Trajectories of emotional and behavioral problems from childhood to early adult life

Outcome variables

The parent-reported SDQ (Goodman et al., 2000) measured current co-occurring mental health symptoms broken down into the conduct, emotional and ADHD subscales. The SDQ is widely used as a screening instrument for child mental health problems and its psychometric properties have been established in several samples, including the United Kingdom (Goodman et al., 2000) and the United States (Bourdon et al., 2005). In the three mental health domains, each symptom is scored 0–2, with domain scores ranging from 0 to 10. Cut-points associated with a “high” or “very high” probability of having a diagnosis have been established in typically developing populations of 4–17 years of age (emotional symptoms cutoffs 5 for high,7 for very high, conduct symptoms 4 and 6, ADHD symptoms 8 and 10). The instrument has also been previously used in population studies of autism and ID (Totsika, Hastings, Emerson, Berridge, & Lancaster, 2011). In a previous report of waves 1 and 2 of the SNAP cohort, we demonstrated good domain internal consistency (Simonoff et al., 2013). For wave 3, we employed the young adult version of the SDQ, described on http://www.sdqinfo.org/Adult/, which retains the same domains but includes minor changes in wording (Brann et al., 2018). We focus on dimensional symptoms scores in the models. However, for descriptive purposes only, we estimate the proportion of participants scoring above the cut-off points associated with clinical diagnosis in epidemiological samples of typically developing youth. These should be considered impressionistic only as these cut-offs have not been validated in autistic samples, nor at this age.

Predictor variables

These were selected from the wave 1 assessment. Given that we are examining three mental health outcomes and the sample is of moderate size, our strategy was to test a relatively modest number of predictors, selecting those thought to be methodologically most robust for the study design and/or most meaningful for interpretation of the findings.

OPEN IN VIEWER

Figure 1.

Path diagram of the multivariate growth model for SDQ conduct, emotional and ADHD at 12, 16 and 23 years of age.

Conduct symptoms

In comparing the present findings to previous reports, the SDQ conduct domain used here maps to the Developmental Behaviour Checklist (DBC) disruptive domain used by Gray et al. (2012), and somewhat to both the irritability subscale of the Applied Behavioral Analysis (ABC) employed by Anderson et al. (2011) and the externalizing scale of the Scales of Independent Behavior—Revised (SIB-R; (Bruininks et al., 1996)) used in the Adolescents and Adults with Autism (AAA) study (Seltzer et al., 2003). Higher level of functioning predicted a greater reduction in externalizing problems whether measured by IQ (Anderson et al., 2011) or presence of an ID (Shattuck et al., 2007) in most studies but not all (Gray et al., 2012). Consistent with others measuring IQ, we found that a higher adaptive function as measured on the VABS score was associated with a greater decline in conduct problems, supporting previous findings. Furthermore, we also found that lower language level at baseline (ADOS module 2 versus 3) predicted less decline in conduct problems, lending robustness to the finding but also highlighting that the exact mechanism of effect needs elucidation. It is well recognized that lower verbal skills are linked to greater rates of externalizing behavior in developmental disorders (McClintock et al., 2003). It will be important that future research attempts to distinguish if this effect can be substantially ameliorated by providing alternative, non-verbal means of communication or whether it reflects underpinning factors associated with lower global cognitive ability. In examining language level at an older age, the AAA study did not find a relationship to externalizing behaviors (Shattuck et al., 2007), raising the possibility the effect is more time-limited. Interestingly, we failed to identify a relationship of either baseline level or change in conduct symptoms to severity of autism as measured on the ADOS, suggesting the risk factor for this domain may be communication rather than social communication.

