Coding decisions on state forces, fatality counts, and more
Prior to the invasion, trench warfare between the Ukrainian armed forces and Russian-led separatists was limited to parts of the Donetsk and Luhansk regions (also collectively known as the Donbas) after a hot war in the wake of the Euromaidan protest movement in Ukraine in 2013-14, which Russia had framed as a Western-inspired coup d’état and used as a pretext to annex Crimea in March 2014. The conclusion of the Minsk Agreements in 2014-15 established a fragile ceasefire in the Donbas that held with varying degrees of success for about eight years. Despite hopes of a rapprochement following a change of government in Ukraine in 2019, Kyiv consistently resisted pressure to reintegrate the separatist-held areas on Russian terms. On 21 February 2022, after months of intimidating military buildup along Ukraine’s borders, Russia recognised the independence of its proxies in the Donetsk and Luhansk regions of Ukraine, paving the way for a direct clash between Russian and Ukrainian armed forces.
On 24 February 2022, Russia invaded Ukraine by land, air, and sea from the north, east, and south, with Kyiv being the prime target of the attack.1Mykhaylo Zabrodskyi et al, ‘Preliminary Lessons in Conventional Warfighting from Russia’s Invasion of Ukraine: February–July 2022,’ Royal United Services Institute for Defence and Security Studies, 30 November 2022 Ukrainian forces stopped Russian columns north of the capital and, by April 2022, forced their retreat across the border. In southern and eastern Ukraine, however, Russian forces quickly seized the entire Kherson and much of the Zaporizhia regions, the eastern part of the Kharkiv region and the southern part of the Donetsk region. By mid-summer 2022, Russia had largely occupied the Luhansk region but became embroiled in the northern part of the Donetsk region. In September 2022, Ukrainian forces liberated most of the Kharkiv region in a surprise counter-offensive. In response, Russia annexed the Donetsk, Kherson, Luhansk, and Zaporizhia regions of Ukraine. By mid-November 2022, beset by supply issues, Russian forces withdrew from the western bank of the Dnipro River and left Kherson city – the only regional center Russia managed to capture since the beginning of its invasion. Renewed Russian offensive in the Donetsk region in winter-spring 2023 mostly failed. In summer 2023, Ukrainian counter-offensive was met with stiff Russian resistance and proceeded slowly.
This expansion of armed conflict led to the highest levels of political violence recorded in a single country by ACLED in 2022 and has presented a number of methodological challenges for the tracking and recording of this violence, including how to handle the variety and obvious biases of sources reporting on war and, consequently, how to deal with differing fatality counts that are often aggregated across larger sections of the frontline. This document will seek to outline how ACLED handles each of these challenges in turn.
Russian and Ukrainian State forces
Since the full-scale invasion of Ukraine on 24 February 2022, the main actors operating in Ukraine are Russian and Ukrainian military forces. The following is a list of the main military actors coded in the data:
Ukrainian military forces:
- The Military Forces of Ukraine (2019-) actor, including separate actors for Ukrainian air force, navy, border guards, and marines, are the main actors representing the Ukrainian state forces in the conflict.
- Military Forces of Ukraine (2019-) National Guard refers to Ukrainian military units with law enforcement powers. Volunteer battalions that were incorporated in the National Guard, including the Azov Battalion, are coded as an associate actor to the National Guard actor.
- Military Forces of Ukraine (2019-) Territorial Defense Forces actor represents volunteer militias created across the country and incorporated into the armed forces of Ukraine that organize defense in the area of their operation.
Russian military forces:
- Military Forces of Russia (2000-), including their air and navy units, is a default actor representing Russia since the full-scale invasion. A separate actor Military Forces of Russia (2000-) Chechen Battalion of Ramzan Kadyrov is coded whenever pro-Russian Chechen fighters are involved in an event.
The actor Russian Occupation Government (Ukraine) is used for members of the occupation administration in areas controlled by Russia. They mainly consist of Ukrainian government members who have defected to Russia since 2022 and other people appointed by Russia to government positions in the occupied regions. This actor is not used for the Donetsk People’s Republic (DPR) and Luhansk People’s Republic (LPR) administrations, which have operated before the full-scale invasion on 24 February 2022 (as discussed in the next section).
