Participants

ALSPAC is a prospective, population-based birth cohort study that recruited 14,541 pregnant women resident in Avon, UK, with expected delivery dates between 1st April 1991 and 31st December 1992 [22, 23]. When the oldest children were approximately 7 years of age, an attempt was made to bolster the initial sample with eligible cases who had failed to join the study originally. The total sample size for analyses using any data collected after the age of 7 is therefore 15,454 pregnancies, resulting in 15,589 foetuses. Of these, 14,901 were alive at 1 year of age. The mothers, their partners, and the child have been followed up using clinics, questionnaires, and links to routine data. The study website contains details of all the data that are available through a fully searchable data dictionary: http://www.bristol.ac.uk/alspac/researchers/our-data/. Ethical approval for this study was obtained from the ALSPAC Law and Ethics Committee and the Local Research Ethics Committees, and participants provided written informed consent. This study uses data from pregnancy through to a questionnaire administered to the participants at age 22 years.

We excluded children with data on fewer than 10% of the ACE questions across all time points (n = 2,604). This threshold was used to ensure sufficient data to inform multiple imputation whilst minimising the number of people excluded from analyses because of missing data, particularly because missingness was related to both exposure to ACEs and experience of the outcomes of interest. After this exclusion, 11,935 participants remained. We created separate analysis samples for analysis of educational attainment and health outcomes by restricting these to participants with at least one outcome measure; we also excluded one child from within each twin pair at random (n = 152) to maintain independence of observations. This resulted in final sample sizes of 9,959 for educational outcomes (obtained through linkage to routine data) and 4,917 for health outcomes (assessed at a research clinic).

ACEs

Data on multiple forms of ACEs were reported by both participants themselves and their mothers at multiple time points, primarily through questionnaires. We used data from 582 questions relevant to one or more ACEs. The majority of the early-life ACE data (0–8 years) are parent reported, but when the children were 8 years old, they began self-reporting ACEs. Additionally, in their twenties, the participants retrospectively reported on child maltreatment (several forms of abuse and neglect) as well as whether their parents were violent towards each other. The majority of ACE data were collected prospectively, but we also included retrospective self-report measures because these complement the prospective data. For instance, the sexual abuse rates prospectively reported by parents were much lower than those retrospectively self-reported by the participants. Full details of the derivation of ACE measures, including the exact phrasing of questions, definitions, and ages at which data were collected, have been described previously [12]. Briefly, dichotomous constructs indicating exposure to adversities between birth and 16 years were created for the 10 ACEs that are included in the World Health Organization ACE international questionnaire [24]:

1. ever sexually abused or forced to perform sexual acts or touch someone in a sexual way (sexual abuse)
2. adult in family was ever physically cruel towards or hurt the child (physical abuse)
3. parent was ever emotionally cruel towards the child or often said hurtful/insulting things to the child (emotional abuse)
4. child always felt excluded, misunderstood, or never important to family, parents never asked or never listened when child talked about their free time (emotional neglect)
5. parent was a daily cannabis or any hard drug user or had an alcohol problem (parental substance abuse)
6. parent was ever diagnosed with schizophrenia or hospitalised for a psychiatric problem or, during the first 18 years of the child’s life, parent had an eating disorder (bulimia or anorexia), used medication for depression or anxiety, attempted suicide, or scored above previously established cutoffs for depression (Edinburgh Postnatal Depression Scale (EPDS) >12–13) (parental mental illness or suicide)
7. parents were ever affected by physically cruel behaviour by partner or ever violent towards each other, including hitting, choking, strangling, beating, and shoving (violence between parents)
8. parents separated or divorced (parental separation)
9. child was a victim of bullying on a weekly basis (bullying)
10. parent was convicted of a crime (parental criminal conviction)

Multiple questions feed into each ACE using predetermined criteria; an ACE is assumed to have been experienced if it is reported in one or more of the questions, even if there are inconsistencies between the questions. The definitions are described in S1 Table and S1 Text. S2 Table contains a list of the questionnaires used.

Educational attainment

ALSPAC data were linked to the National Pupil Database (NPD). This is a governmental database providing data on pupil level attainment in state-funded schools in England. General Certificate of Secondary Education (GCSE) examinations are sat during the 11th year of compulsory schooling when children are aged 15–16 (years 2007–2009 for the ALSPAC cohort). Pupils study up to 12 subjects (8 on average). The subjects are graded individually on a scale of A* (highest) to G (lowest). For this analysis, we used a dichotomous indicator of fewer than 5 GCSEs at grades A*–C including English and Mathematics, which is a widely used benchmark of academic achievement in the UK and a requirement for entry into many further education courses. As an additional outcome, following a suggestion from a reviewer, we used a continuous score of ‘GCSE points’; this is the total score of an individual’s top 8 qualifications ranked in terms of points. The capped score is more suitable than the total score because the number of GCSE qualifications taken by a pupil varies, but the majority take at least 8.

