[Warning: this is a linkpost, and the footnote links will teleport you to my personal website.]
The US homicide rate hit 10.2 per 100,000 in 1980, fell to 4.4 by 2014, then spiked to 6.5 in 2020. This looks like a story about violence rising, falling, and rising again. The actual story is simpler and worse: violence tripled in the 1960s and 70s. After adjusting for medical improvement, there is no clear statistical signal that it ever came back down. What changed was not the violence but how often it killed people.
Deaths vs. attacks
The homicide rate measures deaths. Deaths from gunshot wounds depend on two things: how many people get shot, and how many of those people die. Medicine affects the second thing. The homicide rate doesn't separate them.
If you get shot in the abdomen today, you have about a 92% chance of surviving. In 1960, you had about a 70% chance. In 1900, about 28%. The prognoses for other penetrating trauma wounds also improved, though this varies by wound location: head shots remain almost uniformly fatal regardless of era, while extremity wounds were highly survivable even a century ago. Abdominal wounds are a central case of an injury that kills you without good care and doesn't with it.
This problem was identified in the academic literature by Harris et al. (2002), who estimated that without post-1960 medical improvements, the US would have had 50,000–115,000 homicides per year in the late 1990s instead of the observed 15,000–20,000. But Harris presented this as a single-year counterfactual, as though merely to say that things were worse than they looked in 1999.
We have two standard measures of the trend in serious violence: the homicide rate and the aggravated assault rate. Both are distorted. The homicide rate is distorted by improving medicine, which makes the same number of attacks produce fewer deaths over time. The aggravated assault rate is distorted by changing reporting and classification practices: expanded reporting and definitional drift inflated the count through the early 1990s, while CompStat-era downclassification pressure may have deflated it afterward. This post constructs a third trend line by dividing the medical improvement out of the homicide rate. The resulting point estimates aren't very meaningful (this is not a literal count of attempted murders). But the trend is less distorted than either alternative, because it starts from hard death counts and adjusts using clinical data that doesn't depend on crime reporting.
The vanishing decline
I put together some interactive charts illustrating this adjusted trend line. The adjusted trend tells a different story than the raw homicide rate: the 1933–1960 decline largely disappears, and the 1965–1980 increase is steeper than the raw numbers suggest.
The raw homicide rate fell 57% from the 1991 peak to the 2014 trough. After adjustment, the firearm-lethality-adjusted rate fluctuates in a band from roughly 15 to 18 per 100,000 from 1980 through 2020, with no statistically discernible downward trend: an at-best-ambiguous decline of about 10% from the 1980 level, well within the noise of year-to-year variation.1 The celebrated "great crime decline," the one that got attributed to lead abatement, legal abortion, broken windows policing, and approximately forty other causes, is roughly the same size as the improvement in trauma surgery over the same period.
The post-2014 homicide spike is almost entirely explained by increasing weapon lethality rather than increasing attacks. (On the post-2014 lethality reversal, see the Council on Criminal Justice's 2025 report.)
Two different ways to divide out the medicine
About three quarters of US homicides involve firearms: roughly 60% in the 1960s, rising to about 80% by 2021.2 Knives, blunt objects, and fists make up the rest. The survivability of knife wounds and beatings has improved too, but not nearly as dramatically. This is primarily a story about what happens when you get shot.
If you know how much more survivable a gunshot wound is today than it used to be, you can divide that improvement out of the observed homicide rate to get a trend line that isn't suppressed by medical progress:
I apply this to two independent data sources for the lethality curve.
Data source 1: Overall firearm lethality from crime statistics.Harris et al. (2002) computed the ratio of firearm homicides to (firearm homicides + firearm aggravated assaults) using FBI data. This ratio dropped from 15.5% in 1964 to 5.4% in 1999. Harris used this to estimate a counterfactual homicide count for a single year. I use the same lethality curve as an input to the adjustment formula above, applied continuously across 1964–2020. This is the orange line in Chart 1 and the purple line in Chart 3.
This data source has a known problem. The denominator, FBI-reported aggravated assaults, has been inflated over time by changes in reporting practices and definitions unrelated to actual violence, which makes the lethality ratio decline look larger than the actual improvement in survival. Andresen (2007) replicated the analysis for Canada using attempted homicides instead of all assaults and found no lethality decline at all. The Harris data uses firearm assaults specifically, which should be more robust to reporting drift than all aggravated assault -- there's less ambiguity in whether a shooting occurred than in whether a bar fight qualifies as "aggravated." But just to be sure, I also use a second data source that doesn't depend on crime reporting at all.3
Data source 2: Abdominal gunshot wound survival from clinical literature. Hospital data on the mortality rate for abdominal gunshot wounds provides a lethality curve independent of crime reporting. The well-sourced anchor points come from published surgical case series:4
1860s: ~87% mortality (Civil War, pre-antiseptic)
1898–1902: abdominal GSW mortality comparable to Civil War despite aseptic technique; Boer War laparotomy mortality 69%
~1900: ~72% mortality (interpolation: civilian surgeons had adopted laparotomy but lacked antibiotics and blood banking)
~1916: ~65% dropping to ~45% with forward hospitals in WWI
1935: overall assault lethality 11.4% in Memphis (Giacopassi et al.)
