Risk for Child Maltreatment Among Infants Discharged From a Neonatal Intensive Care Unit

Abstract

Studies suggest that neonatal illness may cause increased risk for child maltreatment (CM), but these findings may be biased by observed and unobserved confounding factors (social, family, and maternal characteristics) including increased surveillance by health care providers. This study expands on previous research by examining and controlling for these potential study biases and confounders using a sibling discordance retrospective cohort study design. Infants born in a Level IV neonatal intensive care unit (NICU) were matched with non-NICU born sibling controls. Cox proportional hazard models with shared frailty terms were used to account for clustering and heterogeneity in CM survival time (time to CM event). Potentially key covariates were selected using the directed acyclic graph approach, and surveillance reports were identified and systematically included or excluded from analyses. Managing these sources of bias reduced but did not eliminate the association between neonatal illness and CM report risk. Risk was especially high during the first year of the NICU infant’s life and among families with multiple well-known CM risk factors.

Keywords

child abuse child maltreatment infants survival analysis NICU

Infants with neonatal complications or infants treated/discharged from the NICU are at greater risk for later child maltreatment (CM; Brayden, Altemeier, Tucker, Dietrick, & Vietze, 1992; Famularo, Fenton, & Kinscherff, 1992; Hunter, Kilstrom, Kraybill, & Loda, 1978; Nandyal et al., 2013; Wu et al., 2004). Existing studies that have demonstrated a relationship between neonatal complications and CM have been based on observational research designs and analyses used have assumed that the underlying risk of CM is the same (homogenous risk) among infants belonging to different families (Famularo et al., 1992; Nandyal et al., 2013; Wu et al., 2004). In comparisons across families with differing characteristics, CM risk among infants and children tends to cluster within families, suggesting heterogeneous risk among infants/children belonging to different families (Damashek, Nelson, & Bonner, 2013; Hamilton-Giachritsis & Browne, 2005 Hansen, 1994; Nelson & Martin, 1985). This suggests that study analyses and, thus, inferences could be improved by accounting for both within-family clustering and heterogeneous risk among children belonging to different families. Furthermore, a major limitation of these observational designs is the potential for selection-type biases (systematic differences in factors that determine study participation), confounded cause-and-effect associations, and the effects of unobserved factors.

Selection-type bias could result when the selection of a comparison group (in this case, infants not treated in an NICU) is not representative of the population that produced the cases of interest (i.e., infants treated in an NICU). This could lead to study results (measures of association such as odds ratios or hazard ratios) that are different from what could have been observed if an entire target population or a representative (i.e., randomly chosen) comparison group was enrolled in a study. For example, nonrandom selection of a group of infants not treated in an NICU may result in a comparison group that is systematically different from the NICU infants, such as being of higher socioeconomic status (SES) or more likely to have received prenatal care. Differential surveillance and referral of infants and children to child protective services (CPS) could also produce similar effects as those of the nonrepresentative comparison groups (Greenland, Pearl, & Robins, 1999).

Confounding at the simplest level may be considered as a “distortion of the effect of a suspected risk factor on an outcome of interest due to the effect of extraneous factors mistaken for—or mixed with—the actual risk factor effect (which could be null)” (Rothman & Greenland, 1998). Unobserved factors or confounding (unobserved heterogeneity) is a known threat to cause–effect inference and is rarely ruled out with certainty in observational studies (Flanders et al., 2011; Rothman & Greenland, 1998). In examining risk for CM, unobserved heterogeneity refers to some children being conditionally different in terms of their risks in ways that are unaccounted for during study analyses. Often, as is the case in past research examining CM risk among infants (e.g., Hunter et al., 1978; Wu et al., 2004), researchers have assumed that children with the same values for all covariates are otherwise identical. If this is not true, and the unobserved heterogeneity is ignored or not accounted for in study analyses, then estimates of association between suspected risk factors and CM may be inconsistent or biased, and, therefore, caution should be exercised in drawing substantive inferences about their cause–effect associations. For example, findings of an association between low birth weight and higher risk of CM (e.g., Wu et al., 2004) are limited by the potential unobserved heterogeneity introduced by the unobserved factor of maternal substance abuse, which could create the observed association.

In lieu of randomized exposure or treatment assignments, three plausible explanations can be made for the relationship observed between neonatal complications and CM reported in previous studies. First, this may be a true causal relationship, and infants with neonatal complications (or NICU experience) may be at higher risk of CM than less frail infants or children. However, without adequate controls for both observed and unobserved factors that may be associated with this cause-and-effect relationship, two other plausible explanations cannot be ruled out.

