The present study had two separate samples having very different early childhood contexts: children with a CMR in early childhood (the Child Abuse and Neglect [CAN] sample) and children with AFDC but no CMR in early childhood (the AFDC sample). Specifically, the CAN sample (N = 2,111) included the full population of children from Missouri CPS records who met the following criteria during the sampling period (i.e., 1993–1994): (1) having a first-time CMR for alleged neglect, physical abuse, or sexual abuse; (2) aged 3 or younger; and (3) residing in St. Louis. It is worth noting that the CAN sample included children with an AFDC case (78.6%), as well as children with no AFDC case (22.4%). The large proportion of children with an AFDC case in the CAN sample was consistent with the well-known relationship between poverty and CMRs (Pelton, 2015). The AFDC sample (N = 1,923) included children randomly selected from Missouri AFDC records who met the following criteria during the sampling period (i.e., 1993–1994): (1) having an active AFDC case; (2) having no current or prior CMR; (3) aged 3 or younger; and (4) residing in St. Louis. Before this selection, the parent study stratified AFDC records by child age and residential region (St. Louis City vs. County). Then, the parent study randomly selected the AFDC sample children that were similar in number to the CAN sample children for each stratum. The present study uses “AFDC” to describe a sample, as well as a participation in a welfare program. When “AFDC” indicates a sample, we always affix “sample” (i.e., the AFDC sample). Otherwise, “AFDC” indicates the AFDC program. AFDC was replaced by TANF during the study period, so the term “welfare” is used to refer to either after baseline.

Both groups were followed from January 1995 (after the close of the sampling period) through March 2009. Children born in 1991 were followed from age 4 to 17, children born in 1992 were followed from age 3 to 16, children born in 1993 were followed from age 2 to 15, and those born in 1994 were followed from age 1 to 14. Data were reformatted into age-year observations, leaving 32,348 (age-year) observations for the CAN sample and 25,596 observations for the AFDC sample. Follow-up was ended at the time of a child’s death (0.66% of the CAN and 0.41% of the AFDC group), out-of-home placement due to the desire to follow CMRs in the family of origin (21.25% and 3.70%), and childbirth which may be indicative of emancipation or no longer residing in the family of origin (0.009% and 0.004%). The final sample sizes were 25,258 observations of 2,111 children in 235 baseline tracts for the CAN sample and 24,552 observations of 1,923 children in 210 baseline tracts for the AFDC sample.

Table 1 lists the study variables and data sources. We established temporal order between predictors and CMRs based on dates of service contacts in the data. It should be noted that a contact date of a service (e.g., welfare) may or may not reflect the actual onset of a given problem (e.g., low economic status). It is also worth noting that most administrative records of this study are statewide. Families moving out of St. Louis but within Missouri, therefore, were not lost to the study. Records for special education and juvenile court were limited to the St Louis region. The geographic limitations are mitigated by the fact that more disadvantaged families tend to make shorter distance moves although they are frequent (Coulton, Theodos, & Turner, 2012). Further, census data indicate that when families move they are more likely to move from St Louis City to the County not to neighboring Illinois communities (Jaco, 2019).

Variables that were subject to change over time were added as (yes/no) indicators within each person year. For ease of presentation, only child demographics and the primary poverty and CMR measures are discussed here. Discussion of other child, family, and neighborhood level risk factors (drawn from birth records, child welfare services, emergency room data, special education records, juvenile justice records, criminal records, and Census data) is provided in Appendix A in the Supplement.

The “lme4” package (version 1.1–15) in R (version 3.4.4) was used for the analyses. Multilevel logistic growth curve models were used to estimate the change in likelihood of having any (first or any subsequent) CMR at each age from 1 to 17 years as a function of various predictors measured by administrative records. We used a multilevel design because age-year observations were nested in children and in turn, children were nested in their baseline residential tracts. The degree of dependency in the nested data can be measured by the intraclass correlation coefficient (ICC) (Snijders & Bosker, 2012). The level-2 ICC values were 0.33 for the CAN sample and 0.40 for the AFDC sample and the level-3 ICC values were almost zero for both samples. The ICC values indicated that 33% to 40% of the variance of the CMR likelihood was placed between level-2 units (children/families) while almost no variance was located between level-3 units (baseline tracts). Given the substantial degree of dependency, particularly between level-2 units, applying a traditional model would lead to severe type I errors. Multilevel modeling adjust standard errors according to the nested data structure (Luke, 2004). The model-building process involved making decisions about parsimony versus model fit, following the approach suggested by Snijders and Bosker (2012). We selected a final model by including only parameters with meaningful contributions to model fit (see Appendix B in the Supplement for details).