The occupation administration in Crimea, formally incorporated into the Russian government, is coded as Government of Russia (2000-), with police forces coded as Police Forces of Russia (2000-).
Other pro-Russian or Russian-led actors
Most of the rebel units in the Donbas region operate as part of the United Armed Forces of Novorossiya (NAF), a nominal coalition structure under which their activities are coordinated. The specific units or groups, when reported, are coded as ‘Associated Actor’ to NAF: United Armed Forces of Novorossiya. In almost all cases, such associated actors are either the Donetsk People’s Militia or Luhansk People’s Militia, which are the two largest rebel groups belonging to the separatist DPR and LPR administrations, respectively. Following their formal incorporation into the regular Russian army in 31 December 2022, the DPR and LPR units are coded as units of the Russian military, as Military Forces of Russia (2000-) Donetsk People’s Militia and Military Forces of Russia (2000-) Luhansk People’s Militia, respectively.
In a few exceptional cases where rebel groups are explicitly not part of the NAF (for example, the volunteer Chechen “Death Battalion” (Батальон «Смерть») in Russian 2Maria Tsvetkova, ‘Chechens loyal to Russia fight alongside east Ukraine rebels,’ Reuters, 10 December 2014 ) , these are coded as distinct actors.
Ukrainian sources rarely specify whether Russian forces or Russian-led rebel groups in Donbas are involved in violence, but rather refer to both groups using terms such as “Russian forces” or “occupying forces”. Since Russia had mainly been operating through the NAF rebel formations prior to the full-scale invasion, the NAF was coded as the default Russian military actor up to 24 February 2022. With the deployment of Russian state forces in Ukraine since February 2022, the default actor for Russian military activity in Ukraine is now the military. Where sources explicitly mention the involvement of the DPR and LPR actors in the Donbass, these actors are coded accordingly.
DPR and LPR have their own self-styled government and police incorporated in the ACLED dataset. This is to distinguish the long-standing rebel administrations formed in 2014 from the newer post-invasion Russian occupation governments.
Ukrainian police forces that cooperate with Russian forces are coded as Police Forces of Ukraine (2019-) with Military Forces of Russia (2000-) added as an associated actor. Similarly, civilians that cooperate with Russian forces and are not members of Russian occupation governments or the L/DPR administrations are coded with Military Forces of Russia (2000-) as an associate actor.
The Wagner Group, a Russian government-aligned private military company (PMC), operated alongside Russian forces between April 2022 and June 2023, primarily in the Donetsk region (read more on Wagner here). Prior to the Russian invasion of Ukraine, there was little information about the activities of the Wagner Group, likely due to underreporting. However, due to various developments during the Wagner Group’s involvement in Ukraine, including its self-promotion and standoff with Russian military leadership in June 2023, media coverage of the group’s activities greatly increased. As of September 2023, ACLED data show that the Wagner Group’s activity in Ukraine has been limited to the battle for Bakhmut.
Ukrainian partisan groups are pro-Ukrainian communal militias in Russian-occupied areas of Ukraine that target Russian military sites, commanders, and occupation officials. When a report suggests the involvement of a Ukrainian partisan group, these are most often coded as communal militias. This usually includes remote violence or attacks against Russian forces, Russian occupation authorities, or groups cooperating with Russian forces in the occupied regions. Communal militias are normally coded with the location they operate in (e.g. Melitopol Communal Militia (Ukraine)).
Some partisan groups operate under a more well-structured organization with a more defined political goal. Whenever this information is available, ACLED codes such groups by their own name, e.g. Atesh.
Unidentified actors, or actors with “Unidentified” as a prefix to their name (e.g. Unidentified Armed Group (COUNTRY) are coded whenever there is not enough information available to code a more specific actor.
The Unidentified Military Forces actor is coded whenever the activity of an unidentified actor is akin to that of Russian or Ukrainian armed forces. For example, when an actor is involved in shelling or the laying of mines.
The Unidentified Armed Group (Ukraine) actor is coded when no clear actor is identified by sources and there is insufficient information to suggest that state military forces were involved. Since the start of the full-scale invasion, this actor has most often been coded in events involving the use of improvised explosive devices (IEDs) and attacks on pro-Russian groups in occupied territories.
