Identifying perinatal risk factors for infant maltreatment: an ecological approach
 Yueqin Zhou^{1, 2}Email author,
 Elaine J Hallisey^{2} and
 Gordon R Freymann^{2}
DOI: 10.1186/1476072X553
© Zhou et al; licensee BioMed Central Ltd. 2006
Received: 02 October 2006
Accepted: 04 December 2006
Published: 04 December 2006
Abstract
Background
Child maltreatment and its consequences are a persistent problem throughout the world. Public health workers, human services officials, and others are interested in new and efficient ways to determine which geographic areas to target for intervention programs and resources. To improve assessment efforts, selected perinatal factors were examined, both individually and in various combinations, to determine if they are associated with increased risk of infant maltreatment. State of Georgia birth records and abuse and neglect data were analyzed using an areabased, ecological approach with the census tract as a surrogate for the community. Cartographic visualization suggested some correlation exists between risk factors and child maltreatment, so bivariate and multivariate regression were performed. The presence of spatial autocorrelation precluded the use of traditional ordinary least squares regression, therefore a spatial regression model coupled with maximum likelihood estimation was employed.
Results
Results indicate that all individual factors or their combinations are significantly associated with increased risk of infant maltreatment. The set of perinatal risk factors that best predicts infant maltreatment rates are: mother smoked during pregnancy, families with three or more siblings, maternal age less than 20 years, births to unmarried mothers, Medicaid beneficiaries, and inadequate prenatal care.
Conclusion
This model enables public health to take a proactive stance, to reasonably predict areas where poor outcomes are likely to occur, and to therefore more efficiently allocate resources. U.S. states that routinely collect the variables the National Center for Health Statistics (NCHS) defines for birth certificates can easily identify areas that are at high risk for infant maltreatment. The authors recommend that agencies charged with reducing child maltreatment target communities that demonstrate the perinatal risks identified in this study.
Background
Child maltreatment, including neglect, physical abuse, emotional abuse, sexual abuse, and other types of abuse, is a persistent problem in the world, not only in poor countries, but also in rich, industrialized nations, including the United States [1]. In general, the rate of child maltreatment is inversely related to the age of the child: children from 0 to 3 years of age have the highest rate [2]. Infants, children under age one, have the highest percentage of fatalities. Infants also suffer more serious physical and developmental consequences from maltreatment [3].
In addition to immediate effects, child maltreatment has pronounced longterm negative medical and social consequences. Numerous studies in medical literature confirm the association between childhood maltreatment and adverse adult health outcomes [4]. Examples include smoking [5], drug abuse [6], depression [7, 8], stress disorders [9], and certain chronic diseases [10]. For example, a study on the continuing consequences of maltreatment in the early years using longitudinal data from infancy through late adolescence confirmed the adverse impact of early maltreatment on later antisocial behavior [11].
In addition to medical and social consequences, the economic impact of child maltreatment is immense. Nationwide costs resulting from abuse and neglect are estimated to be as high as $94 billion per year, of which $24.3 billion are used for the immediate needs of abused or neglected children including hospitalization, treatment of chronic health problems, mental health care, child welfare, law enforcement, and the judicial system; the remaining $69.7 billion are costs associated with longterm and/or secondary effects of child abuse and neglect [12]. A study assessing the economic burden of hospitalization associated with child abuse and neglect found that children whose hospitalization was due to abuse or neglect were significantly more likely to have longer hospital stays, with double the total charges of other hospitalized children, and that nearly twothirds of the primary payer were Medicaid [13].
A recent study, focusing on infant maltreatment assessment, found that a set of perinatal risk factors significantly influences the probability that an infant would be maltreated by caregivers [14]. In this cohort study, the researchers studied 15 perinatal risk factors using data of 1,602 infant victims of maltreatment among 189,055 infants born in 1996 in Florida. They found 11 factors were significantly related to infant maltreatment, five of which had adjusted relative risks of two or greater. The five key factors are: mother smoked during pregnancy, more than two siblings, Medicaid beneficiary, unmarried marital status, and birth weight less than 2,500 grams. Infants who had four of these five risk factors had a maltreatment rate seven times higher than the population average. Other significant risk factors include maternal age less than 20, maternal education less than high school, and prenatal care, as measured by the Kotelchuck Index, less than adequate.
