The Nord-Trøndelag Health Study (HUNT) is a large population-based study that invite participation of all inhabitants aged 13 and above in the county of Nord-Trøndelag . Nord-Trøndelag county (total population in 2009: 130 708) is situated in the middle part of Norway and is geographical, demographical and occupational fairly representative of Norway as a whole, but lacking large cities. Between 1995 and 1997, all adolescents attending middle and secondary school (originally ages 13-19, but some participants were 12 and 20 years) were invited to participate in the first survey of the adolescent part of HUNT, Young-HUNT1. Totally 8950 students (90%) completed a questionnaire during class hours. Participants were linked with biological parents through a national family register code in order to identify siblings, and information on parental education was accessed from the Norwegian National Education Database and from parental data from the HUNT 2 study (1995-97) . We linked individual data from the Young-HUNT1 survey with information from the social insurance database (kept by the National Insurance Administration and available in Statistics Norway's events database ). This database contains complete records of social insurance benefit reception and allowed us to follow all the cohort members in the period 1998-2007.
We excluded all participants who died before the end of follow-up (n = 46) and eight individuals with age-school mismatch. We also chose to exclude 101 individuals receiving a disability pension (DP) at age 18 or 19 and those who were already on sickness benefits in 1998 and later ended up with a DP. This group includes individuals with mental retardation, chromosomal abnormalities and extensive medical problems, for whom reading and writing difficulties are common, but of minor importance in relation to work ability. Of the remaining 8, 795 participants, we obtained information on reading and writing difficulties for 8, 498 who were included in the analyses.
Each student signed a written consent form to participate in the study and parents or guardians of students aged less than 16 also gave their written consent. The study was approved by the Regional Medicine Ethical Committee and the Norwegian Data Inspectorate.
Dependent variable - welfare dependence
We constructed two different measures of welfare depencence based on the type of benefits 1) medical benefits (comprising sickness benefit, medical or vocational rehabilitation and disability pension (DP) in the Norwegian social insurance scheme) and 2) all social benefits (adding unemployment benefit and social support). We only included long-term benefits as we wanted our outcome measure to reflect individuals at substantial risk of future work exclusion. We included the benefits which in nature are long-term (DP, medical or vocational rehabilitation), and other benefits received at least 180 days during one calendar year. We constructed a dichotomous variable of having received or not received benefits (medical benefits and all social benefits) each year during follow-up from the year participants turned 19. We also constructed a dichotomous variable of having received or not received benefits (medical benefits and all social benefits) in the 5-year period from age 24 to 28 for use in the regression analyses. The window of ages 24 to 28 was used in order to have sufficiently many cases of benefit receipt combined with maximum follow-up time.
Self-reported reading and writing difficulties
Subjects were classified as having reading and writing difficulties (RWD) if they answered yes to the question "Do you currently receive help for reading or writing problems?" or if they reported major problems with either reading or writing during the last 12 months (options were major problems, some problems and no problems for both reading and writing problems).
Information on age, gender, living situation, somatic health problems and mental health was collected from the questionnaire at baseline. Parental education at baseline was assessed using parental data on education from the Norwegian National Education Database, supplemented by self-reported educational level in HUNT 2. Living situation was categorized as living with both parents, living with one parent and new partner, living with one parent only, living with other adults, living alone or living with a partner.
In order to adjust for a broad range of somatic health indicators, we constructed a propensity score predicting reading and writing difficulties . The propensity score contained questions concerning disabilities (vision, hearing, and movement), diseases (epilepsy, migraine, diabetes, asthma, other disease lasting more than three months), use of health services (contact with medical specialist, hospital admission) and long-term school absence because of sickness. The propensity score was included in the analyses as a continuous variable, ranging from 0 to 1.
Somatic symptoms was included as a continuous scale score based on the self-reported presence during the last 12 months (never, seldom, sometimes or often) of eight different symptoms (headache, neck or shoulder pain, aching of muscles or joints, stomach pain, nausea, constipation, diarrhoea and palpitations) (Cronbach's alpha 0, 74). Anxiety and depression symptoms was measured with the validated 5-item Symptoms Check List (SCL-5) [26, 27]. Conduct and attention problems were assessed using variables from a school adjustment scale containing 14 school-related items, each with four alternative answers (never, sometimes, often and very often) . Six questions related to conduct and attention problems ("quarrels with the teacher", "get into fights", "get scolded by the teacher", "shirks school", "has difficulties concentrating in class" and "can not be quiet/calm in class") were summed up separately, rescaled in the range 0 to 1 and used in the analyses as a continuous variable (alpha 0, 67). Alcohol consumption was categorized as having ever been drunk more than 10 times, or not.
Parental education was measured as primary, secondary and tertiary education. Data were available for 8, 085 (95%) of the mothers and 7, 442 (88%) of the fathers. Maternal education was used in the multivariable analyses due to little missing data and 87% of the adolescents (92% at age 12 to 15) living with their mother. Siblings (having the same mother) in the study cohort were identified through the family register. In total, 3, 000 subjects had at least one sibling in the cohort.
The associations between RWD and benefit reception were explored in complete case data (N = 7, 817). Multivariable logistic regression analyses were performed with benefit reception in the 5-year period from age 24 to 28 as the outcome measure in two conceptual models. In model 1, we adjusted for the confounding of age, living situation, somatic health and parental education. In model 2, we adjusted for mental health issues (including somatic symptoms, anxiety and depression symptoms, conduct and attention problems and alcohol consumption) additionally, as these factors could represent both confounding and mediating factors. Reception of medical benefits and all social benefits was assessed separately.
Logistic regression analyses were used to estimate predicted 5-year risks and corresponding odds ratios (OR), all reported with 95% confidence intervals (CI). Predictions were made using the program predxcat , keeping covariates at their mean and setting follow-up time to 5 years. All analyses included RWD-status and gender interaction. Effect measure modification by school level and maternal education was explored separately by adding interaction terms in the analyses (between RWD-status and school level and RWD-status and maternal education). Longitudinal assessments using all observations in the follow-up period were conducted in population-averaged models, using generalized estimation equations (GEE) analyses . The development over time was explored by including an interaction term between RWD-status and time (years).
Sibling comparison was used mainly as a way of adjusting for family level covariates by comparing individuals with their own siblings (those having the same mother) instead of with all the other individuals in the cohort. We used multilevel mixed-effects logistic regression. Within-siblings comparisons were performed with sibling-mean centring--subtracting the siblings mean RWD from each individual's value on the RWD variable . All Analyses were conducted using STATA 11 software (StataCorp LP, Texas, USA).