Benefits of longitudinally linked national records of child maltreatment report and foster care

Abstract

Background:

The Report and Placement Integrated Data System (RAPIDS) integrates two U.S. national data systems—NCANDS’ child maltreatment report (CMR) records and AFCARS’ foster care (FC) records—into a single longitudinal dataset spanning 2006–2021. This integration enables comprehensive child maltreatment analysis by linking the annual files from these previously separate systems.

Objective:

To explore benefits of RAPIDS data in understanding CMR outcomes.

Participants and setting:

Children aged 0–10 years with CMRs in 2018 (N = 2,371,119).

Methods:

Using logistic regression, we modeled five outcomes: two current outcomes from 2018 index reports (substantiation and foster care entry) and three future outcomes within two years (re-report, substantiated re-report, and foster are entry). For each outcome, we compared models using only index report data without RAPIDS variables against models incorporating RAPIDS-enabled variables that capture longitudinal patterns across reports, placements, and siblings.

Results:

RAPIDS data improved model performance across all outcomes, with greater gains for future outcomes. Overall model fit (Tjur’s R²) increased for substantiation (11.75 % → 12.53 %), FC entry (4.17 % → 6.38 %), rereport (0.94 % → 4.99 %), substantiated rereport (1.04 % → 3.31 %), and future FC entry (0.85 % → 2.51 %). Predictive performance also improved: at 80 % sensitivity, specificity increased for substantiation (54 % → 56 %), FC entry (52 % → 58 %), rereport (27 % → 36 %), substantiated rereport (33 % → 42 %), and future FC entry (37 % → 49 %). Additionally, RAPIDS data enabled analysis of a wider array of predictors and their associations with outcomes, fully utilizing national longitudinal CMR and FC records.

Conclusions:

RAPIDS data enhance explanatory power and predictive accuracy, enabling nationwide, longitudinal analysis of CMR and FC records and offering valuable insights into risk and protective factors.

1. Introduction

Child maltreatment and subsequent involvement with child protective services (CPS) are significant public health and social work concerns in the United States. Establishing a comprehensive historical database in a single data format that accurately captures longitudinal sequences of nationwide records for child maltreatment reports (CMRs) and foster care (FC) cases has the potential to accelerate research in child maltreatment. Such a database will be particularly useful for providing highly generalizable evidence for large-scale initiatives aimed at preventing and addressing child maltreatment and subsequent CPS involvement. Further, it can help develop a clearer understanding of how key events, such as disease outbreaks or changes in economic conditions, have affected child welfare services delivery (Shusterman, Fluke, Nunez, Fettig, & Kebede, 2022).

National CMR and FC records are maintained separately: the National Child Abuse and Neglect Data System (NCANDS) archives CMRs, while the Adoption and Foster Care Analysis and Reporting System (AFCARS) manages FC cases. Both systems store records in separate yearly files (NCANDS Child Files and AFCARS FC Files), limiting their utility for comprehensive analysis. To address these challenges, the Report And Placement Integrated Data System (RAPIDS) integrates NCANDS Child Files and AFCARS FC Files from 2006 onward into a single longitudinal dataset (Drake et al., 2023). Although RAPIDS does not improve the validity or reliability of the original records, it addresses linkage challenges by accurately reconstructing longitudinal sequences and creating variables based on prior (predictor) and subsequent (outcome) events. RAPIDS also incorporates advanced features, such as sibling identification through network analysis and linkages to external data sources, including county-level census data and state-level policy data. As the most comprehensive longitudinal integration of national CMR and FC data in the U.S., RAPIDS enables simultaneous analysis of these records at a population level. This study evaluates RAPIDS empirically by examining its benefits in modeling two current CMR outcomes (substantiation and FC entry) and three future outcomes (rereport, substantiated rereport, and future FC entry), comparing models with predictors available with and without RAPIDS data.

1.1. Existing gaps in the usage of NCANDS and AFCARS

The longitudinal integration of NCANDS and AFCARS poses significant technical challenges. Although prior research has combined them cross-sectionally, there has been no research, to our knowledge, that integrates them across long time periods. The primary hurdle lies in the “year-end record” format of AFCARS data. AFCARS records can be thought of as a status report, given on September 30th of each year, providing information about all children who contacted the foster care system in the prior year. There is always one record per child, regardless of the number of foster care episodes during the year. In contrast, NCANDS provides more straightforward yearly data, featuring one record per child per CMR episode. These records can be readily “stacked” across multiple years. AFCARS, on the other hand, presents “end of year status” records, providing one record per child per year, regardless of their FC activity (National Data Archive on Child Abuse and Neglect, 2023). A child may experience a single FC episode (defined as the period from entry to exit) that spans multiple years, or multiple episodes within a single year. To restructure the data into a more analytically useful format of “one record per child per FC episode,” careful consideration is needed when processing entry and exit variables both within and across years of AFCARS data. This transformation enables the creation of child/event-level data across both the NCANDS and the AFCARS Files. In this sense, an “event” is either a CMR from NCANDS or an entire (or still open) foster care episode from AFCARS. All original variables from both files are carried forward, and additional variables are created.

A missed opportunity is that families, as opposed to individual children, have rarely been studied using NCANDS and AFCARS data. Identifying families can provide insights into child maltreatment and CPS involvement at a broader family level, extending beyond the focus on single children (Cuccaro-Alamin et al., 2021). For instance, examining families can shed light on the similarities and differences in child maltreatment and CPS involvement experiences among siblings. Understanding prior and current sibling circumstances can enhance comprehension of the conditions leading to further CPS involvement, such as entry into foster care and recurrence among non-victim children of victimized siblings.

Another obstacle hindering the utilization of NCANDS and AFCARS data is the necessity for each research team to independently establish links with Census and other sources. Such linkage is particularly crucial since variables related to child, caregiver, and family characteristics/risk factors are mostly unreliable in NCANDS and AFCARS data. For example, income is likely the strongest predictor of child maltreatment risk, although the causal nature of this relationship is certainly debatable (Drake, Jonson-Reid, & Dvalishvili, 2022; Pelton, 2015; Sedlak et al., 2010). Many current studies rely on caregiver risk factors variables from NCANDS data to gauge poverty, a practice prone to yielding erroneous conclusions. These variables substantially underestimate poverty prevalence, suggesting its occurrence in <20 % of families involved with CPS (Briggs, Huggins-Hoyt, Teasley, & Hopps, 2022; Fix & Nair, 2020). However, multiple studies have demonstrated that actual poverty prevalence is several times higher (Jones et al., 2023; Kim, Drake, & Jonson-Reid, 2018; Sedlak et al., 2010). Two strategies exist to circumvent this issue. One approach involves limiting the NCANDS sample to children from states with reliable measures of these variables, typically numbering about five or fewer states, depending on the year (Jones et al., 2023; Kim et al., 2018). Alternatively, contextual (county-level) income variables from census data can be utilized, leveraging county identifiers in NCANDS and AFCARS data (Aron et al., 2010). Several robust studies have demonstrated that these approaches produce good results (Aron et al., 2010; Kim et al., 2018). Recent studies have linked NCANDS data at the county level to an array of external variables drawn from census, health, crime, or other datasets (Fix & Nair, 2020; Maguire-Jack, Font, & Dillard, 2020). However, few studies, to our knowledge, utilize these variables over extended timeframes. Such endeavors are time-intensive and may yield inconsistent results across studies.

