Childhood maltreatment involves physical, sexual and emotional abuse, and emotional and physical neglect, to people under 18 years of age within a relationship of responsibility, leading to harm in their health, survival and development. Reference Warrier, Kwong, Luo, Dalvie, Croft and Sallis1,Reference Baumeister, Akhtar, Ciufolini, Pariante and Mondelli2 Before the COVID-19 pandemic, over 1 billion children experienced childhood maltreatment globally. Reference Park and Walsh3 During the pandemic, this number increased globally despite Target 16.2 of the United Nations’ Sustainable Development Goals (SDGs): ending all forms of abuse against children. Reference Lee and Kim4
Given the high current prevalence of childhood maltreatment, the long-term psychological consequences cannot be ignored. Reference Park and Walsh3 In addition to cross-sectional and longitudinal studies, a recent systematic review found significantly higher risks of depression, anxiety, attention-deficit/hyperactivity disorder (ADHD) and internalising and externalising problems in childhood maltreatment victims. Reference Baldwin, Wang, Karwatowska, Schoeler, Tsaligopoulou and Munafò5–Reference Strathearn, Giannotti, Mills, Kisely, Najman and Abajobir10 However, pre-existing environmental and genetic risk factors for psychopathology may confound these findings. As such, previous evidence needs to be triangulated with other methods to confirm causality. Reference Baldwin, Wang, Karwatowska, Schoeler, Tsaligopoulou and Munafò5
Besides confirming causality, identification of causal mechanisms prioritises areas for intervention. Observational studies showed independent associations among stress, chronic inflammation, socioeconomic status, maladaptive behaviours and cognitive and emotional dysregulation with childhood maltreatment and neuropsychiatric outcomes. Reference Carr, Duff and Craddock6,Reference Macpherson, Gray, Ip, McCallum, Hanlon and Welsh7 However, while some studies supported the mediating role of behavioural, social and cognitive traits, Reference Alameda, Rodriguez, Carr, Aas, Trotta and Marino11–Reference Keyes, Eaton, Krueger, McLaughlin, Wall and Grant13 others found inconsistent or weak mediating effects Reference Macpherson, Gray, Ip, McCallum, Hanlon and Welsh7,Reference Li, Luyten and Midgley14,Reference Kent, Hopfer, Corley and Stallings15 due to inaccurately measured mediators and inadequate power. Besides, unmeasured confounding and the inability to temporally order investigated traits limit the delineation of causal mediators. Reference Carter, Sanderson, Hammerton, Richmond, Davey Smith and Heron16 These justify further research into behavioural, social and cognitive traits linking childhood maltreatment to neuropsychiatric outcomes.
The role of epigenetics such as DNA methylation (DNAm) has also been of recent interest. Reference Yang, Zhang, Ge, Weder, Douglas-Palumberi and Perepletchikova17 DNAm at genes involved in stress response, serotonin signalling, inflammation and cell growth were previously associated with childhood maltreatment in epigenome-wide association studies (EWAS) of childhood maltreatment. Reference Parade, Huffhines, Daniels, Stroud, Nugent and Tyrka18 However, current research on the mediating effects of DNAm is limited to candidate genes. Thus, the relationship among DNAm, childhood maltreatment and neuropsychiatric disorders remains unclear. Reference Smearman, Almli, Conneely, Brody, Sales and Bradley19,Reference Comtois-Cabana, Barr, Provençal, Ouellet-Morin and Provenzi20 This warrants further exploration of the effect of DNAm at other genes on the risk of neuropsychiatric disorders.
Mendelian randomisation uses genetic variants to proxy the effects of an exposure on an outcome. Reference Bowden, Del Greco, Minelli, Davey Smith, Sheehan and Thompson21 Because genetic variants are randomly allocated at conception, Mendelian randomisation overcomes the limitations of observational studies to support their findings. Recent Mendelian randomisation studies found that childhood maltreatment is causally associated with ADHD, major depressive disorder (MDD) and schizophrenia. Reference Warrier, Kwong, Luo, Dalvie, Croft and Sallis1,Reference Alkema, Marchi, van der Zaag, van der Sluis, Warrier and Ophoff22,Reference Baltramonaityte, Pingault, Cecil, Choudhary, Järvelin and Penninx23 However, two of these studies utilised outcome data-sets with samples from the UK Biobank, which overlap with the exposure data, Reference Warrier, Kwong, Luo, Dalvie, Croft and Sallis1,Reference Alkema, Marchi, van der Zaag, van der Sluis, Warrier and Ophoff22 thus introducing sample overlap bias; in addition, one investigated only a composite psychocardiometabolic outcome trait. Reference Baltramonaityte, Pingault, Cecil, Choudhary, Järvelin and Penninx23 Furthermore, none of the existing studies have conducted mediation Mendelian randomisation to delineate causal mediators of this relationship.
Aims
This study uses a mediation Mendelian randomisation study design to first confirm the causal association between childhood maltreatment and neuropsychiatric outcomes, then to evaluate the mediating role of addiction-related, biological, behavioural, cognitive and socioeconomic traits, and the potential role of DNAm.
Method
Study design
This study follows Strengthening the Reporting of Observational Studies in Epidemiology using Mendelian randomisation guidelines (Appendix 1 (found in the Supplementary material)). The Mendelian randomisation assumptions are described in Appendix 2. Figure 1 shows the overall workflow. Univariable Mendelian randomisation (UVMR) investigated the causal effect of childhood maltreatment on neuropsychiatric outcomes. Two-step Mendelian randomisation determined indirect effects acting through mediators. Multivariable Mendelian randomisation (MVMR) determined residual direct effects of childhood maltreatment, adjusting for significant mediators. Reference Carter, Sanderson, Hammerton, Richmond, Davey Smith and Heron16 Epigenetic Mendelian randomisation determined the relationship between DNAm and outcomes. Reference Carter, Sanderson, Hammerton, Richmond, Davey Smith and Heron16 All analyses utilised data from summary statistics of genome-wide association studies (GWAS) or epigenome-wide association studies (EWAS). Different sets of covariates were used for exposures and outcomes. Ethics approval and consent were obtained for each GWAS and EWAS, and can be found in the original publications.
