-epidemic.

Sample coverage and bias form another pillar of the recommendations. Surveillance data are never a random sample of the population: wastewater sensors cover particular catchment areas, testing rates vary by age, geography, and symptom severity, and syndromic data reflect who chooses to seek care. The authors highlight the need for analysts to describe who and what is represented in their data, and to assess how gaps in coverage might distort the apparent leading relationship. Closely related is the question of spatio-temporal granularity. An indicator aggregated at the national level may hide regional waves that begin weeks apart, while fine-grained data may be too sparse to support stable estimates. The review recommends that analyses report results at the spatial and temporal resolution at which decisions are actually made, and examine whether the leading relationship holds consistently across regions and time periods rather than only in aggregate.

The paper also addresses a family of statistical choices that can quietly manufacture or destroy evidence of leadership between two time series. Smoothing, for instance, is a near-universal preprocessing step in epidemic analysis, but aggressive smoothing introduces autocorrelation and can shift the apparent timing of peaks, making an indicator look more or less leading than it truly is. Transformations such as logarithms or growth-rate conversions change what a relationship means: an indicator may lead the level of an epidemic curve but lag its growth rate. The authors recommend that analysts justify these choices, test the sensitivity of their conclusions to them, and report uncertainty honestly, including the uncertainty introduced by nowcasting, smoothing, and model selection, rather than presenting point estimates as if the underlying data were exact.

Another subtle pitfall the review identifies concerns multiple event measurement. Epidemic time series often contain multiple waves, and a naive analysis that pools all waves together may find a leading relationship that is driven entirely by differences between waves rather than by any within-wave predictive power. Conversely, focusing on a single wave risks overfitting to one idiosyncratic episode. The authors recommend that studies explicitly handle the multiplicity of epidemic events, testing whether an indicator’s lead time and predictive strength are stable across successive waves and across different pathogens or variants, since a signal validated on ancestral SARS-CoV-2 may behave very differently against a variant with a different generation interval or age profile of severity.

Time-varying relationships receive particular attention as well. The lag between an indicator and the outcome it is supposed to predict is rarely fixed. It can shorten when a health system is under strain, lengthen when testing behavior changes, and drift gradually as populations adapt to an endemic pathogen. A leading indicator analysis that assumes a constant lag across the entire study period may therefore misrepresent the very quantity that decision makers care about. The recommendations encourage analysts to model relationships as dynamic, to report how lead times evolve over the course of an epidemic, and to flag periods during which the indicator-outcome relationship appears to weaken or invert, since those are exactly the moments when blind reliance on an early signal could mislead response efforts.

Perhaps the most consequential recommendation, in practical terms, concerns predictive utility. The authors argue that demonstrating a statistical association between an early signal and a later outcome is not the same as demonstrating that the signal improves decisions. A useful evaluation must ask whether incorporating the indicator into a forecasting or trigger framework would have produced earlier, more accurate, or more actionable warnings than the decision maker would otherwise have had, and it must do so under realistic real-time conditions. This reframing shifts the benchmark from statistical significance to operational value: an indicator with a modest correlation but a reliable two-week lead, quantified uncertainty, and stable behavior across settings may be worth far more than a stronger statistical association that collapses once delays and revisions are modeled. The paper’s workflow is built to walk analysts through these questions in sequence, from mechanism and data provenance through delay handling, granularity, sensitivity analysis, and finally an honest assessment of whether the signal would have changed a decision.

The authors position their contribution as a synthesis and extension of existing methodological work, assembling scattered good practices into a single coherent framework with contextual examples from the literature illustrating both exemplary analyses and common failure modes. The stakes of this housekeeping are considerable. Wastewater monitoring, syndromic surveillance, and digital traces expanded enormously after the COVID-19 pandemic demonstrated their potential, and health agencies worldwide are now deciding which of these signals to institutionalize. If leading indicator studies overstate the reliability of these signals, agencies may build early-warning systems that fail quietly during the next crisis; if the studies are appropriately rigorous, the same signals could buy the days or weeks that turn an uncontrolled outbreak into a manageable one. By giving analysts a practical checklist grounded in the realities of real-time data, the PLOS Digital Health review aims to ensure that the next generation of early-warning research produces evidence that decision makers can act on with confidence, at the speed that epidemics demand.

Subject of Research: Methodological recommendations for analyzing leading indicators in real-time epidemic surveillance

Article Title: Recommendations for the analysis of leading indicators in epidemic surveillance

Article References: Mellor, J., Tang, M., Paton, R. S., & Ward, T. (2026). Recommendations for the analysis of leading indicators in epidemic surveillance. PLOS Digital Health, 5(10), e0001759. https://doi.org/10.1371/journal.pdig.0001759

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001759

Keywords: epidemic surveillance, leading indicators, time series analysis, early warning systems, public health decision making, reporting delays, wastewater surveillance, nowcasting, predictive utility, PLOS Digital Health, data revisions, methodology