Childhood Maltreatment and Mental Health Problems: A Systematic Review and Meta-Analysis of Quasi-Experimental Studies

1Childhood maltreatment is a well-established risk factor for mental health problems. For example, systematic reviews and meta-analyses consistently show associations between childhood maltreatment and a range of psychiatric disorders, such as depression, anxiety, psychosis, suicidality, non-suicidal self-injury, attention-deficit hyperactivity disorder (ADHD), conduct disorder, and substance abuse. Understanding the causal nature of these associations is critical for informing preventative interventions.

2Maltreated children might be more likely to develop mental health problems by virtue of their exposure to abuse and neglect. Alternatively, their risk might be higher because they have other genetic and environmental risk factors for mental health problems, which confound previously observed associations. For example, evidence suggests that maltreated children are likely to have family histories of mental illness and may have higher polygenic scores for psychiatric disorders (e.g., schizophrenia and depression). Maltreated children are also likely to experience other environmental risks for mental health problems, such as socioeconomic disadvantage and bullying victimization. However, these co-occurring risk factors have not been not fully accounted for in the majority of previous research on maltreatment and mental health, which used classical epidemiological methods (e.g., multiple regression) with limited ability to account for genetic influences or other unmeasured confounders.

3Quasi-experimental methods offer a solution to help disentangle the causal effects of child maltreatment on mental health from confounding. Four broad categories of quasi-experimental methods have been used to examine the relationship between maltreatment and mental health. First, family-based designs (e.g., twin or sibling differences, Children of Twins [CoT], and adoption designs; 14) capitalize on varying genetic and environmental relationships between family members to examine the effects of maltreatment independent of familial confounding. Second, panel data designs (e.g., fixed-effects methods, random intercept cross-lagged models) leverage longitudinal data to test whether within-individual changes in maltreatment exposure predict changes in mental health, independent of stable individual factors. Third, natural experiments examine the effects of maltreatment that is not related to family or individual risk factors, but rather occurs due to wider social or political processes. For example, the English and Romanian Adoptees (ERA) Study examined the mental health consequences of institutional neglect that occurred during the Romanian Ceauşescu regime, in which widespread poverty and bans on contraception led to thousands of babies being abandoned in orphanages. Fourth, propensity score methods (e.g., propensity score matching; inverse probability weighting) aim to mimic a randomized experiment by statistically removing confounding by (measured) pre-existing differences between maltreated and non-maltreated individuals. Although propensity score methods do not directly account for unmeasured confounding, they have been found to be more effective than multiple regression in reducing confounding.

4While a number of individual studies have applied these quasi-experimental methods to strengthen causal inference about the relationship between maltreatment and mental health, there has been no systematic summary of this evidence. Such a systematic evaluation is important to estimate the pooled magnitude of a potentially causal relationship between maltreatment and mental health, examine whether findings triangulate across quasi-experimental methods with different assumptions and sources of bias, and examine whether findings differ according to key moderators, such as the type of mental health outcome, form of maltreatment, or assessment method. To address this research gap, we conducted the first systematic review and meta-analysis of quasi-experimental studies on childhood maltreatment and mental health problems.

5We performed a systematic review and meta-analysis following a pre-registered protocol on Prospero (CRD42020187520), adhering to PRISMA and MOOSE reporting guidelines (Tables S1-2).

6We searched Embase, PsycINFO, and MEDLINE for peer-reviewed studies written in English and published from database inception until January 1, 2022. Search terms are reported in Methods S1 of the Supplement. Two investigators (J.R.B and B.W) independently screened titles and abstracts of all articles retrieved from the search, and reviewed the full texts of potentially eligible studies. Inter-rater agreement was high (88%) and remaining disagreement was resolved through discussion with a third investigator (J.B.P).

7We included studies that: Included measures of childhood maltreatment in humans, defined as any of the following experiences before age 18: physical abuse, sexual abuse, emotional abuse, physical neglect, emotional neglect, institutional neglect/deprivation, harsh physical discipline/corporal punishment, or broader measures of victimization/adverse childhood experiences (ACEs) that included the above forms of maltreatment. As such, the primary focus of this meta-analysis was on maltreatment rather than ACEs more generally, as in (as in 17).Examined associations with mental health outcomes, defined as diagnoses or symptoms of internalizing, externalizing, neurodevelopmental, or thought disorders, or general psychopathology, assessed concurrent to or after the observational period for maltreatment.Used a quasi-experimental design (see Tables S3-S4 for an overview of included and excluded designs).

