Key Words: Unemployment, Child Maltreatment,Prevention, Budget
Economic Conditions and Child Maltreatment: Delayed Effects of Rising Unemployment
Robert D. Sege, MD, PhD*
Jay L. Zagorsky, PhD**
Michelle Schlesinger*
* Division of Ambulatory Pediatrics, Department of Pediatrics, Boston University School of Medicine and Boston Medical Center, Boston, MA 02118
** Center for Human Resource Research, The Ohio State University, Columbus, Ohio
Abbreviations: National Child Abuse and Neglect Data System, NCANDS; Child Protection Services, CPS; Bureau of Labor Statistics, BLS, Bureau of Economic Analysis, BEA.
Word count: 3,108
Abstract: 201 words
ABSTRACT
Background: Child maltreatment is associated with poverty, and has long-term medical and economic consequences. Previous reports have identified individual risk and protective factors. Earlier reports suggested an association between local unemployment rates and child maltreatment reporting.
Objective: To investigate whether state-level unemployment rates predict child maltreatment rates.
Methods: Ordinary least squares regression was used to evaluate the relationship between unemployment and maltreatment rates. Rates of substantiated cases of child maltreatment by state, and state-level economic data were obtained from US federal data sources: Data on child maltreatment was extracted from the National Child Abuse and Neglect Data System and economic data from the Labor Department’s Current Population Survey.
Results: During the period from 1990 to 2010, regression analysis demonstrated the relationship between unemployment rates and the following year’s substantiated maltreatment rates was between 0.21 and 0.46. This suggests that each 1% increase in unemployment rates is associated with at least a 0.21 per 1,000 increase in substantiated child maltreatment reports one year later. These results are independent of other known social risk factors included in the analyses.
Conclusions: State-level unemployment predicts substantiated cases of child maltreatment. Policy-makers should consider indexing expenditures on child maltreatment prevention and treatment to the unemployment rate.
Background:
Approximately 1.25 million American children are maltreated annually.1 Overall, one in four children suffers from neglect; one in ten suffers from physical abuse; and one in twenty-five suffers from sexual abuse.2,3 Some of these children experience immediate physical consequences of child maltreatment—serious injuries, disability, or death. Recent evidence has demonstrated persistenteffects of maltreatment on children’s nervous and endocrine systems, interfering with healthy adult development and resulting in health consequences many decades after the initial victimization.4,5
Poverty is a risk factor for child maltreatment.6-9 Perpetrators of child maltreatment are more likely to be single parents, low-income, unemployed, and/or on welfare.2,10-18 The Family Stress Modellinks poverty and child maltreatment by noting that poverty leads to stress, in response to which parents exhibit negative or harsher parenting strategies.19-21 The strengthening families approach emphasizes the importance of concrete supports in times of need. In contrast, evidence-based maltreatment prevention programs largely focus on psycho-educational interventions directed at high-risk families, and only tangentially address economically-based stress.22,23
High quality data on unemployment and child maltreatment rates has been available since 1990, allowing further research on the macro level economic conditions affecting child maltreatment patterns.24-28 This study uses these data to explore the observation made by Krugman in 1986 that demonstrated a relationship between local unemployment and child physical abuse.29 More recently, Berger et. al. report a relationship between abusive head trauma and increased unemployment.30 Despite economic buffering provided by unemployment insurance, many families that experience unemployment have less financial resources and more stress. Family stress may increase with time after unemployment, as both unemployment insurance and personal financial resources become exhausted, and family relationships become more strained. Economic and family stress both increase risk for child maltreatment.4,17,18,25,31-35
This study investigated whetherchanges in unemployment rates is a useful predictor of changes in the number of substantiated child maltreatment reports.
METHODS
This research combines information from three US government data sources: the National Child Abuse and Neglect Data System, the Current Population Survey, and the Bureau of Economic Analysis’s Regional Economic Accounts Data. The raw data were adjusted and then used in a regression framework to estimate the impact of unemployment and other factors on child maltreatment rates.
Data Sources
National Child Abuse and Neglect Data System (NCANDS)
Statistics about child maltreatment are derived from case level data collected by each state’s child protective services agencies. These cases are input annually into the federal “National Child Abuse and Neglect Data System” or NCANDS. NCANDS data have been summarized and reported annually since 1990.36 Two of the key pieces of data this research uses from these reports are the estimates of the population under 18 years old and number of child victims.
The number of child victims reflects reports that have been received, screened in, investigated and substantiated by the state Child Protection Services (CPS) agency, and subsequently included in NCANDS.37 State-level analyses reported here account for the large state-level variation in the process of referring, screening, and substantiating cases of child maltreatment.
