Trajectory research in children with an autism diagnosis: A scoping review

Abstract

Researchers increasingly employ longitudinal trajectory methods to understand developmental pathways of people on the autism spectrum across the lifespan. By assessing developmental or health-related outcome domains at three or more timepoints, trajectory studies can characterize their shape and varying rates of change over time. The purpose of this scoping review was to identify and summarize the published breadth of research that uses a trajectory study design to examine development in children (to age 18 years) diagnosed with autism. Using a systematic search and screening procedure, 103 studies were included. This review summarizes methodological characteristics across studies including the varying statistical approaches used. A series of figures maps where published research is available across 10 outcome domains and the ages over which children have been followed. Evidence gaps, informed by the perspectives of the autistic and caregiver stakeholders that were engaged in this review, are discussed. We recommend that future trajectory research addresses the absence of studies from low- and middle-income countries, considers longitudinal assessment of outcome domains that caregivers and autistic people consider meaningful, and plans follow-up periods with assessment timepoints that cover the gaps in ages where more outcome-specific data are needed.

Lay Abstract

The types of outcomes studied in children on the autism spectrum include clinical characteristics, such as social functioning, communication, language, or autism symptoms. Research that measures these outcomes at multiple timepoints is useful to improve our understanding of what to expect as children develop. In trajectory studies, researchers assess outcomes at three or more timepoints. This method has advantages over two-timepoint studies because it allows researchers to describe changes in the speed of development, such as accelerations, plateaus, or slowdowns. We identified and reviewed 103 published trajectory studies in children (to age 18 years) with an autism diagnosis. Importantly, we did not include studies of treatments or their effects, nor did we summarize the results of studies. Instead, this review summarizes the characteristics of the available published research, including the methods used, the many different outcomes that have been studied over time and the ages over which they have been studied. This summary may be of interest to autistic people and caregivers (parents) who want to know about the existence of research that provides answers about what to expect during an autistic child’s development. We have recommended that future trajectory research efforts try to make up for the lack of studies from low- and middle-income countries; that more attention is given to the following outcomes that are meaningful to caregivers and autistic people; and to try to fill in the age gaps where more outcome-specific data are needed.

Background

It has long been recognized that the characteristics used to define autism vary developmentally over the life course, and that there is a need for rigorous longitudinal cohort studies to understand changes in relevant outcomes over time, and identify prognostic factors associated with improvement in developmental pathways (Howlin & Moss, 2012). Trajectory methodology, featuring longitudinal analysis at three or more timepoints, has emerged as more informative for understanding the development of children and adolescents (hereafter children) on the autism spectrum compared to traditional cohort studies that use only two timepoints. The approach has led, for example, to characterizing the shape of the developmental function in autism, understanding the timing of changes and the discovery of different possible subgroups of people on the autism spectrum characterized by different trajectories on the same outcome (Seltzer et al., 2004). Non-trajectory approaches, meanwhile, which simply report or graph average measures across individuals at a given timepoint, have limited use for studying development in people on the autism spectrum because they obscure the fact that data collected from the same person over time are correlated and should not be considered in isolation (Seltzer et al., 2004).

As another example of the utility of trajectory methods, some large-scale longitudinal autism cohort research groups have published serial reports on the development of their cohorts, increasing the number of assessments as they age over time (e.g. Gotham et al., 2012; Szatmari et al., 2015). Such research has expanded our understanding of how changes over time in autism can vary (e.g. Charman, 2018; S. H. Kim et al., 2018). Described as chronogeneity (Georgiades et al., 2017), this variation exists both between trajectory groups or clusters, and between children within those clusters. Recognizing its utility for studying development, autism researchers have increasingly turned to trajectory methods.

Given the expansion of published trajectory studies in autism research, there is a need to understand and characterize the scope of this literature, including the statistical analytic approaches that have been used, the variety of outcome domains (i.e. measurable outcomes, which potentially vary over development) that have been studied and the age intervals over which outcome domains have been followed.

Statistical approaches

The types of statistical approaches used in trajectory research are numerous. Moreover, similarities in statistical terms used for substantively different tests can make delineating between these approaches challenging for non-expert readers. For purposes of this review, we use the acronyms for different approaches since, in many cases, they are the more familiar label used to reference these methods. To facilitate understanding, each individual approach can be understood as belonging to one of the three fundamental categories:

• Traditional approaches

• Growth curve modelling (GCM) approaches

• Group-based modelling approaches

Traditional approaches include statistical methods, such as repeated-measures/multivariate analysis of variance (ANOVA). These more basic methods have less flexibility for complex data structures (e.g. missing data, varying time metrics) and restrictive assumptions (e.g. normality and variance homogeneity), which do not allow for disaggregating within-person and between-person variability. GCM approaches include any variable-centred method that estimates between-person variability in within-person patterns of change over time (usually in terms of intercepts and slopes) for the overall cohort or for pre-defined cohort groups (e.g. autistic vs non-autistic). These include multi-level modelling (MLM) and latent GCM (or latent curve modelling, hereafter LCM). Group-based modelling approaches, meanwhile, are considered person-centred approaches because they estimate distinct trajectories of latent clusters or subgroups formed on the basis of similar trajectories of individual participants within a cohort population (Nagin & Tremblay, 2001). They include growth mixture modelling (GMM) and latent class growth modelling (LCGM).

