Child maltreatment and cardiovascular disease: quantifying mediation pathways using UK Biobank

Between 2007 and 2010, the UK Biobank recruited 502,506 participants from the general population. Participants attended one of 22 assessment centres across England, Scotland, and Wales where they completed a self-administered, touch-screen questionnaire and a face-to-face interview, and trained staff took a series of measurements including height, weight, and blood pressure. Ethnicity, education level, sleep duration, television viewing time, smoking status, and alcohol intake were self-reported. Townsend area deprivation index was derived from the postcode of residence using aggregated data on unemployment, car and home ownership, and household overcrowding [13]. Blood pressure was measured by a trained nurse. Physical activity was self-reported using the validated International Physical Activity Questionnaire [14]. Grip strength was measured to the nearest 0.1 kg using a Jamar J00105 hydraulic hand dynamometer and the mean value from both hands used in the analyses. Height was measured to the nearest centimetre, using a Seca 202 stadiometer, and body weight to the nearest 0.1 kg, using a Tanita BC-418 body composition analyser. Body mass index (BMI) was calculated as weight/height², and the World Health Organization’s criteria were used to classify BMI into underweight (< 18.5 kg/m²), normal weight (18.5 to 24.9 kg/m²), overweight (25 to 29.9 kg/m²), and obese (≥ 30.0 kg/m²). Abdominal obesity was defined as waist-hip ratio > 0.85 for women and > 0.90 for men. Biomarkers were performed at a dedicated central laboratory between 2014 and 2017. In this study, we have selected high-density lipoprotein [HDL] cholesterol, low-density lipoprotein [LDL] cholesterol, glycated haemoglobin, cystatin C, and gamma-glutamyltransferase as potential mediators. HDL and LDL are cholesterols related to CVD and lifestyle factors such as smoking and obesity; glycated haemoglobin is a marker for diabetes and related to obesity; cystatin C is a marker of kidney function and related to diet, smoking, and body weight; gamma-glutamyltransferase is a marker of liver function and related to alcohol drinking. Details of these measures and assay performances can be found online in the UK Biobank showcase and protocol [15].

The participants were subsequently invited to complete a web-based questionnaire [16] on mental health-related items: child maltreatment using the Childhood Trauma Screener, depressive symptoms using the Patient Health Questionnaire 9-items (PHQ-9) [17], anxiety symptoms using the Generalised Anxiety Disorder 7-items (GAD-7) [18], and other mental health problems using the Composite International Diagnostic Interview–Short Form [19]. Approximately one third (31.31%, n = 157,348) of the participants completed the online follow-up. These participants were generally younger, more likely to be female and White, and had healthier lifestyles (Additional file 1: Table S1). Of these, 3714 (2.36%) who did not complete all of the child maltreatment items and 1594 (1.01%) who had missing sociodemographic data were excluded from the analyses. Therefore, this study comprised 152,040 participants.

Child maltreatment was assessed using the Childhood Trauma Screener (CTS) [20], a shortened version of the Childhood Trauma Questionnaire (CTQ) [21]. It consists of one 5-point Likert scale item for each of five types of child maltreatment: physical abuse, physical neglect, emotional abuse, emotional neglect, and sexual abuse, and has been validated against the CTQ with good overall (r = 0.88) and satisfactory type-specific correlations (r = 0.55–0.87) [22]. The CTQ is a widely used instrument for measuring child maltreatment and has been validated against the actual record of abuse and neglect [20, 23]. The threshold values on the Likert scale derived from the validation study [20] were used to define the presence or absence of each type of child maltreatment (Additional file 1: Table S2). In this study, the primary exposure variable was the number of types of child maltreatment (range 0 to 5) as it reflects the dimensions of maltreatment.

