A longitudinal twin study of the association between childhood autistic traits and psychotic experiences in adolescence

Participants

Participants were taking part in the Twins Early Development Study (TEDS), a community
sample of twins born in England and Wales between 1994 and 1996 24]. When twins were aged 16, 10,874 families were invited to take part in the Longitudinal
Experiences And Perceptions (LEAP) study, a study of psychotic experiences in the
general population. In total, 5128 parents returned LEAP data, while 5074 twin pairs
both returned questionnaires. Participating and nonparticipating families in LEAP
were similar on various demographic characteristics (shown in Additional file 1). Autistic traits were assessed four times: at ages 8, 12, 14, and 16. All participants
in LEAP who had autistic trait data at any prior age were included in the study. Exclusions
were conducted for genetic syndromes, chromosomal abnormalities, perinatal/postnatal
complications, and missing first contact or zygosity data. The final sample sizes
are shown in Table 1. Zygosity was determined through parent report of twin resemblance and DNA testing
25]. TEDS has full ethical approval from the King’s College London, Institute of Psychiatry
Ethics Committee. All participants provided informed consent before completing questionnaires.

Table 1. Interpretation of twin correlations

Measures

Autistic traits

Autistic traits were assessed at all four ages. At ages 8 and 12 years, parents completed
the Childhood Autism Spectrum Test (CAST 26]), a 31-item questionnaire (30 items were used at age 12 years due to the removal
of an age-inappropriate item). Each question was answered ‘yes’ or ‘no’; possible
ranges of scores are shown in Table 1. The measure was reliable (see Table 1) and stable between ages 8 and 12 years (r?=?.55, p??.001). Scores over 15 have 100 % sensitivity and 97 % specificity in detecting
individuals at risk of ASC, with a positive predictive value of 50 % 27]. The CAST was divided into three previously published subscales 28]: social difficulties, communication atypicalities, and repetitive behaviours and
interests.

The CAST is a measure for elementary school-aged children 26]; hence, parents completed the shortened Autism Spectrum Quotient (AQ 29], 30]) at ages 14 and 16 years. The AQ comprised 38 items at age 14 years and 28 items
at age 16 years. Parents rated their agreement with each item on a 4-point scale.
The AQ was reliable (see Table 1) and stable between ages 14 and 16 years (r?=?.67, p??.001). Individuals with ASC scored significantly higher on the AQ at ages 14 (t₂₅₄₆
?=??9.43, p??.001) and 16 years (t₄₈₉₄
?=??15.00, p??.001). The AQ was divided into subscales 29]: social difficulties, communication, imagination, attention to detail, and attention
switching. Communication items were omitted from the scale at age 16 owing to space
constraints.

ASC diagnoses

Parents of twins scoring over 15 on the CAST were invited to complete the Diagnostic
and Well-Being Assessment (DAWBA 31]), a structured interview for establishing psychiatric diagnoses. Thirty-eight ASC-relevant
items were administered over the telephone. The DAWBA effectively distinguishes between
ASC cases and controls
³⁰
, and has good inter-rater reliability (??=?0.83) and internal consistency (??=?0.92) in this sample 32]. If either twin in a pair met DAWBA ASC criteria, families were invited to take part
in the Social Relationships Study 33]. Two trained researchers visited the families to complete the Autism Diagnostic Observation
Schedule 34] with the twins and Autism Diagnostic Interview-Revised 35] with the parents; these semi-structured assessments are considered the ‘gold standard’
in diagnosing ASC in research. In total, 32 participants with a confirmed diagnosis
had Specific Psychotic Experiences Questionnaire (SPEQ) data available (29 males,
3 females).

Psychotic experiences

Psychotic experiences at age 16 were assessed using the SPEQ 8]. The SPEQ was constructed from six existing measures of psychotic experiences in
adults. The wording of these measures was adapted for use with adolescents, and the
age appropriateness of the items was established in terms of expert clinical opinion.
Positive and cognitive psychotic experiences were assessed using four self-report
measures: paranoia (15 items), hallucinations (9 items), cognitive disorganisation
(11 items), and grandiosity (8 items). Negative symptoms were assessed via a self-report
measure of anhedonia (10 items) and a parent measure of negative symptoms (9 items).
The use of both a parent- and self-report measure of negative symptoms was designed
to reflect recent recommendations that assessment of negative symptoms should include
multiple informants 36]. One item (‘Has few or no friends’) was removed from negative symptoms owing to overlap
with the autistic trait measures. Table 1 shows ranges of scores and internal consistencies. All subscales were stable across
a 9-month period (r?=?.65–.74). Principal-components analysis supported the division of SPEQ into six
subscales 8]. The measure has been validated against a similar measure, the Psychosis-Like Symptoms
(PLIKS) scale 8], 37]. Individuals reporting having ‘definitely’ had psychotic experiences on the PLIKS
scored significantly higher on all SPEQ subscales than those who did not (all p??.001). Individuals with relatives with schizophrenia or bipolar disorder (N?=?420, 227 males, 193 females) scored significantly higher on all SPEQ subscales
(p??.05), except for anhedonia and hallucinations (which showed a nonsignificant trend).

Data analyses

Skewed measures were log transformed (Table 1). Sex and age effects can inflate twin correlations (described below). As a result,
all scores were regressed for sex and age, as is standard behavioural genetic procedure
38]. Two sets of analyses were performed. Multivariate twin models were fitted to data
on continuous autistic traits and psychotic experiences in the whole sample. Mean
group differences in continuous SPEQ scores were then compared in individuals with
a diagnosis of ASC and those without such a diagnosis.

