Independent Reanalysis: COVID/Shingles Vaccine Signals Amid FDA Publication Blocks
Bayesian Disproportionality Across mRNA, Viral-Vector, and Zoster Products in VAERS
Author
Global Patient Safety
Published
May 13, 2026
Show code
library(arrow)library(dplyr)library(tidyr)library(ggplot2)library(forcats)library(gt)library(scales)library(stringr)# Resolve the VAERS signals parquet from a list of candidate paths so this# QMD renders both on developer workstations and on the VPS.# Need the FULL parquet (with prr/ror/ic/observed/etc.), not the 4-column# aggregated splash version that powers the gps-patient app's headline.SIGNALS_CANDIDATES <-c("/srv/shiny-server/gps-patient/data/signals_vaers_full.parquet","/home/harlan/data/signal-compute/vaers/signals_vaers_v2026-05-03.parquet")SIGNALS_PATH <- SIGNALS_CANDIDATES[file.exists(SIGNALS_CANDIDATES) |dir.exists(SIGNALS_CANDIDATES)][1]stopifnot(!is.na(SIGNALS_PATH))vaers_raw_quarterly <-open_dataset(SIGNALS_PATH) |>collect()# The full parquet stores ONE ROW PER (drug, event, quarter) — the# disproportionality stats recomputed at each quarterly snapshot.## For each (drug, event) pair we keep the FIRST quarter at which it# crossed the project's consensus signal threshold (EB05 > 2 AND# flagged by >= 2 of 4 methods). This is the regulatory-relevant moment:# "when would real-time surveillance have first detected this?". Stats# attached to the row are exactly those at first detection — not the# peak, not the latest, not an average across quarters. This is the# same convention the existing COVID vaccine article uses via the# pre-computed covid_first_seen.parquet table.vaers_raw <- vaers_raw_quarterly |>filter(eb05 >2, n_methods_flagged >=2) |># signal-grade rows onlygroup_by(drug, event) |>arrange(quarter, .by_group =TRUE) |>slice_head(n =1) |># earliest signaling quarterungroup() |>rename(first_quarter = quarter)cat("After collapse to first-detection quarter:",format(nrow(vaers_raw), big.mark =","),"unique (drug, event) signals\n")
After collapse to first-detection quarter: 69,423 unique (drug, event) signals
Context
This analysis is prompted by recent reports of delays and blocks on the publication of vaccine safety studies, including FDA-funded analyses, in May 2026. Because the underlying spontaneous-reporting data — CDC’s VAERS — remains publicly accessible, an independent reanalysis using standard pharmacovigilance methods is feasible regardless of the publication status of any particular paper.
This article applies the same disproportionality framework used elsewhere on the site — GPS / PRR / ROR / BCPNN-IC, with a multi-method consensus threshold of “flagged by at least 2 of 4 methods” — to two vaccine families: COVID-19 vaccines (nine products) and shingles vaccines (three products: Shingrix recombinant, Zostavax live, and unbranded Zoster reports). The focus is on three pre-specified event domains: cardiac, neurological, and thrombotic outcomes. Pre-specifying systems of interest narrows multiple-comparisons exposure and matches the prompt for this reanalysis.
Note
This is a signal-detection reanalysis, not a causal study. Disproportionality methods identify pairs where the reporting rate is unusually high relative to the rest of VAERS. Such signals are hypotheses, not evidence of causation. Confirmation requires controlled study designs, denominator-aware cohort or self-controlled methods, and clinical adjudication.
Methods
All four methods come from the safetysignal R package developed for this project (see project-wide methodology notes). For each (vaccine, event) pair counted in VAERS:
GPS (Gamma-Poisson Shrinker) — DuMouchel 1999 dual-Gamma posterior. EB05 is the 5th-percentile credible lower bound on the linear relative-reporting-ratio scale (a direct posterior quantile, not the EBGM geometric mean which would apply a log transformation).
Each row of every table that follows is a unique (vaccine, event) signal counted once at its first-detection quarter. “Signals detected” therefore counts distinct pairs that crossed the consensus threshold at some point in the data window, not the number of quarterly observations.