8Included measures of childhood maltreatment in humans, defined as any of the following experiences before age 18: physical abuse, sexual abuse, emotional abuse, physical neglect, emotional neglect, institutional neglect/deprivation, harsh physical discipline/corporal punishment, or broader measures of victimization/adverse childhood experiences (ACEs) that included the above forms of maltreatment. As such, the primary focus of this meta-analysis was on maltreatment rather than ACEs more generally, as in (as in 17).

9Examined associations with mental health outcomes, defined as diagnoses or symptoms of internalizing, externalizing, neurodevelopmental, or thought disorders, or general psychopathology, assessed concurrent to or after the observational period for maltreatment.

10Used a quasi-experimental design (see Tables S3-S4 for an overview of included and excluded designs).

11Two investigators (J.R.B and either L.K or A.T) independently extracted data from all relevant studies (see Methods S2 of the Supplement for details). For all studies, we extracted adjusted effect sizes from quasi-experimental methods. Where multiple quasi-experimental estimates were presented (e.g., effect sizes from dizygotic [DZ] and monozygotic [MZ] twin difference analyses), we selected the most stringent estimate (e.g., MZ twin difference estimates). For comparison, we also extracted unadjusted effect sizes, where this was reported or appropriate (e.g., it was not possible for natural experiment studies). We coded study quality using an adapted version of the Newcastle-Ottowa scale (see Methods S3 of the Supplement for details).

12We converted the individual study effect sizes to Cohen’s d values, reflecting the standardised mean difference in mental health problems between maltreated and non-maltreated individuals. Methods and formula used to convert effect sizes are reported in Methods S4 and Table S5. We excluded a subset of effect sizes from an included study presented as negative binomial regression coefficients, because they could not be converted to Cohen’s d without additional unavailable information. We were also unable to convert unstandardised coefficients from a Mendelian randomisation study to Cohen’s d because the nature of the underlying data meant that the conversion was not appropriate. Excluded effect sizes are reported in Table S6.

13We conducted all analyses in R (version 4.1.1) using the metafor and dmetar packages. The script and dataset are available at https://github.com/jr-baldwin/maltreatment_MH_meta. We performed multi-level random-effects meta-analysis models, which account for dependencies among multiple effect sizes per study and/or cohort (e.g., where multiple mental health outcomes or maltreatment subtypes were examined). We specified four levels of variance in effect sizes: random-sampling variance, within-study (i.e., within-article) variance, between-study variance, and between samples variance, and applied robust standard errors to the main meta-analytic models. We first examined the meta-analytic association between maltreatment and mental health in quasi-experimental adjusted estimates, and then examined the meta-analytic association in studies reporting unadjusted effect sizes, for comparison. We estimated heterogeneity through the I2 statistic, which reflects the proportion of the observed variance that is due to variation in true effect sizes if sampling error was eliminated.

14We conducted sensitivity analyses testing for publication bias, undue influence of individual studies, cohorts, or effect sizes, and robustness of the findings to the definition of maltreatment used. To test for publication bias, we used an Egger’s test for multi-level meta-analysis models, which tests whether the meta-analytic effect size is moderated by the study variance. We also inspected funnel plots and conducted a p-curve analysis based on an average effect size per study to test whether statistically significant findings were likely to reflect selective reporting and/or p-hacking. To test for undue influence of individual studies, cohorts, or effect sizes, we conducted three leave-one-out analyses testing changes in the estimate across permutations omitting in turn each study, cohort, and effect size, respectively. To test the robustness of the findings to the definition of maltreatment, we re-ran the meta-analysis including only studies assessing child maltreatment in the home, and excluded studies that focused on institutional neglect or assessed maltreatment along with other forms of victimization or ACEs.

15Lastly, we used meta-regression to test whether the quasi-experimental association between childhood maltreatment and mental health was moderated by a-priori selected factors, including type of quasi-experimental method, mental health outcome, maltreatment subtype, measurement characteristics, sample characteristics, and study quality. Sensitivity analyses were performed to test whether the moderation results differed after excluding a potentially influential cohort.