Table 1 presents descriptive statistics. During the follow-up period, the CMR probability at a given age was higher for the CAN sample (11.4%) than the AFDC sample (5.5%). Most of the CAN sample children were also living in families receiving AFDC at baseline (only 22.4% not on AFDC). Both CAN and AFDC samples often received welfare (AFDC/TANF) at later ages, 35.6% and 30.7% across ages, respectively. Compared to the proportion of Black children in the St. Louis child population from Census data (35.1%), Black children were overrepresented by 1.9 times in the CAN sample (66.4%) and 2.3 times in the AFDC sample (81.5%). This was consistent with studies examining representation in poor or maltreated populations at the national level (Drake et al., 2011). Other risk factors controlled were generally more common in the CAN sample than the AFDC sample.

Table 2 reports the final models. The final models showed fairly high levels of accuracy in predicting future CMRs. The area under the ROC curve (AUC) indicates the overall level of predictive accuracy in balance between sensitivity and specificity. The final models predicted CMRs with 74.73% accuracy for the CAN sample (AUC = 74.73) and with 73.35% accuracy for the AFDC sample (AUC = 73.35). This suggests that using the estimated coefficients of the final models (γ in Table 2), one can predict a child’s CMR risk with an acceptable level of accuracy based on the child’s age and the child’s current state and prior history of risk factors. For example, Figure 1 presents the predicted risks by child age, current welfare, and prior welfare while fixing all other predictors to their grand mean.

In the CAN sample, a one-year increase of child age decreased the CMR likelihood by 13% (odds ratio [OR] = 0.87, 95% confidence interval = 0.86–0.88). This relationship did not significantly vary by current welfare, prior welfare, or child race (see Appendix B in the Supplement). In the AFDC sample, child age alone was not significant (OR = 0.99, 0.96–1.02), whereas the child age × current welfare interaction was significant (OR = 0.93, 0.90–0.97). Due to the significant interaction, we computed odd ratios and confidence intervals of child age by current welfare (on welfare vs. not on welfare) from the model results (see Appendix C in the Supplement). The results showed that in the AFDC sample, there was no significant association between child age and the CMR likelihood while currently not on welfare (OR = 0.99, 0.96–1.02), whereas the CMR likelihood decreased by 8% per one-year increase of child age while currently on welfare (OR = 0.92, 0.89–0.95). These relationships did not differ by prior welfare history or child race (see Appendix B in the Supplement).

The CMR likelihood was over two times higher while currently on welfare than not on welfare in both CAN sample (OR = 2.32, 1.98–2.71) and AFDC sample (OR = 2.08, 1.61–2.68). The relationship between prior welfare and CMR likelihood varied by current welfare status in both samples (see the Appendix B in the Supplement). This required additional computation to estimate odds ratios and confidence intervals of prior welfare by current welfare (see Appendix C in the Supplement). Among children in families not currently on welfare, a 10-percentage-point increase in prior welfare increased the CMR likelihood by 8% in the CAN sample (OR = 1.08, 1.05–1.11) and by 12 % in the AFDC sample (OR = 1.12, 1.08–1.17). While children were currently on welfare, this relationship disappeared for both the CAN (OR = 1.00, 0.97–1.03) and AFDC samples (OR = 0.97, 0.93–1.02).

Due to the complex interactions among child age, current welfare, and prior welfare, we plot prediction graphs in Figure 1. Overall, the CMR likelihood was much higher for the CAN sample than for the AFDC sample, especially at younger ages. The CAN sample children showed substantially higher CMR likelihoods at younger ages regardless of current welfare. The AFDC sample children also showed higher CMR likelihoods at younger ages, but only when currently on welfare. When currently not on welfare, the AFDC sample children showed no higher CMR risks at younger ages. The CMR likelihood was notably higher while currently on welfare. This gap persisted across ages in both samples, but was substantially diminished at older ages in the AFDC sample. Regarding prior welfare, a longer stay on welfare (i.e., 75% vs. 25% of the time) moderately increased CMR likelihoods only while currently not on welfare. In the CAN sample, the effect size of prior welfare was practically “0” when currently on welfare, and accordingly the corresponding prediction lines were almost completely overlapped. Among the AFDC sample children currently on welfare, those with a longer history of prior welfare had a slightly lower CMR likelihood. This difference, however, was not statistically significant.