Another study linked premature births to substantiated infant abuse [15]. The researchers concluded that the cry of the premature infants was perceived to be more aversive than the cry of fullterm infants, and elicited greater arousal and subsequent abuse. Bugental and Happaney also identified low 5minute Apgar scores and premature births as predictor variables of infant maltreatment [16].
Given the negative consequences associated with infant maltreatment, in conjunction with the need to efficiently allocate resources, the goal of our study was to develop a populationbased model that enables public health agencies to identify areas at high risk for infant maltreatment. Many studies fail to consider an ecological, populationbased method that explores why these individual risk factors occur in the larger context of the environment in which they are found. The aforementioned risk factors are themselves outcomes and an ecological approach may help better demonstrate and understand the presence of underlying social and spatial processes that predispose certain caregivers to maltreat. The Health Field Concept [17] and the Health Field Theory [18] consider the interactions of biology, environment, lifestyle, and health system effects and capacities as the major determinants of health. Both take a holistic view of community assessment that supports the ecological approach.
To achieve this goal, we examined the geographic distribution of infant maltreatment in relation to relevant perinatal risk factors, as identified in previous research, using Georgia data aggregated by census tract. Two major components of the goal were: 1) to determine if infant maltreatment, including all types of abuse and neglect combined, is significantly related to a set of individual and composite perinatal risk factors, and 2) to identify a set of risk factors that best predicts infant maltreatment rates to ultimately aid public health agencies in identifying geographic areas for intervention.
Special attention was focused on coping with spatial autocorrelation, a wellknown spatial phenomenon in which data collected from a particular location is often similar to data collected in nearby locations. The presence of spatial autocorrelation violates the key assumption of independence for regression analyses. Researchers working with spatially aggregated data have noted that when spatial autocorrelation exists in the data but is ignored in analyses, such as in classic ordinary least squares (OLS) regression, the results are biased. To derive reliable results, spatial analytical techniques were used to control for the effects of spatial autocorrelation [19].
Methods
Data
Two sources of data were used in this study: substantiated neglect and abuse data from the Division of Family and Children Services (DFCS), and vital record births from the Division of Public Health (DPH), both of the Georgia Department of Human Resources (DHR). The State of Georgia complies with the U.S. Standard Certificate of Live Birth, developed by the National Center for Health Statistics (NCHS) [20], with regards to the variables acquired for birth certificates. This research was undertaken with the approval of the DHR Institutional Review Board (IRB). Given that the research involved the use of existing data sets, with confidentiality preserved, there were no risks to individual human subjects. The DHR Project Number is 060807.
Data on substantiated neglect and abuse were collected from 2000 through 2002 between January 1^{st} and December 31^{st} of each year. In 2000, 2001, and 2002, respectively, 2,642, 2,869, and 3,205 infants were victims of one or more types of maltreatment including neglect, physical abuse, emotional abuse, sexual abuse, and other types of abuse. DFCS declined to provide identifiers to link substantiated maltreatment cases to births, so we were unable to conduct the study at the individual level; instead, we conducted an ecological study. That is, we examined child maltreatment in relation to perinatal characteristics of the communities in which the maltreatment victims lived [21]. The census tract is a surrogate for the community. Tracts are subcounty areal units designed to be demographically homogeneous by the U.S. Bureau of the Census. The outcome variable is the rate of infant maltreatment by census tract. The hypothesis is that a census tract having a higher percentage of births with perinatal risks is more likely to have higher rate of infant maltreatment.
We followed the Standard Geocoding Procedures, designed by and applied within the Office of Health Information and Policy (OHIP), DPH, Georgia DHR, to assign latitude and longitude values to each of the records (F. Millard and G. Freymann, unpublished work, December 2001). The individual records were then aggregated to determine the number of maltreatment events (all types combined) in each census tract.
We extracted the birth records for 1999 through 2002, from which we derived risk factors and calculated the counts of births.
 1.
Medicaid beneficiary
 2.
Unmarried mother
 3.
Maternal age less than 20 years of age
 4.
Maternal education less than high school
 5.
Three or more siblings
 6.
Prenatal care less than adequate
 7.
Birth weight less than 2,500 grams
 8.
Mother smoked during pregnancy
 9.
Gestation less than 37 weeks
 10.
5minute Apgar scores less than 7
 11.
Child_risk: Any of risks 7, 9, or 10 (representing neonatal difficulties) is present.
 12.
Composite_risk1: Of risks 1, 2, 5, 7, or 8, three or more are present.
 13.
