Testing a QuantCrit-Informed Approach to the Empirical Study of Race/Ethnicity and Child Maltreatment

Abstract

The past several years have seen calls from QuantCrit scholars to “disaggregate” samples into same-race groups. To date, however, there has been no attempt to empirically evaluate the benefits of disaggregation within a child welfare sample. Using a child maltreatment dataset derived from the National Child Abuse and Neglect Data System and Census data, we empirically evaluate the utility of employing sample disaggregation (in which separate records are created for White, Black and Latino populations in each county) as well as variable creation disaggregation (in which we avoid using “full county” economic measures, but instead employ “same race/ethnicity” measures). Using model fit and convergence with findings from individual-level studies as evaluation metrics, we find that both kinds of disaggregation are demonstrably beneficial. We recommend that sample and variable disaggregation be considered by any future researchers using national geographically structured child maltreatment data.

Keywords

child maltreatment economics racial trauma

Introduction

In the past decade, QuantCrit perspectives have challenged traditional methodological and statistical approaches to understanding and analyzing race/ethnicity (Carter & Hurtado, 2007; Covarrubias, 2011; Gillborn et al., 2018). This paper operationalizes and empirically tests some of those suggestions quantitatively using national child maltreatment data. We demonstrate that a disaggregation approach employing racial/ethnic sample disaggregation and race/ethnicity-specifically disaggregated variable creation can be a useful quantitative method for better understanding race/ethnicity in child maltreatment. We apply this disaggregation approach to a geographic analysis of merged child maltreatment and Census data. We compare a disaggregation approach against currently prevalent methods (e.g. Fluke et al., 2019; Smith et al., 2021), using statistical models, scatterplots, and correspondence with the best available existing individual-level research (Maloney et al., 2017; Putnam-Hornstein et al., 2013; Slopen et al., 2016).

The focus of this paper is purely methodological. We are only interested in discussing the advantages or disadvantages of using different design approaches in child maltreatment research (sample disaggregation and Race/Ethnicity specific variable creation). While the substantive findings may be of interest to some readers, our goal is not to add to the substantive literature, but purely to evaluate two innovative methodological approaches to see if they warrant more general use in child welfare research. To the degree that disaggregating samples and variables can be shown to be empirically preferable to non-disaggregation approaches, quantitative precision can be advanced and QuantCrit concerns about carefully considering separate racial/ethnic groups can be simultaneously addressed.

The innovations suggested in this paper are consistent with a QuantCrit perspective. Although QuantCrit is very broad, key concerns include the centrality of racism and the difficulty of quantifying it, an assertion that numbers are not neutral, that categories should be critically evaluated, and that data must be interpreted critically (Gillborn et al., 2018; Dettlaff et al., 2021). Some authors apply QuantCrit principles in a directly statistical way, using insights from a critical perspective to rigorously refine existing empirical quantitative methods in an attempt to better capture the different experiences of racial/ethnic groups (Carter & Hurtado, 2007; Covarrubias, 2011).

Many QuantCrit researchers are concerned with “how participants are categorized into racial/ethnic groups, with less attention paid to how researchers then use these categorizations in their analyses” (Viano & Baker, 2020). By focusing only on the labeling per se, and not the analysis of the labels, researchers create a gap in knowledge. Another methodological tension often identified involves “comparing groups or highlighting variability in a single group” (Carter & Hurtado, 2007). Generally, a QuantCrit perspective prefers sample disaggregation (Covarrubias, 2011; Sablan, 2019). In simple terms, they see value in running separate models for different subpopulations, rather than including all subjects in a single sample and statistically controlling for race/ethnicity.

In this respect, child welfare epidemiology has room for refinement. Although some studies do present separate side by side analyses considering different races in separate models (Freisthler et al., 2007; Klein & Merritt, 2014), it is still common for single models including all races to be used. Among studies considering child maltreatment at the county level of observation (the lowest geographic level possible with national data), independent variables representing race/ethnicity are often simply percentages of White residents (Fluke et al., 2019; Smith et al., 2017) or population percentages for multiple racial/ethnic groups (Maguire-Jack, Font, et al., 2020a; Morris et al., 2019). At the dependent variable level, child maltreatment rates are generally not disaggregated. Instead, maltreatment is generally measured as a total (all-race/ethnicity) rate. This non-disaggregation approach with racial/ethnic compositions as independent variables and total child maltreatment rates as a dependent variable has three notable limitations for community-level research on race/ethnicity and child maltreatment.

First, a non-disaggregated approach cannot parse child maltreatment among different racial/ethnic groups within communities. At best, we can assess overall child maltreatment experiences of Black-majority communities (or other racial/ethnic group-majority communities), but this presents obvious limitations in understanding the child maltreatment experiences of Black children (or other racial/ethnic groups) in different community contexts (e.g., Black-minority communities).