Finally, the size of the source data files (NCANDS and AFCARS Files) creates a further barrier for researchers. These data are very large, with yearly NCANDS Files containing about 4 million records and 150 variables. Over time, and with the addition of the (fortunately much smaller) AFCARS Files, datasets can easily comprise dozens of gigabytes. This makes loading and analyzing data memory and processor-intensive. The RAPIDS programs avoid this problem and permit data to be extracted in minutes instead of hours. This is achieved by creating a relational database where each of the over 1300 final RAPIDS variables is stored as its own dataset. This allows researchers to quickly assemble new datasets across the entire 2006–2021 timeframe, taking about 10 s to capture each needed variable. This transforms the current NCANDS and AFCARS Files into a real-time consultable data source.

1.2. Current study

This study demonstrates the benefits of RAPIDS data in estimating the risks of further CPS involvement for children with a CMR (index CMR). It assesses two current outcomes (substantiation and FC entry from the index CMR) and three future outcomes (rereport, substantiated rereport, and future FC entry within two years of the index CMR). Models estimating these outcomes were compared using predictors available without and with RAPIDS data. Without RAPIDS data, only predictors derived from index (current) CMR records are readily available for modeling. While most outcomes (e.g., FC entry, rereport, substantiated rereport, and future FC entry) would also not be readily available, we assumed short-term linkage between NCANDS and AFCARS Files could be conducted by individual researchers. RAPIDS data expands this scope, providing variables not only from index CMR records but also from the full longitudinal range of CMR records, FC records, and siblings’ records. It further enables the creation of additional variables derived from these longitudinal data sources.

This study evaluates how the inclusion of RAPIDS-enabled predictors improves model performance for estimating outcomes, with the assumption that their inclusion enhances both model fit and predictive accuracy. Although RAPIDS data also integrates county-level census and state policy data, these sources were not assessed due to space constraints and the need for a separate analysis. Leveraging the extensive national and longitudinal scope of RAPIDS data, the study also investigates specific predictors to identify their risk and protective associations with the outcomes, contributing to the development of generalizable principles for understanding these relationships. It is assumed that all predictors are associated with the outcomes.

2. Methods

2.1. RAPIDS data

The RAPIDS project established longitudinal linkage between NCANDS Child Files (national CMR records) and AFCARS FC Files (national FC records) from 2006 to 2021. Before 2006, the absence of common child identifiers between Child Files and FC Files made earlier record linkage not feasible. The 2021 Files served as the latest available data for this analysis. This section summarizes the key accomplishments of the RAPIDS project.

The project combined Child Files and FC Files into a longitudinal dataset, containing child/case-level data. Each row within this dataset represents either a CMR episode or an FC entry-to-exit episode. The project addressed two significant challenges during this process. First, each year’s FC File serves as a year-end snapshot of FC cases, complicating the identification of an FC episode from entry to exit, particularly when spanning multiple years. Additionally, discerning all events and records within an FC entry-to-exit episode, as well as establishing the longitudinal sequence between FC episodes, posed challenges. To overcome this obstacle, the project extracted longitudinal information from each FC File, linking and assessing multiple years’ FC Files to accurately identify individual FC entry-to-exit episodes. Subsequently, the project recorded each FC episode as one data element per row, while storing all events and records from entry to exit in distinct variables (columns). Second, to maintain confidentiality, the source Child Files and FC Files had already applied date jittering techniques, albeit in different manners. This inconsistency could potentially lead to variability of CMR dates (Child Files) and foster care entry dates (FC Files) by a maximum of 23 days. The project assessed and resolved this issue, ensuring the optimal longitudinal sequence of CMR episodes and FC episodes for each child.

RAPIDS data have several additional notable features. First, the project employed network analysis to identify sibling groups. Siblings may be dispersed across various CMRs and FC episodes (e.g., Child 1 and Child 2 in Report A; Child 2 and Child 3 in Report B). By identifying each interconnected network of siblings (e.g., Child 1 – Child 2 – Child 3), the project assigned a unique family ID to each connected network, delineating distinct sibling groups. This facilitated the identification of siblings and enabled the measurement of predictors based on their records. Second, the project augmented county-level variables by integrating RAPIDS data with county-level census data, using county identifiers corresponding to a child’s residential county during an episode. Finally, the project links RAPIDS data with state policy data using state identifiers associated with a child’s residential state during an episode. However, it is important to note again that county-level and state policy variables were not included in this study.

Technical details of these data management processes are provided in the RAPIDS programs and the accompanying User Guide available on the project website (Drake, Kim, et al., 2023). The website offers free and publicly accessible RAPIDS programs, along with comprehensive guidance on obtaining NCANDS and AFCARS files, understanding RAPIDS data and variables, and executing RAPIDS programs. It also provides recommendations for the required computing power, such as a computer with at least 32 GB of RAM and 4 TB of storage. Once users obtain the NCANDS and AFCARD files, they can construct RAPIDS data by running the programs with the support of annotations and instructions provided in the programs and User Guide. The process may take several days, depending on the available computing resources (e.g., on a desktop with 32 GB of RAM and a 3.9 GHz processor, the programs typically complete within a few days). However, the procedure requires minimal user intervention, as the programs are designed to run autonomously.

2.2. Sample

From RAPIDS data spanning 2006 to 2021, we selected all children aged 0–10 years with CMRs in 2018 across 50 states and the District of Columbia (N = 2,371,119). We designated CMRs occurring in 2018 (for children with multiple CMRs in 2018, their first CMR in that year) as index CMRs to ensure a two-year follow-up for outcomes without right censoring. The age limit was set to ensure comprehensive coverage of their records from birth onwards with no left censoring. To establish the temporal sequence between predictors and outcomes, all predictors were measured based on records at or before the index CMRs, while outcomes were measured using records subsequent to the index CMRs.

2.3. Measures

The RAPIDS data enabled the measurement of a wide range of variables. We measured five outcome variables (Table 1), including two current outcomes (substantiation and FC entry) and three future outcomes (rereport, substantiated rereport, and future FC entry). Substantiation was coded as 1 if the index CMR received a disposition of substantiated, indicated, or reason to suspect, and 0 otherwise. FC entry was coded as 1 if there was a FC entry within 90 days from the index CMR’s disposition, and 0 otherwise. Rereport was coded as 1 if there was a subsequent CMR within two years of the index CMR, and 0 otherwise. Substantiated rereport was coded as 1 if there was a substantiated rereport within two years of the index CMR, and 0 otherwise. Future FC entry was coded as 1 if there was an FC entry >90 days after the index CMR’s disposition but within two years, and 0 otherwise. Rereports occurring within 90 days of the index CMR were excluded from rereport and substantiated rereport measurements to avoid capturing echoes of the index CMR.