Overall study design with the aims, data-sets, statistical methods and expected outcomes. CM, childhood maltreatment; EWAS, epigenome-wide association study; GWAS, genome-wide association study; PGC, Psychiatric Genomics Consortium; ADHD, attention-deficit/hyperactivity disorder; MR, Mendelian randomisation; MR-PRESSO, Mendelian randomisation pleiotropy residual sum and outlier; MR-Egger, Mendelian randomisation-Egger; IL6, interleukin-6; IL6R, Interleukin-6 receptor; CRP, C-reactive protein; BMI, body mass index; IVW, inverse-variance weighted; MVMR, multivariable Mendelian randomisation; DNAm, DNA methylation; mQTL, methylation quantitative trait loci; GoDMC, Genetics of DNA Methylation Consortium.
Data sources
Exposure
Genetically predicted childhood maltreatment information was obtained from a GWAS of 143 473 participants of European ancestry in the UK Biobank. Reference Warrier, Kwong, Luo, Dalvie, Croft and Sallis1 Childhood maltreatment was retrospectively self-reported through a five-item Childhood Trauma Screener, which rates each trauma subtype out of four, with zero indicating never experienced and four indicating often experienced. Scores for each subtype were summed to obtain a composite score for childhood maltreatment. The quantitative phenotypic trait in this GWAS is a log-transformed score of the Childhood Trauma Screener, with effect sizes representing a log-score increase per effect allele. Additional questionnaire and GWAS details are reported in the original study. Reference Warrier, Kwong, Luo, Dalvie, Croft and Sallis1
Outcomes
Six neuropsychiatric outcomes were included: ADHD (N = 55 374), anxiety disorder (N = 21 761), autism spectrum disorder (N = 46 350), bipolar disorder (N = 413 466), MDD (N = 500 199) and schizophrenia (N = 275 249). Summary statistics from European-ancestry participants were obtained from the Psychiatric Genomics Consortium, and excluded UK Biobank data. Reference Sullivan, Agrawal, Bulik, Andreassen, Borglum and Breen24 Controls were participants without neuropsychiatric disorders or associated symptoms. Study details are shown in Supplementary Table 1 available at https://doi.org/10.1192/bjp.2025.10433.
Mediators
Each mediator was proxied by genetic instruments obtained from their respective GWAS summary statistics (see details in Supplementary Table 1). A total of 27 mediators were selected based on previously hypothesised pathways from observational studies. Reference Baldwin, Wang, Karwatowska, Schoeler, Tsaligopoulou and Munafò5,Reference Strathearn, Giannotti, Mills, Kisely, Najman and Abajobir10,Reference Scott, Malacova, Mathews, Haslam, Pacella and Higgins25 These mediators have also been reported to be associated with childhood maltreatment and/or neuropsychiatric outcomes in observational studies, indicating potential mediating pathways. To evaluate the role of systemic inflammation and chronic stress, four biological mediators were included (serum C-reactive protein (CRP), plasma cortisol, plasma interleukin-6 and interleukin-6 receptor), along with body mass index (BMI) as a general health indicator. Reference Baumeister, Akhtar, Ciufolini, Pariante and Mondelli2 For behavioural traits, seven lifestyle and behavioural traits relating to sedentary lifestyle and sleep problems were evaluated (self-reported: moderate/vigorous physical activity, TV watching hours, sleep duration, leisure screen time, propensity for speeding; accelerometer measured: average acceleration, sleep duration). Childhood maltreatment has known negative effects on different domains of cognitive functioning in children, which further manifests as risk-taking behaviours later in life, including engaging in addiction-related behaviours. As such, three cognitive traits (executive functioning, intelligence and internalising behaviour in childhood/adolescence), two risk-taking measures (risk tolerance and propensity for speeding) and nine risky health behaviours (addiction-related traits: smoking initiation, age of smoking initiation, pack years smoked, smoking cessation, cigarettes per day, cannabis use disorder, alcoholic drinks per week, alcohol dependency and ICD-coded substance abuse) were also assessed as mediators in this study, because these traits have been observationally associated with both childhood maltreatment and mental disorders. Reference Lunding, Ueland, Aas, Høegh, Werner and Rødevand26,Reference Aas, Etain, Bellivier, Henry, Lagerberg and Ringen27 Finally, to consider the potential protective effect of socioeconomic factors, household income and educational attainment were also included as mediators in the analyses. Additional details and references supporting the inclusion of these mediators are shown in Appendix 2.
CpG sites and mQTLs
On 13 June 2023, five childhood maltreatment-related phenotypes were searched in the EWAS Atlas (adverse childhood experience, child abuse, childhood adversity, childhood sexual victimisation and childhood stress) to identify 2929 childhood maltreatment-associated CpG sites at significance P < 1 × 10−4 from previous EWAS. Reference Li, Zou, Li, Gao, Sang and Zhang28 These were measured in blood, buccal cells, leukocytes, saliva and sperm samples. Methylation quantitative trait loci (mQTLs) were obtained from the Genetics of DNA Methylation Consortium (GoDMC) to proxy DNAm levels at these CpG sites. Reference Min, Hemani, Hannon, Dekkers, Castillo-Fernandez and Luijk29 DNAm data were generated using the Illumina Infinium HumanMethylation450 BeadChip from whole-blood samples of 27 750 European participants. Summary statistics of 36 cohorts were combined in an inverse-variance, fixed-effect meta-analysis to determine mQTLs for each CpG site. Reference Min, Hemani, Hannon, Dekkers, Castillo-Fernandez and Luijk29 Details can be found in Supplementary Table 1.