Second, a non-disaggregated approach is not ideally suited to assess community risk/protective factors for child maltreatment experiences for specific race/ethnicity groups (e.g., Latino poverty) (Freisthler et al., 2007; Merritt, 2021; Mixon-Mitchell & Hanna, 2017). This can be problematic given that risk factors, protective factors, and family perceptions of child protective services may vary by race/ethnicity (Maguire-Jack, Lanier, et al., 2020b). Individual-level research can address this limitation by inclusion of interaction terms between race/ethnicity and all variables of interest, which can estimate the potential effects of variables of interests (e.g., poverty) on child maltreatment risk among a certain racial/ethnic group. However, this is not possible for community-level research because interaction terms between racial/ethnic compositions (e.g., percentages of Black residents) and variables of interest (e.g., poverty rates) still estimate the potential impacts of community risk/protective factors on child maltreatment rates among total children rather than within a specific race/ethnicity group. Even for individual-level research, considering all interactions between race/ethnicity and variables of interest is methodologically burdensome. Further, when other interactions exist which do not include race/ethnicity are of interest, we would now be required to look at three-way interaction terms, which are notoriously difficult to interpret (Dawson & Richter, 2006). Put simply, presentation of multiple, racially disaggregated models is generally easier to interpret in individual-level research, and it is the only way to estimate race/ethnicity-specific relationships between community risk/protective factors and child maltreatment rates in community-level research.

Additionally, a non-disaggregated approach also uses total (all-race/ethnicity) independent variables (e.g., poverty rates among all children). However, experiences of community contexts are almost always very different between racial/ethnic groups in the same community. All-race/ethnicity independent variables do not measure race/ethnicity-specific community characteristics and therefore may have limited usefulness in revealing race/ethnicity-specific child maltreatment experiences (Thurston & Miyamoto, 2020). For example, in Hennepin County Minnesota, the White median family income is $121,239, while the Black median family income is $44,383. The total county median family income of $105,514 is fairly close to the White income but is more than twice as high as the Black income. Such distortions between single-race and all-race income measures are necessarily higher as the proportions of minorities in a population is lower, causing a systematic misrepresentation of income metrics by race/ethnicity. These limitations precisely mirror philosophical and moral concerns that minority groups deserve unique and tailored understanding, not merely characterization of how they vary from majority populations (McCoy, 2020).

Finally, a disaggregation approach acknowledges emerging QuantCrit concerns that different racial/ethnic groups deserve and require separate analyses. These concerns arise from a desire to acknowledge different groups as worthy of study and not as mere departures from “normative” dynamics found within majority groups (Carter & Hurtado, 2007; Viano & Baker, 2020). Consistent with a QuantCrit perspective, disaggregation encourages researchers and readers to consider groups separately and attempt to reach a more nuanced understanding of dynamics within each racial/ethnic group.

Current Study

This study proposes a disaggregation approach with sample disaggregation and race/ethnicity-specific economic indicators (“disaggregated”, if you will, from other groups). We empirically test this approach by comparing county-level associations between overall (aggregated) economic indicators (i.e., median incomes and child poverty rates) and child maltreatment report (CMR) rates with race/ethnicity-specific (disaggregated) economic indicators and CMR rates. Economic status is widely understood to be among the strongest predictors of child maltreatment incidents and reports (Berger & Waldfogel, 2011; Cancian et al., 2013; Drake et al., 2021; Pelton, 2015). Unfortunately, the only universal national dataset capturing CMRs—The National Child Abuse and Neglect Data System (NCANDS) Child File—includes no reliable economic indicators. Researchers have attempted to remedy this lack by utilizing contextual economic factors, such as geographic-level median incomes or poverty rates (Fluke et al., 2019; Kim & Drake, 2018; Krieger, 1992; Maguire-Jack, Font, et al., 2020a; Morris et al., 2019). This poses an ideal venue for developing and empirically testing a disaggregation approach for community-level research on child maltreatment by race/ethnicity. Some studies have disaggregated their samples by race/ethnicity (e.g. Jonson-Reid et al., 2013; Kim & Drake, 2018; Wulczyn et al., 2013), and there has been some preliminary use of race/ethnicity-specific variables in child maltreatment research (Drake et al., 2009; Edwards, 2019; Kim & Drake, 2018; Wulczyn et al., 2013), but this is not the norm and these approaches have not been rigorously compared to more conventional methods.

Sample and Variable Disaggregation

We first create a single dataset for all children. We then disaggregate samples by race/ethnicity and create three more datasets, one each for each Black, White, and Hispanic/Latino/a/x (hereafter Latino) children. Correspondingly, we also disaggregate economic variables. Instead of using only the common overall (all persons in county) economic variables, we use both overall and race/ethnicity-specific variables. For example, in 2017, the overall child poverty rate for St. Louis County, MO was 13.5%, while the race/ethnicity-specific values were quite different—29.4% for Black residents and 5.4% for White residents (US Census Bureau, n.d.). CMR rates also widely differ by race/ethnicity in a county (Kim & Drake, 2018). We model both overall and race/ethnicity-specific variables based on overall and racially/ethnically-disaggregated analysis datasets.

In our analyses, we examine three types of relationships. The first type is the relationship between overall economic indicators and overall CMR rates. This is a conventional approach with neither variable creation disaggregation nor sample disaggregation.

The second type of analysis we employ examines the relationship between overall economic indicators and race/ethnicity-specific CMR rates. This is a partial disaggregation approach to examine how overall (all-race/ethnicity) community contexts explain race/ethnicity-specific CMR rates. That is, we use sample disaggregation by creating separate analysis datasets for each of race/ethnicity-specific dependent variables (i.e., race/ethnicity-specific CMR rates) but did not use variable creation disaggregation for community contexts.