1960: ~30% abdominal GSW mortality (pre-EMS, pre-trauma systems, but with antibiotics and blood banking); Memphis assault lethality 5.5%
1968–1975: Vietnam-era KIA ratio fell from 29.3% (WWII) to 19%; helicopter medevac cut wounding-to-surgery time from 4–6 hours to ~35 minutes
Late 1980s: ~12% abdominal GSW mortality (regionalized trauma centers, damage control surgery emerging); Memphis assault lethality 3.2% by 1985
1990s: ~9.5% abdominal GSW mortality (consensus across multiple series)
2010: ~11% abdominal GSW mortality (ACS national database, 16,866 patients -- higher than the 1990s, reflecting worsening injury severity: more shots per victim, higher caliber, more critical-region hits)
Between these anchors, the chart interpolates linearly. The steepest improvements were 1900–1960 (antibiotics, blood banking, hospital proliferation) and 1960–1980 (Vietnam-era surgical advances, 911/EMS, regionalized trauma).
Overall gunshot mortality blends abdominal, thoracic, head, and extremity wounds with very different survival profiles; the abdominal series is not measuring that. Abdominal GSW mortality tracks the medical capability curve: the rate of improvement in trauma care's ability to save a specific wound type.
The steepest improvement in the cited series, from 30% (1960) to 12% (late 1980s), happened in roughly twenty-five years. Before 1966, American ambulances were predominantly operated by funeral homes, which seems like a conflict of interest.5 Two men would arrive with a stretcher and no medical equipment; treatment happened at the hospital or not at all. The 1966 Highway Safety Act set the first federal standards for emergency medical care, and most funeral homes exited the business.
I guess if multiple members of your family got shot, you'd pay a lot of attention to trauma medicine. Both my grandfathers had serious heart attacks, so I grew up paying attention to heart attack propaganda: call 911 immediately, because the patient needs to reach the hospital within sixty minutes of symptom onset. That's how I learned about the "golden hour." But the golden hour didn't come from cardiology. It came from people getting blown up in Vietnam.
Helicopter medevac in Vietnam cut the time from wounding to surgery from 4–6 hours (the Korean War average) to about 35 minutes, and the killed-in-action ratio fell from 29.3% in WWII to 19% in Vietnam. The combat medics who developed these techniques became the model for civilian EMTs; the first standardized EMT training program was published in 1967, and the military's helicopter evacuation model was directly adapted into civilian trauma centers in the early 1970s.
Around the same time, the US government declared war on cancer, drugs, and Vietnam, all three of which still exist despite billions of dollars poured into their respective wars. What made trauma deaths different? I think there are two answers. First, in the war on cancer, progress looks hard to distinguish from grift on 1 year timelines, while rushing someone to the hospital in 1 hour rather than 2 seems verifiable at lower levels of fidelity, like 30 minute pizza delivery. Second, there's money to be made in illegal drugs, but gunshot wound mortality wasn't a revenue stream for anyone involved in the process except the undertakers, and it was never a large percentage of their revenue, so there was no institutional resistance to actually solving the problem. Oh, and the Vietnamese fought back.
While the abdominal series reflects these improvements, it does not capture changes in injury composition: if the average shooting victim in 2020 arrives with more wounds to more body regions than in 1970, the abdominal CFR overstates the population-level rescue effect. The firearm lethality data (Data source 1) implicitly captures this because it measures deaths per assault incident, not deaths per wound type. For this reason, the firearm lethality curve is the primary data source; the abdominal GSW curve is a secondary confirmation of the trend.
Plugging this curve into the same adjustment formula gives the green line in Chart 3 and extends the picture back to 1933 (before FBI data begins) using vital statistics homicide rates.
This adjustment is an upper bound. Some fraction of homicide victims (those with instantly fatal head wounds, those dead on scene before any care is possible) die regardless of the era's medical capability. If 10–15% of homicides are medically invariant, the true adjustment for recent decades is roughly 60–80% of the displayed values.
Where the two adjustment series overlap (1964–2020), they agree on the direction and rough magnitude of the trend.6 From 1964 to 1985, the two adjusted series track within 10% of each other.7 From 1986 to 1999, the abdominal series shows a spike (peaking at ~1.5× the firearm series around 1990) then returns to parity by 1999. The Demetriades 1988 anchor (11.7% mortality at a single urban Level I center, probably below the national average) accounts for roughly a quarter of this divergence; the rest likely reflects the limited resolution of the abdominal series, which has no anchor points between 1960 and 1988 and so cannot track the crack-era surge and subsequent decline in real time the way the annual firearm data can. After 2000, they diverge in the opposite direction: the abdominal series runs below the firearm series through the 2000s (reaching 0.72× at 2010, as worsening injury severity pushed abdominal mortality back up), then crosses back above around 2016 and reaches 1.32× by 2019. The abdominal series is noisier because it has fewer anchor points and the anchors come from individual case series rather than national data. The firearm lethality series is more trustworthy for year-to-year movements. The abdominal series is useful as a check that the overall direction is real and not an artifact of assault-count contamination.