Regarding race, Black children and other minority children had lower CMR likelihoods than White children in both samples while controlling for other variables. Compared to White children, the CMR likelihoods were lower by 16% to 35% for Black children (OR = 0.84, 0.74–0.95 in the CAN sample; OR = 0.65, 0.53–0.80 in the AFDC sample) and by 59% to 57% for other minority children (OR = 0.41, 0.21–0.81 in the CAN sample; OR = 0.43, 0.21–0.88 in the AFDC sample). These relationships did not significantly differ by child age, current welfare, and prior welfare (see Appendix B in the Supplement). Figure 2 shows that while controlling for other variables, Black children and other minority children have lower CMR likelihoods than White children at all ages from early childhood to adolescence regardless current welfare status.

An interesting pattern of CMR risk emerged while comparing the current state and the prior history of child health and behavioral risk factors as well as parental arrest records. Generally, a currently noted risk factor had a larger effect size than the prior history of the same risk factor. For example, in the CAN sample, indication of a child’s current mental health problem in a given child/year increased the CMR likelihood by 3.96 times (OR = 3.96, 2.80–5.59). In comparison, child prior mental health problems increased the CMR likelihood by 1.38 times (OR = 1.38, 1.09–1.76).

This study has several strengths. First, we estimated future CMR risks of very young high-risk children frequently encountered by the state service systems (i.e., the CAN and AFDC samples). Our findings therefore may be useful for secondary prevention efforts by further stratifying risk levels among high-risk children. Second, we examined both current state and prior history of risk factors while considering their changes over time. This allowed for improving predictive accuracy, but also for identifying an interesting longitudinal pattern. We found that the recent status of risk factors was more strongly linked to current CMRs than prior history, suggesting that prompt supports for families’ present needs may reduce CMRs. Using linked administrative data allowed for measuring the dates associated with CMRs, welfare, and other relevant risk factors while reducing bias associated with measuring risk through retrospective recall. Finally, we used multilevel growth curve modeling, a powerful and flexible way to handle missing or censored observations and to examine longitudinal changes and cross-level influences simultaneously (Luke, 2008). This method allowed us to continue following up children even after the first event and examining the risk of subsequent CMRs at specific ages. Most prior studies estimate only the risk of the first event (e.g., the first-time CMR or the first recurrence of CMR). Instead, we provide models that are useful to predict a child’s risk at any age and at any stage of CMR (i.e., first, second, third, and any subsequent CMRs). Traditional approaches using baseline information may be more useful for early prevention efforts. However, our approach using both baseline information and over-time changes can be useful to identify families at imminent risk for ongoing maltreatment at different child ages.

There are, however, limitations to the present approach. Despite the established temporal ordering, we are unable to establish causal relationships as there may be unmeasured factors that could influence the parameter estimates. This is not an uncommon problem in research focused on risk conditions that are not ethically amenable to randomization (e.g., poverty or usual care services in response to referral). Further, observing families in context is often vital to understanding complex longitudinal and socio-ecological processes related to engaging with multiple child, family and adult serving systems.

Another limitation concerns the use of welfare receipt as a proxy of economic status and child maltreatment reports as an outcome rather than actual child maltreatment events. Welfare receipt does not capture the full range of economic status and material need that may impact a family’s functioning (e.g., Berger, 2004). On the other hand, prior studies have consistently found that welfare receipt (Irwin, 2009; McDaniel & Slack, 2005; Slack et al., 2003) and other economic indicators (Courtney et al., 2005; Dworsky, Courtney, & Zinn, 2007; Maguire-Jack & Klein, 2015) were strongly associated with child maltreatment while controlling for marital status and other variables. Additionally, welfare receipt and the possible implications for families present in both welfare and CPS systems are arguably very relevant for policy (Berrick, Frame, Langs, & Varchol, 2006; Slack et al., 2007).

The second concern relates to whether or not a CMR reflects actual maltreatment concerns. Much of this was addressed in the introduction and background provided with the research question and is not repeated here. As stated, CMR have been found to be sufficient markers of a number of later outcomes (Hussey et al., 2005; Putnam-Hornstein, 2011), concurrent outcomes like child injury that may be indicative of maltreatment (Spivey et al., 2009), and child outcomes across domains controlling for poverty (Jonson-Reid et al., 2009). Further, data measuring maltreatment across multiple years have been found to predict similar incidence rates as lifetime self-report studies (Kim et al., 2017). Thus, the present of repeated measures of CMR over time decreases the likelihood of detection. It is, however, not possible to know whether a report reflects actual onset of abuse and neglect or simply when the concerns were sufficient to be noticed by others.

