← Back to the atlas

Methods and sources

How the atlas counts, what every view shows, where each figure comes from, and what is still uncertain. Every figure on the site can be traced to a named source, and no death is meant to be counted twice.

How deaths are counted

The atlas covers armed conflicts since 1900: wars, rebellions, uprisings, communal violence and violence against civilians. Deaths are recorded per conflict, year, place and category, then rolled up to today’s countries.

One source per conflict and year. Where several datasets cover the same conflict-year, only the most detailed one counts: UCDP’s geocoded events (from 1989), then Correlates of War (COW) war totals, then PRIO’s yearly battle deaths. COW gives one figure per war and participant, which the atlas spreads evenly over the years that participant fought. When two datasets record the same fighting under different conflicts (COW splits some wars into phases of one UCDP conflict), only the better source counts.

This year. UCDP’s annual data runs to last year. The current year comes from UCDP’s Candidate Events, preliminary monthly releases in the same format, which UCDP revises before its annual release. The conflict view says how far into the year they run. Conflicts that first appear there are added provisionally. UCDP’s preliminary files don’t name the sides, so the atlas names such a conflict after its country and any armed group its events name, marked “provisional”, until UCDP’s annual release.

Who a death counts for. A government’s soldiers and security forces count for that government’s country, wherever they died (US soldiers in Vietnam count for the United States). Rebels, militias, civilians and deaths of unknown status count where they died.

Today’s countries. Historical territories are mapped to today’s countries by their borders at the time (CShapes). A figure counts for a single country only when it clearly belongs to it; otherwise it is shared and never added to any country’s total: Austria-Hungary’s First World War dead, or a coalition’s combined losses. Countries whose figures are mostly shared are hatched in grey on the globe.

Where the fighting was, before 1989. From 1989 every event has coordinates. Before then the datasets name only the states or colonies a conflict was fought in, which would put a revolt in Ottoman Arabia in Turkey, or a rising in British India in Pakistan and Burma as well. Where the atlas holds curated battle sites or theatres for a conflict and year, the conflict is placed only in today’s countries that contain them, and the timeline marks the sites. Conflicts without curated sites are still placed by territory.

Categories. State forces (army and police together; no source separates them), armed groups, civilians, and unspecified (battle deaths no source splits by side, and deaths of unknown status).

Matching datasets

UCDP and COW often record the same fighting differently: UCDP keeps one long conflict per dispute (“India vs Pakistan”, 1948 on), COW splits it into wars (the Kargil War). The atlas proposes links between their records from shared years, places and states, and each link is decided by a person: the same conflict (merged into one entry), part of it, overlapping, or not related. Records stay separate with their own sources; a confirmed or undecided link only stops the same deaths from counting twice, so links between different wars are rejected explicitly. Wikidata items are matched the same way, for names and for which larger war a conflict belongs to.

Curated estimates

No dataset covers civilian deaths before 1989, wounded, or conflicts before 1946 below COW’s 1,000-death threshold. The atlas fills these from published works, one row per event, country and category, each with its citation: civilian deaths in wars and genocides, wounded and military dead, repression, and 247 minor conflicts 1900–1945 (some still without a sourced toll).

Curated figuresChecked in sourceVia secondaryUnverifiedOther
Civilian deaths, wounded (420)188832147
Minor conflicts 1900–1945 (tolls) (170)61160105
Repression and famine (903)600291111

“Other”: not yet tagged, such as official records taken directly from the publisher (most wounded figures) and minor-conflict tolls from the first research pass.

Ranges

The atlas leads with a range, low to high, and gives the best estimate beneath it. A range sums each source’s own low and high; for curated figures it spans every estimate the atlas holds for the same event and country, counted or not (the Great Leap Forward famine: 15 to 45 million). Where a source gives a single number, as COW does for each war, the range narrows to that number, so a narrow range means the sources give one figure, not that the figure is certain. Where the atlas picks the midpoint of a published range, the figure is marked as an estimate.

Repression

What governments did to their own populations outside war: political imprisonment and labour camps, deportations, executions, deaths in custody, mass killings, and famines caused by policy. It is kept apart from the conflict view so that war deaths and repression are never mixed; a famine caused mainly by war tactics (Biafra) is a war death toll, one caused mainly by state policy (the Holodomor, the Great Leap Forward, Ethiopia 1983–85) is repression.

