We conducted a case-control study nested within the Stockholm youth cohort, which comprises all young people aged 0 to 17 years, residing in Stockholm County between 2001 and 2007 (n=589
The cohort contains prospectively recorded data on the probands and their first degree relatives collected by record linkage with national and regional healthcare, social, and administrative registries using unique national identity numbers assigned to all Swedish residents.2
Sweden has a well developed system of publicly funded screening, diagnostic, and follow-up services relevant to autism spectrum disorders, with national and regional registers recording information about diagnosis and other details.2
Assessments for autism spectrum disorder are typically carried out by child neuropaediatric or mental health services, and, as per local guidelines, include diagnostic evaluations covering the child’s social, medical, and developmental history after interviews with the parents, observation of the child, and a structured neuropsychiatric assessment including cognitive testing.2
We identified children with autism spectrum disorders in the Stockholm youth cohort using a multisource case ascertainment method, with registers covering all pathways of autism diagnosis and care within Stockholm County.2
Diagnoses recorded in these registers (codes from the international classification of diseases, ninth and 10th revisions, ICD-9 (299) and ICD-10 (F84), respectively, or Diagnostic and Statistical Manual of Mental Disorders
, fourth edition, (299)) were supplemented by a record of care in specialist centres for autism with and without intellectual disability, where an autism diagnosis and cognitive testing is a prerequisite. We also identified comorbid intellectual disability status using ICD-9 (317-319), ICD-10 (F70-79), and DSM-IV (317-319) in the child or adult mental health registers or the national patient register.2
As of 31 December 2007, over 5000 cases of autism spectrum disorder have been identified in the Stockholm youth cohort, almost 43% of whom have a comorbid intellectual disability.2
Two validation procedures—a case note validation study by a consultant child psychiatrist and a neuropaediatrician and a cross validation study with a national twin study—both found a high validity of the diagnoses for autism spectrum disorder recorded in the registers used for case ascertainment.2
Figure 1 shows the derivation of the sample for the present analyses. To ensure completeness of diagnostic data for parent and children in the registers we excluded from the study sample those with missing maternal identification numbers, adopted children, those living in Stockholm County for less than four years (thus also excluding all children aged 0-3 years who would be too young to have a reliable diagnosis). In the remaining population of the Stockholm youth cohort, we matched each case of autism spectrum disorder to 10 living controls without autism by date (month and year) of birth and sex (fig 1).
Fig 1 Derivation of analytical sample
Parental history of depression
We identified the psychiatric history of parents using two sources: the Stockholm County adult psychiatric outpatient register, which records the dates and diagnoses for any contact with specialist outpatient psychiatric services in Stockholm County since 1997,12
and the Swedish national patient register, which contains the dates and discharge diagnoses of all inpatients (since 1973) and specialist outpatients (since 2001, although with incomplete psychiatric outpatient data) in Sweden.13
Using these sources, we identified mothers and fathers with depression if they had a registered diagnosis of a depressive episode, recurrent depressive disorder, persistent mood disorder, and other or unspecified mood disorder (see supplementary table S1 for ICD codes). To avoid the possibility of reverse causality we considered only diagnoses recorded before the birth of the child participating in the study.
We used two approaches to handle the presence of more than one recorded diagnosis for a parent. In our primary strategy, we used a hierarchy based on ICD-10,14
adapted for a greater relevance to autism and our research question. This (from higher to lower priority) included schizophrenia or non-affective psychoses or bipolar disorder; neurodevelopmental disorders or personality disorders; alcohol and drug disorders; and depression, anxiety, and somatoform or other disorders (see supplementary table S1 for ICD codes). Depression was therefore coded conservatively, only higher than anxiety or somatoform disorders in case of multiple diagnoses. We grouped other diagnoses into anxiety disorders, psychotic disorders (including schizophrenia and bipolar disorder), and other non-psychotic disorders for use as potential confounders in analysis. In an alternative strategy we allowed participants to be included in the different diagnostic groups if more than one diagnosis had been recorded (and adjusted for these in our regression models).
