How does childhood maltreatment influence cardiovascular disease? A sequential causal mediation analysis

Abstract Background Childhood maltreatment has been consistently associated with cardiovascular disease (CVD). However, the mechanisms of this relationship are not yet fully understood. We explored the relative contribution of anxiety/depression, smoking, body mass index (BMI) and inflammation (C-reactive protein, CRP) to the association between childhood maltreatment and CVD in men and women aged 40–69 years in the UK. Methods We used data from 40 596 men and 59 511 women from UK Biobank. To estimate the indirect effects of childhood maltreatment (physical, sexual and emotional abuse, and emotional and physical neglect) on incident CVD via each of the mediators, we applied a sequential mediation approach. Results All forms of maltreatment were associated with increased CVD risk [hazard ratios (HRs) ranging from 1.09 to 1.27]. Together, anxiety/depression, smoking, BMI and inflammation (indexed by CRP) mediated 26–90% of the association between childhood maltreatment and CVD, and the contribution of these mediators differed by type of maltreatment and sex. Anxiety/depression mediated the largest proportion of the association of sexual abuse, emotional abuse and emotional neglect with CVD (accounting for 16–43% of the total effect), especially in women. In men, BMI contributed the most to the indirect effect of associations of physical abuse and physical neglect with CVD; in women, anxiety/depression and BMI had similar contributions. Conclusions These findings add to the understanding of how childhood maltreatment affects CVD risk and identify modifiable mediating factors that could potentially reduce the burden of CVD in people exposed to maltreatment in early life.

considered as ever smoking (current tobacco smoking occasionally or most/all days, or past tobacco smoking most days, occasionally, or tried once or twice). BMI (kg/m 2 ) was calculated using weight and height measured at recruitment. Serum CRP (mg/L) was measured by immunoturbidimetric -high sensitivity analysis on a Beckman Coulter AU5800. Given the positive skewed distribution of CRP, it was log-transformed for analyses and back-transformed (exponentiated) for presentation of results.

Confounders
Information on age, year of birth, ethnicity, maternal smoking around birth, number of siblings, family history of CVD, Townsend deprivation index, and education were assessed at recruitment. Age (complete years) was calculated based on date of birth and date of attending the initial assessment centre. Year of birth was assessed in years and classified into <1950, 1950-1959 and 1960-1970. Ethnicity was classified as White, Mixed, Asian, Black, Chinese and Others. Maternal smoking around birth (yes/no) was assessed by the question: "Did your mother smoke regularly around the time when you were born?", and classified into yes, no, or do not know. Number of siblings was based on the number of participant's brothers and sisters and categorised into 0, 1, 2, 3, 4 or 5+. Family history of CVD was based on self-report of high blood pressure, stroke or heart disease of the mother, father or siblings. Townsend deprivation index was calculated based on preceding national census output areas, and each participant was assigned a score corresponding to the output area in which their residing postcode was located; the index was divided into quintiles. Educational attainment was assessed in 5 categories: College, university or other professional degree; Advance levels/ advance Subsidiary levels or equivalent; Ordinary levels/ General Certificate of Secondary Educations or equivalent; Certificate of Secondary Educations or equivalent/ National Vocational Qualifications/ Higher National Diploma, Higher National Certificate or equivalent/ other professional qualification; and None.

Sequential mediation analysis
To estimate direct and indirect effects, the following assumptions regarding confounding are needed: a) no unmeasured confounding for the exposure-outcome relationship, b) no unmeasured confounding for the mediator-outcome relationship, c) no unmeasured confounding for the exposure-mediator relationship, and d) no effects of the exposure that confounds the mediator-outcome relationship (VanderWeele & Vansteelandt, 2014). With the approached used in our study (parametric g-computation using Monte Carlo simulations), there is therefore no need to make assumption d, which is needed when using standard methods for mediation (Daniel, 2011). We considered the natural direct effect (NDE) and natural indirect effect (NIE) as decompositions of the total causal effect (TCE); gformula implements logistic regression for time-to-event outcomes and therefore estimates correspond to log odds ratios (ORs). We exponentiated the results and present ORs.
To estimate the combined and individual contribution of the mediators, we performed four models using sequential mediation analysis. In Model 1, we estimated the NIE through anxiety/depression; this includes pathways that act through anxiety/depression and any of its effects (descendants) but does not include those that act solely through smoking, BMI and/or CRP (or other pathways). In Model 2, the NIE through both anxiety/depression and smoking was estimated, which includes their causal descendants, but does not include any other pathways, including paths that act only via BMI and/or CRP; the difference between the proportion mediated in Model 2 and the proportion mediated in Model 1 corresponds to the proportion mediated by smoking beyond anxiety/depression alone (i.e. the additional contribution of smoking). In Model 3, the NIE through anxiety/depression, smoking and BMI was estimated, which includes their causal descendants, but does not include the paths that act solely through CRP; the difference between the proportions mediated in Model 3 and Model 2 corresponds to the mediating effect of BMI beyond the effect of both anxiety/depression and smoking. Finally, in Model 4, the NIE of all potential mediators was estimated, and the difference between the proportion mediated in Model 4 and the proportion mediated in Model 3 corresponds to the mediating effect of CRP beyond anxiety/depression, smoking and BMI.
Exposure-mediator interactions were included sequentially (i.e. interaction between maltreatment and anxiety/depression was included in Model 1, interactions between maltreatment and anxiety/depression and between maltreatment and smoking were included in Model 2, and so on). We did not include mediator-mediator interactions, as these would hinder comparability across models (VanderWeele & Vansteelandt, 2014