| CONNOR'S HEALTH NOTES How a Drug Death Gets Counted Connor Hill · InsightfulWord · August 18, 2026  Numbers describing deaths caused by medication circulate constantly, they vary by an order of magnitude between sources, and almost nobody who repeats one can say where it came from. The variation is not because anyone is lying. It is because at least four entirely different systems produce such figures, they measure different things, and the results are not interchangeable. Understanding which system produced a number is the difference between reading a fact and reading an artifact. The first is spontaneous adverse event reporting. The Food and Drug Administration maintains a database into which manufacturers, clinicians and members of the public submit reports of problems experienced by someone taking a drug. The agency is explicit about what the resulting records are and are not. Reports are not verified. The existence of a report does not establish that the drug caused the event. Duplicate submissions occur. And there is no denominator — the system records reports without knowing how many people took the drug — so no rate can be calculated from it at all. That last point is the one that disposes of most claims built on this data. A count of reports, divided by nothing, cannot be compared with a count of reports about a different drug taken by a different number of people. A widely prescribed medication will generate more reports than a rarely prescribed one for reasons that have nothing to do with either one's safety. The second is death certificate coding. When a person dies, a certifier records an underlying cause and contributing conditions, which are translated into standardized codes and aggregated into national mortality statistics. Adverse effects of therapeutic drugs have their own codes, but recording one requires that the certifier recognized the connection and chose to record it — and certifiers are frequently working without the full medication history. The third is chart review research. Investigators read a sample of hospital records and judge, case by case, whether a death was attributable to a medication. This produces the most careful estimates and the smallest samples, and it depends on reviewer judgment, which is why studies using the same method reach different figures. The fourth is modeling. Take a rate from a chart review, apply it to national admission or prescription volumes, and report a projection. This produces the largest and most quotable numbers, and each of its inputs carries the uncertainty of everything upstream. What follows is what each system can honestly support, why attribution is genuinely hard, and what the question is when the number matters to an individual rather than a population. Why Attribution Is Difficult Even Honestly The hard part is not counting. It is deciding what caused what, in people who typically had several conditions and several medications at once. Consider an older adult who takes a medication, falls, fractures a hip, has surgery, develops pneumonia during recovery, and dies. Did the drug kill them? A reasonable case exists for recording the cause as pneumonia, as the fracture, as the fall, or as the medication that contributed to the fall. Different certifiers make different choices, and all of them are defensible. Now add the counterfactual, which is the part that never appears in a headline. The medication was prescribed for something. The relevant comparison is not between the patient's outcome and a hypothetical patient who took nothing and was otherwise identical. It is between the outcome with treatment and the outcome without it, in a person who had the condition being treated. Attribution figures count events on one side of that comparison and are silent about the other. There is also confounding by indication, which runs through this entire literature. People prescribed a medication differ systematically from people not prescribed it — they have the condition the drug treats, which is itself associated with worse outcomes. A raw association between taking a drug and dying is contaminated by the reason the drug was given, and separating the two requires methods that observational death counts do not employ. | 📈 Number of the Day No denominator The FDA's adverse event reporting system records submissions without knowing how many people took the drug, contains unverified and duplicate reports, and by the agency's own statement does not establish that a drug caused the reported event. No rate can be computed from it, and counts between drugs cannot be compared. Source: U.S. Food and Drug Administration, FAERS public dashboard documentation. | | Support or oppose: should adverse event data be public in this form? The reporting database is open to anyone, which means unverified reports of serious harm are searchable by the public and routinely quoted as if they were confirmed cases. Supporters of open access argue that transparency is the point, that researchers and journalists find genuine signals in it, and that a regulator holding safety data privately is worse. Critics answer that publishing unverified counts without a denominator predictably produces misleading claims, frightens patients away from treatment, and that the same data could be released to researchers with the caveats enforced. Should it stay open in its current form? Hit reply — one line is enough. | What the Systems Are Good For Dismissing all of this would be as wrong as over-reading it, and each system has a job it does well. Spontaneous reporting is a signal detector, and it is genuinely valuable in that role. Its purpose is to surface rare adverse effects that clinical trials — with a few thousand participants over a few years — could never detect. A cluster of unusual reports about a specific drug is a reason to investigate, and several important safety findings originated exactly this way. It generates hypotheses. It does not test them. Death certificate data is good at trends over time within a consistent coding framework, which is what it was designed for. Comparing this year with last year using the same codes is informative. Reading an absolute total as a measure