Two numbers nobody measured

The two claims

Claim one. Some large share of school shooters were taking antidepressants or other psychiatric medication. The share itself is unstable, moving between tellings from a specific percentage to “most” to “almost all”, which is the first sign that no arithmetic sits underneath it.

Claim two. Up to 4% of mothers with untreated postpartum psychosis commit infanticide. This one is remarkably stable between tellings, because it has a citation attached.

The two claims serve opposite arguments. The first is used to argue that psychiatric medication drives violence. The second is used to argue that leaving psychiatric illness untreated does. Both fail on the same structural point: neither number came from a study that measured what the claim asserts, and in both cases the source can be named exactly.

The federal record does not contain what people think it contains

Read this before the figures in this section, because it governs all of them. The numbers below are diagnosis counts and treatment-contact counts. They are not medication counts. Treatment contact means somebody saw a clinician, which is a different variable from being prescribed a drug, and a different variable again from taking one. Every report in the table says which of these it measured, and none of them measured the third.

The Secret Service is the most explicit. Its 2002 Safe School Initiative studied 37 incidents of targeted school violence involving 41 attackers between December 1974 and June 2000, and it states its own denominator convention in the text: “Unless indicated otherwise, when the finding pertains to total attackers all N’s are out of a total of 41.” On medication, in full [1]:

“The only information collected that would indicate whether attackers had been prescribed psychiatric medications concerned medication non-compliance (i.e., failure to take medication as prescribed). Ten percent of the attackers (n=4) were known to be non-compliant with prescribed psychiatric medications.”— US Secret Service and US Department of Education, Safe School Initiative Final Report, 2002 [1]

That sentence is the thesis of this page, written by the agency whose dataset people reach for. The 10% (n=4 of 41) is a non-compliance count. It cannot be inverted into a compliance rate, it cannot be used as a floor for prescription prevalence, and it cannot be used as a ceiling. A study that does not ask a question produces no denominator for it.

The same report’s headline mental-health finding, in its own words, is that “a history of having been the subject of a mental health evaluation, diagnosed with a mental disorder, or involved in substance abuse did not appear to be prevalent among attackers”, with 34% (n=14 of 41) ever evaluated and 17% (n=7 of 41) diagnosed before the attack [1].

Two of these are not shooter samples at all and should never be quoted as if they were. The 2021 report studies plots that were disrupted, so its subjects are students who did not attack. The 2023 report covers mass attacks in all public and semi-public spaces, not schools. Both are marked “LIMITED TO OPEN SOURCE INFORMATION” by their authors.

The FBI study is the cleanest negative in the set. Term counts over the full extracted text, re-run for this page on 2026-09-06: medication 0, medications 0, prescribed 0, antidepressant 0, SSRI 0, psychotropic 0, antipsychotic 0, pharmac 0. The single occurrence of “prescription” is in the glossary of concerning behaviours and refers to illicit use of prescription drugs [5]. The FBI also published something most downstream citations drop: alongside the 25% (n=16 of 63) it could verify as diagnosed, it published the 37% (n=23 of 63) where it could not determine whether a diagnosis had been given. And it attached this warning to its own finding:

“In light of the very high lifetime prevalence of the symptoms of mental illness among the U.S. population, formally diagnosed mental illness is not a very specific predictor of violence of any type, let alone targeted violence... In short, declarations that all active shooters must simply be mentally ill are misleading and unhelpful.”— FBI, Pre-Attack Behaviors of Active Shooters, 2018 [5]

The one database that does code medication cannot produce a rate

There is an exception, and it is instructive. The Violence Project mass shooter database codes prescribed psychiatric medication as an explicit variable, and has since its first version in 2019. It is the only source in this question that produces a real number. It also shows exactly why that number cannot be used the way people use it.

The caveat first, because it changes what the figure means. The variable takes two values: No evidence = 0 and Yes = 1. There is no unknown value on it. The project’s own methodology page, read live on 2026-09-06, defines the negative code this way [7]:

“Our researchers investigated this variable and found no evidence in available sources. This does not mean the answer is definitively ‘No’ — only that no supporting evidence was found in court records, media reports, interviews, or other available documentation.”— The Violence Project, methodology page, on the meaning of “No evidence of” [7]

The same page defines a separate “Unknown” code for variables where information could not be determined, and distinguishes it from “No evidence of” on exactly this ground [7]. The medication variable does not use it. Every case where medication status could not be established therefore sits in the denominator as a no.