The third type of relationship we examine is the relationship between race/ethnicity-specific economic indicators and corresponding race/ethnicity-specific CMR rates (e.g., Black child poverty rates and Black CMR rates). This is a full disaggregation approach (separate analysis datasets for race/ethnicity-specific dependent variables with economic independent variables specific to that race/ethnicity). We compare these three types of relationships to empirically demonstrate whether a quantitative method with a disaggregation approach can provide valid findings for race/ethnicity-specific understanding of child maltreatment. If there is consistency between the first and the third types, it may be unnecessary to pursue race/ethnicity-specific analyses as overall understanding is applicable to race/ethnicity subgroups. We compare the second and third types of relationships to determine if overall and race/ethnicity-specific community contexts differentially explain race/ethnicity-specific CMR rates. We examine and compare these relationships visually using scatterplots, as well as numerically using statistical models.

In addition, we take advantage of an opportunity to evaluate construct and comparative validity by comparing present community-level findings to prior individual-level work. Large scale individual-level studies (Maloney et al., 2017; Putnam-Hornstein et al., 2013) using disaggregation by race/ethnicity, have shown that while child maltreatment rates among Black children are higher than among White children, this difference disappears or even very slightly reverses after adjusting for economic controls. In addition, many prior studies have found that the relationship between poverty and child maltreatment differs by race/ethnicity, with the relationship slope being “steeper” for White children, implying that the difference between poor and wealthy White CMR rates is more pronounced than the difference between poor and wealthy Black CMR rates (Bywaters et al., 2016; Drake et al., 2009; Kim & Drake, 2018; Putnam-Hornstein et al., 2013; Slopen et al., 2016). For a discussion of possible underlying reasons for this observed effect, see Drake et al. (2009) and Putnam-Hornstein and Needell (2011). We use these findings as benchmarks to determine which approach creates findings more consistent with individual-level work.

Our research questions are straightforward: (1) Do race/ethnicity-specific relationships between economic indicators and CMR rates differ from overall relationships? (2) Do overall or race/ethnicity-specific economic indicators better explain race/ethnicity-specific CMR rates? (3) Do overall or race/ethnicity-specific economic indicators generate findings which converge better with prior individual-level findings?

Methods

Data

We obtained CMR data from the NCANDS Child File (National Data Archive on Child Abuse and Neglect, 2020), which is an annually produced child-level dataset of all CMRs screened-in to each state’s CPS agency. We used all CMRs from 50 states and the District of Columbia (DC) received during the 2017 federal fiscal year. Overall and race/ethnicity-specific population, income, and poverty data were obtained from the U.S. Census Bureau’s American Community Survey (ACS) 2015-2019 5-Year file (U.S. Census Bureau, n.d.), centered on 2017 to match the Child File.

The 2017 Child File contained 4,279,060 screened-in reports (each a unique child-report pairing). We aggregated these reports into counties. For confidentiality reasons, the Child File suppressed county identifiers of reports if reports were from counties with <1000 reports. State identifiers of these reports were available. Consistent with prior work (Kim & Drake, 2018), we aggregated these reports from suppressed counties into one pseudo-county per state. Among 3142 counties in 50 states and DC, 839 counties were not suppressed, and 2303 counties were suppressed in the 2017 Child File. We aggregated the suppressed counties into 49 pseudo-counties (two states had no suppressed county data in the 2017 Child File). The not-suppressed 839 counties covered 81.7% of U.S. children and the 49 pseudo-counties covered 17.4% of U.S. children. Altogether, we used 888 counties (and pseudo-counties) for analysis of overall (i.e., all children) CMR rates. We lost no data during this aggregation, and all 888 counties had ≥1000 resident children.

We further created race/ethnicity-specific datasets. For reliable measures of race/ethnicity-specific CMR rates, we excluded counties with <1000 race/ethnicity-specific children. For example, counties with <1000 White children were excluded from the White-specific dataset. We used 887 counties for White-specific analysis, 574 counties for Black-specific analysis, and 698 counties for Latino-specific analysis.

The population for each pseudo-county was obtained by summing the populations of counties in the ACS 2015-2019 5-Year file that were suppressed in the 2017 Child File. Income and child poverty metrics for each pseudo-county were obtained by taking a weighted (by population) average of median incomes and child poverty rates.

Measures

We created four outcome variables: one overall and three race/ethnicity-specific (White/Black/Latino) CMR rates (see Table 1). The overall CMR rate measured the number of reports per 1000 children per county. The Latino-specific CMR rate, for example, measured the number of reports per 1000 Latino children per county. We divided racial/ethnic groups as non-Latino White alone (hereafter White) children, any-ethnicity Black alone (hereafter Black) children, and any-race Latino (hereafter Latino) children. Because the ACS did not provide ethnicity-specific income data for Black alone children, we were unable to create a non-Latino Black alone group, as we did for White children. Our any-ethnicity Black alone and any-race Latino groups slightly overlapped, with children who identified as Latino and Black alone categorized into both. This was done to be consistent with Census race/ethnicity-specific data.

Table 1.

Descriptive Statistics, United States Counties, 2017.