Before the full-scale invasion
Before the full-scale invasion, there were several key sources reporting on ceasefire violations in the Donbass region. The operational commands of the Ukrainian armed forces, the LPR militia, and the DPR militia each provided daily operational reports that systematically detailed ceasefire violations targeting territories under their control. The daily reports of the OSCE’s Special Monitoring Mission to Ukraine, which focused on ceasefire violations at various points along the front line, was another key source reporting on the conflict. Other media sources tended to repeat information from these key sources.
After the full-scale invasion
After the full-scale invasion, the media landscape of Ukraine changed drastically. Due to the rapid changes in the dynamics of the conflict, including the emergence of new actors and the movement of the frontlines, it was no longer sufficient to rely primarily on military reports to adequately cover events in Ukraine. Additionally, the OSCE mandate in Ukraine expired on 31 March 2022 (OSCE, 28 April 2022), and the publishing of disaggregated daily reports of ceasefire violations stopped on 23 February 2022.
Current sources include:
- Military reports of the Ukrainian Ministry of Defence, LPR and DPR militias, and Ministry of Defence of Russia are covered.3ACLED treats the number of fatalities claimed by the Russian MoD differently, see the Fatality Reporting section below.
- Ukrainian news media that re-report information from regional military administrations and other Ukrainian governmental and military officials, military bloggers, activists, as well as do their own investigations
- Ukrainian NGOs and human right groups
- Reports of international organizations
- Ukrainian political parties and groups whose members joined Ukrainian forces
International news media, international organizations, and Ukrainian and international non-governmental organizations and human rights groups that produce longer and often delayed reports are not covered in real-time. These reports are covered at least a week later and used to verify previously coded events, as well as add additional information and new events.
Disaggregation and Event Types
The information available in the daily operational reports provided by the sides is often limited to a list noting which of their positions were fired upon with certain weapon types by the opposite side on a particular day. The reports generally do not report the direction of fire or any return-fire on their part. Because this information is unknown, the weapon type is used to distinguish between the event types Battles (small arms indicating close range combat and exchanges of fire) and Explosions/Remote violence (for one sided engagements using remote weapon types – as per general ACLED methodology).
Where small arms fire is reported (which includes sniper rifles and light machine guns, among others), with no other information present to indicate a particular interaction type, 4E.g. firing at civilians would be coded as Violence against civilians with the sub-event type Attack as per regular coding rules. ACLED interprets this as a firefight between the military forces of Ukraine and either NAF rebels (prior to 24 February 2022) or Russian forces (after 24 February 2022) by default, and hence codes this as event type: Battles, sub-event type: Armed clash. Unless other information is available, it is assumed that small arms fire reported along the frontline is directed at the opposite actor.
Where the sole use of heavy weapons is reported (which includes automatic grenade launchers [AGL], recoilless guns, heavy machine guns [HMG], weapons mounted on armored combat vehicles, mortars, cannons, artillery, and various missile systems, among others), it is coded under event type: Explosions/Remote violence, sub-event type: Shelling/artillery/missile attack. If civilians are injured or killed by the shelling or there is another indication civilians were directly affected by the shelling (e.g. a functioning hospital or an evacuation column was hit), civilians are coded as an actor as well: either as Actor 2 when solely targeted or as Associated Actor 2 if they are affected in addition to an armed group. If the shelling is directed ‘at a town’, or if some other general description is given that does not indicate actors being specifically targeted, no second actor is coded. But in all cases the event type remains: Explosions/Remote violence with sub-event type: Shelling/artillery/missile attack.
Where unknown weapons are reported, these are assessed according to any additional context available. For example, unknown weapons causing explosions or described as projectiles or missiles, are coded with the Explosions/Remote violence event type and the Shelling/artillery/missile attack sub-event type. If no such additional information is available, the coding will default to the most prevalent type of arms used in the conflict Prior to the full-scale invasion these were small arms, and hence the default event type was: Battles, sub-event type: Armed clash.5Assessment based on a sample of OSCE SMM-Ukraine ceasefire violation observation statistics. After the full-scale invasion, the use of heavy weapons became much more prevalent, with reports mentioning small arms or sniper fire on very rare occasions, therefore events involving unknown weapons are currently coded as event type: Explosions/Remote violence, sub-event type: Shelling/artillery/missile attack.