Table 1.

Descriptive statistics of outcome variables.

Outcome variables	%
Substantiation of the 2018 index child maltreatment report (CMR)	18.18 %
Foster care entry within 90 days from the disposition of the 2018 index CMR	5.73 %
Rereport within 2 years of the 2018 index CMR	29.11 %
Substantiated rereport within 2 years of the 2018 index CMR	8.88 %
Future foster care entry (after 90 days from the disposition) within 2 years of the 2018 index CMR	3.98 %

Note. N = 2,371,199 children with child maltreatment reports (index reports) in the United States in 2018.

Predictors available without RAPIDS data were derived solely from index CMR records, including child age, sex, race/ethnicity, report source, and maltreatment type (Table 2). For the three future outcomes, substantiation of the index CMR was also included as a predictor.

Table 2.

Descriptive statistics of predictor variables available without RAPIDS.

Predictor variables	%	Predictor variables	%
Based on index CMR records
Child age		Report source
0 year	12.86 %	Social services personnel	9.45 %
1 year	8.34 %	Medical personnel	11.62 %
2 years	8.54 %	Mental health personnel	3.94 %
3 years	8.61 %	Legal/law enforcement/criminal justice	19.68 %
4 years	8.64 %	Education personnel	18.84 %
5 years	9.03 %	Child daycare personnel	0.80 %
6 years	8.97 %	Substitute care personnel	0.28 %
7 years	8.87 %	Alleged victim	0.15 %
8 years	8.87 %	Parent	6.43 %
9 years	8.72 %	Other relative	6.86 %
10 years	8.55 %	Friends/neighbors	4.38 %
Child sex		Alleged perpetrator	0.03 %
Male	51.69 %	Anonymous reporter	7.78 %
Female	47.47 %	Other/unknown/missing	9.76 %
Missing	0.84 %	Maltreatment type
Child race/ethnicity		Neglect	50.34 %
White alone, non-Latino	42.42 %	Physical abuse	11.93 %
Black alone, non-Latino	22.75 %	Sexual abuse	3.59 %
AIAN alone, non-Latino	1.08 %	Emotional abuse	2.77 %
Asian alone, non-Latino	0.97 %	Sex trafficking	0.01 %
NHPI alone, non-Latino	0.20 %	Other	5.43 %
Two or more races, non-Latino	4.72 %	Multiple types	15.85 %
Latino, any race	20.24 %	No alleged maltreatment	9.97 %
Missing	7.61 %	Unknown/missing	0.10 %

Note. N = 2,371,199 children with child maltreatment reports (index reports) in the United States in 2018. CMR = child maltreatment report. AIAN = American Indian/Alaska Native. NHPI = Native Hawaiian/Pacific Islander.

RAPIDS-enabled predictors (Table 3) were derived from prior CMR records (prior CMRs, prior substantiated CMRs, and prior CMR maltreatment type), FC records (prior FC episodes, last FC discharge reason, and termination of parental rights), and siblings’ records (siblings’ recent/prior CMRs and FC episodes). For the three future outcomes, FC entry from the index CMR was also included as a predictor.

Table 3.

Descriptive statistics of predictor variables available using RAPIDS.

Predictor variables	%	Predictor variables	%
Based on prior CMR records		Based on siblings’ records
# prior CMRs		# siblings with recent CMRs (near index CMR)
0	58.07 %	0	36.95 %
1	17.62 %	1	25.43 %
2	9.30 %	2	19.34 %
3	5.51 %	3	10.58 %
4	3.38 %	4	4.64 %
5	2.13 %	5	1.90 %
6	1.36 %	6	0.75 %
7	0.88 %	7 or more	0.42 %
8	0.58 %	# siblings with recent FC episodes (near index CMR)
9	0.38 %	0	97.45 %
10	0.25 %	1	1.35 %
11	0.17 %	2	0.69 %
12 or more	0.39 %	3	0.32 %
# prior substantiated CMRs		4 or more	0.20 %
0	84.23 %	# siblings with prior CMRs (before index CMR)
1	11.36 %	0	60.09 %
2	3.00 %	1	12.58 %
3	0.91 %	2	9.99 %
4	0.30 %	3	6.73 %
5 or more	0.20 %	4	4.19 %
Having a prior neglect report, yes	33.66 %	5	2.66 %
Having a prior physical abuse report, yes	14.02 %	6	1.72 %
Having a prior sexual abuse report, yes	3.69 %	7	1.09 %
Having a prior emotional abuse report, yes	4.81 %	8	0.65 %
Having a prior sex trafficking report, yes	0.00 %	9 or more	0.29 %
Having a prior other-type report, yes	7.19 %	# siblings with prior FC episodes (before index CMR)
Having a prior multiple-type report, yes	13.44 %	0	92.88 %
Having a prior no-alleged-maltreatment report, yes	6.87 %	1	3.01 %
		2	1.93 %
Based on FC Records		3	1.10 %
# prior FC episodes		4	0.57 %
0	93.78 %	5	0.27 %
1	5.41 %	6 or more	0.25 %
2	0.71 %
3 or more	0.10 %
Last FC discharge reason ≠ reunification, yes	2.21 %
Termination of parental rights before index CMR, yes	1.11 %
Having index CMR while placed in FC, yes	1.68 %

Note. N = 2,371,199 children with child maltreatment reports (index reports) in the United States in 2018. CMR = child maltreatment report. FC = foster care. “Last FC discharge reason” was coded as 1 if the discharge reason for a child’s most recent foster care episode was something other than reunification and as 0 otherwise. For predictors based on siblings’ records, events occurring within 30 days of the index CMR were classified as “near index CMR,” while those occurring >30 days prior to the index CMR was classified as “before index CMR.”

2.4. Analysis

We used logistic regression with cluster-robust standard errors to account for the nesting of children within families. However, incorporating the broader hierarchical structure of families within counties and counties within states was not feasible due to computational constraints associated with the large sample size (N = 2,371,199). Consequently, this omission may lead to an underestimation of standard errors. Nevertheless, given the substantial sample size, the potential impact on type I error rates is likely minimal. To further address this potential issue, we applied a more stringent alpha level of 0.01 instead of the conventional 0.05.

We fitted separate models for each of the five outcomes. To demonstrate the benefits of using RAPIDS data, we evaluated model improvements in estimating each outcome by starting with an intercept-only model (Model 0) and then incrementally adding predictors. First, we incorporated predictors available without RAPIDS data (Model 1), followed by the inclusion of RAPIDS-enabled predictors (Model 2). Logistic model equations for Models 1 and 2 across all five outcomes are provided in the Supplement.

