Assessing Selection Bias In Ncands Restricted-Release Files: A Comparison of Masked and Unmasked Counties

Background and Purpose: When the National Data Archive on Child Abuse and Neglect (NDACAN) National Child Abuse and Neglect Data System (NCANDS) provides Child File data under a restricted-release license, identifiers for small counties (fewer than 700 maltreatment reports) are masked for privacy. Researchers conducting county-level analyses need to understand how excluding these counties may bias findings or limit generalizability. This study compares masked to unmasked counties across several socioeconomic indicators to assess systematic differences and examine potential biases introduced by masking in county-level research using the restricted-release NCANDS file. While researchers have investigated how county-identifier masking influences reidentification risk for children with uncommon characteristics, our objective is to assess the masked dataset’s analytic utility for county-context child maltreatment research.

Methods: We compared known counties (N=976) to masked counties (N=2,168) in the 2022 NCANDS Child File. County-level characteristics were sourced from the 2022 County Health Rankings, the American Community Survey, and the Robert Graham Center’s Social Deprivation Index (SDI) dataset. Indicators included total child population, poverty rates (all ages and 0 to 17), child and overall food insecurity, SDI score, frequent mental distress, drug overdose mortality, and percentage lacking social support. Masked counties were identified by cross-referencing the universal list of U.S. counties against the unmasked counties in the NCANDS file. Group differences were evaluated using descriptive statistics and independent t-tests. We summarized group differences using means and standardized effect sizes (Cohen’s d). Effect sizes were interpreted using common benchmarks (small ≈ 0.2, medium ≈ 0.5, large ≈ 0.8).

Results: As expected, masked counties had substantially smaller child populations than known counties (mean child population = 5,226 vs. 62,623; d = 0.85; large). Effect sizes for differences in poverty indicators fell in the small-to-moderate range: masked counties had higher poverty for all ages (15.11% vs. 13.54%; abs(d) = 0.28; small) and for children age 0 to 17 (21.00% vs. 18.02%; abs(d) = 0.36; small-to-moderate). Differences across remaining contextual indicators were mostly small or smaller, indicating very little difference between masked and unmasked counties. Differences for child food insecurity (14.77% vs. 13.41%; abs(d) = 0.24; small), overall food insecurity (11.68% vs. 11.01%; abs(d) = 0.19; small), and frequent mental distress (18.84% vs. 18.33%; abs(d) = 0.22; small) were small. Differences were very small for SDI (46.55 vs. 49.38; abs(d) = 0.10) and drug overdose mortality (33.34 vs. 32.23; abs(d) = 0.06), and essentially null for percentage lacking social support (25.80 vs. 25.74; abs(d) = 0.02).

Conclusions and Implications: When county-level models focus on socioeconomic indicators alongside child maltreatment outcomes, the omission of masked counties may have a limited impact for many covariates beyond county size. However, poverty-related differences indicate potential for bias when analyses are sensitive to economic context or rural-urban disparities. Researchers using restricted-release NCANDS for county-level work should document masking rules, consider sensitivity analyses such as stratification by county size or rurality, reweighting, or alternative geographic aggregation, and interpret results cautiously when generalizing findings to all U.S. counties.