Outcomes were ascertained through individual-level record linkage of the UK Biobank cohort. Date and cause of death were obtained from death certificates held by the National Health Service Information Centre (England and Wales) and the National Health Service Central Register Scotland (Scotland). Date and cause of hospital admissions were obtained through record linkage to Health Episode Statistics (England and Wales) and Scottish Morbidity Records (Scotland). Detailed information about the linkage procedures can be found at http://​content.​digital.​nhs.​uk/​services. At the time of analysis, hospital admission data were available up to 31 March 2017. We defined incident CVD as death or hospital admission for ischaemic heart disease (including myocardial infarction), atrial fibrillation, heart failure, or stroke (ICD-10 [international classification of diseases, 10th revision] codes I20, I21, I25, I48, I50, I60, I61, I63, I64) after baseline assessment. These outcomes were chosen as they comprised the majority of the CVD events.

Cox proportional hazard models were used to analyse the association between child maltreatment and incident CVD, with the results reported as hazard ratios (HRs) and 95% confidence intervals (CIs). The outcome variable comprised both the event status and time-to-event. The models were adjusted for age at baseline assessment, sex, ethnicity, deprivation index, and education level as potential confounders. Subgroup analyses were conducted for sociodemographic factors: sex, age group (38–50, 51–60, and 61–72 years), education level (with vs. without university degree), and area-based deprivation index (≥ vs. < median). We selected these factors because age could affect the reporting of maltreatment, sex could affect the type of maltreatment, and socioeconomic status might affect the ability to cope with these maltreatment experiences.

We studied four groups of potential mediators: lifestyle (sleep duration [categorical variable]; any smoking and alcohol drinking > 14 units a week [binary variables]; physical activity and television viewing [continuous variables]), physical measurements (BMI categories [categorical variable], abdominal obesity [binary variable], and systolic blood pressure and grip strength [continuous variables]), mental health (diagnosed depression, anxiety, and schizophrenia [binary variables]; depressive and anxiety symptoms [continuous variables]; any psychotic experience, behavioural addiction, drug addiction, or self-harm behaviours [binary variables]), and biomarkers (HDL cholesterol, LDL cholesterol, glycated haemoglobin, cystatin C, and gamma-glutamyltransferase [continuous variables]). Sleep duration and alcohol drinking variables were categorised based on the UK recommendations because of their potential nonlinear association with CVD. All potential mediators were selected because of their association with child maltreatment and/or CVD. These groups of mediators were adjusted in the Cox models to examine whether, and to what extent, the HRs between child maltreatment and CVD were attenuated. In addition, the model-based mediation analysis framework by Tingley et al. was also conducted [24]. Firstly, CVD incidence was regressed by the primary exposure variable (number of child maltreatment), all potential mediators, and sociodemographic factors in a Weibull regression model with a robust standard error. Weibull regression was chosen for mediation analysis because of its superior statistical properties over proportional hazard models [25]. Only the potential mediators reaching statistical significance (α = 0.05) were further investigated. The filtered potential mediators were then regressed by child maltreatment and other covariates (mediator model) in either logistic (for binary mediators) or multiple linear (for other mediators) models. The results of the outcome and mediator models were then combined to estimate the proportion of mediation. Potential mediators were mutually adjusted to avoid overestimation of mediating effects. Quasi-Bayesian estimation with 1000 iterations was used for estimating the P values of the mediating effect.

Four sensitivity analyses were conducted. Firstly, the standardised CTS score was used in a penalised regression spline to examine if its relationship with CVD was consistent with that based on the categorised maltreatment variables. Secondly, the association between maltreatment and CVD by different adjustment models was re-analysed after excluding participants who had prevalent CVD at baseline assessment (n = 6684). Thirdly, in the mediation analysis, because biomarkers can be dependent on sex, the mediation analysis was replicated with sex-specific z-scores of biomarkers. Lastly, because biomarkers can lie on the pathways between lifestyle and CVD, the mediation analysis was replicated without them. This helped to identify lifestyle factors that might be masked by adjustment of biomarkers. Proportional hazard assumptions were verified by statistical tests based on Schoenfeld residuals. Distributional assumption of Weibull regression was examined using the Kaplan-Meier estimate of the residuals. Missing data was handled using complete case analysis. All analyses were conducted using R version 3.5.3 with packages survival and mediation.