3. Results

3.1. Descriptive statistics

Table 1 presents the descriptive statistics for the outcomes. For current outcomes, 18.18 % of children experienced substantiation of their index CMR, and 5.73 % had an FC entry from index CMR. Among future outcomes, 29.11 % underwent a rereport, 8.88 % had a substantiated rereport, and 3.98 % faced a future FC entry within 2 years from their index CMR.

Table 2 reports the descriptive statistics for predictors available without RAPIDS data. Children aged 0 accounted for 12.86 %, higher than other age groups (ranging from 8.34 % to 9.03 %). Male children (51.69 %) slightly outnumbered female children (47.47 %). White children made up the largest group (42.42 %), followed by Black children (22.75 %), Latino children (20.24 %), multi-racial children (4.72 %), and others. The most common report sources were legal, law enforcement, and criminal justice personnel (19.68 %), followed by education personnel (18.84 %), medical personnel (11.62 %), social services personnel (9.45 %), and others. Neglect was the most reported type of maltreatment (50.34 %) among index CMRs.

Table 3 presents the descriptive statistics for RAPIDS-enabled predictors. Among children with index CMRs, 41.93 % had one or more prior CMRs, and 15.77 % had one or more substantiated CMRs. Neglect was the most common maltreatment type among prior CMRs (33.66 %). In terms of FC records, 6.22 % had one or more prior FC episodes, 2.21 % had a last FC discharge reason other than reunification, 1.11 % experienced termination of parental rights, and 1.68 % had their index CMR while placed in FC. For siblings’ records, 63.05 % had one or more siblings with recent CMRs (within ±30 days of the index CMR), 2.55 % had one or more siblings with recent FC episodes, 39.91 % had one or more siblings with prior CMRs (>30 days before the index CMR), and 7.12 % had one or more siblings with prior FC episodes.

3.2. Comparisons of model fit, calibration, and predictive performance

Table 4 presents model fit, calibration, and predictive performance statistics for the intercept-only model (Model 0), the model without RAPIDS data (Model 1), and the model with RAPIDS data (Model 2). For overall model fit, decreases in AIC values >10 indicate strong evidence of significant improvement (Burnham & Anderson, 2004; Kass & Raftery, 1995). The inclusion of RAPIDS-enabled predictors significantly improved model fit across all outcomes. However, while AIC identifies significant improvements, it does not convey the magnitude of improvement, which can be problematic given the study’s large sample size, where even minimal improvements can appear statistically significant. To address this, we used Tjur’s R², which quantifies the closeness between predicted probabilities and binary outcomes (Tjur, 2009). A Tjur’s R² of 1 (or 100 %) represents perfect explanation of outcome variance, although real-world models often yield lower values since predicted probabilities rarely approach 0 % (negative) or 100 % (positive). Adding RAPIDS-enabled predictors increased Tjur’s R² across all outcomes, with larger gains observed for future outcomes: 11.75 % to 12.53 % for substantiation, 4.17 % to 6.38 % for FC entry, 0.94 % to 4.99 % for rereport, 1.04 % to 3.31 % for substantiated rereport, and 0.85 % to 2.51 % for future FC entry.

Table 4.

Comparison of logistic regression models—null model (Model 0), model without RAPIDS data (Model 1), and model with RAPIDS data (Model 2): fit, calibration, and predictive performance.

Model	Overall fit		Calibration		Predictive performance
	AIC	Tjur’s R2	HL test	Calibration slope/intercept	Confusion matrix				Sensitivity = TP/(TP + FN)	Specificity = TN/(TN + FP)	AUC
					TP	FP	TN	FN
Substantiation of index CMR models (observed positives = 431,038; observed negatives = 1,940,081)
Model 0	2,248,259	0.0000	–	–	–	–	–	–	–	–	0.5000
Model 1	1,949,695	0.1175	p < .0001	1.00/0.00	345,046	892,051	1,048,030	85,992	80 %	54 %	0.7466
Model 2	1,933,552	0.1253	p < .0001	1.00/0.00	344,831	860,246	1,079,835	86,207	80 %	56 %	0.7538
FC entry from index CMR models (observed positives = 135,930; observed negatives = 2,235,189)
Model 0	1,041,158	0.0000	–	–	–	–	–	–	–	–	0.5000
Model 1	954,000	0.0417	p < .0001	1.00/0.00	108,744	1,072,698	1,162,491	27,186	80 %	52 %	0.7267
Model 2	918,133	0.0638	p < .0001	0.99/0.00	108,744	930,066	1,305,123	27,186	80 %	58 %	0.7657
Rereport (within 2 years) models (observed positives = 690,165; observed negatives = 1,680,954)
Model 0	2,860,085	0.0000	–	–	–	–	–	–	–	–	0.5000
Model 1	2,836,871	0.0094	p < .0001	1.00/0.00	552,182	1,219,004	461,950	137,983	80 %	27 %	0.5586
Model 2	2,742,417	0.0499	p < .0001	0.99/0.00	552,132	1,067,447	613,507	138,033	80 %	36 %	0.6401
Substantiated rereport (within 2 years) models (observed positives = 210,617; observed negatives = 2,160,502)
Model 0	1,421,787	0.0000	–	–	–	–	–	–	–	–	0.5000
Model 1	1,396,522	0.0104	p < .0001	0.98/0.00	168,532	1,443,150	717,352	42,085	80 %	33 %	0.6037
Model 2	1,350,150	0.0331	p < .0001	0.98/0.00	168,494	1,252,115	908,387	42,123	80 %	42 %	0.6710
Future FC entry (within 2 years) models (observed positives = 94,427; observed negatives = 2,276,692)
Model 0	793,776	0.0000	–	–	–	–	–	–	–	–	0.5000
Model 1	774,039	0.0085	p < .0001	0.97/0.00	75,553	1,441,723	834,969	18,874	80 %	37 %	0.6317
Model 2	741,786	0.0251	p < .0001	0.95/0.00	75,542	1,150,034	1,126,658	18,885	80 %	49 %	0.7120

Note. CMR = child maltreatment report. FC = foster care. AIC = Akaike information criterion. HL = Hosmer-Lemeshow. TP = true positive. FP = false positive. TN = true negative. FN = false negative. AUC = area under the curve. Observed positives = number of children with the observed outcome. Observed negatives = number of children without the observed outcome. To facilitate comparisons, sensitivity was set at 80 % (i.e., true positives and false negatives were set at 80 % and 20 % of observed positives respectively). In other words, the threshold was adjusted so that 80 % of children with the observed outcome are included among the model-predicted positives. This approach allows for comparing model prediction performance based on specificity values, as well as examining true negatives and false positives. Details of predictors are provided in Tables 1–3, and the full results of Model 1 and Model 2 are presented in Table 5.

For calibration, we assessed the consistency between predicted probabilities and observed event rates across probability levels (bins). The Hosmer-Lemeshow test, which formally assesses calibration (Hosmer & Lemeshow, 2000), was significant for all models, likely due to the large sample size, where even minor differences become statistically significant. To address this, we examined calibration slopes and intercepts (Table 4) and calibration plots (Figs. S1–S5 in the Supplement). Slopes were near 1, and intercepts were close to 0, indicating strong agreement between predicted and observed probabilities. Calibration plots confirmed good calibration for Models 1 and 2 across all outcomes.