Health and health-related behaviours at age 17

During a research clinic at age 17 years (mean 17 years and 9 months, SD 4 months), weight was measured to the nearest 50 g using the Tanita Body Fat Analyzer (Model TBF 401A; Tanita Corporation, Amsterdam, the Netherlands), with participants in underwear or light clothing and footwear removed. Height was measured using a Harpenden stadiometer to the last complete mm, with participants unshod. Body mass index (BMI) was calculated as weight in kilograms divided by height in metres squared. Substance use and mental health were assessed using self-administered computer-assisted interviews. We derived the following dichotomous indicators:

• Obesity: BMI was converted into sex- and age-specific Z-scores relative to UK 1990 population reference data. These Z-scores were used to define obesity based on published BMI Z-score cutoffs from the International Obesity Task Force (BMI Z ⩾ 2.212 for boys and BMI Z ⩾ 2.195 for girls).
• Regular smoking: we created a dichotomous indicator of smoking weekly or more versus no smoking or smoking less than weekly.
• Harmful drinking: ⩾16 on the 10-item alcohol use disorders identification test (AUDIT) [25].
• Depression: based on the clinical interview schedule—revised (CIS-R), defined as meeting the depression diagnosis criteria of the international classification of diseases, 10th revision [26].
• Illicit use of drugs: problematic cannabis use or, in the past 12 months, any use of any of the following substances: cocaine, amphetamines, inhalants, sedatives, hallucinogens, or opioids. Problematic cannabis use was measured using the six-item cannabis abuse screen test [27], which assesses cannabis consumption in the previous 12 months and focuses on difficulties controlling use and associated health and social impairment. All items are answered on a 5-point scale (0 never, 1 rarely, 2 from time to time, 3 fairly often, and 4 very often). A response of fairly often or very often to any of the 6 items was used to indicate problem cannabis use.

Additionally, following a suggestion from a reviewer, we used BMI as a continuous variable.

Confounders

At enrolment and prior to delivery, several self-report questionnaires were administered that measured socioeconomic, family, and (mental) health variables. Based on these parental questionnaires, the following covariables were included in the analysis: mother’s home ownership status during pregnancy (mortgaged/owned/council rented/furnished private rental/unfurnished private rental/housing authority rented/other), mother and partner’s highest educational qualification (CSE/vocational/O-level/A-level/degree), household social class (highest of mother and partner social class according to the Registrar General’s Social Classes: professional/managerial and technical/skilled nonmanual/partly skilled/unskilled), parity, maternal report of child’s ethnicity (white/nonwhite), mother’s age at delivery (in years), mother’s marital status during pregnancy (never married/widowed/divorced/separated/first marriage/marriage 2 or 3), mother’s depression score (EPDS) at 18 and 32 weeks gestation, and mother’s partner depression score (EPDS) at 18 weeks gestation. The variables were chosen by searching the ALSPAC variable catalogue for variables related to key concepts determined a priori. The first and last author reviewed the list and decided on a set of variables we felt captured all relevant concepts, using what we felt to be the best measure of each based on the level of missing data and likely degree of measurement error. More details on these variables and their distributions are available in S3 Table.

