Methods and data sources

2×2 table

For an exposure (a drug or a class) and a MedDRA preferred term, reports in the openFDA FAERS extract form:

The proportional reporting ratio is (a / (a+b)) / (c / (c+d)). The reporting odds ratio is (a/b) / (c/d). 95% intervals use the normal approximation on the log scale with 1.96 as the quantile. If any cell is zero, every cell is increased by 0.5 before the ratios and chi-square are calculated (Haldane–Anscombe). The displayed report count stays the observed count.

An Evans signal requires at least 3 reports in cell a, a PRR of at least 2, and a chi-square of at least 4 (Evans, Waller, and Davis, Pharmacoepidemiology and Drug Safety, 2001). A ROR signal requires at least 3 reports in cell a and a lower 95% bound above 1.

What “exposed” means here

openFDA cannot require the ingredient and drugcharacterization = 1 to sit on the same drug element. Counts are therefore ingredient mentions: the report lists the generic name and lists the reaction. Concomitant mentions dilute signals toward the null, so a signal that still appears has survived a conservative case definition. Class totals use one OR query, so a report naming two drugs in the class is counted once.

The database-wide comparator is the rest of the openFDA extract (about the public FAERS file, not a treatment registry). There is no exposure denominator, so these ratios are not incidences.

Why JAK events are compared with TNF inhibitors

The 2021 FDA boxed warning for tofacitinib, upadacitinib, and baricitinib describes major adverse cardiovascular events, thrombosis, malignancy, and death. ORAL Surveillance compared tofacitinib with TNF inhibitors in rheumatoid arthritis. A FAERS-wide background is full of venous thrombosis from hospitalized patients, which can hide that contrast. The prespecified JAK checks therefore use TNF inhibitors as the active comparator. The same reactions are also scored against the whole extract and shown as a sensitivity result in the validation file; that sensitivity result is not the pass/fail rule.

Tuberculosis with TNF inhibitors, and gastrointestinal bleeding or ulcer with NSAIDs, are large enough historically to test against the rest of FAERS.

A known signal is recovered when the prespecified 2×2 table has at least 3 reports in cell a and either an Evans signal (PRR ≥ 2 and χ² ≥ 4) or a ROR whose 95% lower bound is above 1. JAK thrombosis and MACE are compared with TNF inhibitors. TNF–tuberculosis and NSAID–gastrointestinal bleeding are compared with the rest of the openFDA FAERS extract.

Weekly spikes

Publication counts, trial starts, and dated FDA notices are bucketed by complete Monday–Sunday weeks. The week that contains the refresh day is left out, because it is still open. A point is a spike when the trailing baseline has at least four earlier weeks, the z-score is at least 2 (sample standard deviation), and the count is at least 1.5 times the baseline mean. A flat baseline has no z-score; a count at least 1.5 times that flat mean is still flagged. This is a screen for the weekly job, not a seasonal model.

Sources and licenses

data.cdc.gov does not publish an IBD dataset comparable to Chronic Disease Indicators. Arthritis is taken from hksd-2xuw. Crohn’s disease and ulcerative colitis are followed with PubMed and ClinicalTrials.gov. openFDA’s own terms say not to rely on it for medical care; the extract can lag and counts are unvalidated.

Limitations

• Spontaneous reports under-count events, and publicity or litigation can suddenly increase them.
• Duplicate reports are not removed beyond whatever openFDA has already done.
• Indication confounding remains. People on these drugs are not the general population.
• Reports that mention both a JAK inhibitor and a TNF inhibitor stay in both sides of the active comparison.
• Product-use terms such as “drug ineffective” are computed when openFDA returns them and hidden in the default table.
• PubMed lists show titles and links only. Abstracts are not stored or shown. NCBI disclaimer.