Uses and limitations of ACLED data
Published on: 27 February 2023 | Last updated: 7 May 2024
Q. Does ACLED have a minimum fatality criterion for events?
A. No – ACLED does not have a fatality threshold for an event’s inclusion in the dataset. This means that political violence and demonstration events in ACLED do not need to produce any fatalities in order to be included as valid events. No arbitrary number of deaths is used to define a conflict.
Q. How does ACLED code fatalities?
A. ACLED codes fatalities from the most reliable possible sources and/or partners in a given conflict environment. Further, researchers seek out information to triangulate – where and when possible – the fatality numbers from reports.
The total number of fatalities for each event is found in the ‘Fatalities’ column. The ‘Fatalities’ column refers to the estimated number of reported fatalities associated with a single event. ACLED does not collect data on injuries, nor are fatalities attributed to specific groups (for more, see this section below), but such information may be mentioned in the ‘Notes’ column. All fatalities recorded are ‘reported fatalities’ and, as such, are estimates only. ACLED does not independently verify details of reported fatalities; instead, it reflects the content of the source reports.
If source reports differ or a vague estimate is provided, ACLED uses the most conservative estimate available and seeks to note in the ‘Notes’ column when there has been a dispute. When better information becomes available, these totals will be revised and corrected – upward or downward. When sources report estimates such as “tens” or “dozens” or “hundreds,” ACLED codes 10 for “tens,” 12 for “dozens,” 100 for “hundreds,” and so on (see the following section for how ACLED codes unknown fatality numbers).
When summarized fatalities are reported but events occur across several days or in multiple locations simultaneously (e.g. “12 fatalities result from fighting over a span of three days in one location”), the total number of fatalities is divided equally across days and locations of the event (4 fatalities per battle day for the one location, in the example above). If an odd number (including 1), the proportion of fatalities is divided by assigning the first day the additional fatality and distributed as evenly as possible. Such disaggregation of fatalities is recorded in the ‘Notes’ column. It is important to remember that though the ‘Notes’ column entries may match across these events, these are not duplicates but rather are unique events with different locations or dates coded.
Q. What about when the number of fatalities is not known?
A. When a report does not indicate whether or not any fatalities occurred or notes that it is unknown whether fatalities occurred at all, ACLED defaults to coding 0 as the fatality estimate. ACLED distinguishes between ‘fatalities’ and ‘casualties.’ Fatalities are deaths, whereas casualties are assumed to be injuries or fatalities. As such, if a report only indicates “casualties” (or similar ambiguous wording in other languages), the conservative approach that ACLED takes is to assume all casualties are injuries and hence code 0 fatalities.
When a report indicates that an event did indeed lead to fatalities, yet there is no additional information on how many fatalities may have occurred, ACLED codes this unspecified number of fatalities as either 3 or 10 fatalities depending on the circumstances of the event. If the information in the event and the general understanding of the conflict context indicates the event has a high death toll (likely over 10), then 10 fatalities are coded. For example, “Houthi forces launched a large-scale attack against IRG forces in Yemen resulting in an unspecified number of fatalities” is coded as 10 fatalities, as similar battles in the area have led to high (more than 10) fatalities. In all other cases where the circumstances point to a death toll likely less than 10, a more conservative estimate of 3 fatalities is coded. This approach allows ACLED to avoid undercounting fatalities when it is known that some occurred, yet also prevents overcounting fatalities when the actual number is unknown. For example, “a bomb explodes at a checkpoint in Afghanistan, resulting in deaths” is coded as 3 fatalities because these checkpoints are often manned by fewer than 10 people, making it likely fewer than 10 fatalities occurred.
Q. Does ACLED methodology result in underestimating fatalities?
A. Possibly – given that ACLED defers to the most conservative number of fatalities reported and treats uncertainty around unspecified numbers of fatalities conservatively as well (for more, see this section above). But ACLED aims to provide the best, if lower, estimate of fatalities, rather than entirely arbitrary ranges built on assumptions.
Q. Does ACLED track fatalities according to which groups caused or suffered fatalities?
A. ACLED does not code reported fatalities according to which group suffered fatalities, nor according to the number of people killed by a specific group. Most source reports do not provide this level of detail; when they do, the claims can be highly biased. Instead, when available, the total estimated number of deaths arising from an event is coded in the ‘Fatalities’ column of each event. For this reason, the data cannot generally be used to estimate the number of deaths caused or suffered by one actor or another in a conflict, as a single event may contain information on fatalities caused or suffered by both parties in a battle.
The exception to this is events targeting civilians and protesters, who are by definition not engaging in violence themselves.1See the section on Actors in the ACLED Codebook. Therefore, the number of fatalities reported for each event involving civilians or protesters as ‘Actor 2’ can be understood as the number of civilians or protesters killed.2In rare cases, deaths of violent actors are inadvertently included in civilian fatality totals. For example, in suicide bombings against civilians, the perpetrator’s death is also included in the fatality number of the event.
As such, aggregate estimates of “civilian fatalities” in ACLED’s curated data 3ACLED provides a curated dataset that includes all violence targeting civilians and protesters into a single file, available for download here.do not include civilians that may have died as ‘collateral damage’ during fighting between armed groups or as a result of the remote targeting of armed groups (e.g. an air strike hitting militant positions that also kills civilians). These collateral damage fatalities are still recorded in both the ‘Fatalities’ and ‘Notes’ columns — meaning that any analysis of the total impact on civilians, including battles, will require additional modification on the part of the user, along with a number of assumptions as to how to attempt to disaggregate these total event fatality numbers.