Missing data

Because of the derivation of ACE measures from multiple questionnaires and clinics over a long time period (birth–23 years), no participants had data on all of the individual questionnaire items, necessitating the use of multivariate multiple imputation. Ideally, we would impute missing values of each questionnaire item, but the lack of complete cases in combination with the high number of variables (>500 separate questions relating to ACEs) led to convergence errors. Therefore, we adopted a pragmatic approach to imputation, adapted from the scale-level imputation method proposed by Enders [28]. We derived a dichotomous construct indicating presence or absence of each ACE. If a participant responded to 50% or more of the questions related to a given ACE, we used these data to create the dichotomous indicator. If the participant responded to less than 50% of the questions, we set the dichotomous indicator to missing. We derived a cumulative adversity measure (ACE score) by summing exposure to the 10 classic ACEs, defining 4 categories (0, 1, 2–3, and more than 4 ACEs). Because of the considerably larger sample size available for educational attainment than for health measures, these groups of outcomes were considered in separate imputation models. Because there are some sex differences in ACE prevalence (S4 Table) and potentially higher-order interactions between sex and adversity that we want to preserve, males (education n = 5,023, health n = 2,163) and females (education n = 4,936, health n = 2,754) were imputed separately before appending the two data sets before analysis. The dichotomous ACE indicators and the ACE score were included in multiple imputation models, along with all outcome variables, all covariates detailed above, and additional auxiliary variables likely to predict missingness, ACE exposure, or health status (sociodemographic indicators, adversity measures from before the child’s birth, and additional education and health variables; additional details in S5 Table and full details in previous publication [12]). For both males and females, 90 imputed data sets were created using the mice package [29] in R version 3.3.1 with 30 iterations per data set. For secondary analyses exploring interactions between ACEs and parental social class or maternal education, imputation models were rerun stratified by dichotomous indictors of i) parental social class (manual versus nonmanual, highest social class of mother or partner) and ii) maternal education (CSE, vocational education, or lower versus O-level, A-level, degree, or higher).

Statistical analyses

All statistical modelling was done in R version 3.3.1 unless otherwise specified, using binary logistic regression models for all binary outcomes and linear regression for the continuous outcomes of GCSE point score and BMI. Associations of each separate ACE and an ACE score (categorised to give comparability with previous studies) with each outcome were assessed in an unadjusted model and a basic model (adjusted for sex), as well as a fully adjusted model (adjusted for home ownership, maternal and partner education, household social class, parity, ethnicity, maternal age, maternal marital status, and maternal and partner depression during pregnancy). The individual ACEs are analysed separately in order to provide detail on whether certain ACEs are driving any associations of the ACE score or whether the patterns of associations differ across types of ACEs. For the imputed data, the logistic regression results and numerators and denominators for descriptive analyses were obtained by averaging across the results from each of the 90 imputed data sets using Rubin’s rules. This procedure appropriately modifies the standard errors for regression coefficients (used to calculate p-values and 95% confidence intervals [CIs]) to take account of uncertainty in both the imputations and the estimate. Likelihood ratio test statistics were combined using an approximation proposed by Meng and Rubin [30]. As a sensitivity analysis, we replicated these analyses in people with ‘complete’ data, i.e., participants who responded to more than 50% of the questionnaire items for all ACEs and who had data on the outcomes.

To examine whether the associations differed according to sex, we used likelihood ratio tests for interaction and, if applicable, report the results of sex-stratified analyses. We also used likelihood ratio tests to assess interactions between ACEs and manual versus nonmanual parental social class and low versus high maternal education.

Prevalence of the outcomes across exposure categories and risk differences (RDs) and ratios were estimated in the imputed data using the ‘mim: glm’ command in Stata version 15. PAFs were estimated using the formula , where Ppop is the proportion of exposed participants and RR is the risk ratio. The PAF estimates the percentage of cases of the outcome that would be prevented if the exposure was eliminated (assuming causality and absence of bias). Alternatively, it could be conceptualised as the percentage of people who go on to develop the adverse outcome that would be included in interventions targeted at the risk factor of interest. This analysis was performed for each of our binary outcomes for the following exposures, which were selected to represent a range of potential ways of identifying high-risk groups, most of which have a prevalence broadly similar to 4 or more ACEs (prevalences reported are in the imputed data used for analysis of educational attainment): 4 or more ACEs (19%), low maternal education (CSE, vocational qualifications, or lower; 30%), manual social class (classes IIIm, IV, or V in the 1991 UK Office of Population Censuses and Surveys classification; 23%), maternal depression in pregnancy (score of 12 or more on the EPDS on either of the two pregnancy questionnaires; 21%), any self-reported maternal smoking during pregnancy (26%), social housing during pregnancy (15%), and maternal age less than 20 years (4%).

A brief project plan was submitted to the ALSPAC study executive prior to starting analyses (see S2 Text). The following changes were made after this initial plan was written: i) our initial proposal focused only on educational attainment as an outcome, but we decided to broaden the focus to include health outcomes; ii) we initially proposed to include only socioeconomic factors as potential confounders but subsequently realised that other family-level variables could also be important confounders; and iii) after discussing preliminary results amongst co-authors, we decided to include calculation of PAFs. Following reviewer suggestions, we also added analysis of BMI and educational attainment as continuous variables and described the associations between health outcomes. As an additional sensitivity analysis suggested by a reviewer, we repeated the main analysis recoding maternal age as a categorical variable (less than 20 years, 20–34 years, or over 35 years). A STROBE (STrengthening the Reporting of OBservational studies in Epidemiology) checklist is available in S1 Checklist.