Composite_risk2: Of risks 1, 2, 5, 8, or 11, three or more are present.
The child risk composite (Child_risk) is a single variable that represents the presence of one or more neonatal difficulties. Composite_risk1 characterizes extreme risk infants who had three or more key risk factors present (identified in [14]). Composite_risk2 is equivalent to Composite_risk1 except the child risk composite replaces low birth weight.
To adjust for the effects of different lengths of exposure in each year, we applied the persontime concept to calculate the rate of infant maltreatment for each tract by dividing the total number of maltreatment victims by the total number of weighted births during a threeyear period [22]. Births were weighted by the length of time spent, as an infant, in each target calendar year of data collection. In doing so, we calculated for each record three weighting fields, each denoting the proportion of time, in infancy, over a oneyear period. The number of weighted births for a census tract was the sum of all three weights of all the births residing in that tract. Then the numbers of maltreatment victims and weighted births for individual years for each tract were summed to obtain the total number of maltreatment victims and total number of weighted births during a threeyear period.
All birth records for the years 1999 through 2002 were processed to create Boolean fields, 1 meaning present and 0 not present, for each risk factor. The Child_risk factor was coded 1 if any of the three previouslymentioned neonatal difficulties were present. Composite_risk1 or Composite_risk2 were coded 1 if three or more of the related 5 risk factors were present. For any record, if any of the factors were unknown, the record was omitted for the calculations of risk factors. The individual birth records were then aggregated to the tract, from which we obtained the percentage of births coded 1 in each tract for each of the risk factors. Figures 1b and 1c display the percentage of births coded 1 by tract for two composite risk factors (Composite_risk1 and Composite_risk2).
Analysis
In any spatial analysis, reviewing mapped data is recommended to determine if the distribution suggests any patterns or relationships among mapped features [23]. Qualitative, visual analysis of the mapped data, or cartographic visualization (Figures 1a, 1b, and 1c), suggests that spatial autocorrelation, meaning that similar data values tend to cluster geographically, is more pronounced with regards to the risk factors (Figures 1b and 1c) in contrast to substantiated maltreatment (Figure 1a). In Figure 1a, tracts in the lowest classification are more often immediately adjacent to tracts in the highest classification (note that all maps use the quantiles classification method). For instance, tracts in Montgomery and Wheeler, adjacent counties in the southeastern portion of the state, are at opposite ends of the classification scheme in terms of maltreatment, whereas the risk factor maps show both Montgomery and Wheeler in the mid to upper midrange of the classification scheme.
This visual analysis noting the presence of spatial autocorrelation and the correlation between infant maltreatment and risk factors is tenuous at best and must be validated with robust quantitative spatial analysis. The following paragraphs describe the quantitative spatial analysis of perinatal risk factors for substantiated infant maltreatment.
We used bivariate and multivariate linear regression methods for the quantitative analysis. The dependent variable was the maltreatment rate, and the predictor variables were the individual and composite risk factors. To ensure the normal distribution of the dependent variable, we transformed the rate to its natural logarithmic form using the following formula [24]:
where ln() is the natural logarithmic transformation function;
Y_{ i }is the number of infant maltreatments in tract i;
WN_{ i }is the weighted births in tract i;
and z_{ i }is the transformed rate.
This formula not only gives valid values for those tracts with Y_{ i }= 0, but also helps discriminate the tracts with Y_{ i }≤ 1 but with different WN_{ i }[24].
Let x_{1}, x_{2}, ... x_{ k }denote k risk factors chosen to be included in the regression equation; z the dependent variable; and e the error term. A multivariate regression equation is expressed as:
z = b_{0} + b_{1}x_{1} + b_{2}x_{2} + ... + b_{ k }x_{ k }+ e = Xβ + e (2)
in which X = (1 x_{1} x_{2} ... x_{ k }), β = [b_{0} b_{1} ... b_{ k }]^{T}, b_{0} is the constant, and b_{ j }(j = 1, ..., k) is the slope reflecting the relationship between x_{ j }and z. Equation (2) becomes the bivariate regression expression if only retaining a single risk variable.
In traditional statistics with nonspatial data, the OLS method is used to estimate parameters b_{0}, b_{1}, ..., b_{ k }(or b_{0}, b_{1} in bivariate regression). In order for the statistical inference about parameter estimates to be valid, some assumptions about the data must be satisfied [25]. The key assumption is that the error term is independent and normally distributed with a constant mean of zero and constant variance of σ^{2} (homogeneity). For multivariate regression, an additional key assumption is that there is no multicollinearity among the predictor variables.