With that established, here are the published figures. The project’s 2021 final report to the National Institute of Justice, on 168 mass shootings involving 172 mass shooters, 1966 to 2020, states: “23.3% had a known history of taking psychiatric medication (comparable to rates among the U.S. general population; Moore & Mattison, 2017)” [6]. The project’s earlier 2019 public report, on 171 shooters through 2019, gave 20%. Its co-founder told FactCheck.org in October 2025 that across 1966 to 2024 the figure is 24% for any prescribed psychiatric medication, and that 22 perpetrators, 11%, were known to have taken an SSRI [9].

Note the wording the researchers themselves use: a known history of. That is the correct description of what the variable holds, and it is a floor. Anyone quoting 23.3% as “the percentage of mass shooters on psychiatric medication” gets it wrong in both directions at once. It understates prescription prevalence, because cases where nothing was found are counted as noes. And it overstates certainty, because it presents a lower bound as a point estimate.

Under-ascertainment here is not random, either. Medication status lives in medical records that privacy law protects, so it surfaces only when a trial, a lawsuit or sustained journalism exposes it. Ascertainment is therefore systematically better for recent, high-casualty, heavily covered cases and worse for older and smaller ones. The database’s own published limitations say the sample is purposive with no comparison group, that data is missing due to privacy laws and varying media coverage, and that “the majority of people with mental disorders are not violent” [7].

Two research groups, three-fold apart

The Violence Project is not the only group to have tried. The Columbia Mass Murder Database, assembled at Columbia University and the New York State Psychiatric Institute, has coded overlapping populations from the same kind of public sources.

The caveat on the Columbia figures is substantial and comes first. We could not obtain the paper. It is a letter to the editor in Psychiatry Research (2025), PubMed carries no abstract for it, and the publisher blocks retrieval [10]. The figures below are as the database’s curator gave them to FactCheck.org in October 2025, which is a secondary source: about 4% of US mass shooters over roughly three decades with a history of ever taking antidepressants, and about 7% with a history of any psychotropic medication [9]. Treat them as unverified at the primary-source level. They are reported here for one purpose only, which is the size of the gap.

Two serious research groups, coding overlapping populations of American mass shooters from public records, arrive at roughly 7% and roughly 23% for any psychotropic medication. That is a three-fold disagreement on the same construct.

That spread is itself the answer to the question. When two competent groups measuring the same thing differ three-fold, the spread is a property of the instrument rather than of the world. It means the variable is not reliably ascertainable from public records, and that the answer you get depends on your inclusion criteria, your source set, and how you treat cases you could not resolve. Any real difference between shooters and the general population would have to be larger than that spread before either dataset could detect it, and nothing in this literature is.

A systematic review published in Frontiers in Psychiatry reached the same conclusion from the outside. Searching PubMed, PsycINFO and Web of Science from 1900 to 2024, screening 1,862 records and including 13 studies, its headline result is that “approximately 75% of the characteristics we aimed to analyze were not available in the reviewed literature”. Its extraction tables contain no psychotropic-medication row at all. The closest variable it could populate across studies was psychiatric treatment preceding the attack, at a mean of 38% across the five studies that reported it. The reviewers also warn that their 13 study samples overlap heavily, so pooling across published studies double-counts the same people [8].

Their explanation for the gap is structural rather than a criticism of anyone’s effort [8]:

“In the United States, juvenile criminal records are confidential by law in all states unless the minor is tried as an adult... This legal protection of juvenile records, while serving important privacy and rehabilitation purposes, creates structural obstacles to systematic data collection on school shooter characteristics... This legal barrier represents a fundamental challenge that cannot be overcome simply through improved research protocols.”— Minelli et al., systematic review of school shooter characteristics, 2025 [8]

There is one peer-reviewed paper designed specifically to answer this question: Hall and colleagues, “The myth of school shooters and psychotropic medications”, in Behavioral Sciences & the Law (2019), covering 49 school shooters from 2000 to 2017. Its stated conclusion is that “most school shooters were not previously treated with psychotropic medications, and even when they were, no direct or causal association was found”. We could not obtain the full text: the publisher blocks automated retrieval and there is no open-access version. So this page reports no percentage from it [13]. Its sample size, years and conclusion are verified. Its numerator is not, and will not appear here until somebody reads it.

The comparison trap, before the base rates

Any version of this argument ends up comparing a shooter figure to a population figure. Two structural problems make that comparison misleading unless they are handled explicitly, and they are the reason this page prints the caveats before the table.

