Sensitive Periods for the Effect of Child Maltreatment on Psychopathology Symptoms in Adolescence

Childhood maltreatment is one of the most potent though preventable risk factors for psychopathology throughout the lifespan [1]. One of the dominant approaches to operationalizing the relationship between maltreatment and psychopathology is the cumulative risk model, which states that the number of exposures to any risk factor will determine its impact above and beyond the intensity or type of risk factor [2]. However, there are mixed findings in support of this additive assumption of risk because current research has not fully explored if the cumulative effect of multiple instances of maltreatment, for example, could be explained by a single, more intense experience out of many or the context of when and how the maltreatment occurred in the child’s life [2, 3]. Emerging research, particularly from animal studies, suggests maltreatment may not have consistent effects throughout childhood, but rather there may be sensitive periods when the developing brain is particularly vulnerable to adversity [4]. The sensitive period model presumes the developmental timing of exposure is most important for determining the effect the exposure will have on later outcomes such as psychopathology symptoms, positing that the occurrence of exposure coincides with peak periods of brain plasticity, making exposure to maltreatment in one time period more likely to produce a greater effect than exposure to that same exposure occurring earlier or later in development [5‒8]. Yet, in humans, such sensitive periods have been largely unidentified due to a scarcity of research and mixed results (for a comprehensive review of this literature, refer to Schaefer et al. [9]). Some prospective studies have found early maltreatment (before age of 5 years) is more strongly associated with psychopathology risk [10‒12]. Prospective studies have found later maltreatment (after age of 10 years [13] or during adolescence [14]) is most harmful. Some find no developmental timing differences [15‒17]. Well-powered prospective research in large and diverse samples is needed to determine if and when sensitive periods occur. Such research can increase understanding of the processes linking maltreatment to mental health problems and suggest optimal time points for screening and prevention efforts to reduce the negative consequences of adversity exposure [4].

The current study addresses this need by analyzing data on child maltreatment from a population-based cohort of high-risk children followed from birth through adolescence. Children from “fragile families,” defined as unmarried parents, were oversampled. This dataset contained repeated measures of children’s exposure to neglect and harsh physical discipline, the 2 most common types of child maltreatment [18, 19], and measures of child’s internalizing and externalizing symptoms. Although the effects of more extreme forms of physical abuse on psychopathology risk are well established, numerous studies suggest physical discipline practices, such as spanking, slapping, and hitting [20], are associated with various negative developmental outcomes, including increased child aggression [21], slower cognitive development [22], and poor mental health in childhood [23] and adolescence [20]. Indeed, informed by this considerable body of research, the American Academy of Pediatrics released a policy statement urging parents to desist from physical and harsh verbal discipline [24, 25].

We brought three key innovations to these analyses. First, we examined psychopathology outcomes during adolescence, which allowed us to evaluate the longer term impacts of maltreatment on psychopathology in youth. If there was a latency or time lag between the onset of maltreatment to presentation of behavioral symptoms, studies like ours, which follow children for longer, are likely to detect sensitive periods, as compared to studies of shorter duration. Second, among prospective studies examining the role of child maltreatment timing on psychopathology risk [10‒17], the size of this study (n = 3,474) was considerably larger (average for prior studies was n = 826). Thus, we had greater statistical power to detect potential sensitive period effects, if they existed. Third, we evaluated the strength of evidence to support the existence of sensitive periods compared to an accumulation of risk model [2], in which the effect of maltreatment on psychopathology symptoms is presumed to increase with the number of occasions exposed, regardless of timing. Few prior studies [26] have compared the sensitive period model to alternative life-course models to determine which model better explains risk for psychopathology. This dearth of comparison is a major limitation as it could explain mixed results among prior studies.

Data came from the Future of Families and Child Wellbeing Study (FFCWS), formerly known as the Fragile Families and Child Wellbeing Study, a birth cohort study following a random sample of nearly 5,000 families in 20 large cities (populations over 200,000) [27] in the USA. FFCWS oversampled families with unmarried parents in an attempt to capture a nationally representative sample of fragile families who may be vulnerable to risk factors associated with nonmarital childbearing, such as poverty. Between 1998 and 2000, mothers and fathers in 75 hospitals were interviewed after their child’s birth, referencing approximately 4,898 births to 3,711 unmarried and 1,187 married parents; nonmarital births and families who were socioeconomically disadvantaged were oversampled. A description of sociodemographic characteristics of the sample and key study variables is presented in Table S1 (for all online suppl. material, see www.karger.com/doi/10.1159/000530120), available online. Families were interviewed again when the child was 1, 3, 5, 9, and 15 years of age. Follow-up interview completion rates were excellent (89% of mothers at age of 1 year, 86% at age of 3 years, 85% at age of 5 years, 76% at age of 9 years, and 73% at age of 15 years).