Instrument selection
Single-nucleotide polymorphisms (SNPs) were used as genetic instruments for childhood maltreatment at genome-wide significance (P < 5 × 10 −8). Linkage disequilibrium clumping (R 2 < 0.01, 1000 kb window), using the European 1000 Genomes reference panel, Reference Auton, Abecasis, Altshuler, Durbin, Abecasis and Bentley30 obtained independent SNPs. F-statistics were calculated to evaluate instrument strength, with F > 10 indicating strong instruments.
Independent cis-mQTLs (mQTL < 1 Mb from CpG site) at P < 1 × 10−8 were extracted from the GoDMC database to minimise potential horizontal pleiotropy. Reference Min, Hemani, Hannon, Dekkers, Castillo-Fernandez and Luijk29,Reference Richardson, Zheng, Davey Smith, Timpson, Gaunt and Relton31 This resulted in 1091 CpG sites with viable cis-mQTLs for epigenetic Mendelian randomisation.
Statistical analysis
Mediation Mendelian randomisation leverages two-step Mendelian randomisation and MVMR to separate the total causal effect of childhood maltreatment on each outcome into indirect and direct effects, respectively (Supplementary Fig. 1). Statistical analyses were conducted in R (version 4.3.1 for Windows) using these R packages: TwoSampleMR (version 0.6.8) Reference Hemani, Zheng, Elsworth, Wade, Haberland and Baird32 and MendelianRandomization (version 0.10.0) Reference Broadbent, Foley, Grant, Mason, Staley and Burgess33 for UVMR, ieugwasr (version 1.0.3) Reference Elsworth and Rasteiro34 for linkage disequilibrium clumping, MVMR (version 0.4) Reference Sanderson, Spiller and Bowden35 for MVMR, RMediation (version 1.2.2) Reference Tofighi and MacKinnon36 for calculation of indirect effect 95% confidence intervals and coloc (version 5.2.3) Reference Giambartolomei, Vukcevic, Schadt, Franke, Hingorani and Wallace37 for colocalisation. Total and direct effects are presented as odds ratio per unit increase in the Childhood Trauma Screener score. Indirect effects acting through mediators are presented as a percentage of the total effect of childhood maltreatment on each outcome.
A priori statistical power was calculated for Mendelian randomisation analyses with an online tool (https://sb452.shinyapps.io/power/; Supplementary Table 2). Calculation details can be found in Appendix 3. For ADHD as an example outcome, we would have sufficient power (>80%) if the expected odds ratio is ≤0.92 or ≥1.12. The power estimates for all other UVMR and epigenetic Mendelian randomisation associations can be found in Supplementary Table 2.
UVMR: total effects
The random-effects inverse-variance weighted (IVW) method was used as the main way to determine total effect (Supplementary Fig. 1A) to investigate the effect of childhood maltreatment on neuropsychiatric outcomes. Various tests were conducted to determine potential violations of the Mendelian randomisation assumptions. Reference Burgess, Davey Smith, Davies, Dudbridge, Gill and Glymour38 Steiger testing assessed whether genetic instruments explained more of the variance in childhood maltreatment than the outcome, and evaluated violations of assumption 1. This supported the use of selected instruments to investigate the direction of the causal association between childhood maltreatment and neuropsychiatric outcomes. F-statistics are a test for instrument strength, and were used to evaluate violations of assumption 1. Cochran’s Q test for instrument heterogeneity and the Mendelian randomisation-Egger intercept test evaluated violations of assumptions 2 and 3. Further pleiotropy-robust estimates were derived using the weighted median and Mendelian randomisation pleiotropy residual sum and outlier methods. Additional details of each sensitivity analysis are shown in Appendix 2.
Given that outcomes were selected based on prior knowledge, and therefore considered hypothesis-based analyses, UVMR effects under the IVW method that passed the Benjamini–Hochberg-corrected P-value threshold (P < 0.05) were considered significant. Of these, those with directionally concordant sensitivity analyses at P < 0.05 were considered robust and were included in subsequent mediation Mendelian randomisation analyses.
To address the key limitation in previous studies, Reference Warrier, Kwong, Luo, Dalvie, Croft and Sallis1 we further calculated pairwise genetic correlation between childhood maltreatment and each outcome, and their corresponding linkage disequilibrium score regression intercepts. The intercept estimates the inflation in genetic correlation between two traits due to sample overlap. Reference Bulik-Sullivan, Finucane, Anttila, Gusev, Day and Loh39,Reference Lee, McGue, Iacono and Chow40 Bivariate genetic analyses were carried out using linkage disequilibrium score (LDSC) regression software (https://github.com/bulik/ldsc). Intercepts for each pairwise regression were considered. Those with confidence intervals that do not overlap with zero suggest that there may be minimal or residual confounding of the genetic relationship between the two traits because of sample overlap.
Two-step Mendelian randomisation: indirect effects
Indirect effects through each mediator were estimated using two-step Mendelian randomisation (Supplementary Fig. 1B). Reference Carter, Sanderson, Hammerton, Richmond, Davey Smith and Heron16 In step one, UVMR was conducted to determine the causal effect of childhood maltreatment on each mediator. In step two, MVMR was conducted to determine the causal effect of mediators on each outcome after adjusting for childhood maltreatment. Standard errors and corresponding 95% confidence intervals were calculated using the Delta method. Reference Carter, Sanderson, Hammerton, Richmond, Davey Smith and Heron16 The proportion mediated by each mediator is the indirect effect divided by the total effect. IVW estimates for each step were multiplied to obtain indirect effects when nominally significant associations in the same direction were obtained in both steps, with significant and directionally concordant sensitivity analyses. Significant mediators were defined as traits with indirect effects with confidence intervals excluding the null Reference Carter, Sanderson, Hammerton, Richmond, Davey Smith and Heron16 (Supplementary Fig. 2).