Whenever both sides report that the opposing forces fired at the same location during the day, the event is coded as a single Armed clash event. For example, for the event with the notes “On 1 August 2022, Russian forces launched an airstrike and shelled with artillery and tanks near Novoluhanske, Donetsk. On the same day, Ukrainian forces fired from tanks at DPR positions in Novoluhanske,” it is assumed that all actors – Russian and DPR forces on one side and Ukrainian forces on the other side – interacted (exchanged fire), even though each side only reported shelling or airstrike conducted by their adversary.
High-fatality region-wide events are usually coded as Armed clashes (e.g., “On 4 August 2022, Ukrainian forces killed 39 Russian soldiers on the Southern front (coded to Kherson, Kherson)”), unless other more specific information is available, as these events are likely aggregations of various event types where armed clash would be the highest on the coding hierarchy.
Tracking fatalities is one of the most difficult aspects of conflict data collection in general, as fatality counts are frequently the most biased, inconsistent, and poorly documented components of conflict reporting. This is especially true of active conflict environments impacted by high levels of mis/disinformation and severe access constraints. ACLED defaults to conservative estimates based on the best available information at the time of coding in line with our specific event-based methodology and review process. ACLED estimates are restricted to fatalities reported during individual events, meaning that these estimates may be particularly conservative in comparison with sources that do not use an event-based methodology. This means that aggregate tallies provided by sources such as hospitals and government agencies, for example, which cannot be broken down and connected to individual conflict incidents, are not included in the ACLED dataset. When and where possible, ACLED researchers seek out information to triangulate the numbers from any report, but ACLED does not independently verify fatalities. ACLED is also a ‘living dataset’, so all fatality figures are revised and corrected — upward or downward — if new or better information becomes available; in some conflict contexts, this can be months or years after an event has taken place. These figures should therefore be understood as indicative estimates rather than definitive fatality counts (see this article for more on ACLED’s approach to coding fatalities). ACLED additionally only captures fatalities that are directly caused by political violence; indirect conflict-related fatalities caused by disease or starvation, for example, are not included in these estimates. Other sources may come to different figures due to differing methodologies and catchments.
Due to the unreliable nature of fatality estimates in high-intensity conflict environment
ss, including the conflict in Ukraine, ACLED advises users to use metrics other than fatalities to assess the intensity of a conflict, such as the number of violent events.
Civilian fatalities are generally well-reported on in territory controlled by the Ukrainian government. ACLED codes civilian fatalities as reported by news sources and updates fatality numbers based on newest reports. When territory is regained by Ukrainian forces, information regarding civilians killed during the occupation period surfaces. This information is often aggregated or not detailed enough to cross reference and attribute to specific events. In such cases, ACLED codes fatalities when the information available allows us to create an ACLED event or distributes the reported fatality numbers over existing ACLED events based on an approximate location (up to the admin1) and time period (up to two months), taking into account fatalities that have already been recorded to prevent double-counting).
Information on military fatalities became very scarce since the start of the full-scale invasion on 24 February 2022. Most military fatality estimates come from reports about ‘enemy’ fatalities, which can often be exaggerated for operational and propaganda reasons. For instance, Russian Ministry of Defense’s allegations of the Ukrainian military equipment loss and troops killed in action are not verifiable but likely wildly off the mark. Meanwhile, the Ukrainian General Staff daily tally claims under 250,000 “liquidated” personnel since the beginning of the Russian full scale invasion in February 2022 until the beginning of August 2023.6 Ministry of Defense of Ukraine, ‘The total combat losses of the enemy from 24.02.2022 to 03.08.2023,’ 3 August 2023 Furthermore, while assessments of losses on the Russian side are abundant, with most of them estimating Russian military fatalities to have reached at least 50,000 by July 2023 (excluding forcibly mobilized residents of the occupied Ukrainian territories and assault detachments formed of prison recruits), 7Re: Russia, ‘Death Toll: The number of Russian soldiers killed in the war with Ukraine is likely in the range of 50,000-65,000 individuals,’ 19 July 2023; Meduza, ‘Bring out your dead,’ 10 July 2023 independent estimates for Ukrainian military fatalities are exceptionally hard to come by. One approach is to assume that the figure is comparable to assessed Russian combat deaths given the heavy use of artillery on both sides, with some arguing that it could have reached over 65,000 in the first year of the war.8Small Arms Survey, ‘Russia’s War: Weighing the Human Cost in Ukraine,’ 15 May 2023 This contrasts with the Western estimate of 100,000 killed and injured proffered in late 2022.9 The Kyiv Independent, ‘EU clarifies Ukrainian casualty estimates in von der Leyen’s speech: 100,000 figure referred to dead and wounded,’ 30 November 2022 As explained in the sections above, ACLED does not incorporate these aggregated fatality statistics into its own data and instead uses fatality claims it can trace to a specific event by time and place or approximation thereof. While this results in granular distributions of fatalities over time and place and enables deeper analysis, but it also means ACLED reported fatalities can be lower than some of the aggregated counts for the entire conflict as reported by other sources.