A second possible explanation is that family context may have masked (or distorted) the true association between neonatal complications and CM in general infant population studies. Family context is a well-researched risk factor for various outcomes among children such as academic achievement (Kunz & Peterson, 1977; Zajonc & Sulloway, 2007), behavioral conduct problems (Bank, Patterson, & Reid, 1996; Farrington, 2005), and CM (Ohlander & Chew, 2008; Stith et al., 2009). Family context or structure refers to the composition, membership, and organization of the family, including the patterning of relationships among family members (Dawson, 1991; Ohlander & Chew, 2008). Existing research on risk factors for CM have generally focused on perpetrator or victim characteristics with little attention to sibling relations (birth order, age differences, and sex). Recently, however, some researchers have begun examining how and to what extent these factors are associated with CM (Berger, 2005; Hamilton-Giachritsis & Browne, 2005; Ohlander & Chew, 2008; Radhakrishna, Bou-Saada, Hunger, Catellier, & Kotch, 2001). Two themes emerge from these studies: (1) family structure is dynamic and context dependent and (2) there is a positive correlation between sibship size and CM. These studies show that siblings may affect adverse outcome risk between each other in a variety of ways (Zajonc & Sulloway, 2007). Where NICU infants are involved, it is also important to note that neonatal complications or the burden associated with caring for NICU-treated infants changes over time as an infant/child grows, and so does the risk of CM (Nandyal et al., 2013). Study designs and analyses need to account for these changing contexts in order to get valid measures of association when examining the relationship between neonatal problems and CM risk. Without the adequate control for both stable family context indicators (such as parental characteristics) and dynamic family context (mostly sibling characteristics), the association between neonatal complications and CM is likely biased. The direction or magnitude of this bias on the measures of association examining CM risk among medically frail versus healthy siblings is unknown.

Sibling comparison strategies offer a possible remedy to estimate and attenuate or eliminate bias accruing from indicators of family context or structure (Lahey & D’Onofrio, 2010; Rodgers, Cleveland, van den Oord, & Rowe, 2000). The power of sibling designs, especially sibling discordance studies, resides in the ability to vary one aspect of a child’s experiences (having neonatal complications) while keeping other aspects of the home environment the same. These studies allow for the comparison of outcomes among siblings who encounter different family contexts, effects of siblings upon one another, and the resulting changes in family context created by siblings. Sibling study designs provide a major advantage over studies of unrelated infants in terms of having a more realistic comparison group.

The third possible explanation for previously reported NICU-CM associations is that perceived high risk or poor prognosis of NICU-treated infants is often an indication for medical surveillance. Within an NICU sample, medical surveillance refers to the increased opportunity for identifying parents of NICU infants as CM perpetrators because of their increased contact with medical care service providers. This indication is a potential confounding factor (produces an imbalance in prognostic factors among infants being compared) because it is correlated with medical surveillance and is itself a risk factor for CM. Therefore, without accounting for the effects of surveillance, the association between neonatal complications and CM may be overestimated or biased away from the null by the artificial inflation of CM risk among medically compromised infants. The control for surveillance effects is not conceptually different from confounding by other factors; therefore, matching, stratification, restriction, and multivariate adjustment are acceptable remedial options (Rothman & Greenland, 1998).

The simplest approach to managing possible confounders is covariance or stratification using whatever observed, and theoretically important covariates are available. However, even after the adjustment for known risk factors, residual confounding may still occur because of unmeasured or unknown risk factor and measurement error. Moreover, within the context of sibling comparisons, shared family contexts may also infer a shared frailty (i.e., some siblings might be more prone to CM than others for unobserved reasons). To evaluate the contribution of such unobserved variations (heterogeneity) on observed CM hazards, frailty models are warranted (Wienke, 2003). The frailty model approach provides a means to examine the variations (unobserved heterogeneity) among siblings from different families and to estimate the distribution of future failure time (time to CM event) in family clusters. Unfortunately, the effect of such unobserved heterogeneity on CM risk in sibling comparisons or even general infant population comparisons has limited research.

While previous studies have found that neonatal complications increase risk for CM, alternate hypotheses to a true, causal relationship remain to be tested. This study expands on previous research by attempting to manage some of these possible biases. The study’s specific aims were to determine whether CM risk varies among siblings belonging to different families and examine the associations of family context, maternal risk factors, and medical surveillance on the association between NICU experiences (neonatal problems) and CM risk. Specifically, we posited that (1) there would be significant variability among siblings belonging to different families, (2) dynamic family context indicators such as increased sibship size and infancy would function to increase CM risk, (3) stable family context indicators such as young maternal age and low family SES (measured by Medicaid enrollment status) would increase CM risk, and (4) medical surveillance would accentuate the magnitude of the association between suspected risk factors and CM away from the null (exaggerate measures of association–hazard ratio estimates).

Method

Study Sample

Medical record archival data for 2,463 neonates discharged from the NICU (NICU infants), Oklahoma University Children’s Hospital, Oklahoma City, between January 1, 2005, and December 31, 2008, were examined. Demographic and medical data, including health problems, discharge equipment/supports, future specialty referrals, and diagnoses at discharge, for NICU infants were obtained from electronic medical records. All medical and demographic data were initially obtained for clinical purposes. Data for family context and maternal risk factors available from the clinical data set included Medicaid insurance status (as a proxy for lower SES), mother’s age (young maternal age coded as <21 years), and family size (data on the number of children at the time of birth of study child as supplemented with additional data described subsequently).

The sibling information used in study analyses was determined based on the number of live births reported by the NICU infant’s mother at the birth of the NICU infant, plus Oklahoma Department of Human Services (OKDHS) CM report data, and imputed data where no information was available in the NICU and OKDHS databases. For family dyads with at least one CM report, data from the OKDHS database regarding the number of children in the household at the time of the report were used to update sibling control information. For instance, if at birth, a mother reported two previous live births, and several years later at the time of a CM report, the report indicated four total siblings, data for this family dyad would be revised to reflect that there are five total children (two siblings older and another two siblings younger than the NICU infant). Four hundred thirty-eight (18%) mothers had missing data for the number of living children. The number of living children and the age of each child (if not provided in the NICU and Department of Human Services [DHS] data sets) was imputed using MI and MIANALYSE SAS 9.3^® procedures (Ake & Carpenter, 2002; Horton & Lipsitz, 2001; Little & Rubin, 1987; Rubin, 1996). This process produced a total of 2,732 sibling controls. Siblings born after and before the NICU child were on average 2 (SD = 1) and 5 (SD = 3) years old at the end of study follow-up (date of CM outcome data was procured). In total, there were 2,463 family dyads yielding clusters with 1 to 11 children and a mean and median of 2 children per family cluster.