MVMR: direct effects
MVMR simultaneously estimates the direct effects of multiple exposures on an outcome (Supplementary Fig. 1C). Reference Carter, Sanderson, Hammerton, Richmond, Davey Smith and Heron16 MVMR serves two purposes: first, to generate direct effects of mediators on the outcome for step two of two-step Mendelian randomisation; second, to obtain the direct effect of childhood maltreatment accounting for mediators.
As the main method, MVMR-IVW was used to obtain the direct effects of each exposure. Modified Q-statistic (Qa min), conditional Sanderson–Windmeijer conditional F-statistic (F SW) and the MVMR-Egger intercept test were conducted to evaluate MVMR assumptions. MVMR-weighted median was conducted to obtain direct effects under relaxed assumptions on horizontal pleiotropy. Qhet_mvmr was conducted to consider the effects of weak instruments. Reference Sanderson, Spiller and Bowden35 Assumptions and details of MVMR are provided in Appendix 2.
Robust findings are indicated through directional concordance between MVMR sensitivity analyses and MVMR-IVW estimates, and significant MVMR sensitivity analyses with confidence intervals excluding the null.
Epigenetic Mendelian randomisation and colocalisation
UVMR analysis was conducted to determine the causal effect of DNAm at each childhood maltreatment-associated CpG site on significant outcomes (Supplementary Fig. 1D). Statistical significance was determined by Bonferroni-corrected P-values (ADHD: P < 6.68 × 10−5 (748 sites); MDD: P < 6.68 × 10−5 (748 sites)); schizophrenia: P < 6.59 × 10−5 (759 sites)). The stringent Bonferroni-corrected P-value thresholds were used to reduce false positives because this was a hypothesis-free analysis.
Colocalisation was conducted as a sensitivity analysis for associations passing the Bonferroni-corrected P-value threshold. Reference Zuber, Grinberg, Gill, Manipur, Slob and Patel41 colocalisation determines the probability that CpG sites and outcomes share causal variants (PP.H4). This was done within a ±200-kilobase (kb) window of either the nearest gene mapped using the Infinium HumanMethylation450 Manifest file 42 or the CpG site position. Default prior probabilities (Prior.H3 = 1 × 10−4, Prior.H4 = 1 × 10−5) were used to calculate posterior probabilities with the approximate Bayes factor. Reference Giambartolomei, Vukcevic, Schadt, Franke, Hingorani and Wallace37 Strong evidence of colocalisation was defined as PP.H4 > 0.8. Reference Rasooly, Peloso, Pereira, Dashti, Giambartolomei and Wheeler43 CpG sites with statistically significant epigenetic Mendelian randomisation estimates and passing colocalisation (PP.H4 > 0.8) were included in downstream pathway analysis. CpG sites were mapped to the nearest gene, combined into a gene set and searched for enriched pathways in https://string-db.org (version 12.0). Reference Szklarczyk, Kirsch, Koutrouli, Nastou, Mehryary and Hachilif44
Results
Genetic liability to childhood maltreatment on the risk of neuropsychiatric disorders
There were six genetic variants used as instruments for childhood maltreatment, with an average F-statistic of 32.63 (Supplementary Table 3), suggesting strong instruments. The genetic liability to childhood maltreatment on each outcome risk is shown in Supplementary Fig. 3 and Supplementary Table 4. Using the IVW method, genetic liability to childhood maltreatment was associated with higher risk of ADHD (odds ratio 10.09, 95% CI: 4.76–21.40), MDD (odds ratio 1.89, 95% CI: 1.32–2.70) and schizophrenia (odds ratio 5.89, 95% CI: 1.46–23.78) at Benjamini–Hochberg-corrected P-values. There was no causal evidence for the effect of childhood maltreatment on the risks of anxiety disorder, autism spectrum disorder and bipolar disorder. All associations passed Steiger testing. Generally, sensitivity estimates were consistent with the IVW estimate, and sensitivity tests did not suggest violated Mendelian randomisation assumptions. Detailed UVMR results are given in Appendix 4. To assess sample overlap, LDSC regression analysis was conducted for all childhood maltreatment to outcome pairs (Supplementary Table 5). Across all bivariate LDSC regression analyses, only childhood maltreatment and depression had an intercept of 0.065 (95% CI: 0.052–0.078) that did not overlap with zero, suggesting residual confounding possibly due to sample overlap. This suggests that all UVMR analyses, except childhood maltreatment and depression, were robust to residual confounding due to sample overlap.
Two-step Mendelian randomisation: indirect effect of mediators
In step one, 11 mediators were found to have causal associations under the IVW method (Supplementary Fig. 4 and Supplementary Table 6). Among biological traits, childhood maltreatment was associated with higher BMI and serum interleukin-6 levels. Among addiction-related traits, childhood maltreatment was associated with increased risks of smoking initiation, smoking cessation and substance abuse. Among cognitive and socioeconomic traits, childhood maltreatment was associated with increased internalising behaviour in childhood and adolescence, risk tolerance and leisure screen time, and negatively associated with educational attainment, executive functioning and intelligence. All associations passed Steiger testing. Seven mediators passed sensitivity analyses: smoking initiation, smoking cessation, educational attainment, executive functioning, intelligence, leisure screen time and risk tolerance (Supplementary Table 6).
In step two, the MVMR-IVW method determined the effect of mediators, independent of childhood maltreatment, on each outcome (Supplementary Table 7).
Smoking initiation, substance abuse and increased leisure screen time were significantly associated with higher ADHD risk, while higher educational attainment, executive functioning and intelligence were significantly associated with lower ADHD risk (Table 1). These effects are directionally consistent with step 1. Smoking initiation (18.98%) and substance abuse (20.38%) mediated the largest proportion of the effect of childhood maltreatment on ADHD.
