← All Japan heatmap pages

Pfizer — Jan 2022 (2022-01) — cumulative from Jan 2021

VAERS Japan (upper(splttype) LIKE 'JP%') · cumulative: all reports received Jan 2021 through this month (recvdate) · treemaps after BMJ 2026;394:e100806: tiles grouped by MedDRA SOC, sized by distinct reports (a), coloured by score (0–5).

Gradient view

PRR (Proportional Reporting Ratio) · MGPS (Multi-item Gamma Poisson Shrinker; score = EBGM) · RGPS (Regression-Adjusted Gamma Poisson Shrinker; score = ERAM) — side by side, in that order.

Bootstrap colors

The same three algorithms — PRR · MGPS (EBGM) · RGPS (ERAM) — side by side, in that order, in Bootstrap colors.

Top signals

All PTs with at least 5 reports for Pfizer through Jan 2022 (2022-01) — cumulative from Jan 2021 (cumulative), as a DataTable with a Custom Search Builder (including the custom ≥ 2, ≥ 5 and < 1 numeric conditions — e.g. build EB05 ≥ 2 AND ERAM ≥ 2 via the built-in conditions, or pick a score column and apply ≥ 2 (custom)). Sorted by PRR by default; click any symptom name in the Preferred term column to pin its details in the toolbar above and open its masking-effect explainer (its tile is marked with a white dashed outline in every panel). Hover any tile above for detail. Masking effect = ERAM/EBGM − 1 (Harpaz et al. 2022): how much the regression adjustment raises the score over MGPS, e.g. +102% ≈ a 2-fold rise (its SearchBuilder value is the underlying ratio minus 1, so “greater than 1” finds effects above +100%).

Methods & assumptions

• Japan = upper(splttype) LIKE 'JP%' (exact ='JP' returns 0 rows; splttype is a country-prefixed manufacturer report number). Only Pfizer and Moderna pages are generated. Comparator = all other JP vaccine reports in this cumulative period (i.e. Japan reports not in the Pfizer group, including other brands).
• Cumulative scope: each page uses all reports with recvdate from Jan 2021 through the named month (e.g. “2021-07” = Jan–Jul 2021), matching the article’s cumulative analyses — not just that calendar month.
• Algorithms follow Harpaz et al. (2022), “Signaling COVID-19 Vaccine Adverse Events”, Drug Safety 45:765–780 (the analysis this site replicates), as run there in Oracle Empirica Signal. PRR = [a/(a+b)]/[c/(c+d)], unstratified. MGPS = DuMouchel’s Multi-item Gamma Poisson Shrinker: a two-component gamma-mixture prior fitted by empirical-Bayes maximum likelihood to all PTs for the brand in that cumulative run, expected counts stratified (Mantel–Haenszel) by sex × age (age groups 0–1, 2–4, 5–12, 13–16, 17–45, 46–75, 76–85, >85, unknown — the article’s categories); the score is the posterior geometric mean EBGM (EB05/EB95 in tooltips and jp_scores.parquet). RGPS = DuMouchel & Harpaz’s Regression-Adjusted GPS, in its three published steps: (1) age/sex strata grouped by event rate and other vaccine products (type × manufacturer) automatically selected as masking predictors, with an empirical-Bayes penalised logistic-regression fit; (2) regression-adjusted expected counts; (3) a single gamma prior fitted across products giving the score as the posterior arithmetic mean ERAM (ER05/ER95 likewise). RGPS therefore corrects masking by strongly-associated co-reported products, which plain MGPS cannot — it need not always exceed EBGM. Scores capped at 5 for colour only.
• Adaptations vs. the article (unavoidable here): the article ran cumulative all-VAERS data in proprietary Empirica 9.1; these pages are cumulative Japan-only runs, so priors are fitted within each cumulative month × brand “database run”, products are VAERS vaccine type × manufacturer, each report is assigned its priority vaccine (Pfizer > Moderna > other), and the extended-logistic link is used at α = ½ (ordinary logistic — the case the RGPS white paper singles out; Empirica’s α-optimisation is unpublished).
• Colours: The gradient view uses a gradient colourscale (#00aa44 at 0 → #226622 at 1 → #111111/#222222 at ~1.75–2.5 → #661111 at 3.25 → #cc2222 at 4 → #ff0000 at 5). The Bootstrap colors view uses hard bins: <1 light gray #e9ecef, 1–1.5 blue #0d6efd, 1.5–2 green #198754, 2–2.5 orange #fd7e14, 2.5–3 red #dc3545, >3 dark brown #5c3317 (no blending).
• Navigation: click a tile to select it and pin its details (reports a, expected counts, PRR, EBGM with EB05–EB95, ERAM with ER05–ER95) in the toolbar bar, with the tile marked by a white dashed outline in every panel; double-click a tile — or the “Zoom into SOC” button in the details bar — to zoom into its SOC in all panels (Plotly-style); the “↑ Up one level” button inside each heatmap (top-right) returns to the outermost all-SOC view. Hovering shows the same details in a popup that stays inside the screen edge. Symptom search resets to the outermost view, then highlights the matching PT in every PRR/MGPS/RGPS panel.
• Outcome pseudo-terms: the five VAERS outcome fields died, hospital, l_threat, disable and er_ed_visit are treated as if they had been added to the symptoms file — each report with the field = ‘Y’ counts once for that term (raw field names, so they never merge with real MedDRA terms such as “Death”), all five are grouped under Genrl, and they receive full PRR / MGPS (EBGM) / RGPS (ERAM) scores like any other PT, including in the MGPS prior fit and jp_scores.parquet. One report can carry several outcome terms as well as its real symptoms, so outcome tiles overlap. These are report-level outcome fields, not MedDRA preferred terms — interpret their disproportionality scores accordingly (dedicated Outcomes pages are still generated under the hood but are not linked from the listing for now).
• PT→SOC is a keyword heuristic (no official MedDRA map in VAERS). Tiles: a ≥ 2, at most 900 largest PTs. Disproportionality is hypothesis-generating only, not causality.
• All underlying data are stored as parquet in data/ (jp_cells, jp_totals, jp_scores, jp_pages). Pages are regenerated from these datasets by the PRR/MGPS/RGPS heatmap pipeline described above.