Variable	Mean (SD)
Outcomes
All children¹ (N = 887 counties)
Overall child maltreatment report rate: number of reports per 1,000 children	84.6 (51.7)
White children (N = 887 counties)
White-specific child maltreatment report rate: number of reports per 1,000 White children	74.6 (51.3)
Black children (N = 573 counties)
Black-specific child maltreatment report rate: number of reports per 1,000 Black children	108.4 (61.7)
Latino children (N = 695 counties)
Latino-specific child maltreatment report rate: number of reports per 1,000 Latino children	51.7 (33.0)
Predictors
All children (N = 888 counties)
Overall median family income: Median income among all families (1k US$)	73.4 (18.2)
Overall child poverty rate: percentage of children living in poverty (%)	19.2 (7.7)
White children (N = 887 counties)
Overall median family income: Median income among all families (1k US$)	73.6 (18.1)
Overall child poverty rate: percentage of children living in poverty	19.1 (7.5)
White-specific median family income: Median income among White families (1k US$)	81.8 (22.5)
White-specific child poverty rate: percentage of White children living in poverty	13.4 (6.7)
Black children (N = 573 counties)
Overall median family income: Median income among all families (1k US$)	76.6 (19.3)
Overall child poverty rate: percentage of children living in poverty	19.1 (7.7)
Black-specific median family income: Median income among Black families (1k US$)	52.8 (18.7)
Black-specific child poverty rate: percentage of Black children living in poverty	33.7 (13.7)
Latino children (N = 695 counties)
Overall median family income: Median income among all families (1k US$)	76.2 (18.7)
Overall child poverty rate: percentage of children living in poverty	18.4 (7.2)
Latino-specific median family income: Median income among Latino families (1k US$)	53.9 (13.9)
Latino-specific child poverty rate: percentage of Latino children living in poverty	28.3 (10.8)

¹All children refers to the whole sample of children, including White, Black, Latino, and all other racial/ethnic groups.

We constructed four variables (i.e., one overall and three race/ethnicity-specific variables) for each of two predictors (i.e., median incomes and poverty rates). County-level median incomes were obtained from the ACS measure Median Family Income in the Past 12 Months (2019 Inflation-Adjusted Dollars). County-level child poverty rates were computed based on the ACS measure Poverty Status in the Past 12 Months.

Analysis

We fit 14 models (see Table 2). We estimated two “all children” models: (1) Model 1 for the association between overall median incomes and overall CMR rates and (2) Model 2 for the association between overall child poverty rates and overall CMR rates. We estimated four “White children” models (Models 3-6) for the associations between each pair of four predictors (i.e., overall median incomes, overall child poverty rates, White-specific median incomes, and White-specific child poverty rates) and one outcome (i.e., White-specific CMR rates). Similarly, we also fit four “Black children” models (Models 7-10) and four “Latino children” models (Models 11-14). We estimated all these associations separately at the bivariate level with no control variables. We used multilevel negative binomial models, using the R package glmmTMB (Brooks et al., 2017), to handle nesting of counties in states and the distributions of the outcome variables. Model fit was measured via the Akaike information criterion (AIC).

Table 2.

Multilevel Negative Binomial Models of Median Family Income (MFI), Child Poverty Rates (CPR), and Child Maltreatment Report Rates (CMRR), United States Counties, 2017.

Model	Approach	Predictor	Outcome	IRR (95% CI)	AIC
All Children (N = 888)
Model 1	Conventional	Overall MFI	Overall CMRR	0.982 (0.981, 0.984)	27096.2
Model 2	Conventional	Overall CPR	Overall CMRR	1.028 (1.027, 1.030)	27321.3
White Children (N = 887)
Model 3	Partial disaggregation	Overall MFI	White-specific CMRR	0.977 (0.976, 0.979)	26170.9
Model 4	Partial disaggregation	Overall CPR	White-specific CMRR	1.030 (1.028, 1.031)	26462.4
Model 5	Full disaggregation	White-specific MFI	White-specific CMRR	0.979 (0.977, 0.980)	25960.8
Model 6	Full disaggregation	White-specific CPR	White-specific CMRR	1.055 (1.053, 1.057)	26276.3
Black Children (N = 573)
Model 7	Partial disaggregation	Overall MFI	Black-specific CMRR	0.991 (0.990, 0.993)	16158.4
Model 8	Partial disaggregation	Overall CPR	Black-specific CMRR	1.015 (1.014, 1.017)	16190.2
Model 9	Full disaggregation	Black-specific MFI	Black-specific CMRR	0.988 (0.986, 0.989)	16080.8
Model 10	Full disaggregation	Black-specific CPR	Black-specific CMRR	1.013 (1.011, 1.014)	16139.5
Latino Children (N = 695)
Model 11	Partial disaggregation	Overall MFI	Latino-specific CMRR	0.993 (0.991, 0.994)	18753.8
Model 12	Partial disaggregation	Overall CPR	Latino-specific CMRR	1.011 (1.009, 1.012)	18777.9
Model 13	Full disaggregation	Latino-specific MFI	Latino-specific CMRR	0.992 (0.990, 0.993)	18757.2
Model 14	Full disaggregation	Latino-specific CPR	Latino-specific CMRR	1.009 (1.007, 1.010)	18770.4

N = number of counties. IRR = incident rate ratio. IRR > 1 indicates a positive association, and IRR < 1 indicates a negative association. AIC = Akaike Information Criterion. A lower AIC value indicates a better model fit.