Government and press freedom

Both views let you choose the source, and the panel always shows the source’s own verdict in its own words. For government type: the Lexical Index of Electoral Democracy (rule-based, from observable electoral facts), Boix–Miller–Rosato (democracy or not), and International IDEA’s Global State of Democracy (about half of IDEA’s indicators come from V-Dem, which the atlas does not otherwise use). On top of every source, colonies (CShapes), communist-party rule and ruling monarchies come from the atlas’s own list; any classification that is the atlas’s rather than the source’s is labelled so. A country that was part of another state takes that state’s rating.

Press freedom: Freedom House’s Freedom of the Press (1979–2016) and the Global Media Freedom Dataset (1948–2012).

The government view’s other three layers are figures, not verdicts. Population and the median age come from the UN’s World Population Prospects 2024 (medium variant): estimates to 2024, projections after, every country from 1950; the panel’s age pyramid uses the same source’s five-year bands. The age of the leader uses Archigos 4.1, which names the effective ruler of each state with dates in office and of birth up to the end of 2015, and Wikidata’s head of state and head of government from then on. Which of those two offices continues Archigos’s “leader” is settled per country by matching the person in office on 31 December 2015, so Russia follows its president and India its prime minister. The map shows the leader’s age on 1 July of the chosen year. The panel adds the average age of each chamber of parliament from the Inter-Parliamentary Union’s Parline database (CC BY-NC-SA), usually as of the last election; ruling-party bodies are not loaded yet.

Messages and visitors

The messages page keeps what you type: your name, the country you choose and your message. A message with swearing, slurs or a web link is turned away and not kept. Every other message waits until a person has read and approved it, and an automatic check marks the ones that need a careful look. Visits are counted by country (the map of them is not open yet): each time the atlas is opened it adds one to that day’s count for the visitor’s country, which the hosting service works out. The atlas keeps the date, the country and the number, and nothing else: no address, no cookie, and nothing that tells one visitor from another. The question box under the globe can pass a question to Claude, an AI model made by Anthropic. The question, the conversation in that box and the name of the view on screen are sent to Anthropic to be answered; the atlas does not keep them. Claude looks figures up in the atlas’s own data, but it can still be wrong, so check what it says against the panels. The sources page lists, country by country, the datasets and cited works behind the conflict figures.

Religion

Every country is filled with horizontal bands, one per religious group, largest on top, from Pew Research Center’s estimates of religious composition in 2020 (by self-identification, regardless of practice or belief). A band covers the same share of the country’s land as the group has of its people: the depth of each band is worked out from the country’s outline, so a tenth of the people is a tenth of the area, and each island or separate territory carries its own stack. Groups under about one per cent are too thin to see and are in the panel, which lists every share and the change since 2010.

Pew gives Christians and Muslims as single groups, so their branches come from the Correlates of War World Religion Project (Maoz and Henderson, v1.1), the latest year of which is 2010: the atlas takes each branch’s fraction of its religion in that year and applies it to Pew’s 2020 total. Christianity is split into Catholic, Protestant (with Anglican), Orthodox and other; Islam into Sunni, Shia and other (Ibadi in Oman, Alawite, Ahmadiyya and the rest). The split is therefore ten years older than the totals, and in countries without a census question on denomination it is an estimate. Hinduism is one band: neither this project nor any other source divides Hindus by tradition country by country.

Religious persecution comes from two sources, which measure different things and are shown side by side. Pew Research Center’s Government Restrictions Index and Social Hostilities Index (0–10, 198 countries, 2007–2022) count restrictions and hostile incidents affecting any religion, coded largely from the US State Department’s religious-freedom reports and USCIRF; its levels are bands of the 2007 scores (top 5% very high, next 15% high, next 20% moderate), and the panel lists the questions that drive each country’s score. Because the social index counts the kinds and spread of incidents, not their rate per person, large and diverse countries can score high. The Religion and State project (Jonathan Fox, Bar-Ilan University; 183 countries, 1990–2014) measures government discrimination against minority religions (0–108) and societal discrimination against them (0–81), from its own coding; the atlas rescales both to 0–10 for the map.

