All records included in CANDID were generated from publicly available data sources. Primary sources are: (1) decedent or vital death records; (2) fatality reports issued by state or local government agencies, including child welfare agencies, children’s ombudsman or advocate offices, and police departments or prosecutor’s offices; (3) media articles; and (4) archives produced by other entities. We continue to update source listings as older media reports are identified and new media coverage emerges, and as we continue to input existing and new state case notifications and reviews. Given the use of publicly available information, this project was deemed exempt by the UNC Chapel Hill’s Institutional Research Board (IRB).
We searched for states and jurisdictions in which basic decedent information is considered part of the public record (and thus accessible and disclosable). In some jurisdictions, this information is available for download on the Medical Examiner or other government website, and in other jurisdictions we had to file a request under the state’s public disclosure laws. Disclosable information varies by jurisdiction, but generally includes decedent age, date of death, and cause/manner of death. Availability of child name and other demographic or contextual information varies by state. Note: In some jurisdictions, requests must be made county-by-county, rather than statewide. In those states, we have primarily requested records from large urban counties thus far.
A minority of states publish case notifications, summaries or reviews of child maltreatment deaths. Specifically, Arizona, Colorado, Florida, New Jersey, New York, Oregon, Pennsylvania, Tennessee, Washington, and Wisconsin publish reviews of certain suspected or confirmed fatality cases. The set of cases covered varies widely from all suspected maltreatment fatalities to confirmed maltreatment fatalities in families that had contact with child welfare within a year of the fatality. A few other states (Arkansas, Nevada, and Rhode Island) publish notifications of suspected maltreatment fatalities but no case reviews. Indiana and Virginia include limited case-specific information in annual child maltreatment fatality reports. (See the individual state pages for links to the information that each state provides.) We cross-check each state notification or review against all cases known from media reports so as to avoid duplication and add to the information already available on a case for which we previously had media sources only. Many of these cases have not been covered in the media. We are in the process of adding these cases to CANDID with the help of volunteers who are interested in serving as a conduit for information from their states. We welcome offers of assistance from advocates and scholars who want to help enter information from their states.
The approach to identifying media sources has evolved over time and continues to be refined as new technologies emerge. The media sources listed on the site to date were identified through a variety of search strategies. Media searches for new case identification were initially generalized, considering a range of search terms but not location-specific. However, as the dataset expanded, there was a need for state/city specific searches as these were found to yield new articles. General searches were more likely to yield articles in national / large news outlets about high-profile cases, whereas geographically targeted searches yielded more local news articles and were more likely to identify deaths that received limited coverage.
We explored a number of alternative search engines (e.g., Lexis) and web-scraping tools, and some media articles were identified or added through that process, but they were ultimately not particularly effective and were discontinued.
In addition, we also used media searches to resolve inconclusive records from other sources. We conducted child-specific searches for children identified in decedent records with deaths that may or may not qualify. For example, when a child is listed as dying of gunshot wounds and it is listed as homicide, we sought to determine whether the shooter was a caregiver or household member, or otherwise implicated parental negligence, as a child killed in a drive-by shooting or other event involving violence in the community is not considered a maltreatment-related fatality. These searches were largely limited to the types of deaths that tend to be covered by media – gun-related incidents, drownings, hyperthermia (often “hot car” deaths), and motor vehicle accidents. Other types of deaths that may or may not involve parental negligence, such as suffocation (sleep-related) deaths, constitute a large share of deaths in the decedent records but are too infrequently covered by media for child-specific searches to be viable.
We periodically review the following archives for deaths that meet criteria for the LCS project:
Broadly, we rely on the National Child Abuse and Neglect Data System (NCANDS) definition of a maltreatment death as the death of a child as a result of abuse or neglect because of: “(a) an injury resulting from abuse and/or neglect was the cause of death; or (b) abuse and/or neglect were contributing factors to the cause of death” (p.36). Official counting of child maltreatment fatalities is limited to cases in which a person responsible for the death was a parent, adult household member, or otherwise responsible for the care of the child. Cases involving children killed by strangers or individuals not in caregiving roles are not included unless there was evidence of caregiver negligence. Given our reliance on media coverage to identify children who died of maltreatment, most cases in CANDID required little review: the media is most likely to cover deaths in which the manner was classified as a homicide and/or where there was an arrest of a parent or caretaker.