Child Maltreatment Data

Data on CM for the NICU infants, siblings, and their parents/guardians were obtained from the OKDHS database for the period starting from July 1, 1999, through January 7, 2011, with reports of CM post birth of the NICU infant being the primary outcomes of interest. Reports of CM, regardless of substantiation were included, with the exception of reports that were immediately ruled out by OKDHS (i.e., judged to be clearly inappropriate or malicious). Reports rather than substantiations were selected in order to manage variations caused by child welfare differential response practices, and because past research has indicated little meaningful difference between substantiated and unsubstantiated reports (Kohl, Jonson-Reid, & Drake, 2009).

A computerized sequential strategy was used to match participant data in the NICU and OKDHS databases using SSN, and combinations of name (including similar names and spellings), gender, and date of birth. All sets of infant and maternal matches were subsequently examined manually, line by line, in order to rule out inaccurate matches. Reports were aggregated across dates, children, and incidents and within types of CM. Reports of siblings were obtained by searching for records connected to the mother using her identifying information. Maltreatment reports are specific to the child the report is alleged against. Thus, if a caregiver had a report alleging CM against multiple children, each child in the family would have a presence of a CM report in our data set.

Victimized siblings’ time to CM report was determined by using the NICU infant’s date of birth as the reference point (start point for follow-up). For siblings younger than the NICU infant, their birth dates were used as the reference point. The censoring time for the NICU infants and siblings who did not have CM reports was equal to the number of days between the birth of the NICU infant (or the birth date of a younger sibling if identified from the OKDHS data) and the procurement date of the OKDHS CM data set. Censoring time refers to situations in which there is an incomplete observation time to the outcome of interest (in this case, a CM report). Some infants might have had future reports after the date we procured the data; thus, those infants were not observed for the full time to a CM report, because our outcome data were limited to the date we procured it. Information on the reporting source was used to distinguish reports made by health care providers (surveillance reports) from a report by a nonhealth care professional (e.g., a report of general neglect made by a neighbor).

Statistical Analyses

Descriptive statistics were utilized to summarize demographic and clinical characteristics of the NICU infants and mothers. Shared frailty survival regression analysis (proportional hazards) models were used to examine the hypothesis about heterogeneity of CM risk among siblings belonging to different family dyads (clusters). A likelihood ratio–based test (Self & Liang, 1987) was used to test the null hypothesis that there is no CM risk variability among family dyads (i.e., the frailty term has a zero variance). The shared frailty models assume that similar observations (children from the same family) share frailty, even though frailty may vary from family to family. Reports of CM among children from the same family are likely correlated due to child welfare investigation practice (i.e., investigation of the safety of all children in a home once a report on any child is made), maternal, SES, and/or unobserved factors. The shared frailty term accounts for the unobserved heterogeneity and/or statistical dependence between the observed survival times (time to CM event) among siblings using a penalized partial likelihood approach. Frailty (random) effects are incorporated into the frailty models as independent identically distributed random variables and are assumed to follow a gamma distribution for computational convenience and convergence.

To examine the association between family context and CM, the effects of (1) dynamic family context indicators such as a child’s age and sibship size (number of children in the household), (2) stable family context indicators such as Medicaid enrollment (a surrogate indicator of SES), and maternal age, and (3) an indicator of prior CPS involvement were examined as covariates in the shared frailty models. To account for the changing experiences (e.g., reduced caregiving burden over time as children get older) of NICU infants in the short term versus long term relative to their siblings, shared frailty models were examined for a 1-year follow-up period and a longer follow-up period (full-study follow-up). For the 1-year follow-up period, a child remained in the study analyses until his or her exit date, which was the earliest of the following dates: first birth date or date of a confirmed CM report for the 1-year follow-up period. For the full-study follow-up period, a child remained in the study until his or her exit date, which was the earliest of the following dates: January 7, 2011, or date of a confirmed CM report.

Potential pairwise interactions suggested by explanatory Kaplan–Meier survival plots and previous analyses (Nandyal et al., 2013) were investigated using the shared frailty models. That is, the interaction terms between case status (NICU infant vs. sibling) by risk factors (Medicaid enrollment, prior CM reports [before birth of NICU infant] and number of living children) were examined. Where a significant pairwise interaction term was observed (p < .05), stratified analyses based on the identified moderating variable (factor) were examined. The potential confounding effects of investigated covariates were examined in the absence of significant pairwise interaction effects (i.e., an interaction term involving having an NICU experience and the covariate of interest).