Maternal antidepressant use during pregnancy
Since 1995 the Swedish medical birth register15
contains data on current drug use reported by mothers at their first antenatal interview (median 10 weeks’ gestation),16
coded using the World Health Organization’s ATC codes (www.whocc.no/atc_ddd_index/
). Thus for mothers of children born from 1995 onwards we retrieved data on any antidepressant use (ATC code N06A), further divided into the two most commonly used antidepressant classes—SSRIs (ATC code N06AB) and non-selective monoamine reuptake inhibitors (ATC code N06AA), which comprises tricyclic antidepressants (see supplementary table S2 for individual drugs in each group). We could not study other antidepressant categories since their use in pregnancy was rare. For the same reason we did not study individual drugs within any class. The medical birth register has been shown to identify 78% of all antidepressants prescribed during the first trimester,17
and the drug name registered in prescription records and that recorded in the register has been reported to show high concordance (97%).16
We used prospectively collected data on several parental characteristics as potential confounders: maternal age (<20, 20-24, 25-29, 30-34, 35-39, >40 years) and paternal age (<25, 25-29, 30-34, 35-39, 40-49, >50 years) at birth of child, fifths of family income adjusted for year of ascertainment and family size, highest education of either parent (≤9, 10-12, ≥13 years), highest occupational class of either parent (higher professionals, intermediate non-manual employees, lower non-manual employees, skilled manual workers, unskilled manual workers, self employed, or unclassified), maternal region of birth (Sweden, Europe, Americas, Africa, Asia, or Oceania), parity (0, 1, 2, ≥3 previous births). These characteristics were chosen because of their association with autism in the literature.7
We also considered variables with relatively less empirical evidence linking them to autism but which nevertheless may be confounders on theoretical grounds, including maternal smoking reported at the first antenatal interview (non-smoker, 1-9 or ≥10 cigarettes per day), and a diagnosis of maternal diabetes (yes or no) or hypertension (yes or no). We considered these in additional analyses since they had a greater proportion of missing data (18% for maternal smoking and 9% for diabetes and hypertension). We also considered birth weight for gestational age (normal for gestational age, small for gestational age, large for gestational age),19
gestational age at birth (≤32, 33-36, 37-42, ≥43 weeks),19
and Apgar score at five minutes (<7 or ≥7) 7
in separate analyses as these are potential mediators (for example, fetal growth)22
but adjusting for them could result in biased estimates.23
All analyses were conducted using Stata 10.1 for Windows. In descriptive analysis we calculated proportions of individuals with autism spectrum disorder (and autism with and without intellectual disability) with the exposure variables and other covariates. Using conditional logistic regression models, we derived odds ratios and their 95% confidence intervals as estimates of relative risks for the relation between a history of depression in each parent and autism spectrum disorder in offspring as a group, and dichotomised into autism with or without intellectual disability. After estimating crude associations, we adjusted for parental ages, parental income, education, occupation, region of birth, and parity (model 1). In model 2, we further adjusted model 1 for other psychiatric disorders in the parent, and in model 3 we adjusted model 2 for the presence of psychiatric disorders in the other parent. We conducted these analyses in a sample with all available data and those with complete data on all covariates (here we report the latter analysis to ensure consistent numbers across all the models; in the supplementary tables we present analyses using all available data). We used the χ2 test of heterogeneity to calculate P for heterogeneity values to assess the statistical significance of any differences between estimates for paternal and maternal depression and for autism with and without intellectual disability.
We conducted several secondary analyses for a better understanding of the implications and robustness of these findings. Since we did not have data on antidepressants for the full cohort, to assess whether any observed relations existed before the majority of SSRIs were licensed in Sweden and came into use during pregnancy, we restricted the sample to births before 1990 (children born between 1984 and 1989). We also repeated the analysis on children of primiparous women only, to avoid the possibility of observations being due to depression in mothers related to an older sibling with autism or other developmental concerns. Thirdly, we restricted the analysis to parents who were born in Sweden since we have previously observed noticeably different associations of autism with and without intellectual disability in relation to parental immigration to Sweden,9
and a low take-up of psychiatric services in migrant adults has been anecdotally reported. Fourthly, we repeated the main analyses using the alternative coding of psychiatric disorders, allowing for multiple psychiatric diagnoses in the parents. Finally, we repeated the analyses in a sample of children at least 8 years old (born between 1984 and 1999) since the diagnosis of autism spectrum disorder may be most reliable in older children.
To include the antidepressant data, we restricted the sample to births from 1995 onwards and repeated the analysis for maternal depression, adjusting for antidepressant use during pregnancy. We then estimated the associations using a categorical variable to denote mothers with no history of depression and no antidepressant use during pregnancy, a history of depression but no antidepressant use during pregnancy, a history of depression and antidepressant use during pregnancy, and antidepressant use reported during pregnancy but no recorded history of depression. Finally, we estimated the risk of autism spectrum disorder with antidepressant use during pregnancy, irrespective of indication, and repeated these analyses after classifying antidepressants into SSRIs and non-selective monoamine reuptake inhibitors. For these analyses we excluded a small number of mothers reporting multiple antidepressant use. We adjusted these analyses for parental ages, income, education, occupation, migration status, parity, and a variable depicting any psychiatric disorder in the mother. In a separate analysis we further adjusted for any other psychotropic drug use except antidepressants. We used the resulting adjusted odds ratios (OR) and prevalence of exposure to antidepressants in cases (PE) to estimate population attributable fractions (PAF) using the formula: [PAF=PE(OR−1)/OR]. The population attributable fractions denote the proportion of cases with autism that could be prevented if antidepressant use was completely eliminated from the population, assuming the association was causal and all confounders had been accounted for.