Informed consent was obtained for each family at each interview. The Institutional Review Boards at Columbia University and Princeton University approved the FFCWS. Additional details on participation, including attrition at each wave of assessment, are publicly available online [28].

Assessments of neglect and harsh physical discipline were derived from the Parent-Child Conflict Tactics Scales (CTS-PC), a frequently used measure to assess child maltreatment in population-based samples [19, 29, 30]. The CTS-PC was reported by primary caregivers (most often mothers) at child ages of 3 years, 5 years, and 9 years and collected mostly via computer-based in-home assessments with an interviewer and sometimes by a telephone.

The FFCWS CTS-PC measure included a subset of items from the CTS-PC and its supplemental scale on neglect, designed to capture mild and moderate maltreatment (5 neglect items and 5 harsh physical discipline items; see Supplement 1, available online). Items asking about severe physical maltreatment (e.g., “burned or scalded him/her on purpose,” “grabbed him/her around the neck and choked him/her”) were omitted by FFCWS to avoid potentially implicating parents and necessitate involvement of child protective services. Items capturing more mild forms of corporal punishment [31] (e.g., “spanked him/her on the bottom with a hand”) were not included in our measure of harsh physical discipline. Consistent with prior studies using the CTS-PC in FFCWS [32], harsh physical discipline was coded as a dichotomous variable indicating presence versus absence of the following: (1) the parent had shaken the child at any point in the past year and/or (2) the parent had hit the child with an object on 3 or more occasions in the past year. See Supplemental Materials for details on the specific items included.

Children were coded as having been exposed to neglect if their primary caregiver reported at least 1 of these events in the past year at any frequency: (1) parent had to leave child home alone even when an adult was needed; (2) parent was so caught up in own problems that they were not able to express love to child; (3) parent was not able to make sure child got food when needed; (4) parent was not able to make sure child got to the doctor or hospital when needed; and (5) parent was so drunk or high that they had problem taking care of child. These items were also chosen to be consistent with prior studies in the FFCWS [33‒35].

At age of 15 years, child behavior problems were assessed using items from the Child Behavior Checklist (CBCL/6–18), one of the most commonly used measures of psychopathology symptoms in children [36]. Primary caregivers rated their child’s behavior on 34 items using a 3-point scale (0 = not true; 1 = sometimes true; 2 = very true or often true). We analyzed raw total scores from the internalizing (analytic sample α = 0.88; 8 items) and externalizing subscales (analytic sample α = 0.91; 20 items), which were square root transformed prior to analyses to improve univariate normality and then converted to z-scores to aid interpretability. Parent report of psychopathology was analyzed, as it was far more comprehensive than child self-reports (which only focused on measuring depressive symptoms). As we have described elsewhere [37, 38], the occurrence, predictors of, and consequences for discrepancies between parental and child reports are important to study in their own right and could be important in a future investigation.

We adjusted for the following covariates, measured at the time of the child’s birth, to rule out the effects of baseline sociodemographic factors: maternal age; maternal race/ethnicity; maternal marital status; mother-reported receipt of public assistance, welfare, or food stamps; and maternal education. We also adjusted for maternal depression or substance use when her child was 3 years of age because parent psychopathology symptoms were associated with both child psychopathology and maltreatment in our sample and could lead to maternal bias in both the reporting of child abuse exposure and psychopathology symptoms [39, 40] (see Covariate section in Supplement 1, available online). For comparison, we also conducted analyses without adjustment for maternal psychopathology (see online suppl. Table S4, S5, available online).

All analyses were stratified by sex because stress exposure [30] and psychopathology [41‒43] vary between boys and girls and may lead to differences in the effects of these life-course theoretical models [2]. We used sex stratification rather than tests of statistical interaction because interaction terms can only capture differences in magnitude of the exposure effect in the same life-course model, while stratification can capture differences in the structure of the life-course model as well as the magnitude of exposure effects.