Window mismatch. Forensic and open-source samples code “ever prescribed” or “history of”. Federal population surveys code the past 30 days or the past 12 months. Comparing a lifetime numerator against a 30-day denominator manufactures a large spurious excess. There is no federal “ever prescribed” denominator to compare against.

Frame mismatch. NHANES, the survey with the most rigorous measurement, explicitly excludes people who are homeless, incarcerated or hospital inpatients. Those are precisely the populations forensic samples over-represent, and NCHS states its estimates are conservative for that reason [14]. This biases any such comparison toward finding an apparent excess.

A third problem sits inside the population figures themselves. The three federal survey families disagree by roughly two-fold on the same construct in adolescents, because they use different recall windows, different respondents and different verification methods. The gaps below are measurement artefact, not real differences in drug exposure.

The Violence Project’s own lower bound of roughly 23% sits at or modestly above the adult 12-month rate of 16.7%. Columbia’s roughly 7% sits below it. Given the window mismatch, the frame mismatch and the three-fold spread between the two datasets, neither comparison supports a claim of dramatic over-representation, and neither is precise enough to support a claim of under-representation.

Statistic one: where “X% of school shooters were on antidepressants” came from

The claim has no dataset behind it. PolitiFact investigated its origin in August 2019 and traced it to three amplification points, none of which is a study [11].

• A 2018 publication by the Citizens Commission on Human Rights International, an advocacy group co-founded by the Church of Scientology in 1969, presenting psychiatric drugs as a common denominator in mass shootings. That organisation is named here once, because naming the origin of a claim is the reporting; it is not linked, and nothing it publishes is used as evidence anywhere on this site.
• An April 2013 post on a firearms blog listing shooters and alleged medications and asserting that prescription drugs are the single largest common factor.
• A list attributed to a firearms manufacturer that circulated widely on Facebook in early 2013.

Every one of these is a numerator with no denominator. Each is a list of cases in which a medication was reported: no sampling frame, no accounting for cases where medication was absent, and no accounting for the much larger set of cases where nobody knows. A list of N cases in which a factor was present cannot produce a percentage, because there is nothing to divide by. That is why the circulating figure varies so wildly between tellings. There is no underlying calculation to constrain it. PolitiFact could find no percentage of shooters on psychiatric medication in any of the sources it traced [11].

A note on the limits of this account. A fact-check article is a secondary source, and 2013-era social media cannot be independently audited now. The provenance chain is reported here at moderate confidence. What is not in doubt is the structural point, which does not depend on the provenance at all: none of those sources contains a denominator.

The label check, since it travels with the claim

A related claim travels alongside: that SSRIs carry a black box warning for homicidal ideation. We checked the FDA structured product label for fluoxetine directly through the openFDA drug label API on 2026-09-06, across the five label records the API returned [12].

• The boxed warning is titled “WARNING: SUICIDAL THOUGHTS AND BEHAVIORS” in every record, and concerns suicidality.
• The string homicid appears zero times anywhere in any of the records.
• Warnings and Precautions 5.1 does list, verbatim, “anxiety, agitation, panic attacks, insomnia, irritability, hostility, aggressiveness, impulsivity, akathisia (psychomotor restlessness), hypomania, and mania, have been reported in adult and pediatric patients being treated with antidepressants”. That is an instruction to monitor. It is not a causal finding, and it is not in the boxed warning.
• “Violent behaviors” appears once, in the postmarketing list of spontaneously reported events, carrying the label’s own footnote: “These terms represent serious adverse events, but do not meet the definition for adverse drug reactions. They are included here because of their seriousness.”

A spontaneous postmarketing report is somebody telling the FDA that something happened while a person was taking a drug. It has no denominator either.

Statistic two: the same failure, pointing the other way

The claim, in its usual phrasing: up to 4% of mothers with untreated postpartum psychosis will commit infanticide, and 5% will die by suicide. It has propagated through the peer-reviewed literature rather than through social media, which is precisely why it has been so durable.

It traces to one paper: Davidson and Robertson, “A follow-up study of post partum illness, 1946–1978”, in Acta Psychiatrica Scandinavica (1985). We read its abstract verbatim on 2026-09-06 [16]. Three properties of that source break the claim that rests on it.

The route from that paper into modern use is also traceable. Friedman and Resnick’s 2007 narrative review in World Psychiatry carries the sentence, and the reference it attaches to it is Altshuler, Hendrick and Cohen (1998) in the Journal of Clinical Psychiatry [18]. We read that paper’s abstract on 2026-09-06: its subject is the course of mood and anxiety disorders during pregnancy and the postpartum period, and the risk of relapse after medication discontinuation. It reports no infanticide rate [19]. The figure does not appear in the source it is cited to.