As a rule, in case of conflicting military fatality reports, ACLED codes the most conservative (lowest) available fatalities estimate for an event, unless the source providing a higher estimate is considered much more reliable. In some cases, fatality claims are deemed too unreliable to code directly. The Russian Ministry of Defence regularly reports on armed clashes, shelling, airstrikes, and general offensives in Ukraine. While the reporting on these engagements having taken place is generally reliable10There are cases of disinformation relating to air- and missile strikes in particular that ACLED accounts for., the claimed Ukrainian military fatalities inflicted are not reliable. The Russian MoD regularly claims high fatality counts among the Ukrainian military resulting from Russian artillery strikes and offensives, but these claims cannot be corroborated by independent sources or overall expert assessments of the deadliness of the conflict. Possible explanations are that the number is inflated by counting injured personnel as ‘fatalities’, by overestimating the effectiveness and deadliness of a strike in general, and possibly by fabricating fatalities outright. In order to mitigate this problem of exaggerated fatality counts, ACLED limits how it codes Russian MoD claims of Ukrainian military fatalities that cannot be verified through other sources. Claims that exceed 10 fatalities are capped and recorded as being ‘10’ in the Fatalities column for that claim to prevent the propagation of manifestly exaggerated counts. ACLED has not found similar overestimations of Russian military fatalities in Ukrainian government claims regarding specific events – although these might exist for aggregated fatality claims (Radford et al, 2023).
In general, when sources report that fatalities took place during an event but do not specify the number, ACLED will assume the number of fatalities as three or 10, depending on the context. 10 fatalities are coded for events that are usually larger in scale and incur more fatalities. This would normally include armed clashes or shelling/artillery/missile attacks targeting large concentrations of forces. Three fatalities are coded for smaller scale events that would likely incur between three to nine fatalities, e.g. an armed clash involving a smaller reconnaissance group or shelling/artillery/missile attacks targeting a border checkpoint. Whenever a report mentions that there were casualties, without specifying whether somebody was killed or only injured, fatalities are coded as 0.
Sometimes, fatalities mentioned in a report are not tied to a particular location. Instead, the report mentions several targeted locations and an aggregated number of fatalities. In such cases, the reported fatalities are split among the locations reported in that source report, with each event belonging to that split specifically mentioning it in their Notes — for example, “During the day, three Ukrainian soldiers were killed. [3 fatalities split among 4 events]”.
Operational reports generally only report that their own positions are being hit, so if ‘enemy’ casualties are reported, it implies that there must have been return fire to cause these. In such cases, the fatalities are distributed only among the events that could have contained such return fire — i.e. those coded as Battles, given that Explosions/Remote violence events are asymmetrical.
Tracking the Reporting of Ceasefire Violations by the OSCE Special Monitoring Mission in Ukraine (SMM) Prior to February 2022
Prior to the full scale invasion of February 2022, the use of ceasefire violation data published by OSCE SMM enabled ACLED to provide a more comprehensive and detailed overview of violence along the frontlines in Donbass. The OSCE SMM officially discontinued its operations on 31 March 2022. More than half of the Battles and Explosions/Remote violence events coded by ACLED between the start of the dataset (1 January 2018) and the start of the full-scale invasion (24 February 2022) in the Donbass region are uniquely generated from SMM data, providing a more granular overview of areas where fighting occurs. In addition, SMM data is used to cross-reference the reports of the operational commands of the Ukrainian military and the DPR and LPR militias that tend to report primarily on ceasefire violations committed by the other side. SMM data provides third-party verification for over 20% of the events reported by the operational commands prior to the full-scale invasion. ACLED uses this data to update events with information on whether the SMM observed return-fire and the use of particular weapon types.