Indirect effects of significant mediators for childhood maltreatment on neuropsychiatric outcome relationship
ADHD, attention-deficit/hyperactivity disorder; MDD, major depressive disorder; NA, not available.
The effect size in step one is the causal effect of childhood maltreatment on the investigated mediator (reported in beta). The effect size in step two is the causal effect of the mediator on the neuropsychiatric outcome (reported in beta). The indirect effect, the product of effect sizes from steps one and two, was calculated if nominally significant associations in the same direction were obtained in both steps, with significant and directionally concordant sensitivity analyses (reported in beta). The total effect is the causal effect of childhood maltreatment on each neuropsychiatric outcome (reported in beta). The proportion mediated by the investigated mediator is the indirect effect divided by the total effect.
Increased leisure screen time was significantly associated with higher risk of MDD, while higher educational attainment and executive functioning were significantly associated with lower MDD risk (Table 1). The directions of these associations were consistent with step 1. These three mediators mediated a similar proportion of the effect of childhood maltreatment on MDD (educational attainment 6.51%; executive functioning 9.27%; leisure screen time 9.90%).
Higher risk tolerance and educational attainment were significantly associated with higher schizophrenia risk, while increased executive functioning and intelligence were significantly associated with lower schizophrenia risk (Table 1). The directions of these associations for all mediators, except educational attainment, were consistent with those in step 1. Risk tolerance (22.17%) and executive functioning (19.26%) mediated the largest proportions of the relationship between childhood maltreatment and schizophrenia.
MVMR sensitivity analyses assessed the robustness of these findings in the context of Mendelian randomisation assumptions. MVMR-weighted median sensitivity analyses for all models were consistent with MVMR-IVW estimates, suggesting that the findings were robust to unbalanced horizontal pleiotropy. Models with smoking initiation and childhood maltreatment as exposures had conditional F-statistics of 11.55 and 10.27, respectively, suggesting that weak instruments would not bias the results (Supplementary Table 7). For other models, at least one exposure had low conditional F-statistics (F sw < 10). Qhet_mvmr estimates were similar to respective MVMR-IVW estimates, suggesting that weak instruments did not bias the MVMR results (Supplementary Table 7).
MVMR: direct effects of childhood maltreatment
To evaluate the direct effects of childhood maltreatment, each mediator was included in the MVMR model independently, then statistically significant mediators were added to a full MVMR model.
Figure 2 shows the total effect of childhood maltreatment from UVMR, and the direct effects of childhood maltreatment from MVMR on each outcome (Supplementary Table 8). For ADHD, adjustment for the combined effects of all statistically significant mediators resulted in a smaller direct effect (odds ratio 2.57, 95% CI: 1.76–3.77) than when individually including mediators, and compared with the total effect (Fig. 2a). For schizophrenia, only the inclusion of all significant mediators in a MVMR model resulted in a slightly lower direct effect of childhood maltreatment on schizophrenia (odds ratio 5.10, 95% CI: 2.71–9.60), compared with the total effect (Fig. 2c). By contrast, regardless of whether individual or all mediators were included in the MVMR model, the direct effect of childhood maltreatment on MDD (odds ratio 1.95, 95% CI: 1.68–2.26) remained largely unchanged from the total effect (Fig. 2b).
Direct effects of childhood maltreatment (CM) adjusting for individual or all significant mediators. ADHD, attention-deficit/hyperactivity disorder; MDD, major depressive disorder; SNPs, single-nucleotide polymorphisms; OR, odds ratio.
F-statistics for the MVMR model that included all significant mediators ranged from 1.27 to 4.26 with ADHD as the outcome, 1.91–6.93 with MDD as the outcome and 1.77–7.02 with schizophrenia as the outcome (Supplementary Table 8), suggesting potential weak instrument bias. However, Qhet_mvmr estimates were directionally consistent across all analyses, suggesting that weak instrument bias did not affect MVMR results (Supplementary Table 8).
Epigenetic Mendelian randomisation and colocalisation
There were 26 statistically significant associations between CpG sites and ADHD (3), MDD (4) and schizophrenia (19) at Bonferroni-corrected P-value thresholds (Supplementary Table 9). Most CpG sites were associated with only one outcome, except for two sites. Cg23337620 (not near any gene) was associated with ADHD (odds ratio 0.55, 95% CI: 0.42–0.71) and schizophrenia (odds ratio 0.69, 95% CI: 0.58–0.82). Cg15790214 (HCG11) was associated with MDD (odds ratio 0.91, 95% CI: 0.88–0.93) and schizophrenia (odds ratio 0.72, 95% CI: 0.68–0.77). Significant CpG sites were included in downstream colocalisation and pathway analysis.
Colocalisation analysis was performed to determine whether CpG sites and outcomes shared the same causal variant. There was strong evidence of colocalisation with 10 CpG sites (PP.H4 > 0.8) (Supplementary Table 10). For these, the mQTLs used to proxy DNAm of two CpG sites in Mendelian randomisation were found to have 100% posterior probability of sharing causal variants with schizophrenia: cg04585390 (SLC5A10) and cg11329058 (RPTOR).
For these ten CpG sites with strong evidence of colocalisation, nine were situated near one gene whereas one was not near any (Supplementary Table 10). These genes are ANGPLT2, CLU, H2BC6, HNRNPK, MAPT, NDFIP2, RPTOR, SLC5A10 and TSSC1. Although no enrichment of protein–protein interactions was observed, three genes – CLU, MAPT and HNRNPK – were implicated in the formation of neurofibrillary tangles.