The homicide-assault ratio
The adjustment formula isn't the only evidence that medicine is suppressing the homicide trend.
If violence were genuinely declining at the rate the homicide numbers suggest, the ratio of homicides to aggravated assaults should be roughly constant: the same attacks should produce deaths at the same rate. Instead, the ratio fell by a factor of 3.5, from about 59 homicides per 1,000 aggravated assaults in 1960 to about 17 per 1,000 in 2000, where it roughly stabilized.8
Part of this reflects expanded assault reporting. In the early 1960s, most police departments kept paper records and relied on victims to walk into a station. The rollout of 911 systems in the late 1960s and 1970s made reporting easier. Police professionalization and computerized records meant more incidents got written up. Domestic violence, largely ignored by police before the 1980s, began generating assault charges. And the legal boundary of "aggravated" drifted outward: the FBI classifies pepper spray, Tasers, and BB guns as dangerous weapons for UCR purposes. O'Brien (2003) documented the broader pattern, arguing that long-term UCR increases in assault rates reflect reporting and recording improvements rather than actual increases in violence. But even generous estimates of reporting drift account for perhaps half the 3.5× decline in the ratio. The remainder is consistent with the medical improvement thesis: the same number of serious assaults producing proportionally fewer deaths over time.
After 1994, the reporting drift may have reversed direction. The NYPD's CompStat system, introduced that year, held precinct commanders accountable for index crime numbers, creating institutional pressure to make those numbers go down. One way to do this legitimately is better policing; another is to reclassify borderline felony assaults as misdemeanors, which don't count in the index. Eterno and Silverman (2012) documented this practice extensively. In NYPD data from 2000 to 2009, felony assaults fell 42% while misdemeanor assaults fell only 9%, a divergence hard to explain without systematic downclassification.9 CompStat spread to most major US police departments during this period.
Recorded aggravated assault fell roughly 30% from the early-1990s peak to the early 2000s. A reclassification rate of 3 percentage points per year, compounding through a ratchet effect where each year's downclassified baseline becomes the next year's starting point, would account for the entire recorded decline. The true contribution is probably smaller, but it need not be large to meaningfully affect lethality ratios.
The Harris firearm lethality data uses specifically firearm aggravated assaults as its denominator, not all assaults. A shooting is inherently less subject to reporting drift than a borderline bar fight that might or might not get classified as aggravated. You don't have the same "was that really aggravated?" ambiguity with "someone shot at me." The firearm-specific denominator is substantially more robust than the all-assault denominator, though not immune to drift from the rollout of 911 systems and improved police recording practices.
Under UCR rules, any assault involving a firearm is classified as aggravated by definition, regardless of injury; simple assault explicitly excludes incidents involving firearms. A precinct commander under CompStat pressure can downgrade a borderline bar fight from aggravated to simple, but cannot do the same with a shooting. The firearm assault count is still subject to changes in reporting completeness (whether the incident gets a report written at all), but not to the felony-to-misdemeanor reclassification that Eterno and Silverman documented. This means the Harris firearm-specific lethality ratio is substantially more robust to CompStat-era distortion than the all-assault ratio discussed above.
Everything behind locked glass
Property crime, car theft, robbery, and survey-measured victimization all declined after 1990. Scott Alexander makes this case in detail in Record Low Crime Rates Are Real, Not An Artifact, arguing that these parallel declines confirm the homicide decline was genuine. Doesn't the coherence across categories suggest the homicide decline was real too?
These categories are consistent with a genuine decline in some forms of criminal activity, and nothing in the medical adjustment contradicts that. But observed crime rates reflect prevention expenditure and enforcement thresholds, not just underlying criminal intent. A pharmacy with nothing behind locked glass is not the same crime burden as one with the same shoplifting rate but everything behind locked glass. A city that stops prosecuting thefts under $950 has fewer recorded thefts without anyone stealing less. As Scott himself acknowledges, the relationship between recorded crime and the experienced burden of disorder is loose. It's unclear how much recorded crime trends in property categories reflect changing behavior versus changing measurement.
Shootings are the category least subject to this. They get reported regardless of policing philosophy or prosecution thresholds, and you can't reduce them by putting them behind locked glass.10
Fewer young people, more violence per young person
The US population aged 15–29, the peak offending demographic, shrank from 27.4% of the population in 1980 to about 20% in 2020, a 27% decline.11 This age group commits roughly half of all homicides (BJS, 1980–2008), so the demographic shift alone predicts about a 14% decline in homicide rates, all else equal.
The raw homicide rate fell 36% from 1980 to 2020, well beyond what demographics explain. But the firearm-lethality-adjusted rate shows essentially no decline at all. After removing the medical artifact, the age-adjusted rate of lethal-intent violence increased above 1980 levels, peaking around 2005.12
And even this understates the problem. The US incarceration rate went from 220 per 100,000 in 1980 to 744 per 100,000 in 2010, a 3.4× increase.13 By 2000, roughly 3.5% of all 15–29-year-olds were locked up, versus 0.8% in 1980.