Assuming there is no serious problem with multicollinearity, the OLS method provides the best linear unbiased estimates only if the regression model is correctly specified so that the error term meets the above assumption. A regression model is considered misspecified in several situations: 1) the dependent variable is inherently spatially dependent; 2) the unit of analysis does not match the unit of actual phenomena; 3) important risk variables are not included in the model; and 4) the observations of the dependent and/or predictor variables are not free of errors [24, 26, 27]. If a regression model is misspecified, the errors after the OLS fitting will not be independent; instead, the error at one location may be correlated with the errors at nearby locations, resulting in the clustering of similar errors among nearby locations, or spatial autocorrelation. When the errors are spatially autocorrelated, the OLS estimates are no longer unbiased and the goodnessoffit measure R^{2} is upward biased [28].
A common method to handle spatial autocorrelation is to minimize its effects by resampling to create a subset of data. This can be done in one of two ways; either manually selecting data locations or using a random process [29]. However, both the manual and the random process have some drawbacks. Manual selection may be subjective, random selection may not be free of spatial dependency, and both methods may result in loss of information, that is, the selected subset may not represent all the characteristics of the dataset.
A less commonly used but more objective method is spatial regression, which considers spatial autocorrelation an additional variable in the regression equation and solves its effect simultaneously with the effects of other explanatory variables [27]. This method uses all available information in the dataset.
We therefore controlled for spatial autocorrelation effects using the spatial regression method. There are two ways to incorporate spatial autocorrelation in a regression model. One is to model spatial autocorrelation in the error term as a spatially lagged dependent variable, and the other as a spatially lagged error term, that is,
e = ρWz + ε (3)
or
e = λWe + ε (4)
where W is the weight matrix characterizing the spatial relationship between every pair of observations;
ρ or λ is the spatial autoregressive parameter characterizing spatial autocorrelation;
And ε is the independent and normally distributed error term with a constant mean of zero and constant variance.
The former is referred to as a spatial lag model and the latter a spatial error model. Substituting the error term in (2) with Equation (3) or (4) and reorganizing the equation lead to the expression of a spatial lag model:
z = (I  ρW)^{1}Xβ + (I  ρW)^{1} ε (5)
or that of a spatial error model
z = Xβ + (I  λW)^{1} ε (6)
where I is the identity matrix.
The OLS method is no longer appropriate for estimating the parameters in Equations (5) and (6); instead, the maximum likelihood estimation method (MLE) or the instrumental variables estimation (IVE) method should be used [26]. The MLE method estimates model parameters by maximizing the Likelihood Function of the observations [26, 27].
When the MLE method is used to estimate the parameters in Equation (5) or (6), the traditional goodnessoffit measure, R^{2}, is no longer valid for assessing model fit [26]. One appropriate measure is the Akaike Information Criteria (AIC). A model is considered the best among a set of alternatives if the model gives the smallest AIC value. An approximate measure that mimics R^{2} is the socalled pseudoR^{2}, which provides a measure of linear association between the observed and predicted values of the dependent variable, but is no longer related to the variance component explained by the model.
The software used for the spatial regression analysis in the present study is GeoDA (Version 0.9.5i_6) [30, 31]. The program provides tools to calculate spatial weights, and run OLS (Classic) as well as spatial regression (Spatial Lag and Spatial Error) models. The output of the OLS models includes the diagnostic for multicollinearity (i.e., multicollinearity condition number or MCN), and the array of test statistics for spatial autocorrelation, which suggest whether spatial autocorrelation is significant to consider, and if so, which spatial model should be used. A value over 30 for MCN suggests problems with multicollinearity [32].
