Sample
The Avon Longitudinal Study of Parents and Children (ALSPAC) is a population-based, extensive prospective study of women and their children, investigating the effects of environment, genetic and other factors on child health and development [15]. All pregnant women living in the geographical area of Avon, UK, who were expected to deliver their baby between April 1, 1991, and December 31, 1992, were invited to take part in the study. Uptake was high and those enrolled represented approximately 85% of the eligible population. ALSPAC recruited 14,541 pregnant women; all women gave informed and written consent. The study website contains details of all the data that is available through a fully searchable data dictionary: http://www.bris.ac.uk/alspac/researchers/data-access/data-dictionary/.
Procedures and measures
Data collection was carried out in two phases between 2009 and 2012. Figure 1 presents the study participation flowchart.
Phase 1
A total of 9233 women who were still alive, enrolled in the study and participating in assessment waves, and were main carers for their ALSPAC child, were enrolled and sent a version of the Eating Disorders Diagnostic Schedule (EDDS) adapted to cover the whole lifespan [16]. Women were invited to complete the questionnaire either online or on paper. Screen positive and a similar percentage of screen negative (~10%) women were selected for interview, based on sample size calculations. Criteria for screening positive were based on a previous study and identified diagnostic cut-offs [16].
Phase 2
Women who screened positive and a subset of those who screened negative were interviewed using the ED section of the Structured Clinical Interview for DSM-IV-TR disorders (SCID-I) (with no skip rules) [17], supplemented with a version of the LIFE interview [18], adapted to EDs [19], aimed at investigating presence, frequency, and duration of ED behaviors (restriction, fasting, excessive exercise, binge eating, and purging), as well as body mass index (BMI) over the lifetime. Women were asked to anchor their responses using major life events, such as the birth of their study child, in order to increase accuracy of reporting and minimize reporting bias. Each ED behavior was recorded over the lifetime from its first occurrence to time of the interview. Diagnoses were obtained supplementing the information for the SCID-I with detailed information from the LIFE to obtain DSM-5 diagnoses for disorders (e.g., BN and BED) reflecting different frequency thresholds in DSM-IV and DSM-5. Questions about access to healthcare and treatment for ED were specifically devised for the purpose of this study; if women reported any ED behaviors or cognitions they were asked if they had ever sought and/or received treatment for these. If they replied yes they were then asked to describe what kind of treatment they had received (inpatient, outpatient, psychological treatment, medication or other); they were also asked if they had received treatment for other psychiatric disorders.
Training and quality control
Interviews were carried out by three trained interviewers (all psychologists). All interviewers practiced the interview amongst themselves and with colleagues, and conducted interviews under supervision prior to interviewing study subjects, including rating available interviews. Interviewers attended a monthly meeting with the first author, where interviews of symptomatic individuals were discussed. All diagnoses were reviewed and confirmed by the first author. A subset of interviews were recorded for inter-rater reliability purposes. Interviewers demonstrated excellent inter-rater reliability on the SCID with 100% agreement; the intraclass correlation coefficient for diagnosis was 1.00.
Diagnostic properties of the EDDS
The adapted EDDS had a sensitivity of 97.3% (95% CI, 94.9–98.8%) and specificity of 74.6% (71.1–77.8%), and positive and negative predictive values of 65.1% and 98.3%, respectively. False negatives were therefore rare, whilst false positives were more common.
ED diagnoses
Diagnoses of DSM-5 ED (AN, BN, BED, sub-threshold BN and BED, purging disorder (PD), and other specified feeding and eating disorder (OSFED)) were obtained using the SCID supplemented with behavioral data (including frequency and duration of each symptom) from the LIFE. ED diagnoses were derived as shown in Additional file 1: Table S1. Given the age of our sample and the diagnostic instruments used, we were unable to ascertain the prevalence of avoidant/restrictive food intake disorder, pica, or rumination.
Risk factors
Data on relevant predictors were obtained as part of routine ALSPAC data collections approximately 20 years prior to the current study.
Obtained during pregnancy (12, 18, 32 weeks gestation)
Childhood unhappiness
Ascertained using women’s rating of their happiness in childhood (up to 16 years). We derived a binary variable (very unhappy, quite unhappy, and not really happy vs. moderately happy, very happy).