While fit and calibration metrics are important, they do not fully capture a model’s ability to distinguish positives from negatives at the individual level. This distinction is critical for identifying high-risk individuals needing services (sensitivity) and excluding low-risk individuals from unnecessary interventions (specificity). To assess predictive performance, we fixed sensitivity at 80 %, ensuring 80 % of true positives were identified, and compared specificity across models. Adding RAPIDS-enabled predictors increased specificity, with larger improvements for future outcomes: 54 % to 56 % for substantiation, 52 % to 58 % for FC entry, 27 % to 36 % for rereport, 33 % to 42 % for substantiated rereport, and 37 % to 49 % for future FC entry. These findings were supported by Area Under the Curve (AUC) values, where higher AUC indicates better fit, with values around 0.7 or higher considered acceptable (Hosmer & Lemeshow, 2000).

3.3. Full results of logistic regression models

Table 5 presents the full results of Model 1 (without RAPIDS data) and Model 2 (with RAPIDS data) for all outcomes. RAPIDS data allowed for the examination of a broader range of predictors and their relationships with outcomes, drawing on national longitudinal records of CMR and FC. It also provided additional context for predictors available without RAPIDS data (e.g., child age, sex, race/ethnicity, report source, and maltreatment type) by accounting for RAPIDS-enabled predictors. Among these variables, child age was the only one to show notable changes after the inclusion of RAPIDS-enabled predictors, while others remained stable. The association between older child age and lower risks became stronger, particularly for future outcomes such as rereport, substantiated rereport, and future FC entry. RAPIDS data also made it possible to examine new predictors introduced by its integration. The remainder of this section summarizes key relationships between predictors and outcomes.

Table 5.

Full results of logistic regression models (N = 2,371,119).