Results
Descriptive statistics of infant maltreatment rates and perinatal risk factors
Variables  Min  Max  Mean  Standard Deviation  Skewness  Number of outliers  

Lower  Upper  
Infant maltreatment rates  
Rate  0.0  186.9  24.4  23.0  1.58  0  62 
Transformed rate  0.4921  5.2844  3.1190  0.8225  0.37  7  0 
Perinatal risk factors: individual  
Gestation weeks <37  2.4  32.7  11.7  3.4  0.69  2  24 
Birth weight <2500 g  0.0  24.6  8.0  3.2  0.78  0  22 
5Minute APGAR Score <7  0.0  7.8  1.3  1.0  1.27  0  46 
Prenatal care less than adequate  0.4  57.1  23.4  11.4  0.40  0  0 
Maternal age <20  0.0  39.3  14.9  7.6  0.10  0  3 
Maternal education < HS  0.0  80.3  25.0  14.0  0.40  0  8 
Unmarried mother  1.3  97.1  40.3  21.7  0.50  0  0 
Three or more siblings  0.0  32.2  9.4  4.4  1.32  0  62 
Mother smoked during pregnancy  0.0  46.2  9.4  6.6  1.15  0  31 
Medicaid beneficiary  0.7  94.8  48.7  21.1  0.34  0  0 
Perinatal risk factors: composite  
Child_risk  2.4  34.7  15.0  4.3  0.55  1  15 
Composite_risk1  0.0  38.8  10.5  6.9  0.79  0  29 
Composite_risk2  0.0  43.3  12.8  8.1  0.70  0  23 
Spatial regression of infant maltreatment rates on perinatal risk factors: bivariate
Models  Variables  Coefficient Estimates  Pvalue  Spatial autocorrelation estimates  AIC  PseudoR^{2}  

Estimates  Standard Error (SE)  
1  Medicaid beneficiary  0.0213  0.0009  0.0000  0.7133*  2648.5  0.541 
2  Unmarried mother  0.0195  0.0009  0.0000  0.7840*  2683.6  0.537 
3  Maternal age <20  0.0440  0.0024  0.0000  0.5653*  2701.2  0.518 
4  Maternal education < HS  0.0246  0.0014  0.0000  0.7896*  2774.1  0.510 
5  Three or more siblings  0.0558  0.0041  0.0000  0.8243*  2887.4  0.477 
6  Prenatal care less than adequate  0.0294  0.0023  0.0000  0.8400*  2907.0  0.473 
7  Birth weight <2500 g  0.0522  0.0051  0.0000  0.7827*  2950.4  0.452 
8  Mother smoked during pregnancy  0.0327  0.0037  0.0000  0.8345*  2981.6  0.446 
9  Gestation weeks <37  0.0393  0.0046  0.0000  0.8070*  2986.3  0.442 
10  5Minute APGAR Score < 7  0.0728  0.0152  0.0000  0.8388*  3036.6  0.428 
11  Child_risk  0.0408  0.0038  0.0000  0.7733*  2941.9  0.454 
12  Composite_risk1  0.0618  0.0028  0.0000  0.7281*  2649.7  0.542 
13  Composite_risk2  0.0545  0.0024  0.0000  0.7219*  2626.3  0.548 
Spatial regression of infant maltreatment rates on perinatal risk factors: multivariate
Variables  Coefficient Estimates  Pvalue  Spatial autocorrelation estimates  AIC  PseudoR^{2}  

Estimates  Standard Error (SE)  
Mother smoked during pregnancy  0.0221  0.003492  0.0000  0.6894*  2573.3  0.564 
Three or more siblings  0.0166  0.004533  0.0002  
Unmarried mother  0.0077  0.002169  0.0004  
Maternal age <20  0.0132  0.004317  0.0023  
Medicaid beneficiary  0.0054  0.002306  0.0183  
Prenatal care less than adequate  0.0046  0.002321  0.0469  
Multicollinearity condition number (MCN) = 20.5 
Table 3 presents the multivariate spatial regression results. This is the best multivariate model among several alternatives (others not shown) with the smallest AIC value. Included in the model are six risk factors: mother smoked during pregnancy, three or more siblings, maternal age <20, unmarried marital status, Medicaid beneficiary, and prenatal care less than adequate. The model has a value of 20.5 for MCN, indicating no serious problem with multicollinearity [32]. The value of pseudoR^{2} suggests the model has moderate predictive ability. All risk variables are statistically significantly with respect to the rate of infant maltreatment at least at the 0.05 level. Also, spatial autocorrelation is a significant factor with Pvalue < 0.0000.
The multivariate model provides slightly better predictivity than any of the bivariate models since it gives the smallest AIC value. However, the decrease of AIC values is only 2% compared with the best bivariate model (Model 13).
Discussion
Several issues must be addressed to properly interpret the results: 1) ecological design, 2) spatial autocorrelation, and 3) the different types of child maltreatment.