Discussion
Main findings
To our knowledge, this is the first study to delineate mediators of the relationship between childhood maltreatment and neuropsychiatric disorders. We replicated causal associations between childhood maltreatment and ADHD, MDD and schizophrenia from previous Mendelian randomisation studies. Reference Warrier, Kwong, Luo, Dalvie, Croft and Sallis1,Reference Alkema, Marchi, van der Zaag, van der Sluis, Warrier and Ophoff22,Reference Baltramonaityte, Pingault, Cecil, Choudhary, Järvelin and Penninx23 We established addiction-related, cognitive and socioeconomic traits as mediators (Fig. 3). The direct effect of childhood maltreatment persisted, particularly for MDD. Epigenetic Mendelian randomisation revealed ten childhood maltreatment-associated CpG sites associated with outcomes, highlighting novel pathways for further investigation.
Breakdown of the mediators of the relationship between childhood maltreatment (CM) and neuropsychiatric disorders, potential mechanisms and suggested interventions. Mediators in red are addiction-related traits, mediators in green are cognitive-related traits and the mediator in blue is a socioeconomic trait. Solid arrows represent causal links and the dashed arrow represents associations, not causality. DNAm, DNA methylation; ADHD, attention-deficit/hyperactivity disorder; MDD, major depressive disorder; OR, odds ratio.
Interpretation of our findings and comparison with the literature
Our results support the associations between childhood maltreatment and mediators reported in observational studies. Childhood maltreatment was reported to reduce cognitive functioning, intelligence Reference Matte-Landry, Grisé Bolduc, Tanguay-Garneau, Collin-Vézina and Ouellet-Morin45 and academic achievement, Reference Fry, Fang, Elliott, Casey, Zheng and Li46 and to increase tobacco and illicit drug use. Reference Petruccelli, Davis and Berman8 The magnitude of these differed slightly compared with our Mendelian randomisation estimates. Genetic instruments used in Mendelian randomisation studies affect the usual levels of exposure over long periods of time, corresponding to a lifetime effect of the exposure. Reference Burgess, Davey Smith, Davies, Dudbridge, Gill and Glymour38 For example, educational attainment in observational studies has larger effect sizes compared with our Mendelian randomisation estimates. This may reflect childhood maltreatment impact during short yet sensitive periods, when brain development is highly susceptible to external stimuli. However, childhood maltreatment occurs during sensitive periods in childhood development, specifically when brain development is highly susceptible to external stimuli. Reference Schaefer, Cheng and Dunn47 This could explain the larger effects on educational outcomes in observational studies as compared with our Mendelian randomisation estimates. Nevertheless, the directionality of our results was consistent with observational associations, validating our findings.
The results in step two of mediation Mendelian randomisation align with literature that has reported a significant causal effect of executive functioning, educational attainment and smoking initiation on ADHD, MDD and schizophrenia. Reference Ahn, Norman, Justice and Shaw48–Reference Treur, Demontis, Smith, Sallis, Richardson and Wiers51
In applying novel Mendelian randomisation methods to delineate mediators, we validated hypothesised mechanisms linking childhood maltreatment and outcomes. Previous studies delineated childhood maltreatment as a predictor of reduced executive functioning, Reference Letkiewicz, Funkhouser and Shankman52,Reference Masson, Bussières, East-Richard, R-Mercier and Cellard53 and Mendelian randomisation studies showed that this increased the risks of ADHD, MDD and schizophrenia. Reference Ahn, Norman, Justice and Shaw48,Reference Burton, Sallis, Hatoum, Munafò and Reed49 We add to this by establishing temporal ordering between childhood maltreatment, executive functioning and neuropsychiatric outcomes. Early life stressors impair executive functioning through atypical development of brain regions including the prefrontal cortex, which influences emotional regulation and reward processing. Reference Kasai, Koike, Matsuoka, Okada, Sakakibara and Sakurada54 Reduced educational attainment exacerbates these cognitive deficits. Reference Fry, Fang, Elliott, Casey, Zheng and Li46,Reference Masson, Bussières, East-Richard, R-Mercier and Cellard53,Reference Puetz and McCrory55 These mechanisms contribute to reduced stress tolerance and the inability to disengage from rumination, resulting in increased risk of adult neuropsychiatric disorders. Reference Rock, Roiser, Riedel and Blackwell56
In observational studies, childhood maltreatment was reported to lower educational attainment. Reference Fry, Fang, Elliott, Casey, Zheng and Li46,Reference Romano, Babchishin and Marquis57 Adding to this, we found that educational attainment mediated the effect of childhood maltreatment on ADHD and MDD. At the individual level, exposure to childhood maltreatment impacts brain and cognitive development, causing mental health problems. Reference Fry, Fang, Elliott, Casey, Zheng and Li46 Challenges associated with mental health further limit educational attainment and, consequently, cognitive impairments persist. At a social level, childhood maltreatment-exposed children are neither provided with social and emotional support nor exposed to environments fostering a sense of agency or self-efficacy. Reference Romano, Babchishin and Marquis57,Reference Slade and Wissow58 As such, they are unable to internalise the value of education, leading to poorer educational outcomes. These suggest that educational, child welfare and care-giving sectors contribute to sustaining limited educational attainment and increased risk of neuropsychiatric outcomes. Reference Romano, Babchishin and Marquis57
Although smoking and alcohol dependency were commonly reported as consequences of childhood maltreatment, Reference Petruccelli, Davis and Berman8 our analyses supported only smoking initiation and substance abuse as mediators of the relationship between childhood maltreatment and ADHD. Non-significant indirect effects of other smoking traits may be due to limited power from smaller sample sizes. Alternatively, the self-medication hypothesis states that victims who experienced trauma use illicit substances as a coping mechanism to reduce awareness and maintain a dissociated state. Reference Schimmenti, Billieux, Santoro, Casale and Starcevic59 Thus, without healthy coping mechanisms, these behaviours can elicit unexpected effects – for example, reducing the risk of neuropsychiatric disorders. Reference Brady, Phalen, Roche, Cowan and Bennett60
We formally delineated novel mediators such as leisure screen time and risk tolerance. The role of risk tolerance can be explained by impaired cognitive functioning, whereas leisure screen time is more complex. Leisure screen time may be a form of avoidant coping hindering emotional regulation development. Reference Westerlund, Alvarsson, Carli and Hadlaczky61 However, its effect on outcomes remains uncertain due to the wide content variety and influences on coping strategies. Reference Westerlund, Alvarsson, Carli and Hadlaczky61,Reference Tang, Werner-Seidler, Torok, Mackinnon and Christensen62 Therefore, further research into the implication of leisure screen time as a mediator is required.