To minimize loss of power and reduce potential bias due to attrition, we performed multiple imputation among children with complete outcome data on psychopathology symptoms measured at year 15 (n = 3,474; see Data Analysis section in Supplement 1, available online). We used an innovative two-stage structured life-course modeling approach (SLCMA) [44, 45] to test the strength of evidence for sensitive periods. The SLCMA was originally developed by Mishra [46] and later extended by Smith [44] to analyze repeated binary exposure data across the life course. The SLCMA allows researchers to compare competing life-course theoretical models simultaneously and identify the most parsimonious explanation for variation in the outcome of interest. Details about the SLCMA are available in Supplement 1.

We considered the sensitive period theoretical model (assessed at ages of 3, 5, and 9 years) and compared it with a second model – the accumulation of risk model – in which the outcome varies with the number of occasions exposed regardless of timing (i.e., under the accumulation model there are no sensitive periods). The SLCMA uses least angle regression (LARS) to identify which theoretical model (or set of theoretical models working in combination) is most supported by the data. When LARS identifies the accumulation model alone, it indicates no sensitive periods are supported by the data; when LARS identifies a sensitive period model, it indicates the specific period when exposure to maltreatment has the greatest effect on the outcome.

A total of 8 LARS models were conducted, corresponding to each type of maltreatment (neglect and harsh physical discipline) and outcome (internalizing and externalizing psychopathology), separately for boys and girls. We regressed the exposures on the covariates and then implemented LARS on the regression residuals to adjust for possible confounding [45]. We used the covariance test [47] to determine whether a set of theoretical models working in combination explained sufficiently more of the outcome variation than a single theoretical model, applying a Bonferroni significance threshold of 0.05/8 = 0.00625. No combination of models met this threshold; hence, all results are reported as the single best fitting theoretical model. Compared with other variable selection procedures, such as stepwise regression, the LARS has multiple benefits including greater statistical power [44], not overinflating effect size estimates [48], and not introducing bias in hypothesis tests [47].

Having selected a theoretical model using LARS, we then estimated the coefficients of the selected model for each type of maltreatment, outcome, and sex, enabling us to determine the magnitude of effect for a selected theoretical model while continuing to adjust for covariates. We calculated confidence intervals (CIs) for the effect estimates that have 95% coverage while adjusting for the selection made by LARS [44]. Although researchers have grouped adversity exposures into clusters – such as deprivation and threat, as a means to characterize the distinct neural pathways they might influence – we analyzed these exposures separately because we wanted to determine if there were unique timing-outcome effects for each form of maltreatment and preserve statistical power (a multiple prediction model with multiple exposures and life course hypotheses would both reduce power and complicate interpretation of results).

The analytic sample was nearly sex-balanced (48% female) and diverse by race/ethnicity (21.9% white; 50.3% black; 24.4% Hispanic; 3.5% other) and socioeconomic background, though skewed toward more disadvantaged families, as indicated by levels of maternal education (11.2% of mothers were college educated or higher) and receipt of public assistance/welfare (36.2%) (see online suppl. Table S1, available online). Exposure to childhood maltreatment was common, with 988 children exposed to neglect at any time point (28.4%), 1,214 exposed to harsh physical discipline (34.9%), and 334 children (9.6%) exposed to both types of maltreatment at any time point. Reports of child exposure to neglect increased with age, more than doubling between ages of 5 and 9 years, whereas reports of harsh physical discipline were stable across time (Table 1).

Exposures were somewhat correlated across time (average correlation across time points for neglect: r = 0.33; harsh physical discipline: r = 0.51; see online suppl. Table S2, available online). Children exposed to harsh physical discipline at any time point were also slightly more likely to be exposed to neglect at any time point (r = 0.28). These correlation values are well below the correlation of 0.80 observed to limit the ability of the SLCMA to identify the correct life-course model [44].

Internalizing and externalizing symptoms were moderately correlated (Pearson r = 0.51). Boys had, on average, greater levels of externalizing psychopathology at age of 15 years than girls (p < 0.001) (see online suppl. Table S3, available online). Girls, in comparison, had greater levels of internalizing problems at age of 15 years (p = 0.006). Children of mothers with less than a high school education, who received public assistance, used substances, or had periods of depression were also more likely to show higher levels of internalizing and externalizing problems at age of 15 years (all p < 0.05). Children born to younger and unmarried mothers showed higher levels of externalizing problems in particular (p < 0.001) compared to their peers.