This research used an ecological study design. Relationships found in ecological studies may suffer from two problems: the ecological fallacy and modifiable areal unit problems (MAUP) [24]. This means relationships found at the census tract level are pertinent to the current configuration of census tracts. They may not be inferred to finer levels such as block group or to individuals who lived in the areas.
We found spatial autocorrelation effects statistically significant in both bivariate and multivariate regression models. Several situations described in the Method section might be present in this study: inherent spatial dependency, missing variables, a mismatch between the unit of analysis and the unit of actual phenomena, and measurement errors. First, the dependent variable might be inherently spatially dependent. Since the observations of the response variable were not acquired through a strict sampling design but a collection of data arranged by the geographic unit, i.e., the census tract, the interdependence between observations of neighboring census tracts might be the rule rather than the exception [27]. Second, it is apparent that many variables were not included in any of the bivariate regression models. Even in the multivariate regression model, there were important variables missing. This is because child maltreatment is a socialpsychological phenomenon that involves a number of risk factors, including the characteristics of the victim, the maltreater, the family, the community, and the society, as suggested by the ecological theory of child maltreatment [33, 34].
The problem of mismatch between the unit of analysis and the unit of actual phenomena might also be present. This study used the census tract as a surrogate of the community. There is no compelling reason at this time to believe that child maltreatment conforms to the configuration of census tracts, except to note that tracts are designed to be demographically homogeneous. Future study using a more meaningful geographic unit may be beneficial.
The spatially correlated measurement errors in this study might result from the geocoding process. Approximately 85% of the maltreatment records were geocoded based on their address information to the accuracy of the census tract level. For the remaining 15% of the records, the county was known, but had inappropriate address information, such as P.O. boxes, or incorrect and/or incomplete addresses. For these records, a method of spatial imputation was employed in which census tracts within the county were assigned based on fertility probability, a method with an adjusted R^{2} accuracy of 0.82 compared with known county fertility rates. Latitudes and longitudes were assigned randomly within selected tracts (F. Millard and G. Freymann, unpublished work, December 2001). Therefore, a record was more likely to be located in a census tract that is spatially close to the correct tract.
A final issue to address is that of the different types of child maltreatment. In this study we aggregated all types of abuse and neglect, which include neglect, physical abuse, emotional abuse, sexual abuse, and other abuse, to reduce the small number problem. It has been suggested that different types of child maltreatment involve different risk factors [35]. Therefore, it may be of benefit in future studies to separate out the types of maltreatment to determine if significant risk factors change or improve the model.
Conclusion
Efficient allocation of resources is a priority. The method described in this paper can augment decision making regarding funding and intervention decisions.
With proper data architecture, addressing some of the shortcomings of the ecological methods used in this research can be easily overcome. For example, adjusting the spatial unit of analysis from the tract to some other form or level of precision, such as cells of varying sizes or block groups, becomes simple when data and quality standards are in place. What is less clear however, is which unit of analysis is most appropriate for the outcome of interest. In the event that "sociological meaningful scale" is unclear, analyzing the data at multiple scales may be beneficial.
Upon consideration, what at first seemed a limitation, namely the lack of unique identifiers to link individual abuse records with birth records, could actually have led to a more suitable method of analysis for the data. The method described should be used to help target interventions towards population subgroups, as this is the purpose of public health – to assure the conditions in which people can be healthy – and does not necessarily require a personbased medical model.
One should use the ecological model to open inquiry into other social and environmental conditions present in areas with the greatest maltreatment risk. Why do some communities demonstrate the characteristics of the multivariate regression model (Table 3), i.e. high percentages of young, unmarried mothers, mothers who smoke during pregnancy, inadequate prenatal care, large families, and Medicaid recipients? What actions should be taken to alleviate these conditions/outcomes? The model lends itself to a holistic approach toward community health assessment that is based on resiliency, rather than merely presence or absence of disease or poor outcome. As stated previously, the Health Field Concept and Health Field Theory support such an approach.
The presence of spatial autocorrelation has at least two important implications: 1) an assessment model must take spatial autocorrelation into account, and 2) poor health outcomes exhibiting spatial autocorrelation indicate the need for communitylevel (ecological) responses. Implicit in 2) is that casemanagement alone will not prevent child maltreatment when larger ecological issues are not identified nor addressed.
Perhaps most importantly, is that this and similar models allow public health to take a proactive stance, and reasonably predict areas where poor outcomes are more likely to occur. U.S. states routinely collecting variables defined by NCHS for birth certificates can easily identify areas that are at high risk for infant maltreatment. This implies that public health need not be burdened by relying only on the current practice of casemanagement, but can put measures in place to target areas before maltreatment occurs.