A six-step directed acyclic graph (DAG) approach was used for covariate selection to reduce the potential for and degree of bias in effect measure estimation in the final chosen adjusted statistical models (Greenland et al., 1999; Shrier & Platt, 2008). With the increased complexity of hypothesized relationships among sibling characteristics, CM outcomes and potential confounding variables, consideration of association based on a priori expert knowledge is necessary to avoid adjusting for variables that may increase the risk of introducing bias where none existed (Shrier & Platt, 2008). The DAG approach allows for the assessment and incorporation of such knowledge and/or assumptions in the selection of a minimally sufficient adjustment set of covariates needed to estimate an unbiased effect of neonatal complications on CM. Informed by the DAG covariate selection process (Greenland et al., 1999) and existing literature (Sedlak, McPherson, & Das, 2010; Stith et al., 2009), race and marital status were not considered as potential confounding variables and, thus, not included in the final models. Selected covariates were included in our final analysis models to also attenuate or eliminate any bias that could potentially be introduced in the person time (time to CM event) available for analysis (Greenland & Robins, 1985; Weinberg, 1985). The relationship between medical surveillance and CM was examined by comparing the results of models that included and excluded surveillance-related reports. Multiple imputation procedures (using SAS procedures MI and MIANALYZE) were used to create plausible imputations of the missing data for (1) the number of living children (18%) including ages of resultant children based on imputations and (2) maternal age (5%) values to accurately reflect observed data relationships (assuming data are missing at random) and data distribution (Ake & Carpenter, 2002; Horton & Lipsitz, 2001; Little & Rubin, 1987). All analyses and statistical tests were set at .05 level for statistical significance.

Results

Family dyad characteristics of the study sample are shown in Table 1. The study sample included 2,463 NICU infants and 2,732 siblings. The descriptive statistics of the NICU infants have been published elsewhere (Nandyal et al., 2013). Briefly, 70% (1,711) of the NICU infants had a preterm birth, 50% (1,238) were classified has having high caregiving burden (i.e., had at least one of the following indications at discharge: oxygen or other respiratory supports, medications, feeding supports, or specialty referrals) and 25% (627) had very low birth weight (<1,500 g). The average length of follow-up was 38.5 months (3.2 years) among NICU infants and 41.8 months (3.5 years) among their siblings. Only 143 (5.2%) of the siblings were younger than their NICU infant counterpart. The average age of NICU infants at discharge was 21 days (SD = 22 days), whereas the average age of siblings at the time of NICU infant’s discharge was about 7 years (SD = 6) for older children. By January 7, 2011, a total of 889 (17%) of the study sample had been reported and investigated by CPS. More than half of the reports were made on NICU infants (523; 59%). Among mothers of the NICU infants, 12% (303) had prior CPS involvement.

Table 1.

Demographic and Clinical Characteristics of NICU Infant Mothers.^a

Maternal Characteristics	M (SD)	n (%)
Young maternal age (<21)^b		585 (24)
Race
Asian		26 (1)
American Indian		122 (5)
Black/African American		408 (17)
White/Caucasian		1,460 (59)
Hispanic		346 (14)
Multiracial		39 (2)
Other		62 (2)
Para (# viable births)	2 (2)
Gravida (# pregnancies)	3 (2)
No. of living children^b	2 (2)
Marital status
Married		922 (37)
Never married		851 (35)
Divorced, separated, or widowed		71 (3)
First time mother		692 (34)
Medicaid enrolled		1,030 (42)
Unemployed (at birth of NICU infant)		1,129 (46)

Note. NICU = neonatal intensive care unit. There was missing data for the following: maternal age (127), race/ethnicity (9), Gravida (20), Para (347), and marital status (619). ^a N = 2,463. ^bBased on imputed data.

Survival models results were based on a sample of 4,139 children (1,407 NICU infants and 2,732 sibling controls); NICU infants without siblings at the time of CM report were excluded from the analysis. Each NICU infant had between 1 and 11 controls (siblings); 675 (48%) had 1 sibling, 392 (28%) had 2 siblings, and 340 (24%) had 3 or more siblings. During the study follow-up period, 26% (360) of the NICU infants had CM events compared to only 13% (366) of the siblings. Table 2 summarizes the observed associations (using hazard ratios) between neonatal problems (NICU experience), investigated covariates, and CM for both the 1-year and full-study follow-up periods and with surveillance cases included and excluded.

Table 2.

Shared Frailty Cox models: NICU Infants (n = 1,407) Compared to Siblings (n = 2,732) for Risk of CM Summarized Using Hazard Ratios and Their 95% Confidence Intervals (CI) for Both the 1-Year and the Full-Study Follow-Up Periods and Including and Excluding Surveillance Cases.

	One-Year Follow-Up		Full-Study Follow-Up
	All Reports	Surveillance Excluded	All Reports	Surveillance Excluded
No. of living children at discharge
None^a	1.65 [0.87, 3.12]	1.66 [0.83, 3.31]	0.84 [0.59, 1.19]	0.78 [0.52, 1.16]
1 Child^b	4.94 [3.18, 7.69]	4.38 [2.61, 7.36]	2.47 [1.88, 3.24]	1.97 [1.42, 2.73]
2 Children^b	3.39 [2.46, 4.67]	3.07 [2.05, 4.60]	2.76 [2.14, 3.55]	2.42 [1.80, 3.25]
3 or More children^b	5.32 [3.65, 7.76]	4.38 [2.61, 7.36]	3.50 [2.74, 4.45]	2.91 [2.14, 3.95]
Medicaid
Yes^c	5.03 [3.79, 6.66]	4.62 [3.07, 6.95]	2.60 [2.07, 3.26]	1.92 [1.41, 2.62]
No^c	2.57 [1.84, 3.60]	2.10 [1.42, 3.10]	1.67 [1.33, 2.08]	1.52 [1.17, 1.96]

Note. ^aAdjusted for a child’s age, Medicaid enrollment and maternal age. ^bAdjusted for a child’s age, Medicaid enrollment, prior CPS involvement, and maternal age. ^cAdjusted for a child’s age, number of living children, prior CPS involvement, and maternal age.

