The Twenty Percent Rule

In early January, El Salvador's justice and public security minister, Gustavo Villatoro, announced that the country had recorded eighty-two homicides in 2025. That is 1.3 per 100,000 people. In 2015 the rate was 105. England and Wales, whose latest figures are its lowest since 1977, sits at 0.86.

All three numbers are real. They just aren't the same kind of number, and the difference is a definition. This post is about that definition, a short document written in Bogotá in 2015 that tries to pin it down, and what happens when you hold national counts up against its test.

We mostly work with American police data, and we have written plenty about how two US cities can file the same event under different labels. Across borders the problem is the same, only nobody is paid to reconcile it.

What eighty-two leaves out

InSight Crime, which compiles homicide figures for Latin America and the Caribbean every year, reported El Salvador's record low in its 2025 round-up and, in the same breath, listed what the government no longer counts. Bodies found in unmarked graves are out. So are killings by police and murders inside prisons.

Those exclusions arrived in stages. According to a 2024 analysis in Foreign Policy, the interagency table that tallies homicides stopped counting clandestine graves in May 2021, and killings in clashes with police or soldiers dropped out in April 2022, days after the state of exception began. Human Rights Watch's 2026 World Report adds that the government restricts public access to homicide and crime data, and cites the Salvadoran group Socorro Jurídico Humanitario's count of at least 458 deaths in custody during the emergency. (Not all of those are homicides. The restrictions are why it's hard to say how many are.)

We aren't in a position to say how far violence in El Salvador actually fell, and nothing in this post depends on it. The point is narrower. The 1.3 counts one set of deaths and England's 0.86 counts another, and a table that ranks them side by side is ranking two different rulebooks.

Four rulebooks, one word

Most national systems use the word homicide. Hardly any mean exactly the same thing by it. Here are four that a global product is likely to touch, plus the one Mexico's statistics agency uses.

The UN definition, from the UNODC metadata note updated this July, is a single sentence: “Unlawful death inflicted upon a person with the intent to cause death or serious injury.” The note tells countries to apply it regardless of their own legislation. It also says, a page later, that the counting unit is the victim, and that certain countries may use other units. Both instructions are sensible. The second one is also an admission.

Notice where the legal line falls. Under the UN rule, a police shooting that a court finds lawful is not a homicide. Under Bogotá, it is. A country with many lawful police killings will look safer under one standard than the other, and the gap has nothing to do with how often people are killed.

The document from Bogotá

The Bogotá Protocol came out of a conference held September 7 to 9, 2015. More than ninety people from twelve countries attended: prosecutors, health ministries, statistics institutes, academics, and civil-society groups (InSight Crime and Brazil's Instituto Igarapé among them). It runs about ten pages in Spanish.

What we like about it is that it doesn't ask anyone to rebuild their records system. It says so directly. The aim is a shared concept that lets the existing sources converge, plus a list of measurable thresholds a civil-society group can check from the outside. Here are the ones that matter most for anyone consuming the data.

The ambiguous-category rule is the sharpest. A system that starts filing bodies under “remains found” instead of “homicide” has to keep that pile small, or it fails a test anyone can run.

Mexico, four ways

Mexico is the best place to watch the convergence rule work, because Mexico publishes both kinds of source. The Executive Secretariat of the National Public Security System (SESNSP) counts victims in prosecutors' investigation files and reports monthly. INEGI, the national statistics agency, codes death certificates and reports once a year. Here is everything we could find for 2025.

The two official sources are the ones the Bogotá test compares. Here is the arithmetic.

death certificates / prosecutor files
27,989 / 23,374 = 1.197   ->  19.7% apart   (January SESNSP figure)
27,989 / 23,609 = 1.186   ->  18.6% apart   (later SESNSP figure)

same test, 2024 (INEGI final vs. SESNSP)
33,550 / 30,060 = 1.116   ->  11.6% apart

implied population behind each 2025 rate
27,989 / 21.4 * 100,000 = 130.8 million   (INEGI)
23,374 / 17.5 * 100,000 = 133.6 million   (SESNSP)
20,677 / 15.4 * 100,000 = 134.3 million   (InSight Crime)

Mexico passes, by somewhere between a third of a point and a point and a half. A year earlier it passed with room to spare. Both 2025 numbers are preliminary, and Mexico News Daily notes that preliminary homicide counts there are usually revised upward, so the final gap could go either way. We would not read much into the widening until both series are final.