Another novelty of our study is the application of epigenetic Mendelian randomisation to establish outcome-specific DNAm signatures. Significant CpG sites were near genes responsible for biological pathways such as apoptosis (HNRNPK, CLU, RPTOR) and transcription regulation (H2BC6, HNRNPK).
Pathway analysis of these mapped genes identified a functionally enriched in neurofibrillary tangle formation pathways, involving CLU, MAPT and HNRNPK. Neurofibrillary tangles are a hallmark of neurodegenerative disorders, including Alzheimer’s and Parkinson’s diseases, acting as a precursor to microtubule destabilisation and dysregulated dynamics of the neuronal skeleton. Reference Metaxas and Kempf63 Our results showed that hypomethylation at the MAPT gene, corresponding to increased expression of the tau protein, was associated with increased schizophrenia risk. Reference Coupland, Mellick, Silburn, Mather, Armstrong and Sachdev64 HNRNPK, an RNA protein, influences the development of early tau pathology, Reference Kavanagh, Halder and Drummond65 and CLU (clusterin) binds and stabilises tau to prevent degradation, thereby accelerating tau aggregation and its neurotoxic properties. Reference Yuste-Checa, Trinkaus, Riera-Tur, Imamoglu, Schaller and Wang66 Interestingly, increased tau levels were previously associated with depression, anxiety and ADHD. Reference Hall, Petersen and Johnson67,Reference Demirci68 Contrastingly, both total and phosphorylated tau levels were lower in patients with schizophrenia. Reference Demirel, Cetin, Turan, Yıldız, Sağlam and Duran69 Nevertheless, tau pathology has a potential mechanistic role in the pathogenesis of neuropsychiatric disorders. We further identified one additional signal at the RPTOR (regulatory associated protein of MTOR complex 1) gene, which is part of the mTORC1 pathway. This stress-related pathway is reported to be downregulated in the prefrontal cortex of humans and rats with schizophrenia. Reference Chen, Guan, Wang and Lin70,Reference Sanacora, Yan and Popoli71 This aligns with our results, which show that childhood maltreatment-induced hypomethylation at RPTOR, triggering the MTORC1 pathway, was associated with increased schizophrenia risk. Considering the two key findings from our epigenetic Mendelian randomisation analysis, we hypothesise that early childhood adversities such as childhood maltreatment are linked to neuropsychiatric outcomes, possibly through changes in DNAm at key genes involved in neurobiological and/or stress pathways.
However, these epigenetic Mendelian randomisation results should be interpreted with caution. First, the full GWAS summary statistics for DNAm at each CpG site were not available, precluding formal two-step epigenetic mediation Mendelian randomisation analyses and delineation of CpG sites as true causal mediators. Nevertheless, the specificity of genetic instruments ensured specificity of the DNAm effect and allowed postulation of biological mechanisms. Significant CpG sites were near genes responsible for biological pathways such as apoptosis (HNRNPK, CLU, RPTOR) and transcription regulation (H2BC6, HNRNPK). Second, DNAm levels were proxied by at most three mQTLs, which explains only a small proportion of the variation in DNAm. This may overlook any true causal associations resulting from larger methylation changes that are not captured by the selected mQTLs. Third, we considered DNAm levels only in whole blood, potentially missing DNAm signatures in other relevant tissues, including the brain. Nevertheless, we considered all CpG sites associated with any adverse childhood exposures in a hypothesis-free approach to identify potential novel epigenetic mediators. We recognise that epigenetic Mendelian randomisation alone is insufficient to establish biological mechanisms. Therefore, these findings warrant further research to confirm the biological role of each gene in the pathogenesis of ADHD, MDD and schizophrenia in victims. This may identify new markers to screen for high-risk victims, or novel therapeutic targets to reduce the risk of neuropsychiatric disorders.
Clinical implications
Despite SDG Target 16.2 aiming to end abuse against children by 2030, Reference Lee and Kim4,Reference Clark, Coll-Seck, Banerjee, Peterson, Dalglish and Ameratunga72 a large proportion of the adult population is expected to have a history of childhood maltreatment following the COVID-19 pandemic. The direct effect on neuropsychiatric outcomes is indisputable, and therefore strengthening care-giving and economic support for high-risk families would prevent childhood maltreatment by creating a safe environment. Reference Malvaso, Pilkington, Montgomerie, Delfabbro and Lynch73 The consequences of childhood maltreatment can be mitigated through early detection and intervention. Screening for childhood maltreatment history in clinical settings may be an effective strategy to identify high-risk individuals for adverse neuropsychiatric outcomes. 74 Through this, timely provision of trauma-informed therapy and increased access to community trauma services can prevent further childhood maltreatment and mitigate negative consequences.