Results

Table 1 reports the descriptive statistics. For all children, the mean CMR rate was 84.6 per 1000 children. Race/ethnicity-specific CMR rates were highest for Black children (108.4 per 1000 Black children), followed by White children (74.6 per 1000 White children), and Latino children (51.7 per 1000 Latino children). The all-children dataset showed a mean overall median income of $73.4k and a mean overall child poverty rate of 19.2%. These mean values were similar to the overall mean values for both median income and child poverty rate across the three race/ethnicity specific datasets, which ranged from $73.6k-$76.6k and 18.4%-19.1%, respectively. For race/ethnicity-specific economic indicators, White families generally had better economic conditions than Black and Latino families in counties. The mean White-specific median income ($81.8k) was much higher than the mean Black- and Latino-specific median incomes ($52.8k and $53.9k, respectively). The mean White-specific child poverty rate (13.4%) was much lower than the mean Black- and Latino-specific child poverty rates (33.7% and 28.3%, respectively).

Table 2 reports the results of the multilevel negative binomial models. First, we compared the relationships based on a conventional approach and a full disaggregation approach to examine whether race/ethnicity-specific relationships differed from overall relationships. We found that they were different. The incidence rate ratio (IRR) of the overall relationship between median incomes and CMR rates (Model 1) was 0.982 (95% CI: 0.981, 0.984), indicating that per $1000 increase in overall median family incomes, overall CMR rates decreased by 1.8%. While 1.8% is a small decrease, it is worth noting that this decrease occurs for each $1000 increase in median family income. The White-specific income-CMR relationship (Model 5; IRR: 0.979; 95% CI: 0.977, 0.980) was stronger than the overall relationship, while the income-CMR relationships for Black children (Model 9; IRR: 0.988; 95% CI: 0.986, 0.989) and Latino children (Model 13; IRR: 0.992; 95% CI: 0.990, 0.993) were weaker than the overall relationship. These models indicate that per $1000 increase in overall median family incomes, White CMR rates decreased by 2.1%, while Black and Latino CMR rates decreased by 1.2% and 0.8%, respectively. The poverty-CMR relationships also differed between overall and race/ethnicity-specific models. The IRR of the overall poverty-CMR relationship (Model 2) was 1.028 (95% CI: 1.027, 1.030), indicating that per one percentage-point increase in child poverty rates, CMR rates increased by 2.8%. The poverty-CMR relationships were stronger for White children (Model 6; IRR: 1.055; 95% CI: 1.053, 1.057) and weaker for Black children (Model 10; IRR: 1.013; 95% CI: 1.011, 1.014) and Latino children (Model 14; IRR: 1.009; 95% CI: 1.007, 1.010).

Second, we compared the relationships based on a partial disaggregation approach and a full disaggregation approach to examine whether race/ethnicity-specific economic indicators explained CMR rates better than overall economic indicators (Table 2). Specifically, we assessed model fit by AIC (lower the value, better the fit) between models with overall economic indicators and models with race/ethnicity-specific economic indicators (Models 3 versus 5, 4 versus 6, …, and 12 versus 14). For example, the model with White-specific median incomes (Model 5; AIC: 25960.8) showed better model fit than the model with overall median incomes (Model 3; AIC: 26170.9). All findings except for the income-CMR relationships for Latino children showed that models using race/ethnicity-specific economic indicators had better model fit (i.e., a lower AIC value) compared with models using overall economic indicators. This supports a full disaggregation approach over a partial disaggregation approach and suggests the importance of considering differential racial/ethnic community contexts while evaluating race/ethnicity-specific CMR rates.

We further compared partial disaggregation and full disaggregation approaches visually using scatterplots for both Black/White and Latino/White distributions of economic indicators versus CMR rates (Figures 1 and 2, respectively). The “+” and “O” shapes represent counties in corresponding race/ethnicity-specific datasets. For example, in Figure 1, the “+” shapes represent counties in the Black-specific dataset, and the “O” shapes represents counties in the White-specific dataset. The prediction lines reflect the IRRs of the models in Table 2.

Figure 1.

Black/White Distributions of Child Maltreatment Report Rates and Economic Indicators for Overall (Top) and Race/Ethnicity Specific (Bottom) Measures, United States Counties, 2017.

Figure 2.

Latino/White Distributions of Child Maltreatment Report Rates and Economic Indicators for Overall (Top) and Race/Ethnicity Specific (Bottom) Measures, United States Counties, 2017.

In both Figures 1 and 2, the X axis distributions for overall and race/ethnicity-specific metrics are very different. The very strong economic stratification by race/ethnicity present in the United States is clearly visible in the race/ethnicity-specific measures, but completely masked in the overall measures. The disaggregated data confirm visually that minorities experience far worse economic conditions, often even within the same county. In this regard, the race/ethnicity-specific income scatterplots are far more informative.