Q. What are the known biases of fatality numbers?
A. ACLED emphasizes that fatalities are often a poor approximation of a conflict’s form and impact. They are often debated and can vary widely. Conflict actors may overstate or under-report fatalities to appear strong to the opposition or to minimize international backlash. Additionally, the numbers can be unreliable simply because it is difficult to collect exact data mid-conflict – especially if doing so in real-time.
The true cost of conflict cannot be measured by deaths alone. Conflicts that result in fewer deaths may still cause instability, which can ultimately result in additional deaths – for example, from food insecurity or lack of access to medical facilities. Further, battlefield deaths are biased toward men’s experiences of armed conflict. While more men may be killed in fighting or may be targeted to reduce populations from which opposition can recruit reinforcements, women and children are often the victims of sexual violence and other forms of violence ‘off the battlefield.’ These may not necessarily result in death, and concentrating on such disregards those experiences if we were to rely on fatality counts alone.
Q. How reliable are fatality numbers generally?
A. Fatality numbers are not consistently reliable from any source. All reported fatalities, from all forms of media and partners, are estimates. Some media are better at estimates, some are worse. Some (e.g. international media) will report on stories only if they reach a certain fatality threshold that elicits audience attention. Other information sources (e.g. in-depth human rights reporting by international non-governmental organizations) will concentrate on attacks on civilians, regardless of fatalities. ACLED incorporates all forms of media and reporting to maximize the accuracy, reliability, and thoroughness of each component of its data, including fatalities if they occur.
ACLED advises caution in using fatality numbers from any conflict data source. Fatality information is the most biased and least accurate part of any conflict report, and extreme caution should be employed when using a fatality number from any source.
Q. What alternatives exist to assessing a conflict besides fatality numbers?
A. ‘Deadliness’ of a conflict is essentially a proxy for the threat it poses to civilians. But the threat of a conflict is multifaceted; it is based on where violence occurs, over which time period, who is targeted, how they are targeted, and why they are targeted. Fatalities alone do not portray these complexities.
Event counts can be one alternative to fatality counts. The relationship between conflict events and fatalities is not consistent. Some conflicts have relatively fewer events yet the number of reported fatalities is high, such as in Afghanistan. Meanwhile, others have many events yet yield lower reports of fatalities, such as in Ukraine, or in India where mob violence especially is active yet events often do not yield fatalities. This variation points to how fatality counts may be important, but they are not a determinative measure of violence.
ACLED analysis draws on a variety of different measures to understand conflict patterns. We use event types, locations of activity, changes in rates of violence and its volatility, new targets of violence and interactions, actor types and changes in modality of conflict, rates of conflict against the state and civilians, infighting amongst groups, etc. We encourage our users to take advantage of the full range of conflict metrics and patterns that can be discerned from the ACLED dataset using our curated datasets and Early Warning Tools.
Q. What is the process for dealing with uncertainty and how does ACLED mitigate it, if possible? Can you just increase the number of sources in order to mitigate this uncertainty?
A. All sources have biases. Some biases are very apparent; propaganda, misinformation, and ‘fake news’ may have an obvious strong right or left political orientation. Other biases, however, may be less apparent. For example, international and national media may be biased toward reporting larger-scale events, like highly lethal events. New media, such as Telegram channels or Twitter, tend to report more regularly in urban and heavily populated spaces. Well-researched reports tend to prioritize specific types of violence, like human rights violations, while leaving out uncorroborated events. Local organizations rely on their established networks to report information, which may be limited to specific subnational spaces.
Simply increasing the number and types of sources will simply perpetuate these patterns. Hence, while it may seem intuitive that more reports lead to increased reliability, ACLED does not seek to simply increase the number of sources as a means to improve reliability. The quantity of information does not ensure quality. In fact, more sources may lead to data of a lesser quality as inherent biases will be amplified (for more on this relationship, see the section titled “Do more sources mean that data are more reliable?” in this FAQ on ACLED’s sourcing methodology).
To account for these various biases of sources, ACLED develops tailor-made sourcing processes per country and region. The goal is to generate particular source combinations that reflect the reality of disorder and fatality counts in each space. One strategy that ACLED has found to be especially fruitful is the prioritization of reports from local sources, such as local partners and subnational media. These sources mitigate the effects of more traditional, international, often English-language media, that often sensationalize the lethality of events for their audience (for more, see this primer on ACLED sourcing methodology).
Additionally, ACLED’s coding and review processes mean that data are continuously updated to reflect the most recent information. For example, if a bombing results in high casualties, ACLED will record the number of fatalities reported, but will amend this in future weeks if the death toll increases as reports arise of individuals succumbing to their injuries. In this way, ACLED is a ‘living dataset.’ This differs from datasets with an annual publication schedule in which events are not regularly updated to reflect the latest information.
Q. Does ACLED have processes in place for quality control over fatality estimates?
A. ACLED takes a number of steps to ensure quality control of fatality estimates. This is to ensure that while fatality estimates remain biased, the estimates reported by ACLED are as reliable as possible. Internal source weighting is also helpful in increasing the reliability of fatality estimates. All sources have biases; these biases vary.
Regular consultations are one means of doing this. ACLED regularly commissions experts to provide feedback on data coverage for various countries/regions, sharing feedback not only on fatality counts but also on other aspects of data collection (e.g. trends, actors, sourcing, etc.). In addition, partnerships with organizations with expertise on specific contexts also ensure that trends in the data, including fatalities, are a fair and reliable representation of local dynamics.