To address the first study aim, the hypothesis that there exists heterogeneity (i.e., the frailty term has a nonzero variance) among siblings belonging to different family dyads (clusters) was tested. The variance of the frailty terms were significantly different from zero for both the short (p < .01) and long follow-up period shared frailty models (p < .01), suggesting that the CM risk was heterogeneous among family dyads. To address the second and third specific aims that (1) dynamic family context indicators such as increased sibship size and infancy may function to increase CM risk, (2) stable family context indicators such as young maternal age and low family SES (measured by Medicaid enrollment status) increase CM risk, the effects (interaction and confounding) of family context indicators on the association between neonatal complications and CM (NICU infant versus sibling) using a short (1 year follow-up) and long (full study) study follow-up periods were examined with and without surveillance-related reports. The association between neonatal problems and CM was modified by the number of children in a household and Medicaid enrollment, in both the short and long follow-up periods. Similar results were obtained with and without surveillance reports (Tables 2 and 3). Therefore, the results of the association between neonatal complications and CM are reported by strata defined by number of living children and Medicaid enrollment status. CM risk among NICU infants relative to their siblings increased with increasing number of children. In the 1-year follow-up period, the higher the number of living children reported, the higher is the risk of CM among NICU infants, with the exception of a lower hazard ratio observed among families with two living children versus one living child. In the full follow-up period, a similar trend was observed, with the exception of nonsignificant findings for NICU infants who were first born (no siblings at birth). Overall, there was increased risk for CM with more siblings in the home. NICU infants whose mothers were Medicaid enrolled were at higher risk of CM than their siblings, and the magnitude of this risk was slightly lower among NICU infants whose mothers were not Medicaid enrolled.

Table 3.

Shared Frailty Cox models: Fixed effects of Investigated Risk Factors on CM Risk Among Family Dyads With NICU Infants Summarized Using Hazard Ratios and Their 95% Confidence Intervals (CI) for Both the 1-Year and the Full-Study Follow-Up Periods and Including and Excluding Surveillance Cases.

	One-Year Follow-Up		Full-Study Follow-Up
	All Reports	Surveillance Excluded	All Reports	Surveillance Excluded
Prior CPS involvement^a	3.64 [2.78, 4.79]	4.62 [3.17, 6.71]	3.84 [3.13, 4.69]	4.36 [3.42, 5.56]
Maternal age^b (<21)	1.53 [1.17, 2.01]	1.59 [1.11, 2.28]	1.43 [1.17, 1.76]	1.49 [1.18, 1.90]
Child’s age (years)^c	0.58 [0.55, 0.60]	0.60 [0.57, 0.64]	0.65 [0.63-0.67]	0.68 [0.67-0.71]

Note. NICU = neonatal intensive care unit. ^aAdjusted for a child’s age, sibling status (NICU infant vs. Sibling), number of living children, Medicaid enrollment, and maternal age. ^bAdjusted for a child’s age, sibling status (NICU infant vs. Sibling), number of living children, Medicaid enrollment, and prior child protective services (CPS) involvement. ^cAdjusted for sibling status (NICU infant vs. sibling), number of living children, Medicaid enrollment, maternal age, and prior CPS involvement.

Infancy (younger child age), prior CPS involvement, and a young maternal age were significant predictors of future CM; however, these factors did not modify the association between neonatal complications and CM (Table 3). On the other hand, failure to adjust for a child’s age, prior CPS involvement, and maternal age appeared to bias the association between neonatal problems and CM risk. For both a child’s age and maternal age, the bias was away from the null (crude hazard ratio overestimated the adjusted (unbiased) hazard ratio by 53% and 1%, respectively). For prior CPS involvement, the bias was toward the null (crude hazard ratio underestimated the adjusted hazard ratio by 12%). The magnitude of the bias on the hazard ratios summarizing the association between neonatal complications (NICU infant vs. siblings) and CM reports was similar for the 1-year and full-study follow-up periods and did not appear to reverse the finding that CM risk was higher among NICU infants versus their siblings.

To address the fourth specific aim that medical surveillance accentuates the magnitude of the association between suspected risk factors and CM away from the null (exaggerates measures of association–hazard ratio estimates). We examined the association between neonatal complications and CM (NICU infant vs. sibling) with and without the inclusion of surveillance-related CM reports. Overall, the inclusion of surveillance-related reports seemed to bias the association between CM and neonatal complications away from the null (overestimated hazard ratios). For example, the risk of CM among NICU infants relative to their siblings was overestimated by a factor of 8% among Medicaid-enrolled mothers for the 1-year follow-up period. The bias (factor increase in hazard ratio due to the inclusion of surveillance reports) was much larger (26%) with a longer follow-up. The risk of CM among older children, family dyads with prior CPS involvement, or young mothers (<21 years) was underestimated with the inclusion of surveillance reports for both the 1-year follow-up period and the full-study follow-up period. Overall, the risk of CM was much higher among the NICU infants in the first year of follow-up than with extended study follow-up period (Tables 2 and 3).

