Building Capacity to Study Child Sexual Abuse Recurrence through Community-Engaged Methods and AI-Enhanced Data Linkage

Background and Purpose: Recurrent child sexual abuse (CSA) is a critical challenge for social work. Effective prevention of CSA recurrence could limit traumatization and life course harm to health. However, research guiding social work practice in this area is impeded by fragmented administrative data. This study aims to strengthen the capacity of social work to prevent CSA recurrence through the creation of a highly innovative linked dataset in Philadelphia from 2013-2021. We leveraged artificial intelligence (AI) and machine learning (ML) approaches, as well as community-engaged research to enhance data linkage and improve data quality.

Methods: An interdisciplinary team of researchers and community stakeholders collaboratively conceptualized the study and secured legal and ethical approvals. A community advisory board (CAB) guided study design, implementation, and interpretation. Data included 3,373 substantiated child-level child protective services (CPS) records and 20,337 Child Advocacy Center (CAC) child-level records. CAC records included additional CPS cases regardless of substantiation status and cases not in CPS data investigated by law enforcement only (LEO). Data preprocessing incorporated AI-assisted methods (e.g., large language models) to standardize, clean, and reconcile inconsistencies in data fields, thereby improving data quality prior to linkage. We implemented a multi-stage linkage pipeline: (1) a deterministic algorithm applying combinations of case ID, name, and date of birth to establish matches; (2) a probabilistic record linkage using the Fellegi-Sunter (FS) framework; (3) supervised ML models to identify additional matches; and (4) a hybrid ensemble strategy. Record-level matching was conducted using a sliding window clustering strategy to identify CSA recurrence, informed by CAB expertise. Following human verification, statistical evaluation included sensitivity and specificity and cross-tabular analyses examining how AI-enhanced data linkage affected the observations of CSA recurrence.

Results: Deterministic methods matched 2,695 of 3,373 (80%) CPS records. Of the remaining 678 records, FS methods matched 355 pairs and ML methods matched 220 pairs (214 pairs were matched by both methods). FS demonstrated higher sensitivity (.98 FS versus .85 ML) but lower specificity than ML (.75 FS vs. 1.0 ML). The hybrid strategy yielded 258 matches beyond deterministic matches (88% linked). Community partners verified the remaining 12% unmatched cases. The availability of LEO cases (in CAC data only) was associated with observations of 40% higher initial CSA cases and of 35% higher recurrent CSA cases. The availability of CPS data was associated with observations of 2% higher initial CSA cases and 14% higher recurrent CSA cases than would be observed in unlinked CAC data.

Conclusions and Implications: This study demonstrates the value of integrating community engaged research with advanced data science to reduce data siloes and improve research capacity to inform CSA prevention. Findings highlight tradeoffs between sensitivity and specificity across linkage methods and support the use of hybrid approaches. Findings show that reliance on unlinked CPS data may significantly underestimate initial and recurrent CSA through the exclusion of LEO cases, although generalizability to other jurisdictions is unclear. This work advances methodological capacity in social work research and provides a scalable framework for leveraging AI to inform social work practice.