Other limitations are related to our use of secondary data. First, access to records for children with neither welfare receipt nor CMR in early childhood were not possible in the parent study. Therefore, findings cannot be generalized to child populations without early risk related to poverty or maltreatment. Second, this study provides no estimate for risk at birth due to the parent study’s sampling design. National data have identified exceptionally high CMR risks prior to the first year of life (Kim et al., 2017; U.S. DHHS, 2019; Wildeman et al., 2014). Future research is needed that can track groups prior to birth and continue throughout adolescence. Administrative data is by design limited to variables of interest for a given system. We were unable to assess need as compared to service receipt or system encounter. Nor are there sufficient indicators of protective factors in most administrative datasets. While the presence of a CMR in early childhood may be a good proxy for risk, the absence of a CMR is not the same as the presence of protective factors. Because the data came from a single state, outstate mobility could not be tracked or estimated. However, as stated in the methods, we believe the nature of the population and the unique context of the study location mitigate this concern. Our ability to examine parent level factors was largely limited to mothers due to the demographics of the families involved in welfare and CPS systems and we were unable to control for parental risk factors such as substance abuse. We also chose to limit community poverty controls to baseline due to the limitations in how addresses are captured over time. Only those with repeated CMR or welfare receipt would have updated addresses. Finally, another limitation has to do with the simple binary measure of most risk factors. While large administrative data can assure strong generalizability, such data commonly miss nuanced contextual information. For instance, the binary measure of child mental health does not provide a measure of functioning or severity.

Our study provides correlational evidence useful in evaluation of several theoretical perspectives of CMR risk. We found higher CMR risks at younger ages, especially among children who experienced a CMR in early childhood. This is consistent with the life course perspective (White & Wu, 2014) and supports the need for early intervention with young families to try to offset ongoing risk. While recent investments have been made in approaches like home visitation, most models have not shown strong effects in regard to preventing CMR (Jonson-Reid, Drake, Dunnigan, Kohl, & Auslander, 2020). Second, ecological transactional perspectives (Belsky, 1980; MacKenzie et al., 2011) suggest that other factors may interact with early childhood context to affect the relationship between child age and ongoing CMR risk. Our data suggest that interrupting ongoing CMR risk requires the ability to attend to current material needs as well as address child needs that may have arisen as a result of earlier experiences with poverty or CMR. The long established strong association of poverty with maltreatment reports (Pelton, 1994, 2015) has now been supplemented by an emerging body of research demonstrating that even small supplements to income can reduce maltreatment rates (Cancian et al., 2013; McLaughlin, 2017; Rostad et al., 2017). In regard to primary prevention, we need to identify ways of promoting protective early childhood context among the welfare population without CMR histories (Berrick et al., 2006; Putnam-Hornstein & Needell, 2011).

Concern continues regarding the overrepresentation of certain ethnic/racial groups in our child welfare system. Our work and others suggests that attempts to address disparities related to race in CPS contact must attend to the differential needs present among families. This includes attention to the resources in a community that need to be in place in order for a family to successfully parent (Maguire-Jack & Klein, 2015). This does not mean that we ignore potential biases in CPS or in other systems that inappropriately lead to under or overserving a population. Careful monitoring of flows in and through systems by race should continue as we improve the availability and quality of services to support families.

Finally, while primary prevention and early intervention are clearly preferable, the continued import of current child level risk, continued CMRs and welfare receipt, suggest that effective intervention at later stages is possible but must attend to a broader set of needs. Several years ago, child well-being outcomes were added to federal child welfare standards, but most states have struggled to meet those metrics. Although evidence-based parenting programs designed to address child behavior problems are believed to hold promise in preventing maltreatment by reducing the challenges to parenting related to the child’s behaviors (Barth, 2009), much less is known about how attending to child well-being (including intermediate effects of early maltreatment) over time may reduce ongoing CMR risk. It is also possible that addressing these more proximal concerns may have benefits in reducing risk for other forms of victimization outside the home (Finkelhor et al., 2007). Improved cross-sector data systems that allow for understanding engagement across systems to address child behavioral, educational and health needs are needed to understand what services families may or may not be accessing and the quality of care given within those systems to support these efforts (Jonson-Reid & Drake, 2016).