Counting Conflict-Related Homicides in Colombia from 1985–2018

Using advanced statistical models and multiple data sources, HRDAG statistician Maria Gargiulo and her colleagues were able to estimate that between 565,200 and 639,900 people were killed in Colombia during the country’s 30+ year armed conflict. This research was part of Maria’s doctoral dissertation and is designed to teach other statisticians how to use these advanced techniques.

That staggeringly large number is more than 30% higher than the number of known victims documented during that time, which speaks to how many people are missing from the data. This research, including details about the data sources and statistical reasoning, was recently published in Demography.

By the 1960s, political tensions in Colombia had crescendoed into armed conflicts between multiple groups that resulted in large-scale violence, including killings of hundreds of thousands of civilians and the disappearance of tens of thousands more. Human rights researchers have continued to study historical records and victim statements from Colombia to seek a true history of what unfolded during that violent period and the decades after.

Maria’s newly published research adds more statistical certainty to the historical understanding of what occurred from 1985–2018.

Maria’s analysis relied on multiple systems estimation, also known as capture-recapture, which is used to characterize populations across varied fields, including ecology, public health, and in conflict settings. Using these methods, she was able to account for not only the victims who were documented but those whose names and identities were never captured in documented data.

Below is a Q&A with Maria about some of the major findings from the study.

1. One of the key features of your research in Colombia is that you relied on so many disparate data sources. Can you give a couple examples of the data sources that you used to document deaths?

In this study and the larger project with the Truth Commission and the Special Jurisdiction for Peace, we leveraged over 110 data files compiled by over 40 different organizations including data compiled by government institutions, civil society organizations, and victims’ collectives. These files consisted of a series of specialized lists that only documented conflict-related violence and non-specialized lists that documented all forms of violence that occurred during the conflict. Examples of specialized lists include data compiled directly by the Truth Commission through interviews, as well as the database of the Centro Nacional de Memoria Histórica. Non-specialized lists include autopsy data from the Instituto Nacional de Medicina Legal y Ciencias Forenses and judicial case data from the Fiscalía General de la Nación. For these non-specialized lists, we used statistical procedures to classify records as to whether they likely documented conflict-related violence.   

2. Before data can be analyzed, you first cleaned and de-duplicated it to see if someone appeared across multiple lists. However, deduplicating had some challenges because one person might be listed differently in different databases. Can you talk about the process you used to de-duplicate data?

This process, called record linkage or database deduplication, is a common first step in multiple systems estimation studies. In the JEP-CEV-HRDAG Joint Project, we received nearly 13 million records of victims meeting the criteria established for our analytic work. Due to the sheer size of the data, we needed to use a series of statistical models to deduplicate the records and identify individual victims. In one of the record linkage steps, we trained a binary classification model using hand-labeled examples to determine whether two records in a record pair refer to the same individual. To do this, we generated over 115 comparisons using data from 45 fields. We considered both exact matches, as well as a variety of distance metrics that can help account for misspellings or data entry errors (e.g., names or dates in wrong order). We also considered phonic representations of names, comparisons that focused on just a few characters (e.g., first three characters or last five characters), as well as numeric comparisons of dates and ages. All of these features served as inputs to a statistical model that predicted whether two records were a match; no single feature made this determination.     

3. You estimate between 565,200 and 639,900 homicides in Colombia during the period of time you studied. Were all of those deaths related to the armed conflict or was that all deaths during that time?

This estimate accounts only for conflict-related deaths and does not describe homicides that occurred during this period as a result of other forms of  violence. This analytic choice was made in line with the remits of the Truth Commission and the Special Jurisdiction for Peace. However, it is worth noting that all violence exists along a continuum and it can be difficult—both conceptually and analytically—to differentiate conflict- and non-conflict-related violence.  

4. As you uncovered in your research, some people are never recorded in any data sources. Why might that be?

There are many reasons that instances of violence might go unrecorded. Surviving witnesses or loved ones of victims may not report violence due to fears for their safety, shame or stigma regarding the violence that occurred, or because they do not trust the organizations collecting reports. Lethal violence may leave no survivors able to report what happened and forced displacement or exile may move survivors far away from reporting channels. Conversely, institutions carrying out documentation work may not be able to access communities that have experienced violence and ongoing violence may also render data collection too dangerous. Documentation work may be further complicated by resource limitations—time, money, personnel—and is also impacted by organizational priorities, which shape what is documented and how documentation is carried out.

5. Could the methodologies you used to estimate deaths in Colombia be used in studying other violent conflicts?

Yes, and they already have been used in a number of other settings. Some examples include applications in El Salvador, Gaza, Guatemala, Peru, Sudan, and Timor-Leste. In Colombia, Valentina Rozo Angel and Patrick Ball have also used these methods to examine killings of social movement leaders. Along with Tarak Shah and Megan Price, we have applied these methods to estimate conflict-related mortality in Syria, including in an analysis cited in Michelle Bachelet’s report to the Human Rights Council in June 2022, as well as a study estimating deaths in custody. This latter study built on previous work by Megan Price, Anita Gohdes, and Patrick Ball. It is also worth noting that in conflict settings, these methods have implications beyond the study of homicides. During our work with the JEP-CEV-HRDAG Joint Project in Colombia we also applied these methods to examine victimization due to disappearance, kidnapping, and forced recruitment. 

6. If a data scientist or college student is interested in these topics, where would you point them?

For a general overview of MSE applications in public policy, I would recommend Sheila Bird and Ruth King’s article in the Annual Review of Statistics and its Application. “The statistics of genocide” by Patrick Ball and Megan Prices provides a deeper dive into how MSE was used to examine the genocide in Guatemala. Finally, Patrick Ball wrote a blogpost that is a great starting point for fitting MSE models. Students should also sign up for HRDAG’s email list and (if you are on any of these networks) follow us on Bluesky, LinkedIn and Instagram.

Image credit: Láminas matriz de la obra Fragmentos – Doris Salcedo (2018). (Rotated) CC-BY-4.0 


Our work has been used by truth commissions, international criminal tribunals, and non-governmental human rights organizations. We have worked with partners on projects on five continents.

Donate