So you had a smaller crime-prone population, vastly more incarceration, and vastly better medicine, and the adjusted rate still went up. The underlying propensity toward lethal violence increased substantially during the "great crime decline."
Caveats
Pre-1933 data is unreliable due to incomplete death registration, so the adjusted series starts at 1933. The abdominal GSW survival curve is interpolated between clinical anchor points, not continuous annual data. The firearm lethality data depends on police-reported firearm assaults as its denominator, which is subject to some reporting drift, though less than for all aggravated assault, since firearm assaults are inherently more likely to generate a police report regardless of era. The agreement with the independent clinical data through 2010 suggests the overall estimated trend is right; the post-2010 divergence is expected and reflects changing injury severity rather than a failure of either data source. The ideal additional check would be a long-run national series on nonfatal firearm injuries (ED visits, hospitalizations) independent of crime reporting; no such dataset currently exists at the national level. CDC's WISQARS nonfatal injury data begins only in 2001 and the CDC itself flags it as unreliable for firearm injury trend estimation.
The medically-invariant fraction (head shots, DOA) means the abdominal GSW adjustment overstates the true counterfactual by perhaps 20–40% in recent decades; this doesn't affect the firearm lethality data source, where such deaths appear in both numerator and denominator.
The adjustment is most informative for the US specifically. Homicide declined in Western Europe too, roughly 33% from 1990 to 2020, and Europe didn't have the same trauma care revolution in gunshot survival, which might seem to undermine the medical-artifact thesis. But American and Western European homicide operate in fundamentally different regimes. American homicide rates are roughly 8× higher than Western European rates and dominated by firearms (~75% of US homicides); European homicide is knife-dominated and occurs at rates so low that a handful of incidents per year visibly moves a small country's statistics. The medical artifact is roughly proportional to the share of violence involving firearms multiplied by the improvement in gunshot survival; both factors are much smaller in Europe. Western European countries also share enough economic integration, demographic trends, and cultural shifts that their "independent" declines are more like one data point measured several times than several independent replications. European trends tell us essentially nothing about the size of the American medical artifact.
If the adjustment is even directionally correct, most of the literature explaining the great crime decline is explaining the wrong thing. The question is not "why did Americans become less violent after 1990?" but "why did Americans become so much more violent after 1965, and why haven't we stopped?"
Footnotes
This is a claim about the adjusted series as constructed: the year-to-year variance swamps the apparent trend. A separate question is how sensitive the adjusted series is to assumptions about denominator drift in the FBI data. If recorded firearm assaults were inflated by reporting improvements, the adjustment overcorrects and the true decline is larger. The convergence with the independent clinical proxy through 2010 constrains how large this effect can be, but does not eliminate it.
FBI Supplemental Homicide Reports via BJS. The firearm share was 57% in 1964, crossed 65% in the early 1970s, and has been 73–80% since the mid-1990s.
Eckberg (2014) challenged Harris using the National Crime Victimization Survey instead of UCR data, finding stable lethality 1973–1999 rather than declining. Scott Alexander cites this as evidence against the medical-suppression thesis. But Eckberg's "lethality" is a hybrid: homicides from vital statistics divided by assaults from a survey of ~240,000 people (yielding roughly 1,000 aggravated assault victims per year). That's a ratio of two quantities measured by completely different instruments with different biases and coverage. Harris at least uses numerator and denominator from the same system, so some systematic biases cancel. Meanwhile the NCVS can't disaggregate by weapon type, can't measure homicide at all (dead people don't answer surveys), and assault is the worst-recalled crime in the survey. In fairness, my own adjustment also crosses data sources (vital statistics deaths adjusted by clinical surgical case fatality rates), but the adjustment factor comes from a different kind of measurement (hospital outcomes for a specific wound type) rather than a different survey trying to measure the same population-level count with different biases.