Discussion

The central aim of the study was to examine the previously established association between neonatal illness and CM risk and manage potential sources of bias and confounding factors that might create an artifactual association. Given the scant research on the relationships between family context indicators and CM risk in neonatal infant populations, the use of a sibling discordant study design to examine the association between neonatal problems and CM risk provides a major advantage over population studies of unrelated children, by controlling for a host of known, unknown, and unmeasured family factors that could potentially bias study inferences. Consistent with the first study hypothesis, the results suggest that the CM risk does vary among siblings from different families (demonstrated by frailty terms with p < .01). The significance of the frailty terms even after controlling for family contextual factors, maternal characteristics, and surveillance effects suggests that a significant amount of variation (heterogeneity) in CM risk among siblings from different family dyads remains unexplained. Additional research is needed to better understand CM risk heterogeneity among siblings and its implications in targeting CM prevention interventions.

Study findings also reinforce results found in prior studies regarding the effects of dynamic and stable family contexts on CM risk (Berger, 2005; Hamilton-Giachritsis & Browne, 2005; Ohlander & Chew, 2008; Radhakrishna et al., 2001). Our study results show that neonatal problems (NICU experience) are associated with increased risk of CM; however, this association varies (is modified) based on the number of children in a household (an indicator of dynamic family context) and SES (a stable indicator of family context). Overall, CM risk among NICU infants relative to their siblings was higher among families with more children and low SES. This finding may be indicative of higher stress among families with an NICU infant and with more children and less resources. This would be consistent with the review findings that parents of children with disabilities experience a greater level of stress across domains (financial and emotional) and that higher parental stress increases the risk of CM (Algood, Hong, Gourdine, & Williams, 2011). Thus, it may be that among families with higher levels of stress, the NICU infant is particularly vulnerable to experiencing CM. The finding that the NICU infants were at higher risk of CM (relative to their siblings) among families with low SES than among family dyads with higher SES can inform prevention efforts. This finding suggests that targeting prevention interventions among Medicaid-eligible or Medicaid-enrolled mothers in an NICU setting may offer a unique opportunity to address CM risk among low SES families who, otherwise, may not receive any support unless they become involved with CPS.

Child infancy (younger age), medical surveillance, prior CPS involvement, and young maternal age were associated with increased CM incidence during both the 1-year and the full-study period time frames. This is consistent with the existing studies (Damashek et al., 2013; Famularo et al., 1992; Hamilton-Giachritsis & Browne, 2005; Hunter et al., 1978; Wu et al., 2004). These factors also seemed to confound the association between neonatal problems and CM risk and, therefore, affirm existing practice regarding the need to control for these factors in studies that examine CM risk factor associations. A child’s age particularly had a large impact on the association between neonatal problems and CM. Failure to account for children’s ages would have underestimated CM risk among NICU infants compared to their siblings by over 50%.

Comparisons between the short and long study follow-up periods showed that CM risk was higher among NICU infants relative to their siblings in the first year of life than with longer follow-up. The reduced CM risk with longer follow-up might be indicative of an underlying relationship between CM and child’s age and possible birth order, all indicators of a dynamic family context (Zajonc & Sulloway, 2007). Caregiving burden and hence the risk for CM may shift from the NICU child to the younger children or infants born during the extended follow-up period. Corroborating this finding are national statistics that indicate that CM risk reduces with increasing age (U.S. Department of Health and Human Services, 2012). In scenarios of family dyads, where there are no children younger than the NICU infant, reduced caregiving burden as the NICU infant matures is also another possible explanation of this finding (Nandyal et al., 2013). Although such inferences are plausible, the observed results could also be a function of the study design. Given the focus on stable and dynamic family context, it was elected to assess CM risk for NICU infants and siblings over the same discrete period in time (ensuring exposure to similar family context) rather than assessing CM risk during defined age ranges for both NICU infant and sibling (which would require examination of the family context during two differing time frames). While this choice was best for the current study, this design introduced an age confound. Future research is needed to determine how and to what extent age differences affected current findings. Nonetheless, the current study findings suggest that CM prevention efforts among at-risk families may be most efficacious during the first year of an NICU infant’s life.

Exclusion of surveillance-related reports seemed to attenuate CM risk among NICU infants relative to their siblings in strata defined by SES and number of children in the home, for both the 1-year and the full follow-up period. This suggests that surveillance may be a confounding factor beyond the time period immediately following an NICU infant’s hospital discharge. Thus, a concern for maltreatment research is the CM risk attributed to the NICU experience may be biased when surveillance cases are included in study analyses comparing NICU infants to a nonmedically involved comparison group. This finding supports the examination and control for surveillance effects in future studies to better identify and understand CM risk factors associated with neonatal or medical problems among children. For clinical purposes, this finding suggests that increased contact with medical providers could serve as a feasible method of screening or monitoring the risk of CM and other adverse outcomes among infants.

