Skip to content

Variant-Composition Surveillance Dashboard

Monitoring dashboard — instantiates Adaptive Barrier-Circumvention Response

Tracks the shifting share of each variant class over time — not just total incidence — so population-weighted protection loss shows up before the surviving forms take over.

A Variant-Composition Surveillance Dashboard keeps a population's composition in view, not just its total size. Its defining idea is that aggregate incidence can hold steady — or even fall — while the mix underneath shifts toward the variants your barrier covers worst; a headline count is blind to that shift. So the dashboard tracks each variant class's share and trajectory beside its estimated coverage, and reports the population-weighted protection that results — raising a flag when that weighted figure is decaying even though per-event performance looks unchanged.

Example

A regional public-health lab runs a dashboard on sequenced samples of a circulating respiratory virus. Total case counts are flat, which on its own reads as reassuring. But the composition panel shows one lineage — the one the current countermeasure covers least — climbing from ≈5% to ≈30% of sequenced samples over roughly two months, its confidence band staying wide because one district's sequencing volume is thin. The coverage-decay panel multiplies each lineage's share by its estimated protection and reports the population-weighted figure sliding, even though the countermeasure's per-sample performance hasn't moved. That juxtaposition — flat totals, shifting mix, decaying weighted protection — is the signal to convene review before the uncovered lineage becomes the population, not after.

How it works

What separates it from an ordinary incidence chart is that it plots share × coverage over time. Three things it keeps explicit are exactly the ones dashboards usually smooth away: the uncertainty band on each share, an unclassified bucket for samples that fit no known class, and the per-variant coverage estimate that converts raw shares into a weighted-protection line. It reports what is changing in the mix; it deliberately does not explain why the change is happening or decide what to do about it.

Tuning parameters

  • Classification granularity — how finely variants are binned. Finer bins catch an emerging class sooner but thin each bin's sample and inflate noise.
  • Decay-trigger threshold — how large a weighted-protection drop, sustained over how many periods, fires a review. Tight is early but flappy; loose is stable but late.
  • Subgroup disaggregation — whether shares are shown per site/subpopulation or pooled. Pooling is cleaner but can hide a local takeover behind a stable overall average.
  • Unknown-category handling — how prominently unclassified samples are displayed; quietly dropping them flatters coverage.
  • Refresh cadence and lookback window — how fresh the picture is versus how much history frames "normal."

When it helps, and when it misleads

Its strength is separating how much from what kind — the distinction aggregate metrics erase, and the one that lets compositional decay be seen while headline numbers still look calm. Its failure modes are sampling and ascertainment bias (you only see the variants you sequence or test for) and false precision on thin bins. Its classic misuse is reading a reassuring flat total as safety while the covered-least share climbs underneath — or gaming the view by dropping the unclassified bucket so coverage looks better than it is.[1] The discipline that guards against this is to carry uncertainty and the unknown category visibly through the display, and to disaggregate by subgroup rather than trusting the pooled line.

How it implements the components

The dashboard fills the composition-and-decay side of the archetype's machinery — the monitoring an analytic surface can hold, not the detection or decision around it:

  • baseline_variant_composition — maintains each class's share against a versioned pre-barrier reference, so "normal" is defined and drift is measurable rather than felt.
  • coverage_decay_indicator — derives the population-weighted-protection signal from share × coverage and surfaces the decay flag when it falls.

It does NOT do the distributed, multi-site early detection that feeds it (that's Escape-Variant Sentinel Network), test whether an observed shift is truly selection-driven (Selection-Differential Cohort Analysis), or decide the response to a decay flag (Coverage-Decay Trigger and Release Gate).

Notes

The weighted-protection line is only as trustworthy as the per-variant coverage estimates fed into it. A confident-looking curve built on stale coverage inputs is false precision, not reassurance — which is why the dashboard displays the coverage estimate and its age alongside the share, rather than collapsing both into a single number.

References

[1] Simpson's paradox — a trend that holds within every subgroup can reverse or vanish once the subgroups are pooled. It is the precise reason a stable aggregate can conceal a decisive compositional shift, and why disaggregation is a discipline here rather than a nicety. ↩