16The study selection procedure is summarized in Figure S1. We identified 35 quasi-experimental studies on the association between child maltreatment and mental health, of which 34 had available data that could be meta-analysed (see Table S7 for study details). These studies were based on 29 distinct cohorts, comprising 54,646 participants in adjusted analyses (56.72% female, with a mean age of 28.2 years at mental health assessment). From these studies, we obtained 156 effect sizes for the association between child maltreatment and mental health based on adjusted analyses and 103 effect sizes based on unadjusted analyses.

17A multilevel random-effects meta-analysis model showed a small association between childhood maltreatment and mental health problems in quasi-experimental studies (Cohen’s d=0.31, 95% CI=0.24-0.37, I2=76.27; Figure 1). This meta-analytic association between child maltreatment and mental health from quasi-experimental studies was 45% smaller than that obtained in unadjusted analyses (k=20; Cohen’s d=0.56, 95% CI=0.41-0.71, I2=97.29 [Figure S2]; p-value for difference=0.004), suggesting that the unadjusted association is inflated by confounding. This difference in effect size was broadly consistent when the quasi-experimental adjusted meta-analysis was restricted to studies reporting both unadjusted and adjusted effect sizes (k=20; Cohen’s d=0.26, 95% CI=0.17-0.35).

18An Egger’s test suggested evidence of small-study bias (Q_moderation=8.52, p-value=0.004; see Figure S3A for funnel plot). We investigated the cause of this potential publication bias by performing leave-one-out analyses and found that it was due to the inclusion of data from one cohort, the ERA Study, which comprised a comparatively small sample (N=90-148) with large effect sizes linked to severe institutional neglect (Q_moderation after excluding ERA=2.92, p-value=0.09; see Figure S3B for funnel plot). Where relevant, we therefore conducted later moderation analyses both with and without effect sizes from the ERA cohort, to ensure that results were not biased. The p-curve analysis (focusing on one averaged effect size per study) provided evidential value that a true effect was present (Figure S4).

19Leave-one-out analyses showed that the overall meta-analytic estimate was not unduly influenced by individual studies, cohorts, or effect sizes. The meta-analytic effect size ranged between Cohen’s d=0.29 to 0.32 (with overlapping confidence intervals) after omitting in turn each of the 29 cohorts (including ERA), 34 studies, and 156 effect sizes (Figures S5-6). After excluding studies which assessed maltreatment as part of broader measures of victimization (k=3) (25, 27, 29) or ACEs (k=5) (30, 34, 39, 44, 49), or focused on institutional neglect (k=3) (15, 45, 46), the meta-analytic effect size was also similar to the original estimate (Cohen’s d=0.34, 95% CI=0.26-0.42).

20As shown in Figure 2, the association between child maltreatment and mental health was present across different quasi-experimental designs (twin differences, sibling differences, fixed-effects, and natural experiment) and analytic approaches (propensity score matching and inverse probability weighting) and was not significantly moderated by the type of quasi-experimental method used (Q_moderation=3.43; p-value=0.75).

21The association between child maltreatment and mental health was generally similar across different mental health outcomes (Figure 3). Though stronger associations were found for autism symptoms, 4 of 5 of these effect sizes were from the ERA Study, and there was no statistically significant moderation after removing this cohort in a sensitivity analysis (Q_moderation=22.61; p-value=0.07).

22The association between child maltreatment and mental health was moderated by the type of child maltreatment (Q_moderation=22.43; p-value=0.0042). As shown in Figure 4, emotional abuse and institutional neglect were more strongly associated with mental health problems than various subtypes of maltreatment and/or broader composite measures of maltreatment and ACEs. Specifically, pairwise comparisons showed that emotional abuse was more strongly associated with mental health than physical abuse, sexual abuse, emotional neglect, and broader measures of maltreatment and ACEs, while institutional neglect was more strongly associated with mental health than ACEs. However, such findings should be interpreted with caution as the estimates for emotional abuse were only based on 3 studies and 7 effect sizes, and institutional neglect was assessed only in the ERA Study.

23The association between child maltreatment and mental health was similar in studies assessing maltreatment using prospective versus retrospective measures (Table 1), with no significant moderation effect (Q_moderation=0.25; p-value=0.61). This was also the case after we conducted a sensitivity analysis excluding the (prospective) ERA Study on institutional neglect (Q_moderation=1.01, p-value=0.31). However, we were not able to test the independent associations between prospective versus retrospective measures of maltreatment and psychopathology, as this was not tested by any quasi-experimental study.