Limitations

Analyses were based on secondary data sets obtained retrospectively from the NICU medical record databases; therefore, observed covariates were limited to available information on infants, their siblings, and mothers. Potentially relevant characteristics that were not available included are as follows: an assessment of the mother’s parenting skills, mental health status, prenatal substance abuse, and others. There is also a concern that by limiting the CM outcome data to reports that were not screened out by DHS workers, the CM events among at-risk families may have been underestimated if some initially screened reports were later verified as CM events but were not corrected as such or captured in the CPS databases. The influence such exclusions had on observed study associations is likely minimal and nondifferential since the exclusion of these cases from study analyses occurred independent of whether an NICU infant or their siblings (controls) was involved. Another limitation is the potential underestimation of the number of an NICU infant’s siblings especially if younger siblings did not have a CM report and were not identified in the OKDHS data sets. However, it can be argued that the use of NICU data regarding the number of previous births, validation of this information with the OKDHS database, and the use of data imputations (to create plausible number of children values where it was missing in both NICU and OKDHS databases) assured a reasonably accurate estimate of children (siblings) was used in the analysis. The potential magnitude of bias due to missing young sibling information is likely minimal during the full-study follow-up period and even smaller for the 1-year follow-up analyses.

Conclusions

In summary, this study supports the hypothesis that the association between neonatal illness or NICU admission and subsequent risk for a CM report may be causal and not an artifact of confounded social, family, and maternal factors, or increased surveillance by health care providers. Although these factors did appear to bias this association to some degree, when managed (or controlled for in study analyses) they did not eliminate this association. The risk of CM among NICU infants relative to their siblings (as summarized by hazard ratios) remained high and was particularly high during the first year of the NICU child’s life and among families with well-known maltreatment risk factors. The findings support providing preventative services during the first year after children in high-risk families are discharged from the NICU.

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.

References

Ake

Carpenter

(2002). Survival analysis with PHREG: Using MI and MIANALYSE to accommodate missing data. Proceedings of the Western Users of SAS Software Tenth Annual Conference, San Diego, CA, 102–107. Retrieved http://www.sas.com/offices/NA/canada/downloads/presentations/Van10/PHREG.pdf

Algood

C. L.

Hong

J. S.

Gourdine

R. M.

Williams

A. B.

(2011). Maltreatment of children with developmental disabilities: An ecological systems analysis. Children and Youth Services Review, 33, 1142–1148.

Bank

Patterson

G. R.

Reid

J. B.

(1996). Negative sibling interaction patterns as predictors of later adjustment problems in adolescent and younger adult males. In Brody

(Ed.), Sibling relationships: Their causes and consequences (pp. 197–229). Norwood, NJ: Ablex Publishing Corporation.

Berger

L. M.

(2005). Income, family characteristics, and physical violence toward Children, Child Abuse & Neglect, 29, 107–133. doi:10.1016/j.chiabu.2004.02.006

Brayden

R. M.

Altemeier

W. A.

Tucker

D. D.

Dietrick

M. S.

Vietze

(1992). Antecedents of child neglect in the first two years of life. Journal of Pediatrics, 120, 426–429.

Damashek

Nelson

M. M.

Bonner

B. L.

(2013). Fatal child maltreatment: Characteristics of deaths from physical abuse versus neglect. Child Abuse & Neglect, 37, 735–744. doi:10.1016/j.chiabu.2013.04.014

Dawson

D. A.

(1991). Family structure and children’s health and well-being: Data from the 1988 National Health Interview Survey on Child Health. Journal of Marriage and the Family, 53, 573–584. doi:10.2307/352734

Famularo

Fenton

Kinscherff

(1992). Medical and developmental histories of maltreated children, Clinical Pediatrics, 31, 536–541. doi:10.1177/000992289203100904

Farrington

D. P.

(2005). Childhood origins of antisocial behavior, Clinical Psychology & Psychotherapy, 12, 177–190. doi:10.1002/cpp.448

10.

Flanders

W. D.

Klein

Darrow

L. A.

Strickland

M. J.

Sarnat

S. E.

Sarnat

J. A.,

Tolbert

P. E.

(2011). A method for detection of residual confounding in time-series and other observational studies, Epidemiology, 22, 59–67. doi:10.1097/EDE.0b013e3181fdcabe

11.

Greenland

Pearl

Robins

J. M.

(1999). Causal diagrams for epidemiologic research, Epidemiology, 10, 37–48. Retrieved from https://www-jstor-org.web.bisu.edu.cn/stable/253064

12.

Greenland

Robins

J. M.

(1985). Estimation of a common effect parameter from sparse follow-up data, Biometrics, 41, 55–68. Retrieved from https://www-jstor-org.web.bisu.edu.cn/stable/2530643

13.

Hamilton-Giachritsis

Browne

(2005). A retrospective study of risk to siblings in abusing families. Journal of Family Psychology, 19, 619–624. doi:10.1037/0893-3200.19.4.619

14.

Hansen

K. K.

(1994). Twins and child abuse. Archives of Pediatric & Adolescent Medicine, 148, 1345–1346. doi:10.1001/archpedi.1994.02170120107021

15.

Horton

N. J.

Lipsitz

S. R.

(2001). Multiple imputation in practice: Comparison of software packages for regression models with missing variables. The American Statistician, 55, 244–254. Retrieved from https://www-jstor-org.web.bisu.edu.cn/stable/2685809

16.

Hunter

R. S.

Kilstrom

Kraybill

E. N.

Loda

(1978). Antecedents of child abuse and neglect in premature infants: A prospective study in a newborn intensive care unit. Pediatrics, 61, 629–635. Retrieved from http://pediatrics.aappublications.org.webproxy.ouhsc.edu/content/61/4/629.full.pdf+html

17.