Children and society

Two views, Children and Society, share one method: twelve measures of how children, women and families live, each the latest figure a country has reported since 2010 (older figures are shown as no data). Six come from the World Bank’s World Development Indicators, which gather them from the agencies that produce them: children under 5 stunted by chronic hunger (the UNICEF, WHO and World Bank joint malnutrition estimates, with wasting and FAO’s undernourishment in the panel); girls of primary-school age in school (UNESCO’s out-of-school rate, shown as the share in school, with secondary enrolment in the panel); people living on under $3 a day at 2021 prices, the World Bank’s international poverty line (with each country’s own line in the panel); one-year-olds with all three doses of the diphtheria, tetanus and pertussis vaccine, the usual measure of whether basic child healthcare reaches a country’s children (WHO and UNICEF, with measles and skilled birth attendance in the panel); and the WHO’s universal health coverage index, 0–100, for families (with doctors per thousand and out-of-pocket spending in the panel). Average annual salary is the ILO’s average monthly earnings of employees in US dollars at the exchange rate of the time, times twelve; it falls when a currency falls, which is why some countries show a few hundred dollars a year. GNI per person sits beside it in the panel.

Poverty is drawn on two lines, with a toggle under the globe. The UN line is $3 a day at 2021 purchasing power, which already allows for prices differing between countries but is an absolute floor set by the poorest countries’ own lines: it says that about one American in a hundred is poor. The second is the World Bank’s societal poverty line, $3 a day or half the country’s median income plus a little, whichever is higher. In poor countries it equals the UN line; in rich ones it becomes relative poverty, the measure Europe and the OECD use, so the United States shows a fifth of its people poor by its own standard. The same line applies to children and to families.

Head counts multiply the rate by the UN population of the same year: children under 5 for hunger, everyone for family poverty, and under-18s for child poverty. No source publishes child poverty by country for the whole world, so the child figure applies the national rate to the under-18s; children are usually poorer than the average, so treat it as a floor. “Families” are people, since the World Bank counts individuals rather than households. Girls out of school is UNESCO’s own count.

People in modern slavery is Walk Free’s estimate of the number in forced labour or forced marriage on any day in 2021, per 1,000 people, from the Global Slavery Index 2023 (“Global Slavery Index 2023 Dataset, Minderoo Foundation, available from: www.globalslaveryindex.org”). It is modelled from national household surveys and the ILO, Walk Free and IOM global estimate, with a vulnerability model filling in countries without a survey, so it measures what happens rather than what the police find: North Korea, Eritrea and Mauritania lead, and the United States shows about three in a thousand. The index is an estimate with wide margins, and Minderoo Foundation makes no warranty about its accuracy; the atlas has no connection with Walk Free. The panel adds Walk Free’s vulnerability and government response scores and the trafficking victims that the authorities detected and recorded in a year, from the UN Office on Drugs and Crime’s Global Report on Trafficking in Persons (2024 edition, 2007–2023), all ages, women and children. Earlier the atlas drew those detections on the map; it no longer does, because detection is not prevalence: a country with good police and courts finds more victims and one that does not look finds few, so the United Kingdom and the United States topped a map on which Russia and India were nearly empty. Counts the UNODC publishes as “fewer than five” are loaded as two.

Child labour and child marriage come from UNICEF’s household surveys through the UN SDG database, so they measure what families report rather than what the police record. Child labour is the share of children aged 5 to 17 in work that is harmful or keeps them from school (SDG indicator 8.7.1, with the broader figure that counts long household chores in the panel). Child marriage is the share of women aged 20 to 24 who were married or in a union before 18 (indicator 5.3.1, with marriage before 15 in the panel); the head count applies it to the UN count of women 20 to 24. Rich countries rarely run these surveys and so are black on both maps.