Previously, we reported in the SNAP cohort that higher levels of neighborhood deprivation were associated with a greater reduction in conduct symptoms between wave 1 and wave 2. Here, we confirm the earlier finding with the addition of data from wave 3. Furthermore, the statistically efficient analysis undertaken here also identified deprivation as being associated with a higher level of baseline conduct symptoms. Our finding is consistent with research in typically developing populations, where indices of deprivation are strongly associated with greater levels of conduct problems, although the effect may become non-significant in multivariate analyses that include other metrics associated with psychosocial disadvantage, raising questions about the mechanism of effect and whether neighborhood deprivation has a direct effect (e.g. Ford et al., 2004). The previous literature with respect to autism is limited. Gray et al. (2012) did not find a relationship of antisocial symptoms to socioeconomic disadvantage in their Australian sample. In a younger Australian cohort, Emerson reported that environmental risk factors, including neighborhood deprivation, increased risk of conduct problems in 3-year-olds with ASD similarly to that in typically developing and intellectually disabled comparison groups, and also that these risk factors led to greater persistence in autistic children at 5 and 7 years, compared to the other groups (Emerson et al., 2014). This contrasts with our finding that neighborhood deprivation was associated with greater decline in conduct symptoms over time, albeit at a different timepoint in development. The analyses by Emerson et al. (2014) were bivariate and did not account for multivariate indices of environmental risk, making it difficult to infer a direct effect of neighborhood deprivation. Neighborhood deprivation is associated with a wide range of potential risk factors for mental health problems, including poverty, parental mental and physical ill-health, lack of social cohesion and opportunities for community participation (Fone et al., 2014) and meta-analysis has shown that accounting for these variables can alter the relationship between deprivation and other outcomes (Nieuwenhuis & Hooimeijer, 2016). In terms of our present findings, the effect of neighborhood deprivation was identified in multivariate analyses, which included correlated factors of parental mental health and parental education but no other wider environmental correlates of neighborhood deprivation. Future research should aim to replicate our finding and more precisely identify the more proximal risk factors driving this relationship.

We found no significant effect of parental mental health on any domains, whether measured historically or at baseline. In relation to externalizing symptoms, the AAA study identified that a better mother–child relationship was linked to lower baseline and maternal praise predicted a greater decline in problems (Woodman et al., 2015). As poorer parental mental health is associated with parenting stress and parent–child relationships in typically developing samples (Deater-Deckard, 2014) and parents of autistic children often experience higher levels of stress and mental ill-health (Hastings et al., 2005; Yorke et al., 2018), future longitudinal studies should explore this area in more detail

Emotional symptoms

We found that both greater autism severity and lower language level were associated with lower baseline parent-reported emotional symptoms. Gotham et al. (2015) found that higher verbal IQ was associated with greater anxiety but not depression scores and Shattuck et al. (2007) found that those with ID had less decline in internalizing symptoms. We include adaptive function as measured on the VABS composite, rather than IQ which was measured on a variety of measures, in the current study. Although these measures are strongly related (r = 0.57 in this study), they are nevertheless different phenomena, with the VABS composite indexing a much wider range of performance. However, these differences also mirror a large body of inconsistent research on the cross-sectional relationship between IQ and levels of affective symptoms in autism. The literature reports both that anxiety is related to higher IQ (Gadow et al., 2005; Hallett et al., 2013; Weisbrot et al., 2005) or is not different according to IQ (Brereton et al., 2006; de Bruin et al., 2007; Simonoff et al., 2008). Furthermore, the role of IQ may vary across subdomains of anxiety (Witwer & Lecavalier, 2008). The reporting of affective symptoms, requiring inference of another’s internal experiences, may be particularly challenging in autism, where impairments in communication and emotional literacy are common. The style of assessment (interview versus questionnaires) and the content may play a greater role in accurately identifying psychopathology. This uncertainty highlights the pressing need to validate measures of mental health symptoms for autistic people.

In relation to family characteristics, we found that higher parental education was associated with reports of greater emotional problems. The AAA study did not show such a link (Woodman et al., 2016) but Gotham et al. (2015) found that mothers with less education reported greater increases in depressive symptoms in the subsample with lower verbal ability. Again, whether these differences reflect true symptomatic variation or biases in measurement needs evaluation in future studies. The findings from Gotham et al. (2015) with respect to females are intriguing. Very few samples, including SNAP, have sufficient numbers of females to test robustly for sex differences, and this should be a recruitment priority for future cohorts.

ADHD symptoms

We found that higher VABS scores at wave 1 predicted a greater decline in ADHD symptoms, but no other relationships met our tests of significance. Previously in this sample, when not accounting for multiple testing, we reported that wave 1 higher IQ and higher VABS both predicted lower wave 2 absolute scores but not change (Simonoff et al., 2013). Consistent with our present findings, Anderson et al. (2011) reported that higher verbal IQ predicted a greater decline in symptoms of hyperactivity on the ABC.