of true incidence is not. Chart review produces the most defensible individual attributions and the least generalizable totals, because samples are small and settings differ. Modeled projections are useful for policy discussion about the scale of a problem and are the least suited to any individual conclusion, because they aggregate uncertainty at every step. The error in almost every viral claim is using one system's output for another system's question — most often taking a report count or a modelled projection and presenting it as a verified death toll. | The question that actually applies to a person Population attribution figures do not answer the question an individual is asking, which is whether a specific medication is doing more good than harm for them specifically. That question has a different shape and a different set of inputs: what condition is being treated and how severe is it, what is the individual's absolute risk from that condition, what is the absolute risk of the relevant adverse effect at their age and with their other conditions and medications, and does an alternative exist within the same therapeutic class with a different profile. All four are answerable at an appointment, from the patient's own record, in a way no national statistic can be. And the standing caution applies without exception: stopping a prescribed medication on the basis of any number found online, without medical advice, is dangerous for several common classes of drug, and the risk of abrupt discontinuation is frequently larger than the risk being avoided. | Why the Reporting System Exists at All It would be easy to read the limitations above as an argument that spontaneous reporting is worthless. It is not, and understanding why explains the whole architecture of drug safety. A clinical trial that supports approval typically enrols a few thousand participants and follows them for a defined period. That design can detect adverse effects occurring in perhaps one participant in a few hundred. It cannot detect an effect occurring in one person in fifty thousand, because the trial does not contain fifty thousand people. Once a medication is prescribed to millions, those rare effects begin to occur. There is no practical way to run a trial large enough to have found them in advance, which means the only mechanism available is to watch what happens after approval and collect reports. That is what the reporting system is: an early warning network, deliberately designed to be sensitive rather than specific. It is supposed to produce false alarms, because a system tuned to avoid them would miss the genuine signals it exists to catch. The trade-off is that its raw output is noisy by design and was never intended to be read as a count of confirmed harms. The agency's own caveats say so explicitly. The failure mode is not the system. It is treating a sensitive detector's raw output as a verdict. What Happens When a Signal Is Real The step that follows detection is the one that almost never appears in content built on these numbers, and it is where the system does its actual work. A cluster of reports triggers evaluation rather than action. Investigators examine whether the reported events exceed what would be expected in that population without the drug, whether a plausible biological mechanism exists, whether the timing fits, and whether the pattern appears in other data sources such as insurance claims or electronic health records — which, unlike the reporting system, have denominators. If the signal survives that scrutiny, the responses are graduated. A label may be revised to add a warning or a contraindication. Prescribing may be restricted to certain patients or certain settings. A study may be required as a condition of continued marketing. In the most serious cases a product is withdrawn. Those actions are public and dated. Label changes are published, safety communications are issued, and withdrawals are announced. Anyone wanting to know whether a concern about a particular medication has been examined can look for the regulatory record rather than the report count — and the record is the informative document, because it represents a conclusion rather than an allegation. Where the Legitimate Concern Sits None of this means medication harm is a manufactured worry. It is real, it is substantial, and the serious version of the concern is more specific than any total. Adverse drug events are a well-documented cause of hospital admission in older adults, and a substantial share of them are considered preventable. That word is doing important work: preventable means the harm arose from a dose, an interaction, a duration or a prescription that a review would have caught, rather than from an unavoidable property of the drug. The concentration is also documented. A relatively small number of drug classes account for a large share of serious adverse events in older adults, and the mechanisms are known — bleeding, low blood sugar, sedation and falls, kidney injury, electrolyte disturbance. These are monitorable, which is precisely why they are described as preventable. That framing supports a specific and achievable response: monitoring, periodic review of whether each prescription is still indicated, and attention to the interactions that arise as a list lengthens. It does not support abandoning treatment, and the people who study this most closely are the least likely to recommend that. The distinction between the serious version and the sensational version is visible in what each asks the reader to do. One says have the list reviewed. The other says be afraid of the list. | The bill, not the debate Numbers about medication deaths circulate detached from the systems that produced them, and the systems disagree by an order of magnitude because they are measuring different things. The figure that applies to a household is not a national total but an individual balance, and it is computable from a person's own record by someone with access to it. Has anyone ever gone through your medication list and told you what each item is protecting you from, and at what cost? Connor Hill reads every reply. | | Connor Hill · InsightfulWord | | 