Violence against women is the share of women aged 15 to 49 who were subjected to physical or sexual violence by a current or former husband or partner in the past twelve months: the WHO’s 2023 estimates (SDG indicator 5.2.1, read from the UN SDG Global Database), modelled from population surveys that ask women directly, so they do not depend on whether anyone went to the police. The head count applies the share to the UN count of women 15 to 49. The panel adds the figure for all women 15 and over, the share of women who say a husband is justified in beating his wife (World Bank WDI), and rapes recorded by the police from the UN Office on Drugs and Crime’s Crime Trends Survey, as a count and per 100,000 people. Recorded rapes measure reporting and the law’s definition as much as the crime: Sweden and England record far more than countries where victims rarely report and the definition is narrow, so the survey estimate paints the map and the police figures sit beside it: sexual violence offences recorded per 100,000 people, rape among them, and women and girls recorded as victims where the police report victims by sex. The United States reports only rape to the UN, so its sexual violence line is blank. UNICEF’s survey of women aged 18 to 29 who suffered sexual violence before 18 is there too. No police count paints a map in these views. All layers share one colour scale, orange where life is hardest and blue where it is easiest; the layer buttons carry a small picture each rather than an abbreviation.

Social media

No source counts real and fake accounts by country: the platforms keep that to themselves, and academic bot estimates cover one platform at one moment. What can be measured is who spreads false information and what governments do online. The map’s first three pills come from the Digital Society Project (V-Dem Institute; 179 countries, 2000–2025), which asks country experts how often the government, the major parties and foreign governments use social media to spread misleading or false information, on a 0–4 scale from “extremely often” to “never”. The atlas shows the nearest answer in the coders’ own words, and darker means more often. The panel lists the project’s other questions for the country: shutdowns of the internet and of social media, monitoring, censorship, arrests for political posts, party hate speech and polarisation.

Internet freedom is Freedom House’s Freedom on the Net score (72 countries, 2024 edition): 0–100 across obstacles to access, limits on content and violations of user rights, with “free” from 70, “partly free” from 40. One of its questions (B5) asks whether pro-government commentators manipulate online discussion. Both sources are expert assessments, not counts of posts or accounts.

Timeline

The timeline shows one year at a time on the borders of that year (CShapes 2.0, 1886–2019; later years keep the 2019 map): the Soviet Union, Yugoslavia, colonial empires and divided states as they were. Each territory shows the deaths recorded for it that year, counted as elsewhere in the atlas, or its government under the chosen source. Curated figures, which are kept by today’s countries, are placed in the territory that ruled that country that year. Whole-war figures are spread evenly over a war’s years, so a single year’s figure is an average for that war.

Conflict names

Datasets name conflicts by code (“Senegal (Casamance)”, “KR: violence against civilians”). The atlas uses the Wikidata name where a conflict is confirmed to match one; otherwise a best guess built from the closest source (the Wikidata name of the larger war, or the dataset’s own territory, sides and full actor names), marked “name: best guess” with the dataset’s name on hover.

Countries carry today’s names, even where a source writes a former one (UCDP’s “DR Congo (Zaire)”, “Myanmar (Burma)”). The former names, and when they applied: Cambodia was Kampuchea (1975–1989); DR Congo was Zaire (1971–1997); Eswatini was Swaziland (until 2018); Iran was Persia (until 1935); Madagascar was the Malagasy Republic (1958–1975); Myanmar was Burma (until 1989); Russia was the Soviet Union (1922–1991, as its successor state); Serbia was Yugoslavia (until 2006, as its successor state); Sri Lanka was Ceylon (until 1972); Thailand was Siam (until 1939); Turkey was the Ottoman Empire (until 1922); Vietnam was North Vietnam (until unification in 1976); Yemen was North Yemen (until unification in 1990); Zimbabwe was Rhodesia (1965–1979). The Soviet Union and Yugoslavia were states of their own, not renamed ones, so conflicts that ended while they existed keep their names (“Soviet Union vs Hungary”); only later ones take the successor’s.

Home view

The globe opens on the visitor’s country. It is taken from the country code the hosting platform attaches to each request, or failing that from the browser’s language setting (“en-GB” is the United Kingdom); no third-party location service is asked and nothing is stored. Without either, the globe opens on Africa, Europe and the Middle East. The crosshair button returns to it.

Known limitations

Sources and licences

The atlas’s own curated data is CC BY 4.0. Some sources restrict use: CShapes and International IDEA are non-commercial and share-alike; Freedom House permits non-commercial use with attribution.