The estimates are synthesized from scattered military and civilian series, not a single longitudinal dataset. Barr et al. 2023, "'Through and Through' History," reviews GSW management from the 14th century to the present and is the best single overview. It gives Civil War abdominal GSW mortality at ~87%; Spanish-American War abdominal mortality was reportedly comparable to the Civil War despite aseptic technique, because antibiotics and blood transfusion did not yet exist. In the Boer War, British surgeons attempted laparotomy on 26 abdominal wound patients; 18 died (69%). In WWI, placement of forward hospitals reduced abdominal wound mortality from ~65% to ~45%. The 1900 estimate of ~72% is an interpolation: civilian surgeons had adopted exploratory laparotomy by the late 1890s, bringing mortality below the military figures, but without antibiotics or blood banking. No single civilian case series from ~1900 reports this exact figure; it sits between the military data and the ~55% observed in 1930s civilian practice. Wilson & Sherman 1961, "Civilian Penetrating Wounds of the Abdomen" (494 cases), documents ~30% mortality, anchoring the 1960 estimate. Giacopassi et al. 1992 examined Memphis police homicide files at 25-year intervals: overall assault lethality was 11.4% in 1935, 5.5% in 1960, and 3.2% in 1985, corroborating the mid-century improvement from penicillin and blood banking. The Vietnam-era improvement is well-documented: helicopter medevac cut wounding-to-surgery time from 4–6 hours to ~35 minutes, and the overall killed-in-action ratio fell from 29.3% in WWII to 19% in Vietnam. Demetriades et al. 1988, 300 consecutive abdominal GSW cases at an urban trauma center, found 88.3% survival (11.7% mortality). By the 1990s, multiple series converged on ~9.5% mortality for abdominal GSWs at trauma centers. The ACS Trauma Quality Improvement Program national database (16,866 abdominal GSW patients, 2010–2014) found 11% mortality in 2010 declining to 9.4% by 2014 -- the 2010 figure is higher than the 1990s consensus, consistent with worsening injury severity (more wounds per victim, higher caliber) partially offsetting continued improvements in care. Post-2020 single-digit abdominal GSW mortality is standard at Level I centers.
See the National Academy of Sciences 1966 report "Accidental Death and Disability: The Neglected Disease of Modern Society," which documented the state of emergency medical services and prompted the Highway Safety Act.
From 1964 to 1985, the ratio of ABD-adjusted to FL-adjusted rates stays within 0.975–1.098. From 1986 to 1999, the ABD series spikes to 1.49× FL at 1990, then returns to 1.02× by 1999. Post-2000 the ratio drops below 1.0, hitting 0.72 at 2010 (as abdominal mortality ticked back up with worsening injury severity), then crosses back above 1.0 around 2016 and reaches 1.32 by 2019.
The strongest objection to the FBI lethality series is that denominator drift (inflation of recorded firearm assaults due to 911 adoption, better record-keeping, etc.) could account for much or all of the observed lethality decline, making the adjustment an overcorrection that erases a real decline. The convergence with the clinical proxy is the main defense: the abdominal GSW survival curve has zero exposure to crime reporting practices, so if denominator drift were dominating the FBI series, there's no reason the clinical series would track it within 10% for twenty years (1964–1985). The agreement doesn't prove both are right, but for both to be wrong, two completely independent measurement systems would have to be biased in the same direction by a comparable amount across the same period.
Computed from FBI UCR homicide and aggravated assault rates: 5.1 homicides and 86.1 aggravated assaults per 100,000 in 1960 (ratio 59.2 per 1,000); 5.5 and 324.0 in 2000 (ratio 17.0 per 1,000).
Eterno and Silverman's data on the felony/misdemeanor assault divergence is from The Crime Numbers Game (2012), drawn from NYPD non-index crime data obtained via lawsuit in 2010. An independent audit commissioned by Commissioner Kelly confirmed that "misclassification of reports may have an appreciable effect on certain reported crime rates." Similar practices have been documented in Atlanta, Baltimore, Dallas, New Orleans, and agencies in the UK, Australia, and France.
If would-be shooters are emboldened by the lower probability of a murder rap, the lower expected penalty could even increase the rate of shootings, making the medical adjustment conservative rather than aggressive. This is probably most relevant to those who shoot people as part of some other vocation such as robbery. True amateurs, who kill from causes like anger, on the other hand, could even be discouraged by the lower probability of the closure provided by the death of one's enemy, or they could try to inflict more severe wounds.
I estimate the age-adjusted trend using the per-young-person (15–29) rate rather than a formal age-standardization across all cohorts. The 30–44 age group, responsible for another ~30% of homicides, also shrank, from about 22% to 19% of the population over the same period. A formal adjustment accounting for both cohorts would likely show an equal or larger increase in underlying violence.
BJS Prisoners Series. Total incarceration (state + federal) per 100,000 US residents.
[Warning: this is a linkpost, and the footnote links will teleport you to my personal website.]
The US homicide rate hit 10.2 per 100,000 in 1980, fell to 4.4 by 2014, then spiked to 6.5 in 2020. This looks like a story about violence rising, falling, and rising again. The actual story is simpler and worse: violence tripled in the 1960s and 70s. After adjusting for medical improvement, there is no clear statistical signal that it ever came back down. What changed was not the violence but how often it killed people.
Deaths vs. attacks
The homicide rate measures deaths. Deaths from gunshot wounds depend on two things: how many people get shot, and how many of those people die. Medicine affects the second thing. The homicide rate doesn't separate them.
If you get shot in the abdomen today, you have about a 92% chance of surviving. In 1960, you had about a 70% chance. In 1900, about 28%. The prognoses for other penetrating trauma wounds also improved, though this varies by wound location: head shots remain almost uniformly fatal regardless of era, while extremity wounds were highly survivable even a century ago. Abdominal wounds are a central case of an injury that kills you without good care and doesn't with it.
This problem was identified in the academic literature by Harris et al. (2002), who estimated that without post-1960 medical improvements, the US would have had 50,000–115,000 homicides per year in the late 1990s instead of the observed 15,000–20,000. But Harris presented this as a single-year counterfactual, as though merely to say that things were worse than they looked in 1999.