Ian Brockington, reviewing suicide and filicide in postpartum psychosis in Archives of Women’s Mental Health in 2017, addresses the American usage of the 4% figure directly and calls it an error traced to a misreading, noting that none of the filicides in this literature was committed by a mother with an acute psychosis without depression [17]. That correction is from the body of a paywalled paper, read in the research memo behind this page rather than re-verified here.

His own published figures deserve the same treatment as the number he is correcting, and he supplies the caveat himself. Reviewing a literature of over 4,000 childbearing psychoses plus a personal series of 321, his abstract reports a filicide rate of 4.5% in depressive psychoses and under 1% in episodes without overt depression, adding that “some of these appear to be accidental, without intent to kill” [17]. The denominator there is the published literature plus one clinician’s caseload, not a population of women with postpartum psychosis. Published case reports over-select the severe and the remarkable, because those are the cases clinicians write up and journals accept. No population-based study has estimated a filicide rate among women with postpartum psychosis, so no figure in this paragraph, or in the claim it corrects, is a rate that applies to any individual.

One more thing about the filicide literature is worth stating plainly, because it is the same negative as the shooter literature. No study codes the psychiatric medication status of a filicide cohort. A full-text term search of the largest population-based study in existence, Flynn and colleagues’ 2013 analysis of 6,144 homicide convictions in England and Wales yielding 297 filicide perpetrators, returns zero hits for medication, prescrib, psychotropic, adheren and complian [20]. The literature codes diagnosis, psychotic status at the time of the offence, and contact with mental health services. It does not code what drug anybody was taking.

That same paper is worth citing for its denominator discipline, which is the opposite of the Violence Project’s. The widely quoted “37% were mentally ill at the time of the offence” has a denominator of 173, not 297: psychiatric reports existed in only 195 of 297 cases, and the authors coded the absence of a report as missing rather than as an absence of illness. They also state, of their own study, “we are unable to draw aetiological conclusions because we do not have a comparison group” [20].

What is actually known

The caveat first, and it applies to everything in this section. All of it is observational register data. None of it is about mass shooting or filicide, because those events are far too rare to appear in cohorts of this size. Where a study uses a within-individual design, each person acts as their own control, which removes every confounder that stays constant for that person over the follow-up, and removes none that varies over time. Symptom severity varies over time.

Antipsychotics and mood stabilisers. Swedish national registers, 82,647 patients prescribed antipsychotics or mood stabilisers, convictions 2006 to 2009, analysed within-individual. Compared with periods when the same people were not on medication, violent crime fell 45% during antipsychotic treatment (HR 0.55, 95% CI 0.47 to 0.64) and 24% during mood stabiliser treatment (HR 0.76, 0.62 to 0.93), with a dose-response relationship and a larger reduction at higher doses. The mood stabiliser reduction was present only in patients with bipolar disorder [22]. These are the drug classes actually used in postpartum psychosis and bipolar disorder.

SSRIs. The same research group, larger cohort, opposite direction and much smaller effect. Among 785,337 Swedish individuals aged 15 to 60 dispensed an SSRI between 2006 and 2013, with 32,203 violent crimes over 5,707,293 person-years, the within-individual hazard ratio was 1.26 (1.19 to 1.34), elevated during treatment and for up to 12 weeks after discontinuation. Roughly 97% of that cohort committed no violent crime at all, and only 2.6% of it was informative to the analysis [24]. An earlier study in the same registers found the association significant only in ages 15 to 24 (HR 1.43, 1.19 to 1.73) and non-significant in every older band [23]. The outcome in both is violent crime convictions of any kind in Sweden, mostly assault and unlawful threats.

The authors state the limit of their own finding [24]:

“Key to interpreting our findings is the acknowledgement that any pharmacoepidemiological study is likely to be subject to confounding by indication. Mood disorders, the main indication for SSRI treatment, are associated with violent crime... suggesting that onset, or worsening, of the underlying disorder that led to the indication for treatment could be driving the associations in our study.”— Lagerberg et al., European Neuropsychopharmacology, 2020 [24]

The elevated hazard for up to 12 weeks after stopping is often quoted as evidence of a withdrawal effect. The authors offer the competing explanation in the same paragraph: people who stop medication are frequently people whose illness is relapsing [24]. Seena Fazel, senior author on both Swedish studies, told FactCheck.org that this work “will not be directly relevant to mass shootings”, and that he was not aware of the question having been studied in a way that would allow definitive conclusions [9].