We encourage interested users to review cases and alert us to any you think we have inaccurately included based on other information that may be available! We welcome feedback and corrections. And, of course, please let us know if there are cases we have missed.
Each child listed on LCS includes one or more of the sources identified to date. We cross reference potential new sources with existing CANDID records to resolve duplicate entries and expand the amount of information available. Because each source may involve different sets of unique information (e.g., one record may include child age, location and date of death, whereas another may include name or cause of death but not age or location), we review all overlapping identifiers for possible exclusion.
We use a Large Language Model (LLM) to read and categorize all child maltreatment-related fatality cases from 2022 onward in our database. The LLM uses all sources of information we have collected for each record.
The LLM is designed to identify features of each death — such as the contribution of torture, inadequate supervision, medical neglect, or other forms of maltreatment. The LLM also characterizes who was involved (e.g., biological parents, relatives) and information known about specific risk factors such as substance use or domestic violence. The list of attributes that the LLM searches for in each CANDID record were developed by the PIs in consultation with the advisory board. Initial runs of the LLM were reviewed manually to determine potential systematic errors in classification and the model prompts were revised and retested.
The LLM coding was run using Opus 4.5 run through Anthropic API.
Many media reports do not include children’s names. In those cases, we record the age and (if available) gender of the child. Most reports do include the name of a parent or adult perpetrator and those are recorded and cross-referenced in the data to make sure unnamed children are not entered multiple times from different sources. For those children whose names are already a part of the public record, we report that information in full. If we identified more than one version of a child’s name in media coverage, we listed the spelling we thought most likely based on attempts to identify additional information.
A child’s age is recorded in CANDID based on their age at the time of death. We have done our best to accurately report a child’s exact age based on information available in source documents. In situations where we were unable to determine a child’s age, we simply defaulted to the age description from the source (e.g., toddler, child).
There are often significant lags between a fatality and the release of information that allows us to categorize the death as maltreatment-related. As a result, earlier years will contain more deaths than later years: this should NOT be interpreted as evidence of a decline in deaths over time.
We update periodically as new information comes available. Children are included based on the date the death occurred, not the injury incident date or the date the death was reported to the public. When the exact date of death was unclear, but death was clearly recent and occurred after 2022, we recorded the date the child was found.
Sources vary in the information they contain. Where the LLM identifies a particular attribute as present, there should be extracted text to support that conclusion. However, where a given attribute is not identified as present, this may reflect limited information available. That is: if our only source of information is a medical examiner record, we cannot discern who was responsible for the child’s death and thus all fields related to persons involved will be empty. Likewise, media reports often contain no or limited information pertaining to the families’ history with child protection agencies, so absent a state fatality report where that information is systematically disclosed, the family’s history is better characterized as “unknown.”
In sum: the LLM indicators should not be used to report prevalence rates (e.g., X% of child maltreatment fatalities involve families without child protection history). Underidentification of specific attributes is least likely for the contexts/conditions of the fatality itself and most likely for attributes under the filter columns for individuals involved, child characteristics, and parent/caregiver factors. Underidentification is most widespread in states where no state fatality reports are issued, because then all records are identified through decedent records (where information is typically limited to cause/manner and demographics) and media (where the depth of reporting varies dramatically).
Our team is actively working to add additional records to our database – and to develop a more comprehensive understanding of each child’s death through additional sources. All data we identify and information we gather will be made public through our site. Please feel free to register to receive periodic notifications of larger project updates and data releases.
The Lives Cut Short project is housed at the University of North Carolina at Chapel Hill and funded by the Center on Opportunity and Social Mobility at the American Enterprise Institute.