Current Population Survey
Monthly U.S. unemployment statistics are derived from the Bureau of Labor Statistics’(BLS) Current Population Survey.38 Each month interviewerscontact slightly more than 50,000 randomly selected households. In these interviews information about the past month’s activities of every person in the householdaged 16 and older are recorded. The Bureau of Labor Statistics’ (BLS) survey classifies individuals as employed, unemployed and not-in-the-labor-force. Individuals working for money, no matter how briefly are counted as employed. The employed also include people not working, but with a job, e.g., those on vacation. Persons who were not employed but who searched for work in the past month and were available to start if hired, are classified as unemployed. All remaining individuals are put in the not-in-the-labor-force category. The unemployment rate is calculated by dividing the number of unemployed by the total of number of people employed and unemployed.
National unemployment rates are taken directly from the Current Population Survey. Because a national sample of 50,000 households is insufficient to create precise unemployment rates for every state, the BLS creates special state level unemployment rates as part of a “Local Area Unemployment Statistics” program.39 The local area program first takes each state’s Current Population Survey figuresand then adjusts the figures using a regression methodology. These adjusted unemployment figures are the key explanatory variable used in this project.
Regional Economic Accounts Data
State-level information from the Department of Commerce Bureau of Economic Analysis (BEA) was used to capture population, personal income and employment figures provided by the BEA at the state level from their “Regional Economic Accounts.”40 BEA regional data track year-by-year the population of each state, and show steady population growth. The BEA’s personal income series is a year-by-year list of the total income earned from all sources for residents of a state. Average personal income was determined by dividing total income by the state population, and then adjusted using the BLS’ Consumer Price Index data41 to remove the impact of inflation. Finally, the year-to-year change in income was computed.
We also use BEA total employment figures to track the growth or shrinkage in a state’s job market. While employment and unemployment figures are related, there are differences between the concepts because some people hold multiple jobs, while others search for work but never find it.
Finally, this project uses educational attainment data taken from theAmerican Community Survey42, and earlier Census surveys.43 These data provide state-by-state information on the percentage of persons 25 years old and over who attained at least a high school degree.40 In 1990, 76.2% of the average state’s adults had a high school degree, which increased to 87% by 2009.
Data Analysis
State-specific maltreatment rates reflect differences in definitions, systems of response to concerns, and resource availability. Primary analyses were constructed from the change in maltreatment over time,; the dependent variable is the change in maltreatment rates for each state from year-to-year. Using year-to-year change, rather than absolute rates, minimizes the effects of differences in reporting and substantiation rates by state.
Ordinary least squares (OLS) regressions were used to estimate the relationshipbetween maltreatment and unemployment rates. In mathematical form our key baseline regression where the subscript t stands for the particular year of time (from 1990 to 2010), t-1 stands for time lagged one year, i stands for a particular state and β the coefficients being estimated is:
1)(Maltreati,t - Maltreati,t-1)=β0 + β1(Unempi,t - Unempi,t-1) + β2(Unempi,t-1 - Unempi,t-2)
We report additional regression results that explicitly account for other factors like changes in education and income, and fixed effects including the state and year. As a secondary analysis, we also analyzed the absolute amount of maltreatment cases using the following equation:
2)Maltreatmenti,t = β0 + β1Unemployment Ratei,t + β2Unemployment Ratei,t-1
We report two types of regressions; weighted and unweighted. Unweighted regressions cause each state to have the same impact, and are useful for predicting what will occur in the typical state. Weighted regressions correct for population differences between small states (e.g., Rhode Island) and more populous ones (e.g., California). Weighting gives larger states more impact, making the results useful for national estimates. Regression weights were constructed by dividing the number of people living in each state during a year by that year’s national population, using data from United States’ Census.
Aggregation bias results when calculations using separate pieces of data, such as individual case reports, are different from results using summary statistics, such as state level rates. Individual abuse case reports do not indicate the labor force status of the family members. Since data are not available it is impossible to directly test for aggregation bias. Nevertheless, we include a comparison of regressions using national information with regressions using state level information to partially investigate aggregation bias.
RESULTS
Since 1990, there has been declining trend in the number of child maltreatment victims (see Figure 1) tracked by NCANDS. In 1990, the year when national data were first released, the[rds1] average state reported 13.5 victims of child maltreatment for every 1,000. Abuse peaked in 1993 with 15.3 cases per 1,000. In 2010, the latest publically available year of information, the average rate had fallen to 9.98 per 1,000 children. The average yearly reduction in victimization is 0.24. Most regressions include a year variable which ensures the unemployment coefficients account for the declining trend in maltreatment.
Trends in unemployment rates from 1990 to 2010 are presented in Figure 2. In 1990, over 5% of the labor force was unemployed. Unemployment peaked at 7.5% in 1992 and then fell steadily until 2000. From 2000 until 2003 unemployment rose from 4% to 6%,fell in the mid 2000s, and peaked at 9.6% in 2010.
Change In Unemployment Regression Results
Table 1 contains the coefficients from using OLS regressions to estimate equation (1) using two decades of public data. In this table the key regression results are on the line labeled “1 Year Lag Change in Unemp.” The key coefficient in the first column is 0.22; when the national unemployment rate changes by 1 percentage point there is an associated rise of about ¼ more substantiated cases of child maltreatment per 1,000 the following year. The R2 figure at the column’s bottom indicates that tracking just the change in unemployment rates and a yearly trend explains about one-third (32%) of the change in national child abuse rates.