Lastly, we compared our community-level findings to prior individual-level findings to assess whether findings from overall or race/ethnicity-specific indicators converge better with prior individual-level findings. In Figure 1 (Black/White), we see that the graphs with overall economic indicators include essentially parallel lines with Black children being reported at higher rates than White children. This is inconsistent with prior individual-level studies which generally do not show higher CMR rates for Black children after controlling for economic conditions (Maloney et al., 2017; Putnam-Hornstein et al., 2013; Slopen et al., 2016). In the race/ethnicity-specific graphs, the lines are no longer parallel, but show that in poor areas, White children are at higher risk of reporting, while in wealthier areas, Black children are at higher risk, a well-known and previously established interaction effect shown in numerous prior studies (Bywaters et al., 2016; Drake et al., 2009; Kim & Drake, 2018; Putnam-Hornstein et al., 2013; Slopen et al., 2016).

In Figure 2, we see that White CMR rates are higher than Latino CMR rates, especially among counties with lower median incomes or higher child poverty rates. This gap is clearly larger in graphs by race/ethnicity-specific metrics than those by overall metrics. Some may be surprised by the higher CMR rates among White children than among Latino children. However, this tendency is well recognized in medical and child welfare research as “Latino paradox” or “Healthy Immigrant Effect”, suggesting that despite low economic conditions, Latino populations have lower rates of medical and child maltreatment problems, perhaps due to cultural protective factors, such as familism, religiosity, and social support (Millett, 2016). In both Black/White and Latino/White graphs, the findings based on race/ethnicity-specific measures converge better with prior individual-level findings compared with the findings based on overall measures.

Discussion

The purpose of this paper was to empirically test two QuantCrit-inspired approaches to data management. The first involves “sample disaggregation” in which race/ethnicity groups are examined in separate, same-race models. The second approach involves using race/ethnicity-specific economic metrics rather than full area metrics (e.g. using Black poverty rates as a contextual poverty metric for Black child maltreatment rates, instead of using total poverty rates). We found that these approaches have merit – that they work better statistically and produce findings which align more closely with prior individual-level studies than do traditional approaches. As an additional benefit consistent with QuantCrit sensibilities, there is not only increased interpretability in examining single-race scatterplots or models, but racial differences can be presented side by side, rather than as a description of how one race diverges from another.

When race/ethnicity-specific data exist, we think that sample and variable disaggregation should be employed. The availability of race/ethnicity-specific data are wide-spread. Most U.S. Census variables are measured by race/ethnicity, including geographic mobility, household compositions, educational attainment, health insurance status, and receipt of public assistance. Race/ethnicity-specific public health data, including data around births and deaths, can be obtained from the WONDER online databases developed by the Centers for Disease Control and Prevention. Certainly, using a disaggregation approach is not onerous. It is simply not that difficult to employ the approach we suggest, mainly requiring a few additional steps during the data management process. The most difficult transition is conceptual – the models currently being employed are, perhaps, imbued with a certain inertia, and moving to a different approach may take some time.

Lastly, we find that our sample disaggregation method yields findings more aligned with prior individual-level studies. Our findings indicate an interaction effect between income and race for Black and White children. In poor areas, White children have higher CMR rates than Black children, while in wealthy areas, Black children have higher CMR rates than White children. More research needs to be done to better understand the causal mechanisms and policy implications of these findings. In particular, it would be helpful to determine if the racial differences we and others have observed in the income*CMR slope are spurious or causal. This would require studies using approaches that would speak to causality, such as instrumental variable, sophisticated control or matching approaches or even randomized designs. Any such designs would ideally include more complex measures of economic status, including, for example, assets and community poverty, which differ radically between Black and White families (Drake & Rank, 2009; Oliver & Shapiro, 2018) and are often not included.

Limitations

We identify four main limitations to our study. Firstly, two of our race/ethnicity-specific categories, any-ethnicity Black alone and any-race Latino groups are not mutually exclusive. Children that identify as Latino and Black alone are placed into both sets, which limits the degree that these two groups can be compared.

While some individuals in our data do identify as having multiple racial identities, the number doing so is too small to allow analysis at the county level of observation in the NCANDS data. Looking at individuals endorsing multiple racial categories in more detail is certainly a worthy endeavor, but would have to be done using large county (only), state or national levels of observation, and cannot be done within the current analysis.

Our second main limitation is that we only test our disaggregation approach using median family incomes and child poverty rates to predict CMR rates. Although our approach yielded better model fit with these two variables as predictors of CMR rates, we are unable to provide evidence if similar improvements in model fit are present, or as pronounced, when examining other relationships. For example, more research needs to be conducted using different community-level predictors of CMR rates, including rates of public assistance utilization or teen pregnancy. Similarly, more research needs to be done examining the utility of our approach for other outcome measures, which might include rates of foster care placement or case management referrals.

Thirdly, our sample and variable disaggregation approach is naturally limited to the degree that each race/ethnicity is adequately represented across counties. If each race/ethnicity is not well representative in a given county, there is little to no data to disaggregate. As mentioned in the Data section, we discarded counties with <1000 children from each race/ethnicity-specific dataset to ensure reliable measures of CMR rates. Doing so created different samples across our three race/ethnicity-specific datasets, placing some limits of their comparability. Unfortunately, we are unable to prescribe a minimum threshold representativeness to warrant our sample and variable disaggregation approach, and think it should be carefully considered based on the research question and domain knowledge of the reader.