24The meta-analytic association was present when child maltreatment and mental health outcomes were assessed through a shared rater (self-reports) or through different sources (Table 1), with no moderation effect (Q_moderation=0.47, p-value=0.49). The findings remained similar after excluding the ERA Study, which included different sources (Q_moderation=1.78, p-value=0.18).

25The meta-analytic association was consistent when childhood maltreatment and mental health were assessed cross-sectionally (i.e., at the same time point) or longitudinally (Table 1; Q_moderation=1.41, p-value=0.23). This was also the case after excluding the longitudinal ERA Study (Q_moderation=0.45, p-value=0.50) and when restricting to longitudinal quasi-experimental studies controlling for pre-existing psychopathology (Table 1).

26The meta-analytic association between child maltreatment and mental health was not moderated by sex (Q_moderation=2.91; p-value=0.09) or age at mental health assessment (Q_moderation=0.02; p-value=0.87). We did not examine moderation by race or ethnicity as the majority of studies (25 of 34) did not report this information.

27Lastly, the meta-analytic association between child maltreatment and mental health was not moderated by an overall measure of study quality (Q_moderation=0.15; p-value=0.70; see Table S8 for coding).

28To our knowledge, this is the first meta-analysis to examine the relationship between childhood maltreatment and mental health in quasi-experimental studies. Across 34 studies that included over 54,000 individuals, our meta-analysis provides novel insights into causality, confounding, and specificity of the relationship between child maltreatment and mental health.

29Regarding causality, we found that child maltreatment had a small association with mental health problems (Cohen’s d=0.31) after stringent quasi-experimental control for confounding. Notably, this association triangulated across multiple types of quasi-experimental methods with different assumptions and potential sources of bias, strengthening causal inference. Although a number of included studies used retrospective measures of maltreatment, findings were consistent among studies using prospective measures, suggesting that recall bias does not fully explain the relationship between maltreatment and mental health. Furthermore, while some studies were cross-sectional, we also observed consistent results in longitudinal studies controlling for pre-existing mental health problems, suggesting that the findings are not due to reverse causation. Taken together, this evidence is consistent with a small causal contribution of maltreatment to mental health. Although small, these effects of maltreatment could have far-reaching consequences, given that mental health problems predict poor outcomes in major life domains, such as occupational attainment, physical health, and mortality.

30Regarding confounding, we found that the association between child maltreatment and mental health in quasi-experimental adjusted models was substantially (45%) smaller than in unadjusted models. Of note, the unadjusted association was moderate in magnitude (Cohen’s d=0.56) which is similar to effect sizes reported in prior meta-analyses of non-quasi-experimental studies (Table S10). This reduction in effect size after quasi-experimental adjustment suggests that a large part of the overall relationship between child maltreatment and mental health is confounded by pre-existing risk factors for psychopathology. Future research is needed to identify the specific factors that elevate risk of psychopathology in maltreated children, which might include environmental adversities (e.g., socioeconomic disadvantage; 9) and genetic liability.

31Regarding specificity of effects, we observed three key findings. First, quasi-experimental evidence suggested small causal effects of maltreatment on a broad range of mental health outcomes, including internalizing disorders (e.g., depression, anxiety, suicidality, self-harm), externalizing disorders (e.g., conduct problems, ADHD, alcohol and drug abuse), and psychosis, rather than a few specific outcomes. This finding supports evidence from non-quasi experimental studies, and suggests that maltreatment may affect broad factors underlying multiple disorders (e.g., altered brain structure and function, emotional dysregulation, or information-processing biases) rather than disorder-specific risk factors.

32Second, all subtypes of maltreatment were associated with mental health problems, including experiences involving threat (e.g., physical, sexual, and emotional abuse) and deprivation (e.g., neglect), consistent with non-quasi experimental evidence. However, emotional abuse and institutional neglect were more strongly associated with mental health problems than some other types of maltreatment. The stronger association between emotional abuse and mental health problems was based on evidence from only 3 studies, but supports findings from non-quasi-experimental studies. If causal, it might occur because parental criticism becomes internalized and directly leads to negative self-views and distress. However, non-causal mechanisms are also possible: for example, because emotional abuse is particularly likely to co-occur with other forms of maltreatment, stronger associations might reflect cumulative effects of other maltreatment types. The finding might also partly reflect recall bias, as two of the three included studies used retrospective reports of emotional abuse, which might be particularly influenced by current psychopathology given the more subjective nature of the experience (compared to other maltreatment types, e.g., sexual or physical abuse). Institutional neglect (experienced in Romanian orphanages) may also be strongly associated with mental health problems because (i) it reflects a particularly extreme, severe, and pervasive form of trauma, and/or (ii) it occurred the first few years of life (e.g., between birth to 3 years), disrupting development of brain architecture and neurobiological systems that can affect psychopathology.