The clearest picture of confounding by indication anyone has published

This table is not about medication. That is exactly why it is here. It is from the Swedish register study of women delivering a first live infant between 1987 and 2001, split by whether the woman had ever been hospitalised for a psychiatric reason before delivery [21]. The denominator throughout is first births.

That is roughly a 230-fold gradient, and it is driven entirely by whether a woman had previously been identified and hospitalised as psychiatrically ill. Having been hospitalised for a psychiatric reason is also one of the strongest markers of having been prescribed a psychiatric medication.

This is the whole problem in one table. A psychiatric prescription is not a random exposure. It is issued because a clinician assessed somebody as unwell, so the treated group differs from the untreated group on exactly the variable that independently predicts the outcome. Any comparison of medicated to unmedicated postpartum women runs across this gradient. Comparing prescription rates among perpetrators to prescription rates in the general population compares people identified as psychiatrically ill to everybody, and would show an apparent excess even if the drug were pharmacologically inert. The co-founder of the Violence Project made the same point about his own database: “people in acute crisis are more likely to be prescribed medication, so simple yes/no comparisons overstate any drug–violence link” [9].

The direction of the bias is not even fixed. Treatment marks illness severity, which raises risk. Treatment reduces symptoms, which lowers it. And being prescribed anything means somebody reached a clinician and was noticed, which may itself be protective. Those three effects run in different directions and no descriptive comparison of rates can separate them.

The study that would answer this, and why one half of it cannot be built

The design is not mysterious. It has four requirements.

1. Exposure from dispensing records, meaning a pharmacy or prescription monitoring database, rather than from news reports and court files. This is the step that removes the ascertainment problem entirely, because a dispensing record exists whether or not a journalist found it.
2. Controls matched on diagnosis and severity, never on age and sex alone. The comparison has to be treated versus untreated people with the same condition. Against a general-population comparator the study answers nothing, for the reason in the table above.
3. A within-individual or self-controlled design, so that each person is their own control and every time-invariant confounder disappears, combined with an active-comparator new-user design in which people starting one drug are compared with people starting a different drug for the same indication.
4. A national register linkage joining births, dispensings, hospital records, cause of death and criminal convictions. The Nordic countries have this. The United States does not: there is no national prescription register linked to criminal records.

For violent crime broadly, this design is feasible and has been executed twice, in Sweden [22][24]. For the two outcomes people actually argue about, it cannot be powered, and the arithmetic is not close.

The study that estimated a hazard ratio of 1.26 for violent crime of any kind used 785,337 exposed individuals and 32,203 events [24]. For comparison, the Violence Project’s database, built on the Congressional Research Service definition of a mass shooting, contains 172 mass shooters across the 55 years from 1966 to 2020 [6]. Filicide runs at roughly 5.09 per 100,000 live births in Finnish national data covering 200 cases over 25 years [26]. Sweden’s complete national autopsy database, linked to the national patient register, yielded 71 filicide victims aged 0 to 14 across 19 years, matched to 355 population controls [27]. England and Wales produced 297 filicide convictions in 10 years from 6,144 homicide convictions [20].

An entire country’s registry, complete, over two decades, yields dozens of cases. The design that would work has an outcome too rare to power it. Pooling every Nordic register would yield a few hundred cases over decades, which is enough to describe a population and not enough to estimate a drug effect with any useful precision. The honest alternative is to substitute a more frequent outcome on the same causal pathway, such as psychiatric emergency or involuntary admission, and to accept that any inference about filicide from it is extrapolation.

The nearest US instrument is the National Violent Death Reporting System, which draws on death certificates, coroner and medical examiner reports including toxicology, and law enforcement reports. Two limits are structural. It is a death reporting system, so a perpetrator who survives is captured only as a circumstance in a victim’s record rather than as a subject with her own history. And medical examiner toxicology panels are not standardised for psychotropics, so the absence of a finding is not the absence of exposure. This paragraph is reported at moderate confidence: it rests on CDC’s summary description of the system, because the coding manual PDF returns a 404.

How to recognise a numerator-only statistic

Both numbers on this page failed the same way, and the failure is recognisable without any subject knowledge. Four questions, in order.