Lastly, for two reasons, we could not conduct race/ethnicity-specific analysis for Asian, American Indian and Alaska Native, Native Hawaiian and Other Pacific Islander, and Two or More Races groups. The first is that the low population counts of these racial groups in counties can create unstable estimates. The second is that the Census censors out low population counts, meaning that we could not measure race/ethnicity-specific median family income or child poverty rates for these racial groups in many counties. Certainly, more research on sample and variable disaggregation needs to be conducted focusing on these racial groups.

Conclusion

This study finds value in using race/ethnicity-specific sample disaggregation and variable specification for community-level research on child maltreatment. Such disaggregation improves statistical analyses, produces more visually credible distributions, and brings findings from geographic-level studies into closer conformity with individual-level studies. We hope this paper can serve as an impetus to incorporate racial/ethnic disaggregation throughout the field of child welfare epidemiology.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Dylan Jones

Hyunil Kim

Brett Drake

References

Berger

L. M.

Waldfogel

(2011). Economic determinants and consequences of child maltreatment. OECD. https://doi.org/10.1787/5kgf09zj7h9t-en

Brooks

M. E.

Kristensen

Benthem

K. J. van

Magnusson

Berg

C. W.

Nielsen

Skaug

H. J.

Mächler

Bolker

B. M.

(2017). GlmmTMB Balances speed and flexibility among packages for zero-inflated generalized linear mixed modeling. The R Journal, 9(2), 378-400. https://doi.org/10.32614/RJ-2017-066

Bywaters

Brady

Sparks

Bos

(2016). Inequalities in child welfare intervention rates: The intersection of deprivation and identity. Child & Family Social Work, 21(4), 452-463. https://doi.org/10.1111/cfs.12161

Cancian

Yang

M.-Y.

Slack

K. S.

(2013). The effect of additional child support income on the risk of child maltreatment. Social Service Review, 87(3), 417-437. https://doi.org/10.1086/671929

Carter

D. F.

Hurtado

(2007). Bridging key research dilemmas: Quantitative research using a critical eye. New Directions for Institutional Research, 2007(133): 25-35. https://doi.org/10.1002/ir.202

Covarrubias

(2011). Quantitative intersectionality: A critical race analysis of the chicana/o educational pipeline. Journal of Latinos and Education, 10(2), 86-105. https://doi.org/10.1080/15348431.2011.556519

Dawson

J. F.

Richter

A. W.

(2006). Probing three-way interactions in moderated multiple regression: Development and application of a slope difference test. The Journal of Applied Psychology, 91(4), 917-926. https://doi.org/10.1037/0021-9010.91.4.917

Dettlaff

A. J.

Boyd

Merritt

Plummer

J. A.

Simon

J. D.

(2021). Racial bias, poverty, and the notion of evidence. Child Welfare, 99(3), 61-89.

Drake

Jonson-Reid

Kim

Chiang

C.-J.

Davalishvili

(2021). Disproportionate need as a factor explaining racial disproportionality in the CW system. In: Racial disproportionality and disparities in the child welfare System (1st ed., pp. 159-176). Cham Springer.

10.

Drake

Lee

S. M.

Jonson-Reid

(2009). Race and child maltreatment reporting: Are Blacks overrepresented? Children and Youth Services Review, 31(3), 309-316. https://doi.org/10.1016/j.childyouth.2008.08.004

11.

Drake

Rank

M. R.

(2009). The racial divide among American children in poverty: Reassessing the importance of neighborhood. Children and Youth Services Review, 31(12), 1264-1271. https://doi.org/10.1016/j.childyouth.2009.05.012

12.

Edwards

(2019). Family Surveillance: Police and the Reporting of Child Abuse and Neglect. RSF: The Russell Sage Foundation Journal of the Social Sciences, 5(1), 50-70.

13.

Fluke

J. D.

Harlaar

Brown

Heisler

Merkel-Holguin

Darnell

(2019). Differential response and children Re-reported to child protective services: County data from the national child Abuse and Neglect data system (NCANDS). Child Maltreatment, 24(2). https://doi.org/10.1177/1077559518816381

14.

Freisthler

Bruce

Needell

(2007). Understanding the geospatial relationship of neighborhood characteristics and rates of maltreatment for black, hispanic, and white children. Social Work, 52(1), 7-16. https://doi.org/10.1093/sw/52.1.7

15.

Gillborn

Warmington

Demack

(2018). QuantCrit: Education, policy, ‘Big Data’ and principles for a critical race theory of statistics. Race Ethnicity and Education, 21(2), 158-179. https://doi.org/10.1080/13613324.2017.1377417

16.

Jonson-Reid

Drake

Zhou

(2013). Neglect subtypes, race, and poverty: Individual, family, and service characteristics. Child Maltreatment, 18(1), 30-41. https://doi.org/10.1177/1077559512462452

17.

Kim

Drake

(2018). Child maltreatment risk as a function of poverty and race/ethnicity in the USA. International Journal of Epidemiology, 47(3), 780-787. https://doi.org/10.1093/ije/dyx280

18.

Klein

Merritt

D. H.

(2014). Neighborhood racial & ethnic diversity as a predictor of child welfare system involvement. Children and Youth Services Review, 41, 95-105. https://doi.org/10.1016/j.childyouth.2014.03.009

19.