33Third, we did not find that effect sizes differed between studies assessing child maltreatment prospectively versus retrospectively. This may seem surprising, as prospective and retrospective measures identify largely different groups of individuals and retrospective reports have been found to be more strongly associated with psychopathology than prospective measures. Nevertheless, the overall pattern of our findings (i.e., visually larger effect sizes for retrospective vs prospective measures) broadly supports these previous results. Such differences may be less pronounced in this meta-analysis because of heterogeneity between samples, as quasi-experimental data were not available to compare effect sizes between prospective versus retrospective measures within the same samples, as in previous non-quasi-experimental research.

34Our findings should be interpreted in the context of limitations. First, each of the quasi-experimental approaches involves potential sources of bias (detailed in Table S3). For example, family-based designs do not account for non-familial confounding, panel data designs can be affected by time-varying confounding, propensity score approaches are liable to unmeasured confounding more generally, and natural experiments may not examine generalisable exposures (e.g., institutional neglect may be qualitatively different to maltreatment in the home). While we therefore cannot entirely rule out unmeasured confounding, convergent findings across multiple quasi-experimental methods provide stronger support for causal inference than findings from any individual method (i.e., triangulation; 57). Second, the type of quasi-experimental method can be related to the type of maltreatment measure used. For example, twin and sibling difference studies primarily used retrospective self-reports to assess maltreatment (e.g.,,,, ), likely because prospectively collected parent reports and official records detect minimal within-pair variation in maltreatment. In contrast, the natural experiment design of the ERA Study meant that official records were used to prospectively assess institutional neglect (as adoption out of the Romanian institutions was handled by the authorities). It was not possible to disentangle the type of quasi-experimental method from these measurement characteristics in a multivariate moderation analysis, because limited data were available for comparisons (e.g., only 4 of 84 effect sizes from twin/sibling studies did not use self-reports to assess maltreatment and mental health). As such, we cannot rule out the possibility that some true moderation effects of measurement or design features were not detected due to suppression by other correlated variables. Third, we cannot draw firm conclusions about the specific effects of maltreatment types, because poly-victimization is common and studies rarely controlled for other co-occurring forms of maltreatment. Fourth, due to a lack of available data, we could not examine whether the findings were moderated by key factors, such as the timing of maltreatment, the time interval between maltreatment and psychopathology, and race or ethnicity. Future quasi-experimental research is needed to address these questions (e.g., utilising cohorts with detailed temporal measures of maltreatment, and racially diverse samples). Finally, although our meta-analysis focused on quasi-experimental studies, other designs can also strengthen causal inference about the effects of maltreatment, such as experimental animal studies or high-quality prospective longitudinal studies.

35Our findings have implications for research, clinical practice, and public health. Regarding future research, these findings highlight the importance of adopting rigorous quasi-experimental methods to test potentially causal effects of childhood maltreatment. Observational studies using conventional multiple regression approaches are likely to over-estimate the causal effects of maltreatment, and caution should thus be applied when interpreting such findings. Of course, quasi-experimental methods may not completely alleviate bias, and researchers should also take steps to further strengthen causal inference, such as using longitudinal designs and prospective measures incorporating multiple informants. Future research capitalising on these methods should provide valuable insights into the effects of child maltreatment on other outcomes (e.g., physical health, socioeconomic outcomes) and can be used to inform estimates of the economic costs of maltreatment.

36Regarding clinical practice, to minimise risk of psychopathology in individuals exposed to maltreatment, interventions should adopt a holistic approach that addresses both the maltreatment experience (e.g., via trauma-focused cognitive-behavioural therapy; 84) as well as wider risk factors for mental illness. Regarding public health, interventions that can prevent maltreatment (e.g., home visitation programs for high-risk families; 85, 86) should also prevent a proportion of cases of mental health problems in the population. As such, investing in these interventions is not only essential for children’s welfare, but could prevent long-term financial costs and suffering due to mental illness.