Variables	Odds ratio
	Models without RAPIDS Data					Models with RAPIDS Data
	Substantiation	FC entry	Rereport	Substantiated re-CMR	Future FC entry	Substantiation	FC entry	Rereport	Substantiated re-CMR	Future FC entry
Predictors based on index CMR records
Child age (ref = 0 year)
1 year	0.63	0.58	1.07	0.99	0.91	0.59	0.52	0.93	0.85	0.74
2 years	0.57	0.51	1.05	0.95	0.83	0.53	0.43	0.86	0.75	0.60
3 years	0.54	0.47	1.07	0.90	0.75	0.49	0.37	0.82	0.66	0.50
4 years	0.53	0.44	1.08	0.88	0.70	0.47	0.33	0.79	0.61	0.44
5 years	0.52	0.41	1.06	0.86	0.66	0.47	0.30	0.75	0.57	0.39
6 years	0.50	0.38	1.06	0.83	0.64	0.45	0.27	0.72	0.53	0.36
7 years	0.50	0.37	1.02	0.78	0.60	0.45	0.25	0.67	0.48	0.32
8 years	0.49	0.35	1.00	0.75	0.58	0.44	0.24	0.64	0.45	0.30
9 years	0.49	0.34	0.98	0.74	0.57	0.43	0.22	0.61	0.43	0.29
10 years	0.49	0.34	0.95	0.72	0.57	0.43	0.22	0.59	0.41	0.28
Child sex (ref = Male)
Female	1.04	1.00	0.99	1.01	1.00	1.04	1.00	0.99	1.02	1.00
Missing	0.48	0.05	0.55	0.47	0.32	0.50	0.05	0.61	0.54	0.38
Child race (ref = White, non-Latino)
Black, non-Latino	0.95	0.78	0.94	0.82	0.78	0.95	0.80	0.95	0.84	0.81
AIAN, non-Latino	1.16	1.94	1.10	1.28	1.63	1.13	1.75	1.10	1.28	1.52
Asian, non-Latino	0.89	0.60	0.43	0.45	0.37	0.95	0.74	0.56	0.60	0.51
NHPI, non-Latino	1.11	1.57	0.55	0.59	0.73	1.14	1.83	0.65	0.73	0.94
Two or more races, non-Latino	1.10	1.33	1.15	1.19	1.31	1.07	1.22	1.08	1.10	1.17
Latino, any race	1.08	0.99	0.83	0.85	0.85	1.08	1.02	0.86	0.89	0.90
Missing	0.67	0.30	0.63	0.57	0.43	0.70	0.34	0.74	0.68	0.53
Report source (ref = Social services)
Medical personnel	1.14	0.98	0.92	0.88	0.92	1.19	1.02	0.98	0.94	1.01
Mental health personnel	0.45	0.42	1.01	0.87	0.85	0.46	0.41	1.01	0.88	0.85
Legal/law enforcement/CJ personnel	1.68	0.91	0.93	0.88	0.81	1.75	0.95	0.97	0.95	0.88
Education personnel	0.44	0.44	1.09	0.96	0.93	0.45	0.43	1.09	0.98	0.95
Child daycare personnel	0.52	0.52	0.91	0.83	0.90	0.52	0.51	0.92	0.85	0.90
Substitute care personnel	0.62	0.66	1.02	1.10	1.11	0.57	0.72	0.97	1.04	1.01
Alleged victim	0.49	0.36	0.75	0.61	0.67	0.52	0.36	0.81	0.69	0.74
Parent	0.35	0.22	1.05	0.77	0.53	0.36	0.22	1.04	0.78	0.55
Other relative	0.51	0.65	1.09	1.01	1.10	0.52	0.65	1.09	1.02	1.12
Friends/neighbors	0.34	0.48	1.06	0.98	1.06	0.34	0.47	1.07	1.00	1.10
Alleged perpetrator	1.57	1.09	1.09	1.53	1.46	1.51	1.00	1.03	1.43	1.36
Anonymous reporter	0.31	0.42	1.02	0.91	1.03	0.31	0.42	1.03	0.94	1.07
Other/unknown/missing	0.51	0.62	0.97	0.96	1.00	0.52	0.61	0.98	0.97	1.01
Maltreatment type (ref = Neglect)
Physical abuse	0.68	0.51	0.74	0.62	0.65	0.71	0.53	0.80	0.69	0.69
Sexual abuse	0.85	0.23	0.59	0.49	0.46	0.91	0.26	0.66	0.56	0.51
Emotional abuse	0.43	0.31	0.91	0.89	0.77	0.45	0.34	1.03	1.05	0.91
Sex trafficking	0.84	1.00	0.89	0.60	0.79	0.84	1.14	0.87	0.58	0.82
Other	0.25	0.52	0.98	0.69	1.05	0.25	0.53	0.95	0.70	1.10
Multiple types	1.55	1.31	1.01	1.03	0.91	1.54	1.33	0.99	1.00	0.91
No alleged maltreatment	0.00	0.27	0.94	0.64	0.69	0.00	0.28	0.91	0.66	0.73
Unknown/missing	0.23	0.79	0.91	0.89	1.01	0.23	0.80	1.00	0.99	1.09
Substantiation of index CMR	–	–	1.00	1.38	1.49	–	–	1.03	1.36	1.49
Predictors based on prior CMR records
# prior CMRs (ref = 0)
1	NA	NA	NA	NA	NA	1.05	1.55	1.38	1.29	1.76
2	NA	NA	NA	NA	NA	1.02	1.84	1.65	1.49	2.30
3	NA	NA	NA	NA	NA	0.98	2.06	1.86	1.67	2.69
4	NA	NA	NA	NA	NA	0.91	2.16	2.09	1.83	3.05
5	NA	NA	NA	NA	NA	0.88	2.27	2.29	1.98	3.35
6	NA	NA	NA	NA	NA	0.80	2.30	2.37	2.08	3.46
7	NA	NA	NA	NA	NA	0.79	2.40	2.55	2.30	4.03
8	NA	NA	NA	NA	NA	0.75	2.58	2.64	2.37	4.44
9	NA	NA	NA	NA	NA	0.73	2.37	2.72	2.44	4.27
10	NA	NA	NA	NA	NA	0.67	2.46	2.77	2.62	4.59
11	NA	NA	NA	NA	NA	0.60	2.22	3.07	2.99	5.32
12 or more	NA	NA	NA	NA	NA	0.62	2.59	3.01	2.91	5.30
# prior substantiated CMRs (ref = 0)
1	NA	NA	NA	NA	NA	1.60	1.13	0.99	1.27	1.01
2	NA	NA	NA	NA	NA	2.10	1.11	0.98	1.50	1.02
3	NA	NA	NA	NA	NA	2.62	1.06	1.01	1.75	0.99
4	NA	NA	NA	NA	NA	3.25	1.00	1.00	1.86	0.96
5 or more	NA	NA	NA	NA	NA	4.50	1.07	1.08	2.28	1.01
Having a prior neglect report	NA	NA	NA	NA	NA	1.04	0.99	1.13	1.26	1.08
Having a prior physical abuse report	NA	NA	NA	NA	NA	0.91	0.97	1.04	0.97	1.00
Having a prior sexual abuse report	NA	NA	NA	NA	NA	0.88	0.93	1.01	0.96	0.94
Having a prior emotional abuse report	NA	NA	NA	NA	NA	0.94	1.11	0.96	0.91	1.06
Having a prior sex trafficking report	NA	NA	NA	NA	NA	1.61	1.25	0.94	0.84	0.30
Having a prior other-type report	NA	NA	NA	NA	NA	1.15	0.91	1.14	1.16	0.90
Having a prior multiple-type report	NA	NA	NA	NA	NA	1.04	0.88	1.04	1.08	0.93
Having a prior no-alleged-CM report	NA	NA	NA	NA	NA	1.00	0.98	1.06	0.94	0.94
Predictors based on FC records
# prior FC episodes (ref = 0)
1	NA	NA	NA	NA	NA	0.84	1.21	1.08	0.98	1.42
2	NA	NA	NA	NA	NA	0.81	1.53	1.05	0.93	1.63
3 or more	NA	NA	NA	NA	NA	0.82	2.00	1.12	0.95	1.81
Last FC discharge reason ≠ reunification	NA	NA	NA	NA	NA	1.03	1.31	0.77	0.72	0.75
TPR before index CMR	NA	NA	NA	NA	NA	0.90	1.93	0.72	0.76	0.92
Having index CMR while placed in FC	NA	NA	NA	NA	NA	0.99	0.03	0.88	0.72	0.43
FC entry from index CMR	–	–	NA	NA	NA	–	–	0.70	0.94	0.67
Predictors based on Siblings’ records
# siblings with recent CMRs (ref = 0)
1	NA	NA	NA	NA	NA	1.00	0.75	0.87	0.84	0.68
2	NA	NA	NA	NA	NA	1.04	0.77	0.90	0.88	0.69
3	NA	NA	NA	NA	NA	1.06	0.79	0.92	0.89	0.71
4	NA	NA	NA	NA	NA	1.07	0.81	0.94	0.87	0.72
5	NA	NA	NA	NA	NA	1.06	0.87	0.94	0.91	0.74
6	NA	NA	NA	NA	NA	1.02	0.88	0.94	0.91	0.75
7 or more	NA	NA	NA	NA	NA	1.03	0.73	0.89	0.75	0.66
# siblings with recent FC episodes (ref = 0)
1	NA	NA	NA	NA	NA			0.87	0.66	1.44
2	NA	NA	NA	NA	NA			0.87	0.60	1.31
3	NA	NA	NA	NA	NA			0.91	0.61	1.36
4 or more	NA	NA	NA	NA	NA			1.03	0.58	1.06
# siblings with prior CMRs (ref = 0)
1	NA	NA	NA	NA	NA	1.06	1.10	1.35	1.24	1.25
2	NA	NA	NA	NA	NA	1.04	1.08	1.36	1.23	1.21
3	NA	NA	NA	NA	NA	1.01	1.04	1.39	1.24	1.25
4	NA	NA	NA	NA	NA	1.01	1.00	1.45	1.28	1.30
5	NA	NA	NA	NA	NA	0.99	0.95	1.46	1.32	1.30
6	NA	NA	NA	NA	NA	0.98	0.92	1.47	1.29	1.28
7	NA	NA	NA	NA	NA	1.00	0.88	1.48	1.31	1.28
8	NA	NA	NA	NA	NA	0.94	0.83	1.40	1.21	1.14
9 or more	NA	NA	NA	NA	NA	0.94	0.77	1.19	0.99	0.91
# siblings with prior FC episodes (ref = 0)
1	NA	NA	NA	NA	NA	1.20	2.46	1.07	1.23	1.72
2	NA	NA	NA	NA	NA	1.23	2.58	1.07	1.22	1.82
3	NA	NA	NA	NA	NA	1.18	2.68	1.05	1.17	1.72
4	NA	NA	NA	NA	NA	1.32	3.15	1.00	1.12	1.72
5	NA	NA	NA	NA	NA	1.25	3.30	0.97	1.07	1.64
6 or more	NA	NA	NA	NA	NA	1.29	3.71	0.97	1.03	1.83
Model fit: Area under the curve	0.7466	0.7267	0.5586	0.6037	0.6317	0.7538	0.7657	0.6401	0.6710	0.7120

Note. CMR = child maltreatment report. FC = foster care. Ref = reference category. AIAN = American Indian or Alaska Native. NHPI = Native Hawaiian or Pacific Islander. CJ = criminal justice. CM = child maltreatment. TPR = termination of parental rights. NA = not available. Cluster-robust standard errors were used for inference on estimated coefficients to account for the clustering of children within families. Significant odds ratios (p < .01) are in boldface.