Krieger

(1992). Overcoming the absence of socioeconomic data in medical records: Validation and application of a census-based methodology. American Journal of Public Health, 82(5), 703-710. https://doi.org/10.2105/ajph.82.5.703

20.

Maguire-Jack

Font

S. A.

Dillard

(2020a). Child protective services decision-making: The role of children’s race and county factors. The American Journal of Orthopsychiatry, 90(1), 48-62. https://doi.org/10.1037/ort0000388

21.

Maguire-Jack

Lanier

Lombardi

(2020b). Investigating racial differences in clusters of adverse childhood experiences. The American Journal of Orthopsychiatry, 90(1), 106-114. https://doi.org/10.1037/ort0000405

22.

Maloney

Jiang

Putnam-Hornstein

Dalton

Vaithianathan

(2017). Black-White differences in child maltreatment reports and foster care placements: A statistical decomposition using linked administrative data. Maternal and Child Health Journal, 21(3), 414-420. https://doi.org/10.1007/s10995-016-2242-3

23.

McCoy

(2020). Black lives matter, and yes, you are racist: The parallelism of the twentieth and twenty-first centuries. Child and Adolescent Social Work Journal, 37(5), 463-475. https://doi.org/10.1007/s10560-020-00690-4

24.

Merritt

(2021). Lived experiences of racism among child welfare-involved parents. Race and Social Problems, 13, 63-72. https://doi.org/10.1007/s12552-021-09316-5

25.

Millett

L. S.

(2016). The healthy immigrant paradox and child maltreatment: A systematic review. Journal of Immigrant and Minority Health, 18(5), 1199-1215. https://doi.org/10.1007/s10903-016-0373-7

26.

Mixon-Mitchell

Hanna

M. D.

(2017). Race matters: Child protection and the communication process. Journal of Ethnic & Cultural Diversity in Social Work, 26(4), 366-381. https://doi.org/10.1080/15313204.2017.1344944

27.

Morris

M. C.

Marco

Maguire-Jack

Kouros

C. D.

White

Bailey

Rao

Garber

(2019). County-level socioeconomic and crime risk factors for substantiated child Abuse and Neglect. Child Abuse & Neglect, 90, 127-138. https://doi.org/10.1016/j.chiabu.2019.02.004

28.

National Data Archive on Child Abuse and Neglect (2020). National child abuse and neglect data system (child file), FFY 2017. Ithaca, New York: Cornell University.

29.

Oliver

Melvin L.

Shapiro

Thomas M.

(2018). Black wealth/white wealth: A new perspective on racial inequality. The inequality reader (pp. 296-303). Routledge.

30.

Pelton

L. H.

(2015). The continuing role of material factors in child maltreatment and placement. Child Abuse & Neglect, 41, 30-39. https://doi.org/10.1016/j.chiabu.2014.08.001

31.

Putnam-Hornstein

Needell

(2011). Predictors of child protective service contact between birth and age five: An examination of California’s 2002 birth cohort. Children and Youth Services Review, 33(11), 2400-2407. https://doi.org/10.1016/j.childyouth.2011.07.010

32.

Putnam-Hornstein

Needell

King

Johnson-Motoyama

(2013). Racial and ethnic disparities: A population-based examination of risk factors for involvement with child protective services. Child Abuse & Neglect, 37(1), 33-46. https://doi.org/10.1016/j.chiabu.2012.08.005

33.

Sablan

J. R.

(2019). Can you really measure that? Combining critical race theory and quantitative methods. American Educational Research Journal, 56(1), 178-203. https://doi.org/10.3102/0002831218798325

34.

Slopen

Shonkoff

J. P.

Albert

M. A.

Yoshikawa

Jacobs

Stoltz

Williams

D. R.

(2016). Racial disparities in child adversity in the U.S.: Interactions with family immigration history and income. American Journal of Preventive Medicine, 50(1), 47-56. https://doi.org/10.1016/j.amepre.2015.06.013

35.

Smith

B. D.

Kay

E. S.

Womack

B. G.

(2017). How can county-level maltreatment report rates better inform child welfare practice? Children and Youth Services Review, 79(C), 341-347. https://doi.org/10.1016/j.childyouth.2017.06.045

36.

Smith

B. D.

Wang

Smith

A. M.

(2021). A national study of child maltreatment reporting at the county level: Interactions among race/ethnicity, rurality and poverty. Children and Youth Services Review, 122, 105925. https://doi.org/10.1016/j.childyouth.2021.105925

37.

Thurston

Miyamoto

(2020). Disparity in child welfare referrals from public schools: An example of Simpson’s Paradox? Child Abuse & Neglect, 102, 104397. https://doi.org/10.1016/j.chiabu.2020.104397

38.

US Census Bureau . (n.d). American community Survey (ACS). Census.Gov. Retrieved January 31, 2022, from https://www.census.gov/programs-surveys/acs

39.

Viano

Baker

D. J.

(2020). How administrative data collection and analysis can better reflect racial and ethnic identities. Review of Research in Education, 44(1), 301-331. https://doi.org/10.3102/0091732X20903321

40.

Wulczyn

Gibbons

Snowden

Lery

(2013). Poverty, social disadvantage, and the black/white placement gap. Children and Youth Services Review, 35(1), 65-74. https://doi.org/10.1016/j.childyouth.2012.10.005