Cohort details
Data for this study were from the 1958 National Child Development study (NCDS, referred to as the 1958 birth cohort hereafter), a birth cohort collecting data on all live births taking place in a single week in 1958 in England, Wales and Scotland34; and the 1970 British Cohort Study (BCS, the 1970 birth cohort hereafter), a birth cohort collecting data on all live births taking place in England, Scotland, Wales and Northern Ireland in a single week in 1970.35 Measurements on all variables in this study, including ages of measurement, are summarised in figure 1. A flow diagram showing the study sample was arrived at is displayed in figure 2.
Figure 1Measurements used in this study. To estimate associations of assault with mortality, participants were followed for mortality from 42 years of age(1958 cohort) and from 30 years of age (1970 birth cohort). ACEs, adverse childhood experiences; SEP, socioeconomic position.
Figure 2Flow of participants through the study
Measurements
Mortality
Data on mortality were taken from the NCDS Deaths Dataset, containing data on known deaths in the 1958 birth cohort occurring from 1958 to 2014, and the BCS Deaths Dataset, with data on known deaths in the 1970 birth cohort occurring 1970–2014. Each dataset was compiled using records maintained by organisations responsible for the two studies over their lifetimes: the National Birthday Trust Fund, the National Children’s Bureau, the Social Statistics Research Unit and the Centre for Longitudinal Studies. Sources for mortality data in both cohorts included death certificates and other information from the National Health Service Central Register, and from relatives and friends during survey activities and cohort maintenance work by telephone, letter and e-mail. The analysis of mortality used only deaths occurring from 2000 to 2014, because of the measurement point for assault. Data were shared by the UK Data Service in 2017, and main analyses were undertaken from 2018 to 2020.
Assault
We made use of identically worded questions measuring assault exposure in the two cohorts. In the 1958 birth cohort, assault data were collected in two consecutive waves: self-report information on assaults occurring since age 25 was collected at age 33, in 1991, and further self-report information on assault occurring since age 33 was then collected at age 42, in 2000. In the 1970 birth cohort, self-report information on assaults occurring since age 21 was collected at age 30, in 2000. As described below, sensitivity analyses assessed the impact of the longer overall reference period in 1958 birth cohort members, on our results.
In both birth cohorts, information on assault was gathered by asking participants if they had since the reference date, received medical attention for a physical or sexual assault. Relevant responses to this item were used to generate a binary variable for any (physical or sexual) assault within the reference period, and for physical and sexual assaults separately. Data on the total number of physical or sexual assaults experienced in the reference period were also collected. Self-report information was used to generate a three-level variable for highest level of medical treatment received for assault, classified into no assault, medical treatment but not overnight and overnight medical treatment.
Measurement of mediators
To assess psychological distress as a potential mediator of the association between assault and mortality, we included similar data on psychological distress based on measurements collected on the Malaise Inventory at age 42 in 2000 (1958 birth cohort) and 34 in 2004 (1970 birth cohort). To measure alcohol, units of alcohol consumed per week were taken at age 42 in the 1958 birth cohort and age 34 in the 1970 birth cohort, using conventional formulae for converting self-reported weekly consumed quantities of different beverages (beer, wine, shandy, sherry, spirits and alcopops (1970 birth cohort only)) to units of alcohol consumption.36 To measure cigarette smoking, we used measurements of current numbers of cigarettes smoked at age 42 in the 1958 birth cohort and age 34 in the 1970 birth cohort. Measurement of covariates is described in the methodological supplement.