Older child age was associated with notably lower risks across all outcomes, more prominently for FC entry and future FC entry, and to a lesser extent for rereport. For example, compared to children aged 0, 10-year-olds had a 78 % lower likelihood of FC entry (OR = 0.22), a 72 % lower likelihood of future FC entry (OR = 0.28), and a 41 % lower likelihood of rereport (OR = 0.59).

While statistically significant differences by child sex were observed in some outcomes, these differences were minor (≤4 %) and practically negligible.

Race/ethnicity showed notable variations. Black and Asian children had lower risks for all outcomes compared to White children, while AIAN children and those of two or more races had higher risks. Latino and NHPI children demonstrated mixed patterns, with higher risks for substantiation and FC entry but lower risks for rereport, substantiated rereport, and future FC entry.

Reports from social services, medical, and legal/law enforcement sources, as well as from alleged perpetrators, were associated with higher risks for immediate outcomes (substantiation and FC entry). However, the differences were less pronounced for future outcomes (rereport, substantiated rereport, and future FC entry).

Alleged maltreatment types in the index CMR were linked to variations in risks. Physical abuse-only and sexual abuse-only cases were associated with lower risks for all outcomes compared to neglect-only cases. Multiple-type cases had higher risks for immediate outcomes but showed no higher risks of future outcomes. Emotional abuse-only, other, no-alleged-maltreatment, and unknown/missing-type cases had lower risks for immediate outcomes but comparable risks for future outcomes. Sex-trafficking-only cases were too rare to yield statistically significant results.

Substantiation of the index CMR was associated with slightly higher risks for rereport (3 %) and moderately higher risks for substantiated rereport (36 %) and future FC entry (49 %).

The number of prior CMRs was associated with varying patterns of risk. Having more prior CMRs, especially four or more, was linked to lower substantiation risks, potentially reflecting the influence of prior unsubstantiated reports (as prior substantiated CMRs were accounted for). In contrast, more prior CMRs were associated with higher risks for FC entry, rereport, substantiated rereport, and future FC entry.

Prior substantiated CMRs were associated with higher risks for substantiation and substantiated rereport but showed limited relationships with FC entry and no significant association with rereport or future FC entry.

Alleged maltreatment types in prior CMRs showed smaller variations in risks compared to the types in the index CMR.

Prior FC episodes were associated with higher risks for FC entry, future FC entry, and rereport, but lower risks for substantiation.

Adjustments for non-parental care factors (e.g., last FC discharge reasons, TPR before the index CMR, placement in FC during the index CMR, and FC entry from the index CMR) showed varied patterns. Most predictors were associated with lower risks, except for higher substantiation risks linked to TPR and non-reunification discharge.

Sibling factors were also associated with outcome risks. Having siblings with recent CMRs was linked to slightly higher substantiation risk but lower risks for FC entry and future FC entry. Siblings with recent FC episodes were associated with lower risks for rereport and substantiated rereport but higher risks for future FC entry. Siblings with prior CMRs showed limited relationships with substantiation and FC entry but generally higher risks for rereport, substantiated rereport, and future FC entry. Siblings with prior FC episodes were mostly associated with higher risks across all outcomes, particularly FC entry and future FC entry.

4. Discussion

This study showcased the advantages of using RAPIDS data to estimate risks for both current and future CMR outcomes, including substantiation, FC entry, rereport, substantiated rereport, and future FC entry. It demonstrated how integrating RAPIDS-enabled predictors improved model performance (fit and predictive performance) and enhanced understanding of key variables in national longitudinal CMR and FC records and their relationships with outcomes.

Predictors derived from current index CMR records (i.e., those available without RAPIDS data) played a significant role in estimating all outcomes, consistent with prior research highlighting the importance of index case factors in decision-making and their potential influence on long-term outcomes (e.g., Baumann, Fluke, Dalgleish, & Kern, 2014; Maguire-Jack & Font, 2014). However, incorporating RAPIDS-enabled predictors—based on full longitudinal CMR and FC records—substantially improved model performance, both in terms of overall model fit and individual-level predictive accuracy (Table 4). These results support previous studies that underscore the value of understanding a child’s longitudinal history in predicting future risks of maltreatment and CPS involvement (e.g., Brown, Cohen, Johnson, & Salzinger, 1998; Kim, Drake, & Jonson-Reid, 2020). This highlights the importance of establishing longitudinal linkages of CPS records, such as those provided by RAPIDS data.

Furthermore, the inclusion of RAPIDS-enabled predictors yielded the smallest gains for substantiation, moderate gains for FC entry, and substantial improvements for future outcomes, such as rereport, substantiated rereport, and future FC entry. These findings carry two key implications. First, the minimal gain for substantiation aligns with prior research suggesting the unreliability of substantiation decisions, which often hinge more on legal evidence of current allegations than on the severity of harms or the level of needs (Cross & Casanueva, 2009; Drake, 1996; Hussey et al., 2005; Kohl, Jonson-Reid, & Drake, 2009; Leiter, Myers, & Zingraff, 1994). Second, the greater improvements for future outcomes are particularly significant for prevention efforts. While immediate outcomes, like substantiation and FC entry, allow little opportunity for intervention, identifying children at risk for future issues offers a chance to intervene before problems escalate. RAPIDS data proved valuable in improving such identification.

From the specific relationships between predictors and outcomes (Table 5), we derive several generalizable principles. Younger children consistently exhibited elevated risks across all CMR outcomes, aligning with prior research on onset of CMRs, substantiated CMRs, and FC entries (Kim, Wildeman, Jonson-Reid, & Drake, 2017; Wildeman et al., 2014; Wildeman & Emanuel, 2014). Notably, this trend, seen in all CMR outcomes, is particularly pronounced in current and future FC entries compared to other outcomes.

Regarding race/ethnicity, once reported, Black children did not exhibit higher risks across all CMR outcomes compared with White children, consistent with prior research (Drake et al., 2023; Morton, Ocasio, & Simmel, 2011). This underscores the importance of addressing the Black-White disparity in CMR onset to mitigate CPS disparities. Evidence suggests that socioeconomic factors likely underlie this racial disparity, as the disparity in CMRs disappear after accounting for these variables (Drake, Jonson-Reid, Kim, Chiang, & Davalishvili, 2021; Kim & Drake, 2018; Maloney, Jiang, Putnam-Hornstein, Dalton, & Vaithianathan, 2017; Putnam-Hornstein, Needell, King, & Johnson-Motoyama, 2013). Conversely, other racial/ethnic groups, including AIAN, children of two or more races, NHPI, and Latino children, exhibited significantly higher risks across various outcomes compared to White children, warranting further investigation into the underlying mechanisms.

Differences in substantiation risks between report sources were significant, notably higher for law enforcement, medical, social services personnel, and alleged perpetrators. This may be attributed to the likelihood of their reports—such as police reports, medical diagnoses, state employee’s reports, and perpetrators’ admissions—being considered as legal evidence. Discrepancies were also observed, albeit to a somewhat lesser extent, in FC entry risks. Interestingly, these differences diminished considerably in future CMR outcomes. These findings align with prior research suggesting that stronger legal evidence for substantiation does not necessarily correlate with more severe harms or greater future risks (Drake, 1996; Hussey et al., 2005; Kohl et al., 2009; Leiter et al., 1994).