Analyses were performed in the OpenMx 39] package of R 40]. OpenMx uses full information maximum likelihood when fitting models, meaning that
missing data can be accounted for. Rather than excluding participants with any missing
data, full information maximum likelihood includes participants with data for at least
one age, meaning that the sample is not limited by the presence of missing data 39].

Phenotypic analyses

Phenotypic correlations (r_ph
) between measures were estimated from twin models.

Mean standardised SPEQ scores were compared across individuals with ASC and controls
using independent t tests. In addition, we tested whether individuals with a family history of a psychotic
disorder would display more autistic traits than those without such a family history,
again using independent t tests. Welch’s degrees of freedom for unequal sample sizes were applied.

Twin analyses

The twin design seeks to partition variance in a phenotype into three components:
additive genetic influences (A); shared environmental influences (C), which are common
to both twins in a pair and cause similarity between them; and nonshared environmental
influences (E), which are unique to each twin and create differences between them.
These parameters are estimated on the basis of the phenotypic resemblance of monozygotic
(MZ) twins, who share all their segregating DNA code, and dizygotic (DZ) twins, who
share on average 50 % of their segregating DNA code.

Cross-twin correlations were used to establish the phenotypic similarity of MZ and
DZ twins. These correlations, estimated from a saturated model of the observed data,
were obtained separately for MZ and DZ twins, and involved correlating one twin’s
score on a trait (e.g. autistic traits) with their co-twin’s score on the same trait.
One can then gain an indication of the extent of A, C, and E influences by examining
the MZ and DZ cross-twin correlations. Table 1 shows how the pattern of cross-twin correlations can be interpreted.

The multivariate twin design allows one to investigate the covariance between multiple
traits by dividing the correlation between them into A, C, and E components. Cross-trait
cross-twin correlations were estimated as a starting point and involved correlating
one twin’s autistic traits with their co-twin’s psychotic experiences score. Again
derived separately for MZ and DZ twins, these correlations can be used to obtain an
initial indication of the extent to which A, C, and E influence the covariance between
two traits. Table 1 shows how cross-trait cross-twin correlations can be interpreted.

Structural equation twin-model fitting formally estimated A, C, and E. Cholesky decompositions
were fitted to measures displaying sufficient covariance with one another (r_ph
??.20 at age 16). These decompositions estimated the degree to which the causes of
autistic traits could account for psychotic experiences. In Fig. 1, the pathways from latent variables (enclosed in circles) A
₁
, C
₁
, E
₁
, A
₂
, C
₂
, E
₂
, A
₃
, C
₃
, E
₃
, A
₄
, C
₄
, and E
₄
to SPEQ subscales at 16 years denote the extent to which the causes of autistic traits
influence psychotic experiences. In squaring these estimates, one can derive the proportions
of variance explained. A
₅
, C
₅
, and E
₅
represent residual variance in SPEQ scores that is unique to psychotic experiences.

Fig. 1. Path diagram of the Cholesky decomposition. Variables enclosed in circles are latent variables. The red arrows connecting A
₁
–A
₄
, C
₁
–C
₄
, and E
₁
–E
₄
to SPEQ age 16 represent the sources of shared variance between autistic traits and
SPEQ subscale scores. The blue arrows, which connect A
₅
, C
₅
, and E
₅
to SPEQ age 16, represent the residual variance in SPEQ subscales scores. The black arrows represent the paths between autistic traits at each age. CAST Childhood Autism Spectrum Test, AQ Autism Spectrum Quotient, SPEQ Specific Psychotic Experiences Questionnaire, A additive genetic influences, C shared environmental influences, E nonshared environmental influences

In the first instance, all estimates were free to differ by sex (quantitative sex limitation), and were then equated across sexes to test the significance of any sex differences.
Parameters were then successively dropped from the ACE Cholesky decomposition by fixing
them to 0. The fit of the Cholesky decompositions was compared against a baseline-saturated
model of observed means and variances. The fit of each model was summarised by the
fit statistic which was ?2 times the log likelihood of the data (?2LL); differences
in ?2LL between two models are ?²
distributed, with degrees of freedom equivalent to the difference in number of parameters.
Significant ?²
results suggest a model is a significantly poorer fit than the saturated model. Model
fit was further assessed using Akaike’s Information Criteria (AIC) and Bayesian Information
Criteria (BIC). Best-fitting models were selected based on the most negative BIC value.

Effects of two potential confounders were then tested: general cognitive ability and
internalising traits. A composite g-score was computed when twins were aged 7, 12,
14, and 16 years 41]. At age 7, twins completed three measures from the Wechsler Intelligence Test for
Children (WISC 42]): picture completion, vocabulary, and similarities. Twins also completed the McCarthy
Conceptual Groups test 43]. At age 12, twins completed the WISC vocabulary, picture completion, and general
knowledge tests 42] as well as the Raven’s Standard Progressive Matrices 44]. At ages 14 and 16, the Raven’s Standard Progressive Matrices and WISC vocabulary
tests were administered. At each age, scores on all completed measures were standardised;
g-score was calculated as the average of the standardised test scores at each age.

Internalising traits were a second potential confounder. They were assessed using
two self-report measures, the Childhood Anxiety Sensitivity Index 45] and Short Moods and Feelings Questionnaire 46]. Analyses were repeated with each of these confounders regressed out of all measures.