Tables 2 and 3 display the life-course theoretical models selected by the LARS procedure separately for each type of childhood maltreatment exposure, psychopathology outcome, and sex. These tables also show the effect estimates and 95% CIs for these selected models, adjusted for covariates.

As shown in Table 2 and Figure 1, an accumulation model best explained the relationship between neglect and internalizing symptoms in both girls (r² = 0.6%) and boys (r² = 0.9%). For exposure to harsh physical discipline, a sensitive period at age of 9 years best explained the association with internalizing symptoms in girls (r² = 0.7%; see Fig. 2). A sensitive period at age of 5 years best explained the association in boys (r² = 0.4%; see Fig. 2). The same life-course theoretical models were chosen in the analyses, omitting the maternal psychopathology variables (see online suppl. Table S4, S5, available online).

As shown in Table 3, the accumulation model best explained the relationship between neglect and externalizing symptoms in girls, although this was the only model in which the CI for the regression coefficient included zero. Similar to the internalizing symptom findings, a sensitive period at age of 9 years was again selected as the best fitting model for explaining the relationship between harsh physical discipline and externalizing symptoms in girls (r² = 1.0%; see Fig. 2). In boys, however, an accumulation model explained the most variation in externalizing symptoms following both neglect exposure (r² = 0.7%) and harsh physical discipline (r² = 1.4%; see Fig. 1). These findings were also consistent when maternal psychopathology variables were removed (see Supplement 1, online suppl. Table S4, S5, available online).

Two primary findings emerged from this study. First, we identified sensitive periods for harsh physical discipline: both internalizing and externalizing symptoms were elevated in girls exposed to harsh physical discipline at age of 9 years; internalizing symptom scores were elevated among boys exposed at age of 5 years. These findings are consistent with work by Teicher and colleagues [49, 50] in showing that physical abuse and neglect may not only have different sensitive periods for psychopathology symptoms but also the timing of these sensitive periods may be sex dependent. Our ability to detect these sensitive period effects was notable because reports of harsh physical discipline were moderately correlated over time in this sample, making sensitive periods harder to discern. The large sample size of FFCWS and the statistical power it therefore afforded enabled us to differentiate sensitive period from accumulation effects.

Second, with neglect, we saw stronger and consistent evidence for the accumulation of risk model for both boys and girls. For each additional time period of exposure to neglect, the z-score for psychopathology increased by one-tenth, on average. Though not a fully direct comparison, these findings align with the Bucharest Early Intervention Project, which found that children randomly assigned to remain in institutional care had greater psychopathology symptoms in adolescence than children who were never institutionalized or randomized to high-quality foster care [51]. However, in the Bucharest Early Intervention Project, sex differences in this accumulation effect have been observed, with girls being somewhat protected from the effects of severe early deprivation [51]. Reports of neglect were weakly correlated across time in our FFCWS analytic sample, suggesting parental reporting was either inconsistent or experiences of neglect were intermittent, perhaps due to changes in work, childcare, and neighborhood conditions [52]. To our knowledge, this is the largest study to test the sensitive period hypothesis for neglect in children or adolescents.

These findings provide important clues for researchers to consider in narrowing the search space to identify mechanisms underlying psychopathology risk. If sensitive periods begin after children reach specific maturational goals, as has been found [53], then our findings hint at what domains to study and when. Our results can also guide clinicians in the assessment of adversity exposure, particularly in pediatric care settings [54]. Experts have discussed which screening measures to use [55], how to address patient- and provider-level barriers to implementation [56], and general challenges of screening [57]. This study sheds light on when these tools should be deployed if universal screening (the ideal scenario) is not an option.