List of abbreviations
 AIC :

– Akaike Information Criteria
 DFCS :

– Department of Family and Children Services
 DHR :

– Department of Human Resources
 DPH :

– Division of Public Health
 IRB :

– Institutional Review Board
 MAUP :

– Modifiable Areal Unit Problem
 MCN :

– Multicollinearity Condition Number
 MLE :

– Maximum Likelihood Estimation
 NCHS :

– National Center for Health Statistics
 OHIP :

– Office of Health Information and Policy
 OLS :

– Ordinary Least Squares
Declarations
Acknowledgements
The authors thank the following persons and agencies within the Georgia Department of Human Resources (DHR) for their support: The Division of Family and Children Services, Stuart T. Brown, M.D., Director of the Division of Public Health, and B.J. Walker, DHR Commissioner.
Authors’ Affiliations
References
 UNICEF: A League Table of Child Maltreatment Deaths in Rich Nations. Innocenti Report Cart No5. 40[http://www.uniceficdc.org/publications/pdf/repcard5e.pdf]
 DHHS: Child Maltreatment 2003. 170[http://www.acf.hhs.gov/programs/cb/pubs/cm03/index.htm]
 Scannapieco M, ConnellCarrick K: Understanding Child Maltreatment: An Ecological and Developmental Perspective. 2005, New York , Oxford University Press, 301View ArticleGoogle Scholar
 Springer KW, Sheridan J, Kuo D, Carnes M: The longterm health outcomes of childhood abuse: an overview and a call to action. J Gen Intern Med. 2003, 18: 864870. 10.1046/j.15251497.2003.20918.x.PubMedPubMed CentralView ArticleGoogle Scholar
 Anda RF, Croft JB, Felitti VJ, Nordenberg D, Giles WH, Willimson DF, Giovino GA: Adverse childhood experiences and smoking during adolescence and adulthood. JAMA. 1999, 282: 16521658. 10.1001/jama.282.17.1652.PubMedView ArticleGoogle Scholar
 Dembo R, Williams L, Berry E, Getreu A, Washburn M, Wish ED, Schmeidler J: The relationship between physical and sexual abuse and illicit drug use: a replication among a new sample of youths entering a juvenile detention center. Int J Addict. 1988, 23 (11): 11011123.PubMedGoogle Scholar
 Bifulco A, Bernazzani O, Moran PM, Ball C: Lifetime stressors and recurrent depression: preliminary findings of the Adult Life Phase Interview (ALPHI). Soc Psychiatry Psychiatr Epidemiol. 2000, 35 (6): 264275. 10.1007/s001270050238.PubMedView ArticleGoogle Scholar
 Kessler RC, Magee WJ: Childhood family violence and adult recurrent depression. J Health Soc Behav. 1994, 35 (1): 1327. 10.2307/2137332.PubMedView ArticleGoogle Scholar
 Widom CS: Posttraumatic stress disorder in abused and neglected children grown up. Am J Psychiatry. 1999, 156: 12231229.PubMedGoogle Scholar
 Felitti VJ, Anda RF, Nordenberg D, Williamson DF, Spitz AM, Edwards V, Koss MP, Marks JS: Relationship of childhood abuse and household dysfunction to many of the leading causes of death in adults: the adverse childhood experiences (ACE) study. Am J Prev Med. 1998, 14 (4): 245258. 10.1016/S07493797(98)000178.PubMedView ArticleGoogle Scholar
 Egeland B, Yates T, Appleyard K, van Dulmen M: The longterm consequences of maltreatment in the early years: a developmental pathway model to antisocial behavior. Children's Services: Social Policy, Research, and Practice. 2002, 5 (4): 249260.View ArticleGoogle Scholar
 Fromm: Total Estimated Cost of Child Abuse and Neglect in the United States: Statistical Evidence. [http://member.preventchildabuse.org/site/DocServer/cost_analysis.pdf]
 Rovi S, Chen PH, Johnson MS: The economic burden of hospitalization associated with child abuse and neglect. Am J Public Health. 2004, 94 (4): 586590.PubMedPubMed CentralView ArticleGoogle Scholar