Show code
both |>group_by(family, vaccine) |>summarise(signals_detected =n(),signals_4_methods =sum(n_methods_flagged ==4, na.rm =TRUE),max_eb05 =round(max(eb05, na.rm =TRUE), 1),earliest_signal =min(first_quarter, na.rm =TRUE),.groups ="drop" ) |>arrange(family, desc(signals_detected)) |>gt() |>tab_header(title ="Distinct signals per vaccine — cardiac, neurological, thrombotic only",subtitle ="One row per (vaccine, event) at its first-detection quarter") |>fmt_number(columns =c(signals_detected, signals_4_methods), decimals =0) |>cols_label(family ="Family",vaccine ="Vaccine",signals_detected ="Signals detected",signals_4_methods ="Flagged by all 4 methods",max_eb05 ="Max EB05",earliest_signal ="Earliest" ) |>tab_options(table.font.size =13)
Distinct signals per vaccine — cardiac, neurological, thrombotic only
One row per (vaccine, event) at its first-detection quarter
Family
Vaccine
Signals detected
Flagged by all 4 methods
Max EB05
Earliest
COVID-19
Janssen (J&J)
71
57
5.4
2021Q1
COVID-19
Pfizer-BioNTech
40
39
4.0
2020Q4
COVID-19
Moderna
37
34
5.3
2020Q4
COVID-19
Pfizer Bivalent
35
32
4.1
2022Q3
COVID-19
Other / unknown
29
20
36.5
2021Q1
COVID-19
Novavax
9
5
2.5
2023Q1
COVID-19
Moderna Bivalent
7
7
2.8
2023Q1
Shingles
Zostavax (live)
50
21
12.0
2006Q3
Shingles
Shingrix (recombinant)
20
18
3.1
2021Q1
Shingles
Zoster (no brand)
15
3
4.3
2015Q4
Top signals — cardiac
Show code
cardiac_top <- both |>filter(category =="Cardiac") |>arrange(desc(eb05)) |>group_by(family) |>slice_head(n =12) |>ungroup()cardiac_top |>select(family, vaccine, event, first_quarter, eb05, prr, ic, n_methods_flagged) |>gt() |>tab_header(title ="Top cardiac signals — COVID-19 + shingles vaccines",subtitle ="Statistics at first-detection quarter · top 12 per family by EB05") |>fmt_number(columns =c(eb05, prr, ic), decimals =1) |>data_color(columns = eb05,palette =c("white", "#a50f15"),alpha =0.7) |>cols_label(first_quarter ="First detected") |>tab_options(table.font.size =12)
Top cardiac signals — COVID-19 + shingles vaccines
Statistics at first-detection quarter · top 12 per family by EB05
family
vaccine
event
First detected
eb05
prr
ic
n_methods_flagged
COVID-19
Other / unknown
Immune-mediated myocarditis
2022Q3
36.5
930.5
8.4
2
COVID-19
Other / unknown
Cardiac ventriculogram left
2022Q2
4.4
82.5
6.5
2
COVID-19
Pfizer-BioNTech
Sinus tachycardia
2020Q4
4.0
17.0
2.6
4
COVID-19
Pfizer-BioNTech
Tachycardia
2020Q4
3.8
6.9
2.1
4
COVID-19
Pfizer Bivalent
Heart rate abnormal
2025Q4
3.8
32.1
5.0
4
COVID-19
Pfizer Bivalent
Ventricular tachycardia
2024Q4
3.3
11.7
3.3
4
COVID-19
Moderna
Tachycardia
2020Q4
3.3
4.5
2.0
4
COVID-19
Pfizer-BioNTech
Heart rate increased
2020Q4
3.1
4.9
1.8
4
COVID-19
Other / unknown
Pericarditis
2023Q1
2.8
4.0
2.0
4
COVID-19
Janssen (J&J)
Cardiac disorder
2024Q4
2.8
5.2
2.4
4
COVID-19
Moderna Bivalent
Heart rate
2024Q4
2.8
3.8
1.9
4
COVID-19
Janssen (J&J)
Cardiac ablation
2025Q3
2.7
18.4
4.3
4
Shingles
Zostavax (live)
Acute cardiac event
2021Q4
6.1
105.7
7.0
2
Shingles
Zostavax (live)
Heart injury
2020Q1
5.5
107.3
5.5
4
Shingles
Shingrix (recombinant)
Nerve conduction studies abnormal
2021Q1
3.1
9.9
3.0
4
Shingles
Zostavax (live)
Myocardial ischaemia
2020Q2
3.1
139.9
5.7
2
Shingles
Zostavax (live)
Myopericarditis
2024Q2
2.9
18.9
4.4
4
Shingles
Zostavax (live)
Cardiac stress test
2007Q2
2.8
27.3
4.4
4
Shingles
Zostavax (live)
Myocardial infarction
2018Q4
2.5
9.4
3.0
4
Shingles
Zostavax (live)
Atrial fibrillation
2009Q2
2.5
8.8
2.9
4
Shingles
Zostavax (live)
Cardiac failure
2019Q2
2.5
11.1
3.3
4
Shingles
Zoster (no brand)
Cardiac discomfort
2017Q1
2.4
Inf
9.0
2
Shingles
Zostavax (live)
Cardiac pacemaker insertion
2008Q1
2.4
Inf
5.0
2
Shingles
Zostavax (live)
Cerebral infarction
2019Q3
2.2
14.2
3.7
4
Top signals — neurological
Show code
neuro_top <- both |>filter(category =="Neurological") |>arrange(desc(eb05)) |>group_by(family) |>slice_head(n =12) |>ungroup()neuro_top |>select(family, vaccine, event, first_quarter, eb05, prr, ic, n_methods_flagged) |>gt() |>tab_header(title ="Top neurological signals",subtitle ="Statistics at first-detection quarter · top 12 per family by EB05") |>fmt_number(columns =c(eb05, prr, ic), decimals =1) |>data_color(columns = eb05,palette =c("white", "#08519c"),alpha =0.7) |>cols_label(first_quarter ="First detected") |>tab_options(table.font.size =12)
Top neurological signals
Statistics at first-detection quarter · top 12 per family by EB05
When did each signal first cross the consensus threshold? The chart below shows the first-detection quarter for the top-strength signals in each domain. Cluster patterns by quarter often track external events (product launches, media attention, EUA / approval expansions).