We have two standard measures of the trend in serious violence: the homicide rate and the aggravated assault rate. Both are distorted. The homicide rate is distorted by improving medicine, which makes the same number of attacks produce fewer deaths over time. The aggravated assault rate is distorted by changing reporting and classification practices: expanded reporting and definitional drift inflated the count through the early 1990s, while CompStat-era downclassification pressure may have deflated it afterward. This post constructs a third trend line by dividing the medical improvement out of the homicide rate. The resulting point estimates aren't very meaningful (this is not a literal count of attempted murders). But the trend is less distorted than either alternative, because it starts from hard death counts and adjusts using clinical data that doesn't depend on crime reporting.
The vanishing decline
I put together some interactive charts illustrating this adjusted trend line. The adjusted trend tells a different story than the raw homicide rate: the 1933–1960 decline largely disappears, and the 1965–1980 increase is steeper than the raw numbers suggest.
The raw homicide rate fell 57% from the 1991 peak to the 2014 trough. After adjustment, the firearm-lethality-adjusted rate fluctuates in a band from roughly 15 to 18 per 100,000 from 1980 through 2020, with no statistically discernible downward trend: an at-best-ambiguous decline of about 10% from the 1980 level, well within the noise of year-to-year variation.1 The celebrated "great crime decline," the one that got attributed to lead abatement, legal abortion, broken windows policing, and approximately forty other causes, is roughly the same size as the improvement in trauma surgery over the same period.
The post-2014 homicide spike is almost entirely explained by increasing weapon lethality rather than increasing attacks. (On the post-2014 lethality reversal, see the Council on Criminal Justice's 2025 report.)
Two different ways to divide out the medicine
About three quarters of US homicides involve firearms: roughly 60% in the 1960s, rising to about 80% by 2021.2 Knives, blunt objects, and fists make up the rest. The survivability of knife wounds and beatings has improved too, but not nearly as dramatically. This is primarily a story about what happens when you get shot.
If you know how much more survivable a gunshot wound is today than it used to be, you can divide that improvement out of the observed homicide rate to get a trend line that isn't suppressed by medical progress:
I apply this to two independent data sources for the lethality curve.
Data source 1: Overall firearm lethality from crime statistics. Harris et al. (2002) computed the ratio of firearm homicides to (firearm homicides + firearm aggravated assaults) using FBI data. This ratio dropped from 15.5% in 1964 to 5.4% in 1999. Harris used this to estimate a counterfactual homicide count for a single year. I use the same lethality curve as an input to the adjustment formula above, applied continuously across 1964–2020. This is the orange line in Chart 1 and the purple line in Chart 3.
This data source has a known problem. The denominator, FBI-reported aggravated assaults, has been inflated over time by changes in reporting practices and definitions unrelated to actual violence, which makes the lethality ratio decline look larger than the actual improvement in survival. Andresen (2007) replicated the analysis for Canada using attempted homicides instead of all assaults and found no lethality decline at all. The Harris data uses firearm assaults specifically, which should be more robust to reporting drift than all aggravated assault -- there's less ambiguity in whether a shooting occurred than in whether a bar fight qualifies as "aggravated." But just to be sure, I also use a second data source that doesn't depend on crime reporting at all.3
Data source 2: Abdominal gunshot wound survival from clinical literature. Hospital data on the mortality rate for abdominal gunshot wounds provides a lethality curve independent of crime reporting. The well-sourced anchor points come from published surgical case series:4
Between these anchors, the chart interpolates linearly. The steepest improvements were 1900–1960 (antibiotics, blood banking, hospital proliferation) and 1960–1980 (Vietnam-era surgical advances, 911/EMS, regionalized trauma).
Overall gunshot mortality blends abdominal, thoracic, head, and extremity wounds with very different survival profiles; the abdominal series is not measuring that. Abdominal GSW mortality tracks the medical capability curve: the rate of improvement in trauma care's ability to save a specific wound type.
The steepest improvement in the cited series, from 30% (1960) to 12% (late 1980s), happened in roughly twenty-five years. Before 1966, American ambulances were predominantly operated by funeral homes, which seems like a conflict of interest.5 Two men would arrive with a stretcher and no medical equipment; treatment happened at the hospital or not at all. The 1966 Highway Safety Act set the first federal standards for emergency medical care, and most funeral homes exited the business.
I guess if multiple members of your family got shot, you'd pay a lot of attention to trauma medicine. Both my grandfathers had serious heart attacks, so I grew up paying attention to heart attack propaganda: call 911 immediately, because the patient needs to reach the hospital within sixty minutes of symptom onset. That's how I learned about the "golden hour." But the golden hour didn't come from cardiology. It came from people getting blown up in Vietnam.
Helicopter medevac in Vietnam cut the time from wounding to surgery from 4–6 hours (the Korean War average) to about 35 minutes, and the killed-in-action ratio fell from 29.3% in WWII to 19% in Vietnam. The combat medics who developed these techniques became the model for civilian EMTs; the first standardized EMT training program was published in 1967, and the military's helicopter evacuation model was directly adapted into civilian trauma centers in the early 1970s.