We observed intriguing patterns regarding prior records. Prior CMRs generally did not elevate substantiation risks but strongly increased the risks of all other outcomes. Conversely, prior substantiated CMRs primarily heightened risks for substantiation and substantiated rereport, with minimal impact on other outcomes. Prior FC episodes notably increased the risks of current and future FC entries, while exerting minimal influence on other outcomes.

Siblings’ records revealed mixed patterns, likely due to their representation of both family-level risks and parental experiences, depending on the presence of absence of siblings. Notably, siblings’ records may have a particularly significant impact for the youngest children. Although the youngest are at the highest risk of maltreatment and CPS involvement (Kim et al., 2017; Rebbe, Martinson, & Mienko, 2021), their brief life histories pose challenges in predicting future outcomes. Using records from their older siblings may help address this gap. Additionally, leveraging siblings’ records empirically reinforces the importance of family systems in understanding the risks associated with child maltreatment and CPS involvement. This perspective has long been recognized, tracing back to Bronfenbrenner’s ecological systems theory (Bronfenbrenner, 1979). RAPIDS provides empirical data to incorporate this family-level insight into analysis.

4.1. Strengths/limitations

The core innate strength of the RAPIDS data lies in the ease of use conferred when using longitudinal national child maltreatment report and foster care data. RAPIDS decreases data access time by about two orders of magnitude. The various quality assurance components of RAPIDS are also valuable, and we hope that the many created and merge variables will be of interest to researchers in the future. With regard to data validity, our findings converge with other studies using NDACAN and AFCARS data is high (e.g., Drake, Jones, et al., 2023). Even more promising, our findings align closely with data not derived from NCANDS and AFCARS data (e.g., Brown et al., 1998; Leiter et al., 1994; Maloney et al., 2017).

Of course, RAPIDS inherits any limitations present in the source data. The Family ID variable, for example, assumes all children reported together or estimated to be placed from the same case are siblings, while it is possible they could be otherwise related. The AFCARS family ID generation algorithm cannot use a “shared report” indicator like the Child File, as none exists. Fortunately, there are far fewer AFCARS foster care entries than Child File reports, and each AFCARS entry includes data on the child (e.g., race/ethnicity), the caretaker’s age and relationship to child, geography (county), and the time (date). We assume that children who match on these variables are brought in together. During recent years, there are about 700 children across fifty states entering care on an average day, and multiple children of the same race in the same county with mothers of the same age certainly are possible, causing some potential overinclusion within families. Final Family ID assignments were plausible, with a median of two children per family and <5 % of families having 7 or more siblings.

Another limitation that has potentially important implications from a longitudinal perspective is the inability to connect families and children who move to a different state with prior events. This suggests that our estimates of prior (and in using retrospective cohorts) subsequent events may be missing from some trajectories. This limitation is inherent in both NCANDS and AFCARS person IDs and the likely consequence is that our observed estimates of system contacts per child or family are lower than would be the case if it were possible to positively identify persons across states. This is an area that for future research and perhaps the identification of methods that could be used to understand the magnitude and nature of the issue and consider possible methods for adjustment.

Analytically, we could have used different variables and more sophisticated techniques considering non-linear relations and interactions, but our intention was to use a relatively simple and transparent approach to demonstrate the benefits of employing different strata of RAPIDS data. Nevertheless, to date, our analysis was based on the most comprehensive national longitudinal data, and several models showed acceptable model fit even based on simple logistic regression modeling (AUC > 0.07). RAPIDS data provide rich longitudinal information that helps innovatively measure predictors (e.g., sequence and longitudinal changes of prior episodes, combination of maltreatment types, time proximity of prior episodes, etc.), which can help improve prediction of future CPS outcomes. Future work may consider longitudinal patterns (e.g., timing, duration) of prior events, specific case characteristics (e.g., post-investigation services, risk assessment), or county- and state-level conditions.

Our analysis is based exclusively on children with a CMR in 2018. While the use of national data mitigates the risk of regional overfitting, the findings may have limited generalizability to other cohorts. A common approach in predictive analytics to enhance validation and avoid overfitting involves splitting data into training and testing data, typically in a 70:30 ratio (Kuhn & Johnson, 2013). The model is then estimated using the training set and validated by evaluating its predictive performance on the testing set. Although we did not adopt this method for the sake of parsimony, future research could benefit from incorporating this approach.

While we established temporal ordering between predictors and outcomes by leveraging the longitudinal structure of RAPIDS data, we did not employ more advanced longitudinal models, such as time-to-event analysis, or causal inference approaches, such as fixed effects models. Our primary goal was to demonstrate the advantages of RAPIDS using a straightforward approach. Future studies could readily apply time-to-event analysis or causal inference models to further enhance the utility of RAPIDS data.

4.2. Recommendations/implications

The results detailed in the analyses presented here provide a basis for recommendations for future research users of the RAPIDS data to explore the impact on child welfare service indicators of critical events (e.g., disease outbreaks, economic performance, etc.) as well as, practice, or policy implementation. That is, our analyses provide a baseline methodology and benchmark results for analyses that intend to incorporate risk adjustment to assess the effects of additional independent variables. Further, for other subsets of the data (e.g., states, maltreatment type, etc.), they provide a starting point for the identification of key independent variables that are most likely to be salient in risk adjustment models for several relevant dependent variables. Thus, we recommend that researchers consider these results as a standard for planning analyses, and to incorporate the range of risk factor (independent) variables that have been shown here to influence child welfare outcome indicators (dependent variables).

In setting these analyses as a type of standard we recognize that both NCANDS and AFCARS are subject to modification over time that may improve the both the ability to construct longitudinal data (Federal Register, 2020) and that over time the lower quality data are subject to improvement. That said, we strongly recommend against the use of child and family risk factor data without a careful review of the state and submission year specific concerns that might attend their appropriate application. In this study we applied alternative methods such as county indicators of poverty for example, and certainly recommend the use of such methods when utilizing the data, and especially if the aim is to estimate national models. At the very least researchers should acknowledge potential limitations in the use of risk factor data and describe any steps taken to assess their quality.

Operationally, we believe that RAPIDS provides a streamlined and useful way to use both existing Child File and AFCARS variables and a range of newly created variables. Further, through the documentation provided in open access the potential user will have access to the documentation and procedures used to build on the analyses and recommendations for appropriate use of the data to tackle their research aims. We see the creation of “episode based” rather than “year-end status” foster care data as a real advantage for future researchers. Substantively, this paper demonstrates that different strata of variables can be combined to increase predictive power for a range of outcome variables. RAPIDS data also provide a potential for real-time interaction with data which was heretofore impossible at the national level. The possible applications of RAPIDS are almost limitless, and we look forward to seeing what uses future researchers find for it.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.chiabu.2025.107262.

Data availability

The authors do not have permission to share data.