Parental divorce or separation, adoption or being under health authority care, death of a carer
Assessed by asking women whether their parents had divorced or separated prior to their 18th birthday; whether they had been legally adopted or had been placed under local authority care (foster care and or group homes); and whether a parent or person who cared for them had died prior to their 17th birthday. These variables were retained as dichotomous (yes or no).
Early sexual abuse
Assessed using a questionnaire about early sexual experiences covering a range of sexual experiences (including noncontact exposure, fondling, oral sex, and sexual intercourse) involving boy/girlfriends, parents, other relatives, family friends, and strangers. Experiences involving physical sexual contact with an individual other than a boy/girlfriend prior to age 16 were defined as sexual abuse. A dichotomous variable (childhood sexual abuse vs. none) was generated.
Life events
Data on life events experienced up to age 17 were obtained from a questionnaire completed during pregnancy, containing a life-event inventory (indicating occurrence of the event), with five response categories for each event (indicating the extent to which the respondent was affected) based on Brown & Harris’s work [20, 21]. In order to take into account both the wide variation of life events severity and their impact ratings, life events were weighted to create a continuous score as previously reported [22].
Bonding with parents
Assessed using the Parental Bonding Instrument [23]. Two scores were derived, (1) parental over-protection (degree to which women felt that their own parents had been over-protective and failed to allow them to make their choices in childhood – higher scores indicate a more oppressive relationship) and (2) maternal care (measured the woman’s perception of the relationship she had with her own mother – higher scores indicate a warmer relationship). We categorized the latter according to the top and bottom quartile, with the interquartile scores being the referent, to determine whether a warm relationship (top quartile) would be protective and a poor relationship (bottom quartile) would be risk-conferring for EDs.
Locus of control (LOC)
Assessed with a shortened version of the Adult Nowicki-Strickland Internal/External Locus of Control Scale [24], measuring external (higher scores) versus internal LOC.
Interpersonal sensitivity
Measured using the Interpersonal Sensitivity Measure [25] a valid and reliable measure assessing sensitivity to interpersonal and social feedback, and interpersonal avoidance [26].
Fixed factor
General intelligence
Measured using the Wechsler Abbreviated Scale of Intelligence [27] 15–16 years after enrolment in the study on 2165 women. We used the total IQ score.
Covariates
Maternal age was obtained during Phase 1; women’s ethnicity and educational status were obtained combining data provided at various time-points between enrolment and child age 18 [28].
Statistical analyses
All analyses were carried out using STATA 13 [29].
Prevalence
Prevalence estimates were calculated allowing for the two-phase sampling procedure by using weights [30, 31]. A sampling weight was generated using information from Phase 1 and the Phase 2 sampling strategy as described by Dunn et al. [29]. The sampling weight indicates how many Phase 1 participants each participant in Phase 2 represents. The weighted prevalence estimates of ED diagnoses from diagnostic interviews in Phase 2 were weighted back to the sample that participated to Phase 1. The survey (svy) set of commands was used in STATA to obtain prevalence estimates and carry out regression analyses, as they allow for stratified sampling and provide robust estimation of 95% confidence intervals (CIs).
Risk factors analyses
Analyses were adjusted for the a priori confounders of maternal age, ethnicity and education. Given high diagnostic crossover [32], women were categorized into mutually exclusive diagnostic groups. Women who only met one diagnosis were assigned to that diagnostic group; for those who had more than one diagnosis over their lifetime a hierarchical approach was used: full diagnoses (AN, BN, BED) trumped OSFED subtypes, BED trumped BN, and BN trumped AN, in accordance to our and others’ previous studies [19, 33], and evidence that diagnostic crossover over the lifetime in ED and in this sample occurs most commonly from restrictive type disorders (anorexia nervosa – restrictive (AN-R)) to binge and/or binge-purge disorders [19, 32].
Sensitivity analyses were carried out by removing from the analytic sample women who met criteria for more than one ED and assessing differences in association estimates with hypothesized risk factors. Women with complete data on exposures and outcomes were included in these analyses. Missing data on covariates (1%) were imputed using multiple imputation by chained equation performed in STATA 13. Imputation models included all variables in the analyses and predictors of missingness. Results obtained after imputation were similar to those found using complete case analyses; therefore, we report results from multiple imputation.
All tests were two tailed and a P value of 0.05 was used as a cut-off for statistical significance.