Prior to February 2022, the SMM monitored ceasefire violations in Donbass through a combination of static and mobile observation teams, automated static acoustic and visual sensors posted mainly along the frontlines, and through reconnaissance drone flights. Monitoring occurs around the clock, with daily reports covering incidents over the previous 24 hours as reported up to 19:30 local time the day before. The ceasefire observations are summarized in the main daily report with the details of each observation published in an annex attached to the report. The annex contains information on eight variables for each observed ceasefire violation, as can be seen in the table snippet below.
In order to code SMM observations using ACLED methodology, ACLED uses the same assumptions that apply to the rest of the reporting on the conflict in the Donbass region. ACLED converts SMM data into a format more readily transcribed using ACLED methodology and then aggregates observations to establish a single ACLED event for each location for each day.
For each line in the annex table shown above, ACLED converts SMM data into the following ACLED format:
Event type: When no other information is available, ACLED uses weapon type to determine the Event type, with small arms interpreted as event type: Battles, sub-event type: Armed clash. If the event contains observations with both actors involved, it is also coded as event type Battles with sub event type Armed Clash. If the event contains only incidents of heavy weapons use by either Ukrainian forces or NAF, then the event type Explosions/Remote violence with sub-event type Shelling/artillery/missile attack is coded.
Actors: The SMM does not attribute ceasefire violations to actors. Given the static nature of the frontline before the full-scale invasion on 24 February 2022, ACLED uses direction of fire and control of territory to assume Actor 1 and Actor 2. For example: A projectile is observed by the SMM camera in Svitlodarsk (government-controlled, 57km NE of Donetsk) with ‘description’ SE to NE. To the southeast of that ‘SMM position’ is NAF-held territory. Therefore, Actor 1 would be coded as NAF.
If there is no direction of fire in the ‘description’, the location of the event is used to determine who is targeting whom. If the location is within government-held territory, NAF is assumed as a perpetrator (unless the observation is described as ‘outgoing,’ i.e. originating from that location) and vice versa.
Because the information provided by SMM does not provide information on whether the targeted location also contained actors — e.g. one of the main armed actors or civilians — the second actor in the event type Explosions/Remote violence will be empty in most cases. Only by relying on information from other sources can a second actor be established. For example, if an operational report for that same day states that NAF positions, civilians, or civilian housing were hit at that same location, then they would be coded as Actor 2. For the Battles event type, the presence of a second actor is assumed, as described above.
Location: One of the more complicated aspects of dealing with SMM reporting is that the ‘SMM position’ is not the location of the actual event. In addition, ACLED codes events to the nearest populated place when no other specific information about a location is reported. This being the case, when coding SMM reporting, ACLED determines the event’s location in three steps. First, the coordinates of the ‘SMM position’ are determined. Second, the compass direction and distance of the ‘Event location’ column are used to establish the exact location. Third, the event is coded to the nearest populated place to those coordinates with geoprecision 2.
For example: An explosion is observed by the SMM camera in Svitlodarsk (government-controlled, 57km NE of Donetsk) from ‘SMM position’ at coordinates 48.4323,38.2218 (A), occurring 4-5 km to the southeast at ‘event location’ coordinates 48.4065,38.2528 (B) which is coded to ACLED location Luhanske located at coordinates 48.4305,38.2554 (C) as the nearest populated place at geoprecision 2. The following image illustrates this example.
Notes: The notes specify the details of all SMM observations aggregated into a single ACLED event. This includes the sum of the number of observed shots/projectiles/explosions, the involved observation types, the involved weapon types, and whether fire was observed coming from both directions of the frontline.
For example: an ACLED event based on three SMM observations, consisting of 5 explosions from artillery and 3 explosions from unknown weapons and 25 shots from small arms, is summarized in the Notes as “33 shots/explosions from artillery/unknown weapons/small arms.”