Association between ACEs and educational attainment

Fig 1 and S7 Table show the results for the associations of the ACE score and each individual ACE with educational attainment (3 models: unadjusted, adjusted for sex, and adjusted for confounders). Higher ACE score was associated with lower educational attainment. Associations were apparent even when comparing those experiencing only one ACE to those with no ACEs (for example, odds ratio [OR] for fewer than 5 GCSEs at grade C or above: 1.37, 95% CI: 1.15–1.63, p > 0.001 after adjustment for confounders) and were stronger for each increasing value of the ACE score. Experiencing 4 or more ACEs was associated with double the odds of obtaining fewer than 5 GCSEs at grade C or above (OR 2.00 [95% CI 1.65–2.43, p < 0.001]) after adjustment for confounders (S7 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Forest plot for the associations of the ACE score and separate ACEs with obtaining fewer than 5 GCSEs at A*–C, including Maths and English.

The reference category for each category of the ACE score (1, 2–3, and 4+) is experiencing 0 ACEs. The basic model is adjusted for sex. The adjusted model additionally includes home ownership, maternal and partner education, household social class, parity, ethnicity, maternal age, maternal marital status, and maternal and partner depression during pregnancy. GCSEs are qualifications obtained at age 16, which was the end of compulsory education for this cohort. ACE, adverse childhood experience; CI, confidence interval; GCSE, General Certificate of Secondary Education.

https://doi.org/10.1371/journal.pmed.1003031.g001

Most of the individual ACEs were associated with lower educational attainment. The strongest associations were seen for emotional neglect, for example, OR for fewer than 5 GCSEs at grade C or above 1.90 (95% CI 1.64–2.20, p < 0.001). Associations were weak or absent for physical abuse. For all ACEs and all measures of educational attainment, associations were considerably weaker after adjustment for confounders. For bullying, there was evidence of a sex interaction, with the associations with fewer than 5 GCSEs at grade C or above being stronger in females (1.78, 95% CI 1.48–2.15, p < 0.001) compared with males (1.25, 95% CI 1.05–1.50, p = 0.01); p-value for interaction 0.01 (S7 Table). Including maternal age as a categorical variable (less than 20 years, 20–34 years, or over 35 years) instead of a continuous variable, as requested by a reviewer, did not materially change the pattern of results, with ORs changing between −2.8% and 5.7%; full details are in S8 Table.

There was no consistent evidence of an interaction between ACEs and parental social class or maternal education (S9 Table and S10 Table; S1 Fig and S2 Fig) in their relationship with educational outcomes.

Association between ACEs and health/health-related behaviours

The ACE score was associated with depression, illicit drug use, and smoking (Fig 2, S7 Table). The group of participants who experienced 1 ACE had higher odds of all of these outcomes compared to those who experienced no ACE, but the CIs included the null apart from for illicit drug use (OR after adjustment for confounders 1.4, 95% CI 1.0–2.0, p = 0.03). People who experienced 2–3 or 4 or more ACEs were more likely to be depressed, use illicit drugs, and smoke; for example, ORs for 4+ ACEs compared with no ACEs after adjustment for confounders were low educational attainment 2.0 (1.7–2.4, p < 0.001), depression 2.4 (1.6–3.8, p < 0.001), drug use 3.1 (2.1–4.4, p < 0.001), and smoking 2.3 (1.7–3.1, p < 0.001) (S7 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Forest plots for the associations of the ACE score and separate ACEs with poor health outcomes.

The reference category for each category of the ACE score (1, 2–3, and 4+) is experiencing 0 ACEs. The basic model is adjusted for sex. The adjusted model additionally includes home ownership, maternal and partner education, household social class, parity, ethnicity, maternal age, maternal marital status, and maternal and partner depression during pregnancy. ACE, adverse childhood experience; CI, confidence interval.

https://doi.org/10.1371/journal.pmed.1003031.g002

There was only weak evidence of associations of the ACE score with obesity and harmful alcohol consumption; the OR for 4+ ACEs compared with no ACEs was 1.4 for obesity (95% CI 0.9–2.2, p = 0.13) and 1.4 for harmful alcohol consumption (95% CI 0.9–2.0, p = 0.10).