The national data in column (1) of Table 1 report the results of a highly aggregated data analysis. Columns (2) and (3) rerun the national regression, with and without using weights, using disaggregated state data. This increases the number of observations from 20 (the number of years in the analysis) to 968 (data from each state each year is one observation; not all states reported every year to NCANDS). Using state level information, instead of national, does not substantially change the results, but does improve the statistical significance of the key coefficients (p<0.1 in all cases). Comparing the national regression with the state regressions suggests that aggregation bias is not a key factor. As expected, the R2 drops, reflecting the observation that there are large state-level variations in the process of child abuse reporting, and that state policies may vary from year to year as well.
In contrast, coefficients in columns (2) to (7) on the “Change in Unemployment” are generally small, sometimes negative and lack statistical significance. This suggests there is no immediate impact of a change of unemployment on maltreatment either statistically or qualitatively.
Examination of possible confounding factors.
Unemployment is only one of many possible social factors that might influence the rate of child maltreatment. We explored the influence of other possible factors. These results, reported in columns 4 through 7 and described below, do not substantially change our primary conclusion.
Columns (4) and (5) demonstrate that including educational attainment, changes in the growth of personal income and the specific year have little impact on the key change in unemployment coefficients. After accounting for these three additional factors, unemployment rates are still significantly related to maltreatment after a one year lag.
Columns (6) and (7) expand the regression further by including the change in population growth rates, the change in employment growth and state dummy variables that check if state fixed effects were important in the analysis. State effects are not reported because the coefficients on these variables typically not statistically different from zero in Table 1. Including these factors results in coefficients in columns (6) and (7) on the lagged change in unemployment rates of 0.46 and 0.26 respectively. This suggests adding these extra explanatory factors do not change the findings.
Absolute Unemployment Regression Results
OLS regressions which estimate equation (2) are found in Table 2. This table’s coefficients are useful for estimating the number of maltreated children in each state given a particular set of economic conditions. For example, regression (9) shows that a state with an unemployment rate this year of 10% and a rate last year of 7% should have about 13.38 cases of maltreatment (7.18 - 0.57*10 + 1.70*7) for every 1,000 children. The coefficients on the “1 year Lag Unemp” line are uniformly positive, large and most are statistically significant. This suggests it is not only the lagged change in unemployment but also last year’s absolute level that influences child maltreatment; higher levels of unemployment appear to raise maltreatment.
To further test the result’s robustness we downloaded from the Center for Disease Control all U.S. birth records ( from 1990 to 2004. Data after 2004 was not included because state identifiers were stripped from the public use files starting in 2005. From these records mothers’ education and marital status by state was calculated for each year. Including these additional variables in regressions (6), (7), (13) and (14) did not change the key results or statistical significance, and are not reported.
Additional Indicators – CPS process measures
To expand on the results, additional outcome variables were tried on the left hand side of equation (2). NCANDS also provides data concerning the number of reports filed in each state.
The overall trend in child abuse reporting in the United States from 1990 to 2010 is shown graphically in Figure 3. In contrast to the results shown in Figure 2, the reporting rate shows an overall increase over the entire period time period 1990 to 2010, and a sharp increase since the nadir in 2001.
Additional regressions, found in Table 3 and illustrated in Figure 3, estimated the impact of unemployment on reports of potential child maltreatment and on rates of substantiation. The coefficients in (15) to (17) shows changes in current unemployment rates are not statistical distinguishable from zero, suggesting current rates have no impact on the number of reports per 1,000 children. However, the coefficients on the “1 Year Lag Unemp” line are uniformly positive, large and two are statistically significant. The coefficients in (15) and (17) suggest each one percentage point rise in last year’s unemployment rate is associated with CPS receiving approximately half (0.44 to 0.69) an additional report of potential child abuse for every 1,000 children. This is a noteworthy rise since the average state CPS gets less than 30 reports for each 1,000 children (mean 29; median 26.8).
Regressions (18) to (20) show the substantiation rate, which is the number of children victimized, divided by the number of initial reports. Approximately thirty percent of reports (mean 30.9%, median 29.6%) are substantiated. The coefficients on the “1 Year Lag Unemp” line show each one percentage point rise in last year’s unemployment rate is associated with up to a 2.4 percentage point increase in substantiation rates. Redoing the table (not shown for space reasons) using unweighted regressions shows qualitatively and quantitatively similar results.
DISCUSSION
Over the two decades from 1990 to 2010, we found statistical evidence that unemployment rates predict changes in the rate of reported child maltreatment one year later. The regression models reported here suggest that each one percentage point increase in unemployment rates is associated with an increase between 0.21 and 0.46 substantiated cases of child maltreatment per 1,000 after a lag of one year. Based on current population estimates, the model suggests that, at the national level, each 1 percentage point increase in unemployment rates is associated with 15,900 to 34,700 more victims of child maltreatment a year later. Moreover, after a lag of one year there are roughly 33,200 to 52,100 more reports of potential abuse.