Summary of results
- Using the reported symptom mix, this analysis did not identify distinct adverse-event clusters among lots; the profiles form a continuum rather than separable groups.
- The two largest sources of between-lot variation that the analysis could identify were reporting era / completeness and manufacturer platform.
- At the manufacturer level, the serious-event pattern matches the previously documented platform differences (higher thrombosis / GBS / TTS reporting for the Janssen adenoviral-vector vaccine than for the mRNA vaccines).
⚠️ Scope and limitations
- VAERS is passive, unverified reporting. Everything here is a reporting pattern, hypothesis-generating only — not incidence, not causation.
- “AE profile” = the mix of symptom categories mentioned in a lot’s reports, normalized to profile shape (not report volume).
- A clustering analysis describes the structure of the reports. Not finding clusters is not proof that all lots are identical, and finding structure would not by itself prove a batch is unsafe. Small differences, or differences confined to a rare event, can fall below what this method can detect.
The question
Some accounts of “hot lots” or “bad batches” propose that particular lots carry a distinctive adverse-event signature. This analysis tests that proposition against the data: it builds a symptom profile for every COVID lot with enough reports, then asks — with no pre-chosen number of clusters — whether lots fall into separable groups, and if so, what distinguishes them. The sections below report what the analysis found and where the data are inconclusive.
Method
Per-lot profile = fraction of the lot’s reports mentioning each of 17 symptom categories (or, in the finer pass, 35 clinical MedDRA PTs), for lots with ≥50–100 reports so profiles are stable. Objectivity comes from using multiple independent validity indices rather than hand-picking an answer.
Finding 1 — Lots form a continuum, not clusters
The indices agreed on k=2, and the split was stable on resampling. On inspection, the two groups differ in reporting completeness rather than symptom type: the category rank-order is the same in both, and one group has richer reports (1.85 vs 0.75 categories flagged per report). That axis aligns with calendar era — the richer group is mostly 2021 lots (and includes all the Janssen lots), the sparser group mostly 2022 lots — while doses are similar. Within a single manufacturer the same k=2 split correlates +0.61 with lot start-year.
When the relative symptom mix is re-clustered with manufacturer and era statistically removed, the separation largely disappears (silhouette ~0.15). A high Hopkins statistic together with a near-zero silhouette is characteristic of data that is non-uniform but not organized into discrete clusters — a continuous distribution rather than separable groups.
Finding 2 — Differences by manufacturer
Manufacturer is associated with the AE profile (PERMANOVA pseudo-F = 39, p = 0.001), accounting for ~13% of the profile variance; this association persists after removing the report-volume axis. The Janssen (adenoviral-vector) profile is the most distinct: relatively more neurological/systemic reporting and relatively less injection-site reactogenicity than the mRNA vaccines.
At the manufacturer level, the serious-event reporting rate per million doses (dose denominators from openVAERS) shows higher Janssen rates for several events, consistent with the platform-specific patterns reported for that vaccine elsewhere:
| Serious event (reports / million doses) | Pfizer | Moderna | Janssen |
|---|---|---|---|
| Myocarditis | 2.39 | 1.72 | 1.72 |
| Pericarditis | 1.77 | 1.71 | 2.46 |
| Guillain-Barre syndrome | 0.57 | 0.67 | 5.08 |
| Anaphylactic reaction | 1.83 | 2.07 | 3.16 |
| Pulmonary embolism | 2.93 | 3.77 | 14.47 |
| Deep vein thrombosis | 2.1 | 2.74 | 13.16 |
| Thrombocytopenia | 0.95 | 0.93 | 3.94 |
| Immune thrombocytopenia | 0.28 | 0.36 | 0.81 |
| Thrombosis with thrombocytopenia syndrome | 0.01 | 0.0 | 0.3 |
| Bell's palsy | 2.55 | 3.3 | 6.16 |
| Seizure | 4.23 | 4.61 | 17.84 |
| Cerebrovascular accident | 2.9 | 3.9 | 10.77 |
| Death | 12.4 | 20.07 | 30.12 |
| Myocardial infarction | 1.26 | 1.63 | 3.4 |
| Cardiac arrest | 1.44 | 1.78 | 3.84 |
| Encephalitis | 0.08 | 0.11 | 0.17 |
| Transverse myelitis | 0.0 | 0.0 | 0.0 |
Finding 3 — Lot-level signals for specific serious events
Testing every lot × serious-event combination against its own manufacturer’s dose-denominated baseline, a naive Poisson model flagged 742 lots. VAERS counts are, however, strongly overdispersed (dispersion φ = 16 overall, 73 for Death), which makes those p-values anti-conservative. Under a quasi-Poisson correction the count drops to 47. The remaining signals are concentrated in the earliest, largest 2021 lots — the lots often described as “hot lots.”
These lots were also disproportionately administered early in the rollout, to older recipients and under intensive surveillance (examined on the rollout page). Age standardization removes much, though not all, of the death excess, and the anaphylaxis signal is not explained by age. The available data are consistent with reporting and population effects but cannot, on their own, exclude a contribution from the lots themselves.
Finding 4 — Density scan for small tight groups
As a sensitivity check for a small cluster that averaging methods could miss, a DBSCAN density scan looked for any tight group of ≥5 near-identical lots in the manufacturer+era-residualized space. The groups it identified corresponded to:
- Lot-number families (consecutive lots such as
EN6204/EN6205/EN6208) — sequential lots distributed and administered together, which share exposure timing and population; - Calendar-era groups (e.g. a 2022 group defined by Cough + Oropharyngeal pain, matching the Omicron-era shift in what was reported);
- Small reactogenicity differences (±1–4 percentage points in Chills / Nausea / Dizziness).
None of the groups was defined by an anomalous serious event.
Summary
Across global clustering, era/manufacturer residualization, a finer PT-level pass, and a density scan, this analysis did not identify distinct adverse-event clusters among COVID vaccine lots: the profiles vary along a continuous reporting-completeness / calendar gradient and, separately, by manufacturer platform. Where individual lots show elevated reporting of specific serious events, those lots are also the earliest and most intensively surveilled, and were given to different populations — factors that the analysis shows account for much of the difference but that passive data cannot fully separate from a possible batch-level effect. Readers should weigh these findings as one line of evidence, subject to the limitations of spontaneous-reporting data described above.
Reproducible pipeline: cluster_lot_ae_profiles.py, pca_manufacturer_signature.py, lot_serious_pt_disproportionality.py, cluster_relaxed_exploration.py, cluster_density_scan.py. Data: VAERS 2006–2024 lot events × openVAERS per-lot doses. Passive surveillance; hypothesis-generating only.