Examining individual ACEs revealed associations of i) physical abuse with depression, illicit drug use, and smoking; ii) sexual abuse with depression and smoking; iii) emotional abuse with illicit drug use and smoking; iv) bullying with depression; v) violence between parents with illicit drug use and smoking; vi) parental substance abuse with harmful alcohol use, illicit drug use, and smoking; vii) parental mental health problems with depression, illicit drug use, and smoking; and vii) parental separation with illicit drug use and smoking. Thus, overall, the patterns mirrored those of the ACE score—associations were observed for depression, illicit drug use, and smoking and were weak or absent for obesity and harmful alcohol use. The strongest associations tended to be seen for physical and sexual abuse, although very strong associations were seen for parental substance use in relation to substance use outcomes in the offspring.

In contrast with the educational outcomes, adjustment for sociodemographic confounders did not markedly attenuate the associations between ACEs and health/behavioural outcomes, and in some instances, adjustment resulted in stronger associations.

The only sex interaction was for exposure to parental substance abuse and illicit drug use (S7 Table), with stratified analyses indicating a stronger association in females than males (females, OR = 2.7, 95% CI 1.8–4.0; males, OR = 1.7, 95% CI 1.1–2.6; p-value for interaction 0.04).

Consistent with the educational outcomes, there was no evidence that associations between ACEs and health and behavioural outcomes differed according to parental social class or maternal education (S9 Table, S10 Table, S1 Fig and S2 Fig). Also consistent with the educational outcomes, including maternal age as a categorical variable (less than 20 years, 20–34 years, or over 35 years) instead of a continuous variable, as requested by a reviewer, did not materially change the pattern of results, with ORs changing between −4.3% and 8.5%; full details are in S8 Table.

PAFs

Differences in the risk of achieving fewer than 5 GCSEs at grade C or above including English and Maths (Table 2) ranged from 11% (95% CI 9%–14%) for maternal depression during pregnancy to 35% (95% CI 32%–37%) for social housing. All but one sociodemographic factor (maternal depression) had RDs that were higher than for 4+ ACEs (RD 18%, 95% CI 15%–20%). The lowest PAF was for maternal age less than 20 years (2%), reflecting the low prevalence of this risk factor. The highest PAF was for maternal education (19%). This compares to 9% for 4+ ACEs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Associations of ACEs and various sociodemographic markers with educational attainment on the RD scale and PAFs.

PAF is the proportion of the people experiencing <5 GCSEs (which are qualifications obtained at age 16, which was the end of compulsory education for this cohort) at grade C or above (including English and Maths) who also experienced the ‘exposure’ (ACE or sociodemographic variable). PAF can be interpreted as the proportion of the cases of <5 GCSEs at grade C or above that could be prevented if the exposure was eliminated, assuming causality. Note that the reference categories in this table differ from those in other parts of the manuscript; here, the reference category is all other participants apart from those with the exposure (for example, the reference category for 4+ ACEs here is <4 ACEs).

https://doi.org/10.1371/journal.pmed.1003031.t002

The highest obesity PAFs were seen for socioeconomic markers; for example, the PAFs for social housing and low maternal education were 14% and 13%, respectively, compared with 5% for 4+ ACEs (Table 3). In contrast, PAFs for illicit drug use and depression were highest for 4+ ACEs, with high PAFs also seen for maternal smoking in pregnancy and relatively low PAFs for socioeconomic and other family measures. For example, the PAFs of depression for 4+ ACEs, maternal smoking during pregnancy, and manual social class were 14%, 10%, and 3%, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Associations of ACEs and various sociodemographic markers with health and health risk behaviours on the RD scale and PAFs.

PAF is the proportion of the people experiencing each outcome (depression, illicit drug use, obesity, harmful alcohol use, or smoking) who also experienced the ‘exposure’ (ACEs or sociodemographic variable). PAF can be interpreted as the proportion of the outcome cases that could be prevented if the exposure was eliminated, assuming causality. Note that the reference categories in this table differ from those in other parts of the manuscript; here, the reference category is all other participants apart from those with the exposure (for example, the reference category for 4+ ACEs here is <4 ACEs).

https://doi.org/10.1371/journal.pmed.1003031.t003

Harmful alcohol use and smoking exhibited a different pattern again; the highest PAF for each of these outcomes was for maternal smoking during pregnancy: 11% for harmful alcohol use and 15% for smoking. PAFs for 4+ ACEs were 6% for harmful alcohol use and 12% for smoking. Socioeconomic markers had lower PAFs for harmful alcohol use, but PAFs for smoking were similar to the PAF for 4+ ACEs, i.e., PAF for manual social class in relation to smoking was 10%.