Previous prospective cohort studies of the association between child maltreatment, or ACEs more broadly, and CVD have been obliged to use intermediate cardiometabolic risk markers, rather than an incident disease, as the endpoints due to insufficiently long follow-up [11]. A number of retrospective cohort studies have shown an association with CVD but have relied on self-reported endpoints [6]. Our findings are consistent with a retrospective cohort study which reported that traditional risk factors (smoking, physical inactivity, BMI, diabetes, and hypertension), as a group, attenuated the association between ACEs and self-reported IHD by 19% and psychological factors (anger and depressed affect) by 38% [6], and the Nurses’ Health Study 2 which found adult health-related factors (BMI, smoking, alcohol use, depression, etc.) collectively accounted 79% and 63% of the association of severe physical and sexual abuse, respectively [12]. This cohort study extends the existing evidence by analysing a wide range of cardiovascular disease endpoints, including heart failure, an increasingly important public health problem [26].

Reviews [15, 27] have suggested that ACEs may increase the risk of CVD through behavioural, emotional, biological, social, and cognitive pathways. Two previous studies have shown that the association between ACEs and CVD was attenuated following adjustment for clusters of behavioural, psychological, and medical factors, but did not study the contribution of individual factors within these clusters [6]. Our study looked at these factors individually and demonstrated that the association was mediated primarily through depressive symptoms. Our findings also suggested that, after adjusting for current symptoms of depression, historical diagnosis of depression was no longer associated with CVD (Additional file 1: Table S4). This could be because the symptom score can capture the severity of depression, or because diagnosed depression is often treated with selective serotonin reuptake inhibitors, a group of medications that was shown to reduce CVD risk in a randomised controlled trial [4].

Because the majority of assessments of the UK Biobank cohort were conducted at baseline, it was not feasible to conduct robust sequential mediation analysis. However, mediators can be causally related to each other. For example, lifestyle/behavioural factors, such as smoking and obesity, may affect serum CVD risk markers, which in turn increases the risk of CVD [28, 29]. We attempted to elucidate additional mediators that were masked by the adjustment of biomarkers through a sensitivity analysis. This revealed obesity to be a mediator of the association between child maltreatment and CVD once HDL cholesterol was removed from the model. Given the close relationship between obesity and HDL cholesterol [30], we hypothesise that children who have been maltreated are at higher risk of becoming obese which, in turn, lowers the HDL cholesterol level and increases the risk of CVD. However, we should note that child maltreatment might be a causative factor in the development of obesity or could be acting through other confounding factors, such as ADHD. ADHD was found to predispose children to being maltreated [31] and to obesity and premature mortality. Notably, we hypothesise that HDL cholesterol may not necessarily be a causal risk factor but more of a risk marker, on the basis that recent trials and genetic studies go against HDL cholesterol being causally linked to CVD per se [32].

As with any observational study, residual confounding, such as genetic traits [33] and ADHD [34], may have occurred. Child maltreatment was recalled by participants rather than captured prospectively, which is a common limitation of studies of childhood experiences with long duration of follow-up. Another common limitation was that information on child maltreatment was limited to type, with no information on severity, even though we identified similar findings when we used the child maltreatment score as the exposure variable. Mediation analysis assumes causality, but it is possible that some mediators, such as depression, influence whether maltreatment is reported since the report of child maltreatment and the assessment of the mediators occurred at the same time [35]. Even though the mediation analysis has considered a wide range of factors, there are still important mediators unexamined, such as post-traumatic stress disorder (PTSD). PTSD was found to have significant overlap with other psychopathologies that we included (such as depression) [36], and therefore, omitting it should not significantly alter the conclusion in this study. We should also note that different types of child maltreatment (for example abuse vs. neglect) may influence health and well-being in distinct ways [37]. Lastly, whilst the UK Biobank cohort is representative of the general population in terms of sociodemographic characteristics, it is not representative in terms of lifestyle [38, 39]. Therefore, whilst effect sizes should be generalisable, summary statistics and estimates of absolute risk should not be generalised.