Around the same time, the US government declared war on cancer, drugs, and Vietnam, all three of which still exist despite billions of dollars poured into their respective wars. What made trauma deaths different? I think there are two answers. First, in the war on cancer, progress looks hard to distinguish from grift on 1 year timelines, while rushing someone to the hospital in 1 hour rather than 2 seems verifiable at lower levels of fidelity, like 30 minute pizza delivery. Second, there's money to be made in illegal drugs, but gunshot wound mortality wasn't a revenue stream for anyone involved in the process except the undertakers, and it was never a large percentage of their revenue, so there was no institutional resistance to actually solving the problem. Oh, and the Vietnamese fought back.
While the abdominal series reflects these improvements, it does not capture changes in injury composition: if the average shooting victim in 2020 arrives with more wounds to more body regions than in 1970, the abdominal CFR overstates the population-level rescue effect. The firearm lethality data (Data source 1) implicitly captures this because it measures deaths per assault incident, not deaths per wound type. For this reason, the firearm lethality curve is the primary data source; the abdominal GSW curve is a secondary confirmation of the trend.
Plugging this curve into the same adjustment formula gives the green line in Chart 3 and extends the picture back to 1933 (before FBI data begins) using vital statistics homicide rates.
This adjustment is an upper bound. Some fraction of homicide victims (those with instantly fatal head wounds, those dead on scene before any care is possible) die regardless of the era's medical capability. If 10–15% of homicides are medically invariant, the true adjustment for recent decades is roughly 60–80% of the displayed values.
Where the two adjustment series overlap (1964–2020), they agree on the direction and rough magnitude of the trend.6 From 1964 to 1985, the two adjusted series track within 10% of each other.7 From 1986 to 1999, the abdominal series shows a spike (peaking at ~1.5× the firearm series around 1990) then returns to parity by 1999. The Demetriades 1988 anchor (11.7% mortality at a single urban Level I center, probably below the national average) accounts for roughly a quarter of this divergence; the rest likely reflects the limited resolution of the abdominal series, which has no anchor points between 1960 and 1988 and so cannot track the crack-era surge and subsequent decline in real time the way the annual firearm data can. After 2000, they diverge in the opposite direction: the abdominal series runs below the firearm series through the 2000s (reaching 0.72× at 2010, as worsening injury severity pushed abdominal mortality back up), then crosses back above around 2016 and reaches 1.32× by 2019. The abdominal series is noisier because it has fewer anchor points and the anchors come from individual case series rather than national data. The firearm lethality series is more trustworthy for year-to-year movements. The abdominal series is useful as a check that the overall direction is real and not an artifact of assault-count contamination.
The homicide-assault ratio
The adjustment formula isn't the only evidence that medicine is suppressing the homicide trend.
If violence were genuinely declining at the rate the homicide numbers suggest, the ratio of homicides to aggravated assaults should be roughly constant: the same attacks should produce deaths at the same rate. Instead, the ratio fell by a factor of 3.5, from about 59 homicides per 1,000 aggravated assaults in 1960 to about 17 per 1,000 in 2000, where it roughly stabilized.8
Part of this reflects expanded assault reporting. In the early 1960s, most police departments kept paper records and relied on victims to walk into a station. The rollout of 911 systems in the late 1960s and 1970s made reporting easier. Police professionalization and computerized records meant more incidents got written up. Domestic violence, largely ignored by police before the 1980s, began generating assault charges. And the legal boundary of "aggravated" drifted outward: the FBI classifies pepper spray, Tasers, and BB guns as dangerous weapons for UCR purposes. O'Brien (2003) documented the broader pattern, arguing that long-term UCR increases in assault rates reflect reporting and recording improvements rather than actual increases in violence. But even generous estimates of reporting drift account for perhaps half the 3.5× decline in the ratio. The remainder is consistent with the medical improvement thesis: the same number of serious assaults producing proportionally fewer deaths over time.
After 1994, the reporting drift may have reversed direction. The NYPD's CompStat system, introduced that year, held precinct commanders accountable for index crime numbers, creating institutional pressure to make those numbers go down. One way to do this legitimately is better policing; another is to reclassify borderline felony assaults as misdemeanors, which don't count in the index. Eterno and Silverman (2012) documented this practice extensively. In NYPD data from 2000 to 2009, felony assaults fell 42% while misdemeanor assaults fell only 9%, a divergence hard to explain without systematic downclassification.9 CompStat spread to most major US police departments during this period.
Recorded aggravated assault fell roughly 30% from the early-1990s peak to the early 2000s. A reclassification rate of 3 percentage points per year, compounding through a ratchet effect where each year's downclassified baseline becomes the next year's starting point, would account for the entire recorded decline. The true contribution is probably smaller, but it need not be large to meaningfully affect lethality ratios.