The last three lines surprised us. The rates don't just use different numerators. They sit on population estimates about three and a half million people apart. That's the denominator problem, at national scale.

InSight Crime's round-up adds two cautions of its own. Mexico's methodology has changed several times in recent years, which makes year-over-year comparison hard, and disappearances, which several outlets have covered rising as homicides fell, aren't in the homicide data at all. UPI reported in July 2025 that missing-person cases in the first half of 2025 ran nearly 18% above the same months of 2024. A disappeared person who was killed is a homicide under every rulebook in the table above. Under every counting system, they're invisible until someone finds them.

Try it on the United States

The US has the same two sources: FBI murder estimates built from police reports, and CDC homicide counts built from death certificates. Jeff Asher compared them last September. Through most of the 2010s the FBI reported about 0.9 murders for every CDC homicide. In 2023 and 2024 it was 0.84.

Some of the gap is definitional and always has been, since the CDC's homicide category is broader than the FBI's murder category. Asher traces the recent widening mostly to agencies that didn't report (Jackson, Mississippi, hasn't sent the FBI data since 2019, and Louisiana has gaps) and to the estimates that fill in for them. When the FBI later revises its numbers, most of the extra gap closes.

So the United States, on its first estimate, sits about where Mexico does: a point or so inside the line drawn in Bogotá. We don't think that makes either country's data bad. We think it means the twenty percent rule is a sensible threshold, and that the first number published is rarely the last. That is also why we keep saying the Real-Time Crime Index and the FBI measure different things well.

A decline in many languages

None of this is an argument that the decline isn't happening. The directions line up across systems that share almost nothing. InSight Crime puts the median homicide rate in Latin America and the Caribbean at about 17.6 per 100,000 in 2025, down more than 5%. The Office for National Statistics recorded 522 homicide victims in England and Wales in the year ending March 2025, the lowest rate since 1977. Among the 590 US agencies in the Real-Time Crime Index, murder in the first half of 2026 ran 17.3% below the first half of 2025.

A decline that shows up at once in countries with different laws, police, and paperwork is, in our opinion, decent evidence for the kind of structural explanation we have written about before, and weak evidence for any single national policy. It is also exactly when you have to be most careful. A definition change and a real decline look identical in a year-over-year column. We watched that happen inside one American city's file when Phoenix changed records systems. A country can do the same thing by memo.

If your product crosses a border

Travel-safety apps, global security teams, and executive protection tools all end up putting countries in the same list. We skipped the ingestion code because every national source needs its own parser anyway. The part that carries over is the metadata.

1. Store the definition next to the number.A rate without its rulebook can't be compared to anything. Something like this, per country, per year:

   {
     "country": "MX",
     "year": 2025,
     "victims": 27989,
     "rate_per_100k": 21.4,
     "source_type": "death_certificates",
     "definition": "ICD-10 X85-Y09",
     "includes_lawful_state_killings": false,
     "includes_prison_deaths": true,
     "includes_found_remains": "unknown",
     "date_basis": "registered",
     "preliminary": true,
     "published": "2026-08-03"
   }

2. Don't rank across rulebooks. If two countries count different deaths, put them on separate rows with a note, not in one sorted column.
3. Carry both sources when they exist, and show the range. Mexico has two. So does the United States.
4. Use the twenty percent test as a flag, lightly.Past 20%, treat the country's figure as low confidence. Under it, remember that two of the largest systems we looked at pass by about a point.
5. Track the companion series. Disappearances, found remains, deaths in custody, undetermined-intent deaths. When homicide falls and one of those rises, the ratio between them is the story.
6. Treat a definition change like a schema change.Draw a vertical rule, write the note, and don't compute a year-over-year change across it. This is the normalization problem again, with governments instead of vendors.
7. Keep every vintage. Preliminary counts move. Store the January number and the August number, and know which one your chart used.

Two

Section seven of the Bogotá Protocol is the least dramatic part of it. Any later change to an official homicide figure gets reported, with an explanation, to the people checking the numbers.

England and Wales did a version of that this year. Of the 524 homicides first recorded for the year ending March 2025, the ONS noted that two were no longer recorded as homicide by December 5. Two. With a date.