This study had four major limitations. First, there were limitations in the measurement of child maltreatment. We were unable to triangulate data from multiple sources (e.g., child self-report; administrative records) as such data were not available in this study. Further, FFCWS investigators did not ask caregivers about their children’s exposure to maltreatment before age of 3 years. Children younger than 3 years of age have the highest rates of reported maltreatment, comprising one-quarter of all documented cases of child maltreatment [58]. Thus, we are likely underestimating the amount of maltreatment and also including children who were maltreated in the unexposed group. Assuming that people are unlikely to self-report abuse if there is none, our estimates are likely underestimates of the effects due to this misclassification of maltreated children. Consistent with our findings, however, a 2018 paper using data from the ALSPAC study found evidence for sensitive periods during middle childhood (defined in ALSPAC as age of 6.75 years) associated with sexual and physical abuse. This finding is striking because the ALSPAC study used the same analytic approach as ours but had maltreatment measures available as early as age of 1.5 years [26]. Nevertheless, analyses of secondary data are always limited by the measures originally collected by the primary study investigators, and for us, the timing of those measurements might not be fully optimized to detect sensitive periods. Second, as with most longitudinal studies, there was nonresponse and attrition over waves, which could bias the results (i.e., 89%, 86%, 85%, 76%, and 73% of baseline families participated at ages of 1, 3, 5, 9, and 15 years, respectively). However, it is notable that a small percentage of the sample appears to permanently attrite, or leave, the study; most families missing a wave return in the following wave [59‒61]. More importantly, there is little evidence that nonresponse in FFCWS for any given wave is predicted by social-demographic factors, including marital status, education, race, health status, and poverty (see attrition tables here https://ffcws.princeton.edu/sites/g/files/toruqf4356/files/documents/attrition_table_1.pdf). These findings imply that the missing may be mostly at random and therefore not significantly biasing the results [62]. Nevertheless, we addressed this attrition through multiple imputation, though some bias from unobserved variables is still likely. Third, the SLCMA does not accommodate time-varying covariates. Thus, we were unable to account for macro-level shocks associated with the Great Recession or other factors. Future studies should investigate these additional risk factors – as well as promotive (and protective) factors – in the etiology of psychopathology. Fourth, as with the measure of maltreatment, psychopathology relied on caregiver reports, which may cause lower reported severity of symptoms. In summary, these findings suggest more nuanced work is needed to assess early exposure to maltreatment, which could lay the groundwork for research, policy, and intervention.

The authors thank Janine Cerutti, Stephanie Gomez, Thomas Soare, and Nitasha Siddique for their assistance in preparing this manuscript for publication. The authors also thank Susan T. Landry for her assistance in copyediting this manuscript.

Informed consent was obtained for each family at each interview. The Institutional Review Boards at Columbia University and Princeton University approved the FFCWS (Approval #5767).

Colter Mitchell received grants R01MD011716 and R01HD076592 to support research on FFCWS. Erin Dunn, Daniel Busso, Kathryn Davis, Andrew Smith, Henning Tiemeier, and Ezra Susser report no conflicts of interest.

This research was specifically funded by grants K01MH102403 and R01MH113930 (Dunn). FFCWS is funded by grants R01HD36916, R01HD39135, and R01HD40421 and a consortium of private foundations. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health, the U.S. Department of Veterans Affairs, or the United States Government. The funding sources had no role in the design or conduct of the study.

All authors made substantive contributions to the work presented in this paper. Erin C. Dunn led the concept and design for this paper and provided supervision. Erin C. Dunn, Daniel S. Busso, Kathryn A. Davis, Andrew D.A.C. Smith, Colter Mitchell, Henning Tiemeier, and Ezra S. Susser were involved in the acquisition, analysis, and/or interpretation of data; performed critical revision of the manuscript for important intellectual content; and had full ability to access to all the data in the study and accepted responsibility to submit for publication. Daniel S. Busso and Kathryn A. Davis conducted statistical analysis. Erin C. Dunn drafted the manuscript along with Daniel S. Busso and Kathryn A. Davis. Daniel S. Busso provided administrative, technical, and/or material support.

The data that support the findings of this study are not publicly available due to the inclusion of confidential information that could compromise the privacy of the research participants. Access can be granted via a contract data use license to faculty and research personnel at institutions with an institutional review board (IRB) or human subjects review committee registered with the US Office for Human Research Protections (OHRP) or the National Institutes of Health (NIH), once they obtain IRB approval for their research purposes. To submit an application, you can contact ffdata@princeton.edu with the required documents listed at https://ffcws.princeton.edu/restricted. Further inquiries can be directed to ffdata@princeton.edu.