Show code
emergence_df <- both |>group_by(family) |>slice_max(eb05, n =30, with_ties =FALSE) |>ungroup() |>mutate(event_short =str_trunc(event, 48),first_date =as.Date(paste0(str_extract(first_quarter, "^\\d{4}"), "-",sprintf("%02d", (as.integer(str_extract(first_quarter, "\\d$")) -1L) *3L +1L),"-01" )) ) |>filter(!is.na(first_date))ggplot(emergence_df,aes(x = first_date,y =fct_reorder(event_short, first_date, .desc =TRUE),colour = vaccine,size = eb05)) +geom_point(alpha =0.8) +scale_colour_manual(values = VACCINE_COLOURS) +scale_size_continuous(range =c(2, 7), guide ="none") +scale_x_date(date_breaks ="6 months", date_labels ="%Y Q%q") +facet_grid(family ~ ., scales ="free_y", space ="free_y") +labs(title ="First quarter each signal crossed the consensus threshold",subtitle ="Point size = EB05 at first detection · top 30 per family by EB05",x =NULL, y =NULL, colour =NULL) +theme_minimal(base_size =11) +theme(axis.text.x =element_text(angle =30, hjust =1),axis.text.y =element_text(size =9),legend.position ="bottom",strip.background =element_rect(fill ="#f0f0f0", colour =NA),strip.text =element_text(face ="bold"))
Manufacturer comparison
Show code
mfr <- both |>count(family, vaccine, category, name ="signals") |>group_by(family, vaccine) |>mutate(total =sum(signals)) |>ungroup()ggplot(mfr, aes(x =reorder(vaccine, -total), y = signals, fill = category)) +geom_col() +scale_fill_manual(values =c("Cardiac"="#a50f15","Neurological"="#08519c","Thrombotic"="#54278f")) +facet_wrap(~ family, scales ="free_x") +labs(title ="Per-vaccine signal counts in the three pre-specified domains",subtitle ="Stacked bars: cardiac + neurological + thrombotic · first-detection count",x =NULL, y ="Distinct signals (first-detection quarter)") +theme_minimal(base_size =12) +theme(axis.text.x =element_text(angle =35, hjust =1),legend.title =element_blank())
Interpretation
A few patterns worth flagging:
COVID-19 cardiac signals are dominated by myocarditis/pericarditis in younger reporters and by mRNA products — consistent with the epidemiological signal that emerged in 2021 and that subsequently appeared in product labelling. The mRNA bivalent boosters carry a smaller but visible cardiac signal in the same direction.
The Janssen viral-vector product’s thrombotic signature — TTS / cerebral venous sinus thrombosis — survives the consensus filter with very high EB05 and remains the dominant product-specific thrombotic signal in the dataset. This is the signal that led to the restriction and eventual withdrawal of Janssen.
Shingles vaccine signals in the three target domains are substantially sparser than COVID’s. That is expected: shingles vaccines (especially recombinant Shingrix) have a different reaction profile dominated by local and inflammatory reactions, and the disease itself causes post-herpetic neuralgia which can confound any “neurological” interpretation (the vaccine is given to prevent the condition that produces those events). Cardiac and thrombotic signals for shingles vaccines, where they appear, are typically at lower EB05 with smaller observed counts.
Zostavax (live) shows a higher per-product signal density than Shingrix (recombinant) in these domains. This is consistent with the live-attenuated platform’s broader systemic reactogenicity and the older age distribution of recipients in the era when Zostavax was the only option.
Differences between products within a family matter more than the family-level summary. Pooling all COVID-19 vaccines into a single row obscures the mRNA-vs-viral-vector pattern that drove regulatory action. Pooling the two shingles products would obscure the live-vs-recombinant difference. The bubble chart above is the appropriate granularity.
Warning
Confounding by indication is a real risk for shingles signals. “Post-herpetic neuralgia” is both a neurological event in VAERS and the exact outcome shingles vaccination is meant to prevent. Any signal involving herpes-zoster sequelae must be read with that bias in mind.
Limitations
VAERS is a passive, voluntary reporting system. Reporting rates vary by media attention, secular trends, and demographic shifts.
Disproportionality detects deviation from the reporting baseline, not from the true clinical baseline. Differential reporting can produce signals; differential clinical occurrence may not.
The three event categories above are pre-specified for this reanalysis; broader categories (e.g., allergic, gastrointestinal, hepatic, renal) are excluded by design.
No demographic stratification (age, sex, dose number) is performed here. Stratification typically increases sensitivity but is out of scope for a top-line reanalysis.
All four methods rely on the same 2×2 contingency. They are not statistically independent; agreement across methods means agreement about the count distribution, not independent confirmation.
Replicable R
The full source of this article is the Quarto file shingles_vaccine_analysis.qmd in the globalpatientsafety repository. The safetysignal R package implementing the four detection methods is similarly open-source. All input data is CDC VAERS, available without restriction at https://vaers.hhs.gov/data.html.