The Harris firearm lethality data uses specifically firearm aggravated assaults as its denominator, not all assaults. A shooting is inherently less subject to reporting drift than a borderline bar fight that might or might not get classified as aggravated. You don't have the same "was that really aggravated?" ambiguity with "someone shot at me." The firearm-specific denominator is substantially more robust than the all-assault denominator, though not immune to drift from the rollout of 911 systems and improved police recording practices.
Under UCR rules, any assault involving a firearm is classified as aggravated by definition, regardless of injury; simple assault explicitly excludes incidents involving firearms. A precinct commander under CompStat pressure can downgrade a borderline bar fight from aggravated to simple, but cannot do the same with a shooting. The firearm assault count is still subject to changes in reporting completeness (whether the incident gets a report written at all), but not to the felony-to-misdemeanor reclassification that Eterno and Silverman documented. This means the Harris firearm-specific lethality ratio is substantially more robust to CompStat-era distortion than the all-assault ratio discussed above.
Everything behind locked glass
Property crime, car theft, robbery, and survey-measured victimization all declined after 1990. Scott Alexander makes this case in detail in Record Low Crime Rates Are Real, Not An Artifact, arguing that these parallel declines confirm the homicide decline was genuine. Doesn't the coherence across categories suggest the homicide decline was real too?
These categories are consistent with a genuine decline in some forms of criminal activity, and nothing in the medical adjustment contradicts that. But observed crime rates reflect prevention expenditure and enforcement thresholds, not just underlying criminal intent. A pharmacy with nothing behind locked glass is not the same crime burden as one with the same shoplifting rate but everything behind locked glass. A city that stops prosecuting thefts under $950 has fewer recorded thefts without anyone stealing less. As Scott himself acknowledges, the relationship between recorded crime and the experienced burden of disorder is loose. It's unclear how much recorded crime trends in property categories reflect changing behavior versus changing measurement.
Shootings are the category least subject to this. They get reported regardless of policing philosophy or prosecution thresholds, and you can't reduce them by putting them behind locked glass.10
Fewer young people, more violence per young person
The US population aged 15–29, the peak offending demographic, shrank from 27.4% of the population in 1980 to about 20% in 2020, a 27% decline.11 This age group commits roughly half of all homicides (BJS, 1980–2008), so the demographic shift alone predicts about a 14% decline in homicide rates, all else equal.
The raw homicide rate fell 36% from 1980 to 2020, well beyond what demographics explain. But the firearm-lethality-adjusted rate shows essentially no decline at all. After removing the medical artifact, the age-adjusted rate of lethal-intent violence increased above 1980 levels, peaking around 2005.12
And even this understates the problem. The US incarceration rate went from 220 per 100,000 in 1980 to 744 per 100,000 in 2010, a 3.4× increase.13 By 2000, roughly 3.5% of all 15–29-year-olds were locked up, versus 0.8% in 1980.
So you had a smaller crime-prone population, vastly more incarceration, and vastly better medicine, and the adjusted rate still went up. The underlying propensity toward lethal violence increased substantially during the "great crime decline."
Caveats
Pre-1933 data is unreliable due to incomplete death registration, so the adjusted series starts at 1933. The abdominal GSW survival curve is interpolated between clinical anchor points, not continuous annual data. The firearm lethality data depends on police-reported firearm assaults as its denominator, which is subject to some reporting drift, though less than for all aggravated assault, since firearm assaults are inherently more likely to generate a police report regardless of era. The agreement with the independent clinical data through 2010 suggests the overall estimated trend is right; the post-2010 divergence is expected and reflects changing injury severity rather than a failure of either data source. The ideal additional check would be a long-run national series on nonfatal firearm injuries (ED visits, hospitalizations) independent of crime reporting; no such dataset currently exists at the national level. CDC's WISQARS nonfatal injury data begins only in 2001 and the CDC itself flags it as unreliable for firearm injury trend estimation.
The medically-invariant fraction (head shots, DOA) means the abdominal GSW adjustment overstates the true counterfactual by perhaps 20–40% in recent decades; this doesn't affect the firearm lethality data source, where such deaths appear in both numerator and denominator.
The adjustment is most informative for the US specifically. Homicide declined in Western Europe too, roughly 33% from 1990 to 2020, and Europe didn't have the same trauma care revolution in gunshot survival, which might seem to undermine the medical-artifact thesis. But American and Western European homicide operate in fundamentally different regimes. American homicide rates are roughly 8× higher than Western European rates and dominated by firearms (~75% of US homicides); European homicide is knife-dominated and occurs at rates so low that a handful of incidents per year visibly moves a small country's statistics. The medical artifact is roughly proportional to the share of violence involving firearms multiplied by the improvement in gunshot survival; both factors are much smaller in Europe. Western European countries also share enough economic integration, demographic trends, and cultural shifts that their "independent" declines are more like one data point measured several times than several independent replications. European trends tell us essentially nothing about the size of the American medical artifact.
If the adjustment is even directionally correct, most of the literature explaining the great crime decline is explaining the wrong thing. The question is not "why did Americans become less violent after 1990?" but "why did Americans become so much more violent after 1965, and why haven't we stopped?"
Footnotes