Other predictors

Of note, we failed to find any significant prediction for a history of parental mental health problems (trend association with emotional symptoms intercept = 0.083) or for school placement. Parental mental health problems is a robust predictor of childhood symptoms in general population studies (Ford et al., 2004) and parents of autistic children appear to be at greater risk, probably both for genetic (Bolton et al., 1998; Piven & Palmer, 1999) and likely environmental reasons, manifesting as parental stress (Estes et al., 2009). Hence, it is particularly striking that there was very little evidence for a role of such difficulties in the level and persistence of mental health symptoms in their offspring. It should be borne in mind that, like all our measures, a history of parental mental health problems was captured at baseline and we did not include measures at later timepoints, as these make interpretation of trajectories more difficult In addition to this direct role, we tested for a possible informant rater effect, as our outcome measures were solely reported by parents. Reassuringly, we failed to find evidence supporting a rater effect suggesting bias due to parental affective symptoms.

The null finding with respect to school placement is of particular interest as we have found in the SNAP cohort that exposure to mainstream schooling predicts a relative decline in autistic symptoms over the same time period (Simonoff et al., 2019). These previous analyses did not examine collateral effects of different schooling experiences and a concern was that mainstream school attendance might increase levels of emotional symptoms, in particular, anxiety in relation to greater actual or perceived performance demands. The present findings do not support this potential concern, although it should be noted that the analytic strategy did not attempt to partial out effects between waves 1 and 2 (the school-age period) and waves 2 and 3. Our earlier report of wave 1 to 2 relationships similarly failed to identify a role for school placement in any of the symptom domains, supporting the present results (Simonoff et al., 2013).

The present study has several strengths. The sample is population-representative, and the use of inverse probability weights means that the estimates reflect the wider population of people with autism from which the sample was drawn. We had high retention levels at wave 3, and the analytic procedure took account of any selective attrition, including the purposive sampling at wave 2, so that estimates are unbiased. In order to avoid false positive results, we employed an efficient analysis strategy of a single model with parsimonious selection of predictor measures, which should increase confidence in the positive findings. We were able, however, to include a broader range of potential predictor variables than that of any previous longitudinal analysis. To increase clarity of interpretation, we aimed to use predictor variables that would not share method variance with the outcome, for example, the ADOS CSS for autism severity, and we accounted for possible rater effects by including baseline teacher ratings on the SDQ and the maternal GHQ score.

An important limitation of the study is the use of the SDQ subscales to measure mental health domains. While a well-recognized and validated measure for research studies in non-autistic populations, like the overwhelming majority of currently available measures, its sensitivity and accuracy in autism are uncertain. As mental health symptoms may have atypical manifestation in autism (Kerns & Kendall, 2012), brief questionnaires may potentially miss important psychopathology. This issue of valid psychopathology measurement in autism needs urgent attention. The SNAP cohort has a limited number of female participants, and we judged there was insufficient power to explore sex differences. However, sex difference in the mental health of autistic people is a particularly important research priority for future studies. The SNAP cohort was established before there was a general appreciation of the importance of mental health conditions in autism, and the early assessments were not designed to measure mental health risk factors. Only variables collected at wave 1 are used as predictors, as the interpretation of the predictive role of subsequently collected data would be unclear. However, this places a limitation on the interpretation of factors that may change over time, such as parental mental health and neighborhood characteristics. We did not obtain systematic information about any interventions received across the entire study period, and it is therefore not possible to take account of how treatment may have influenced individual trajectories. But, even if they had been measured their heterogeneity in terms of type, intensity, duration, and timing would have made estimating any causal effect challenging. Finally, we should bear in mind the modest sample size, increasing likely inferential error rates and the need for replication of findings in other cohorts.

In summary, we report the mental health trajectories in a population-based cohort from middle childhood to adult life for the three common domains of psychopathology. We show that while symptoms in all domains decline over this time period, a substantial proportion still remain above the currently quoted cutoffs for a likely disorder at least one area. The striking consistency in rank order over time lends hope to the expectation that in the near future it should be possible to identify early which individuals will be at the greatest risk for persisting mental health problems. There were few predictors of trajectories across domains, but better adaptive functioning on the VABS predicted a more rapid decrease in both conduct and ADHD symptoms and higher initial language level predicted greater decline in conduct and, at trend level, ADHD symptoms. Interestingly, better initial language also predicted higher emotional symptoms and it is unclear whether this relationship is true or artefactual. Both language and adaptive function are frequent targets for intervention in autism and future research should evaluate the effect of autism-focused interventions on co-occurring psychopathology. Given the prevalence and persistence of additional mental health problems in autism, a further research priority will be to test the role of theory-driven risk factors.