1. Where is the denominator, and is it stated? A percentage is a division. If the source is a list of cases in which something was present, there is nothing to divide by, and the percentage was assembled rather than computed. This is what happened to statistic one.
2. Did the source code the variable at all? Diagnosis is not treatment. Treatment contact is not a prescription. A prescription is not an ingested drug. Reports that coded one of these are routinely quoted as though they coded another. The Secret Service wrote into its own report which of them its instrument captured, which is why that report is the one nobody can misread.
3. How are unknowns handled? This is the one people miss. A variable coded only yes-or-no, with no unknown value, silently converts every unresolved case into a no and turns a floor into what looks like a rate. Compare two conventions on this page: the Violence Project collapses “not found” into the negative, while Flynn and colleagues coded a missing psychiatric report as missing and published the smaller denominator that resulted.
4. Does every word of the claim appear in the source, including the modifiers? “Untreated” carried the entire meaning of statistic two, and it appears nowhere in the 1985 paper the claim rests on. Check the adjectives, not just the number.

A fifth question applies whenever the claim compares two groups. Do the groups differ on the thing being measured, or on the reason they were measured? Everyone holding a psychiatric prescription was judged by a clinician to be unwell. That fact is what such a comparison is actually measuring.

What this page does not claim

• That psychiatric medication does not cause violence. That is a causal claim and this page has no basis for it either. The point is that the question is open and the circulating numbers do not bear on it.
• That anyone lied. A figure that does not appear in the source it cites is a fact about a citation. Statistic two propagated through peer review for decades, which is usually how a misreading survives: in good faith, by repetition.
• That the researchers whose datasets appear here did their work badly. The Violence Project publishes its coding conventions, its limitations and its own warning that most people with mental disorders are not violent. The FBI published its undeterminable share alongside its positive share. The Secret Service wrote down what it failed to collect. The failure lies in what downstream citation does to careful work.
• Anything about any individual. No person is named on this page and no person is linked to a medication.
• Anything about you. This is a page about two statistics.

What we could not verify

Published here because a page about broken citation has to show its own gaps. Each of these is a source we tried to obtain and could not, and each is the reason a figure is absent above.

• Hall et al. 2019, the only peer-reviewed study designed to answer the school shooter question. The publisher blocks automated retrieval and no open-access version exists. Its sample and conclusion are verified; its medication percentage is not, so no percentage from it appears on this page [13].
• Girgis et al. 2025, the Columbia group’s medication paper. No abstract on PubMed, publisher returns 403, no open-access version. The 4% and 7% figures on this page are second hand from a fact-check article and are labelled as such wherever they appear [10].
• Moore & Mattison 2017, the base rate the Violence Project cites when it calls its own medication figure “comparable to rates among the U.S. general population”. A paywalled research letter with no abstract. We could not read the comparison figure they relied on, so this page uses the federal survey estimates in the base-rate table instead.
• Brockington 2017 body text, source of the correction quoted above. The abstract was read directly on 2026-09-06; the correction itself is from the paywalled full text as read in the research memo behind this page [17].
• The Violence Project codebook tab. The methodology definitions quoted here were read live on 2026-09-06. The binary coding of the medication variable itself comes from the codebook distributed inside the public spreadsheet, which is served through a Google Drive folder we could not enumerate without a browser on that date [7].
• NCHS Data Briefs 377 and 135. cdc.gov returns HTTP 403 to non-browser clients, so the base-rate table was not re-fetched on 2026-09-06; those figures were read directly from the PDFs in the research memo behind this page [14].

How this page was built

It began as an internal verification memo whose rule was that no number would be reported that had not been read in a source, with a confidence label attached to every claim. Before publication, the load-bearing items were checked again directly against primaries: the Safe School Initiative medication passage and denominator convention were re-extracted from the PDF; the medication vocabulary counts in the FBI 2018 and Secret Service 2019 reports were re-run over the extracted text; the National Institute of Justice report’s 23.3% sentence was read verbatim; the Violence Project’s coding definitions were read from its live methodology page; the fluoxetine label was pulled from the openFDA API and the term counts re-run; and the PubMed records and abstracts for Davidson & Robertson 1985, Harlow 2007, Fazel 2014, Brockington 2017, Altshuler 1998 and Minelli 2025 were read directly. Advocacy publications are never used as sources here. The single mention of one on this page is as the traced origin of a claim, sourced to a fact-checking organisation’s reporting, with no link and no reproduction of anything it says.

A page like this invites one more provenance question: who pays us to write it. The answer is published in the same spirit as everything above. Who funds veisund states the revenue model and the money we do not accept, with the checkable parts separated from the policy parts.