Kohl

Jonson-Reid

Drake

(2009). Time to leave substantiation behind. Child Maltreatment, 14, 17–26. doi:10.1177/1077559508326030

18.

Kunz

P. R.

Peterson

E. T.

(1977). Family size, birth order, and academic achievement, Social Biology, 24, 144–148. doi:10.1080/19485565.1977.9988274

19.

Lahey

B. B.

D’ Onofrio

B. M.

(2010). All in the family: Comparing siblings to test causal hypotheses regarding environmental influences on behavior. Current Directions in Psychological Science, 19, 319–323. doi:10.1177/0963721410383977

20.

Little

R. J. A.

Rubin

D. B.

(1987). Statistical analysis with missing data. New York, NY: John Wiley & Sons, Inc.

21.

Nandyal

Owora

Risch

E. C.

Bard

Bonner

Chaffin

(2013). Special care needs and risk for child maltreatment reports among babies that graduated from the Neonatal Intensive Care. Child Abuse and Neglect. Retrieved from https://dx-doi-org.web.bisu.edu.cn/10.1016/j.chiabu.2013.04.003

22.

Nelson

H. B.

Martin

C. A.

(1985). Increased child abuse in twins. Child Abuse & Neglect, 9, 501–505. doi:10.1016/0145-2134(85)90059-6

23.

Ohlander

Chew

(2008). Family structure and differential child abuse: The role of siblings. International Journal of Social Inquiry, 1, 121–150. Retrieved from http://www.socialinquiry.org/articles/IJSI-V1N12008%20-%20007.pdf

24.

Radhakrishna

Bou-Saada

I. E.

Hunger

W. M.

Catellier

D. J.

Kotch

J. B.

(2001). Are father surrogates a risk factor for child maltreatment? Focus section I: Fathers and child maltreatment: Findings from the Longitudinal Studies of Child Abuse and Neglect (LONGSCAN). Child Maltreatment, 6, 281–289. doi:10.1177/1077559501006004001

25.

Rodgers

J. L.

Cleveland

van den Oord

Rowe

D. C.

(2000). Resolving the debate over birth order, family size, and intelligence. American Psychologist, 55, 599–612. Retrieved from http://web.ebscohost.com.webproxy.ouhsc.edu/ehost/pdfviewer/pdfviewer?vid=3&sid=9b695eb6-591e-4d37-8ab5-e8282937b41a%40sessionmgr112&hid=118

26.

Rothman

K. J.

Greenland

(1998). Modern epidemiology (2nd ed., pp. 147–161, 283–285). Philadelphia, PA: Lippincott-Raven Publishers.

27.

Rubin

D. B.

(1996). Multiple imputation after 18+ years. Journal of American Statistical Association, 91, 473–489. Retrieved from https://www-jstor-org.web.bisu.edu.cn/stable/2291635

28.

Sedlak

A. J.

McPherson

Das

(2010). Fourth National Incidence Study of Child Abuse and Neglect (NIS–4): Supplementary analyses of race differences in child maltreatment rates in the NIS-4. Washington, DC: U.S. Department of Health and Human Services, Administration for Children and Families. Retrieved from http://calswec.berkeley.edu/files/uploads/doc/CalSWEC/FE_Needell_NIS4_Analyses_052510.doc

29.

Self

S. G.

Liang

K. Y.

(1987). Asymptotic properties of maximum likelihood estimators and likelihood ratio tests under nonstandard conditions. Journal of the American Statistical Association, 82, 605–610. Retrieved from https://www-jstor-org.web.bisu.edu.cn/stable/2289471

30.

Shrier

Platt

R. W.

(2008). Reducing bias through directed acyclic graphs. BMC Medical Research Methodology, 8, 70. doi:10.1186/1471-2288-8-70

31.

Stith

S. M.

Ting Liu

Davies

Boykin

E. L.

Alder

M. C.

Harris

J. M.,

Dees

J. E.

(2009). Risk factors in child maltreatment: A meta-analytic review of the literature. Aggression and Violent Behavior, 14, 13–29. Retrieved from http://hdl.handle.net/2097/14878

32.

U.S. Department of Health and Human Services, Administration on Children, Youth and Families. (2012). Child maltreatment 2011. Retrieved from http://www.acf.hhs.gov/sites/default/files/cb/cm11.pdf

33.

Weinberg

C. R.

(1985). On pooling across strata when frequency matching has been followed in a cohort study. Biometrics, 41, 117–127. Retrieved from https://www-jstor-org.web.bisu.edu.cn/stable/2530648

34.

Wienke

(2003). Frailty models. Max planck institute for demographic research. Retrieved from: http://www.ressources-actuarielles.net/EXT/ISFA/1226.nsf/0/37d8815030298678c125737b002e1b13/$FILE/wp-2003-032.pdf

35.

S. S.

Carter

R. L.

Ariet

Feaver

E. A.

Resnick

M. B.

Roth

(2004). Risk factors for infant maltreatment: A population-based study. Child Abuse & Neglect, 28, 1253–1264. Retrieved from http://mcherdc.sites.medinfo.ufl.edu/files/2010/10/Wu_Risk-factors-for-infant-maltreatment.pdf

36.

Zajonc

R. B.

Sulloway

F. J.

(2007). The confluence model: Birth order as a within-family or between-family dynamic? Personality and Social Psychology Bulletin, 33, 1187–94. doi:10.1177/0146167207303017