Addiction-related, cognitive and socioeconomic mediators may be interdependent. For example, without healthy coping strategies, successful prevention of addiction in the childhood maltreatment victim may lead them to engage in other unhealthy behaviours. Therefore, policies and interventions need to target the multidimensional consequences of childhood maltreatment. Simultaneous addiction prevention and provision of psychoeducation and skills to cope with trauma was proposed as an effective intervention. Reference Bailey, Newton, Perry, Grummitt, Baams and Barrett75 ‘Rise Above’ is an example of an evidence-based trauma-informed prevention programme. This is a comprehensive intervention curriculum for youths covering trauma, education on e-cigarette and drug use, psychosocio-emotional development and prosocial behaviour. Reference Shin76 In providing the necessary coping tools for childhood maltreatment victims, the detrimental impacts on mental health can be mitigated. Moreover, psychotherapies, including trauma-focused cognitive behavioural therapy, facilitate the development of cognitive coping strategies to improve self-regulation, while cognitive remediation focuses on strategies that improves cognitive functioning. Reference Masson, Bussières, East-Richard, R-Mercier and Cellard53,Reference Rhoades, Mitnick, Heyman, Slep, Del Vecchio, Wampler and McWey77 Furthermore, close coordination between educational, child welfare and care-giving sectors is required to enable access to education, and to foster a positive attitude towards education in childhood maltreatment victims. Reference Romano, Babchishin and Marquis57 Therefore, a holistic approach is required to reach the therapeutic potential of trauma-informed interventions.
Strengths and limitations
This study used mediation and epigenetic Mendelian randomisation to provide novel insight into the pathways between childhood maltreatment and neuropsychiatric outcomes. It overcomes the limitations of unmeasured confounding and reverse causation in traditional observational study designs to delineate key causal mediators, effectively prioritising areas for intervention.
Our findings should be taken in the context of several limitations. First, while the Mendelian randomisation framework generates causal evidence, one must satisfy the Mendelian randomisation assumptions. Nevertheless, sensitivity analyses and instrument strength were satisfactory for UVMR analyses, and MVMR analyses accounting for weak instruments showed unlikely biased results. Second, childhood maltreatment and some mediators were based on self-reported measurements. Incorrect group assignment or reporting bias in the original study may have biased our results. Third, childhood maltreatment was reported as a composite measure of childhood maltreatment subtypes. We could not evaluate the effects of childhood maltreatment subtypes on outcomes. Fourth, the direct and indirect effects of binary mediators may be inaccurate given the non-collapsibility of odds ratios. However, a recent study showed that log(odds ratio) could be used as an outcome unit to calculate indirect effects. Reference Carter, Sanderson, Hammerton, Richmond, Davey Smith and Heron16 Fifth, our analyses were limited by the type and availability of data. GWAS data-sets used in Mendelian randomisation analyses were limited to European ancestry, and caution is required when generalising these findings to other populations. As discussed, formal mediation epigenetic Mendelian randomisation analyses were not possible with the current data. Furthermore, mQTLs were also limited to regions within 1 Mb from the CpG site, resulting in very few instruments to proxy DNAm levels. Although this may have limited instrument strength, the specificity of genetic instruments enabled postulation of biological mechanisms. Sixth, the effect sizes presented in this study represent a lifetime average effect of exposures on outcomes. Thus, mediators with large effects acting within shorter time frames may have been overlooked in this study. Finally, the heritability of childhood maltreatment is moderate, and genetic instruments represent gene–environmental interactions such as transgenerational transmission and passive or evocative heritability. Reference Espinosa Dice, Lawn, Ratanatharathorn, Roberts, Denckla and Kim78 Because genetics are more stable and predictable than environmental factors, our findings are based on childhood maltreatment variance, as determined by the myriad of childhood maltreatment risk factors with which the variants are associated. Reference ter Kuile, Hübel, Cheesman, Coleman, Peel and Levey79 While the genetic and environmental causes of childhood maltreatment are not distinguished, our findings determine the consequences of childhood maltreatment regardless of the underlying cause. Such analyses reflect how childhood maltreatment is perpetrated, and present crucial evidence to guide interventions.
In conclusion, our results support the role of addiction-related, cognitive and socioeconomic traits in mediating the effect of childhood maltreatment on neuropsychiatric disorders. Responsive interventions in these areas may be more relevant for ADHD and schizophrenia, but not for MDD. These include addiction prevention and treatment, improving cognitive functioning and supporting socioeconomic attainment. Ultimately, preventative strategies are vital to reducing childhood maltreatment prevalence and its long-term impacts. Reference Ng80
Supplementary material
The supplementary material is available online at https://doi.org/10.1192/bjp.2025.10433
Data availability
All genetic instruments can be obtained from summary statistics from the respective genome-wide association study (GWAS). Outcome estimates were obtained using publicly accessible data made accessible by the GWAS consortia.
Acknowledgements
The authors thank the researchers associated with each GWAS for providing publicly available summary statistics used for the analyses in this study. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975 as revised in 2013. All procedures involving human participants can be found in the original studies. This manuscript includes material adapted from a Master of Philosophy thesis submitted to the University of Hong Kong in 2024. Necessary permissions for reuse have been obtained.
Author contributions
N.Y.T.N. and B.H.Y.C. conceived the study. N.Y.T.N., J.V.Z., C.C.Y.M. and B.H.Y.C. contributed to the study design. N.Y.T.N. and J.Y.T. conducted the main analysis and drafted the manuscript. J.Y.T. conducted sensitivity analyses and generated the figures. All authors were involved in the interpretation of results, helped refine the manuscript and approved its final version. B.H.Y.C. is the guarantor and attests that all authors meet authorship criteria.
Funding
This work was supported by Seed Funding for Basic Research (no. 2202100680) from the University of Hong Kong.
Declaration of interest
None.
Transparency declaration
N.Y.T.N. and B.H.Y.C. affirm that the manuscript is an honest, accurate and transparent account of the reported study; that no important aspects of the study has been omitted; and that any discrepancies from the study as planned have been explained.
Analytic code availability
Code used to carry out statistical analyses can be found in the respective packages and are available upon reasonable request.
Open access
eLetters
No eLetters have been published for this article.