Analysis
Data were analysed by using Stata V.17.37 Individuals with missing data on assault were excluded. Counts and proportions of cohort members experiencing any assault, and the number of total assaults, were described by covariates for each birth cohort separately. To account for confounding, adjustment variables were selected for inclusion based on a review of the literature for each characteristic’s influence on both assault and mortality. A directed acyclic graph displaying pathways underlying the analysis is shown in online supplemental figure 1. Data on continuous variables (alcohol use, cognitive (verbal) ability) from both birth cohorts were z-standardised to have a mean of 0 and an SD of 1, and all were included as linear terms after assessment of difference in goodness of fit versus quadratic terms and indicator terms for each quintile (measured by the Akaike information criteria and the Bayes information criteria, see online supplemental table ST1). We used information on how many years ago each assault occurred to derive a variable for age of assault, describing this in both cohorts (online supplemental table ST2).
To test any association between assault and mortality, Cox regression models were estimated including 1 January 2000 as the start of the at-risk period, defining death as the time-to-event outcome and censoring defined by being unavailable for follow-up at any wave of data collection, or the conclusion of the time-at-risk. Graphical inspection of log/log plots (see online supplemental figure 2) and of Schoenfeld residuals were used to check departure from the assumption of proportionality of hazards. Kaplan-Meier graphs for each cohort stratified by any assault are displayed in online supplemental figure 3 (1958 cohort) and 4 (1970 cohort). Models were estimated for any assault (ie, either physical or sexual assault), physical assault, sexual assault, a count variable for the number of assaults, and highest level of medical treatment for any assault (classified into no assault, medical treatment but not overnight and overnight treatment). All analyses were carried out on data combining the two birth cohorts, and on each birth cohort separately—we also tested for interaction by cohort year. To report the impact of adjustments for different confounders, we estimated the crude association (model I), models only adjusting for gender, adverse childhood experiences and birth year only (model II), then further adjusting for socioeconomic variables (marital status, class, educational attainment, model III) and then further adjusting for model III variables and prior psychological distress, verbal ability and alcohol use (model IV, the final model). For adjusted analyses combining both cohorts, we adjusted for adverse childhood experiences using the restricted definition of adverse childhood experiences (excluding neglect) to ensure uniformity of definitions between the cohorts. For analyses of 1958 birth cohort data, we estimated models with both definitions of adverse childhood experiences (including and excluding neglect). To assess the impact of differing reference periods for assault measurement on our results, combined cohorts analyses were estimated excluding either the early or the late reference period for 1958 birth cohort data (presented in online supplemental table ST3). Based on the final model for the combined cohorts, we tested for differences between men and women, by the presence of prior psychological distress, and by cohort year, using multiplicative interaction terms, setting p≤0.05 as the criterion for determining significant interactions.
To address the impact of missing data, we also estimated the same models based on 10 multiply imputed datasets, using the mi command in Stata, which deploys multiple imputation with chained equations. Imputation commands for each missing variable were specified based on their form in complete case models—imputation models contained: sex, marital status and prior psychological distress, imputed by logistic regression; adverse childhood experiences, social class and educational attainment, imputed by ordinal logistic regression and cognitive (verbal) ability and alcohol use imputed by linear regression. Directions of associations remained after accounting for missing data in imputations (based on the missing at-random assumption).
In addition to Cox models, we used the gformula package to explore mediation.38 According to a counterfactual framework, G formula analyses contrast-specific exposure (or treatment) scenarios—in this case, comparing a scenario where the entire population is assigned to no exposure (eg, no assault), with the scenario where the population is assigned to exposure (assault). To examine mediation of the association between assault and mortality by the candidate mediators (psychological distress, alcohol use and cigarette smoking), we estimated total controlled effects, direct effects (DE) and natural indirect effects (NIE) for any assault. The DE represented the effect of assault on mortality that was independent of each respective mediator. The NIE represented the proportion of mortality that could be explained by its association with changes in each respective mediator over time. To quantify the magnitude of mediation, we estimated the proportion of the association mediated (NIE/[DE+NIE]). G formula models further adjusted for general health status as a postbaseline mediator-outcome confounder (as shown in online supplemental figure 1). For missing data, the gformula package implements single stochastic imputation using chained equations. We included identical specifications of each variable in G formula models as in Cox models described above.