 Wu SS, Ma CX, Carter RL, Ariet M, Feaver EA, Resnick MB, Roth J: Risk factors for infant maltreatment: a populationbased study. Child Abuse Negl. 2004, 28: 12531264. 10.1016/j.chiabu.2004.07.005.PubMedView ArticleGoogle Scholar
 Frodi AM, Lamb ME, Leavitt LA, Donovan WL, Neff C, Sherry D: Fathers' and mothers' responses to the faces and cries of normal and premature infants. Dev Psychol. 1978, 14 (5): 490498. 10.1037/00121649.14.5.490.View ArticleGoogle Scholar
 Bugental DB, Happaney K: Predicting infant maltreatment in lowincome families: the interactive effects of maternal attributions and child status at birth. Dev Psychol. 2004, 40 (2): 234243. 10.1037/00121649.40.2.234.PubMedView ArticleGoogle Scholar
 Lalonde M: A New Perspective on the Health of Canadians: a Working Document. [http://www.hcsc.gc.ca/hcssss/alt_formats/hpbdgps/pdf/pubs/1974lalonde/lalonde_e.pdf]
 Dever GEA: Community Health Analysis: Global Awareness at the Local Level. 1991, Gaithersburg, MD , Aspen Publishers, 3842ndGoogle Scholar
 Anselin L, Griffith DA: Do spatial effects really matter in regression analysis?. Pap Reg Sci Assoc. 1988, 65: 1134.View ArticleGoogle Scholar
 NCHS: Birth Edit Specifications for the 2003 Revision of the U.S. Standard Certificate of Birth. [http://www.cdc.gov/nchs/data/dvs/FinalBirthSpecs3242005.pdf]
 Bland M: An Introduction to Medical Statistics. 2000, New York , Oxford University Press, 4053rdGoogle Scholar
 Simpson DG, Imrey PB, Geling O, Butkus S: Statistical estimation of child abuse rates from administrative databases. Child Youth Serv Rev. 2000, 22 (11/12): 951971. 10.1016/S01907409(00)001183.View ArticleGoogle Scholar
 Hallisey EJ: Cartographic visualization: an assessment and epistemological review. Prof Geogr. 2005, 57 (3): 350364. 10.1111/j.00330124.2005.00483.x.View ArticleGoogle Scholar
 Waller LA, Gotway CA: Applied Spatial Statistics for Public Health Data. 2004, Hoboken, NJ , Wiley, 494View ArticleGoogle Scholar
 Chatterjee S, Hadi AS, Price B: Regression Analysis by Example. 2000, New York , Wiley, 3593rdGoogle Scholar
 Anselin L: Spatial Econometrics: Methods and Models. 1988, Dordrecht , Kluwer Academic Publishers, 284View ArticleGoogle Scholar
 Anselin L, Bera AK: Spatial dependence in linear regression models with an introduction to spatial econometrics. Handbook of Applied Economic Statistics. Edited by: Ullah, Aman, Giles DEA. 1998, New York , Marcel Dekker Inc., 237289.Google Scholar
 Benirschka M, Binkley JK: Land price volatility in a geographically dispersed market. Am J Agric Econ. 1994, 96: 185195. 10.2307/1243620.View ArticleGoogle Scholar
 Mitchell A: The ESRI Guide to GIS Analysis: Volume 2: Spatial Measurements and Statistics. 2005, Redlands, CA , ESRI PressGoogle Scholar
 Anselin L: GeoDa 0.9.3 User's Guide. 126[http://www.sal.uiuc.edu/stuff/stuffsum/pdf/geoda093.pdf]
 Anselin L: GeoDa 0.9.5i Release Notes. 64[http://www.sal.uiuc.edu/stuff/stuffsum/pdf/geoda095i.pdf]
 Anselin L: Exploring Spatial Data with GeoDa: a Workbook. 244[http://www.geoda.uiuc.edu/pdf/geodaworkbook.pdf]
 Belsky J: Child maltreatment: an ecological integration. Am Psychol. 1980, 35 (4): 320335. 10.1037/0003066X.35.4.320.PubMedView ArticleGoogle Scholar
 Belsky J: Etiology of child maltreatment: a developmentalecological analysis. Psychol Bull. 1993, 114 (3): 413434. 10.1037/00332909.114.3.413.PubMedView ArticleGoogle Scholar
 Tzeng OCS, Jackson JW, Karlson HC: Theories of Child Abuse and Neglect: Differential Perspectives, Summaries, and Evaluations. 1991, New York , Praeger, 355Google Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.