Among the three approaches (traditional, GCM and group-based), group-based approaches have been gaining noticeable interest in the field (Georgiades et al., 2017), but it is unknown whether studies employing non-group-based approaches (i.e. traditional or GCM approaches) may have declined or increased in popularity over time by comparison. Characterizing the range of statistical approaches that have been used may provide a useful foundation for advancing methodological standards for trajectory research in future.

Prior reviews of longitudinal research

Early reviews of longitudinal research in autism have provided useful overviews of the results of earlier research, but have not employed systematic search strategies (Henninger & Taylor, 2013; Howlin & Moss, 2012; Seltzer et al., 2004). More recent reviews have focused on adults to varying degrees and included mostly non-trajectory longitudinal studies. For Magiati et al. (2014), an important review focus was childhood predictors of later outcomes. Bieleninik et al. (2017), meanwhile, conducted a systematic review and meta-analysis on two outcome domains – diagnostic stability and autism symptom severity – for all ages ranging to adulthood. Finally, Howlin and Magiati (2017) more broadly summarized all adult outcome-focused research. Further characterization of trajectories into adulthood remains an important area of inquiry. The set of outcome domains that are relevant after the transition to adulthood (e.g. employment status, romantic relationships), however, differs from those relevant to child development (including functional and educational domains). Separate attention to characterizing trajectory research in children on the autism spectrum is warranted to summarize areas where research has been conducted and highlight where gaps still exist (Gentles et al., 2021). We are unaware of prior reviews characterizing the available research employing a trajectory design to study change in outcome domains over time in children on the autism spectrum. Scoping reviews, which ‘aim to systematically identify and map the breadth of evidence available on a particular topic, field, concept, or issue’ (Munn et al., 2022, p. 950), are best suited to address this gap.

The current study

The primary objective of this scoping review was to identify and summarize the scope of published research that uses a longitudinal trajectory study design (with three or more timepoints) to study the temporal progression of different developmental outcome domains – including the shape, timing and subgroups – in children on the autism spectrum (to age 18 years). Notably, we excluded studies whose focus was on adulthood (i.e. where at least half of age timepoints assessed were above 18 years), studies primarily evaluating intervention effects (some studies that followed development for at least three timepoints over more than 18 months post-intervention were included) and trajectory studies of neuroanatomical or physiological development because they have been previously reviewed (Baribeau & Anagnostou, 2013; Lainhart, 2015). Per scoping review methodology, we did not pool or analytically synthesize study findings to inform clinical decision-making as would occur within a systematic review or qualitative evidence synthesis nor did we assess risk of bias (Khalil et al., 2021; Munn et al., 2018; Peters et al., 2020). Rather, we specifically sought to summarize methodological characteristics, including different analytic approaches used, map the research available across 10 of the more commonly followed outcome domains and the ages over which children have been followed, and highlight where further child trajectory research may be warranted.

Methods

We followed scoping review methods described by Arksey and OʼMalley (2005) and informed by subsequent clarifications and enhancements (Levac et al., 2010; O’Brien et al., 2016). We adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR; Tricco et al., 2018). The protocol for this review, which contains some additional detail of the planned methods, is published open access (Gentles et al., 2021).

A librarian-assisted (L.B.) search was developed that combined terms related to the population (autism diagnosis, age up to 18 years), methodology (trajectory study) and relevant outcome domains (developmental, educational). We searched six databases (MEDLINE, EMBASE, CINAHL, PsycINFO, ERIC and Cochrane Database of Systematic Reviews) on 25 May 2021. EndNote reference management software (Clarivate Analytics) was used to manage citations. After removal of duplicates, we selected articles for inclusion by first screening titles and abstracts in duplicate, and second by individually screening the retrieved full text. Title-and-abstract screening disagreements and full-text screening queries were resolved through consensus and team meetings. A statistician (E.D.) helped resolve whether studies qualified as true trajectory study methodology both in terms of the data collection criterion (three timepoints or more) and two analysis criteria (accounting for the correlated nature of within-case data over time; and not simply averaging measures across individuals at each timepoint providing only cross-sectional estimates). The following 12 exclusion criteria, in five areas, were used during screening.

Study design

• Longitudinal studies with fewer than three timepoints (i.e. non-trajectory design)

• Case study, case series of individual trajectories (not a study of group trajectories)

• Evaluations of intervention effects, unless they followed development for at least three post-intervention timepoints for > 18 months

Population

• Most trajectory timepoints at 19 years or older (not a study of child trajectories)

• Focus of analysis was NOT primarily on participants on the autism spectrum (however, studies comparing autism versus non-autism populations were included)

• Lacks at least one trajectory group with ⩾ 90% diagnosed autism

Trajectory outcome

• Neuroanatomical or physiological developmental outcomes only

• Dichotomous outcomes only (e.g. diagnostic status)

Analysis

• Non-trajectory analysis (does not consider WITHIN-case trajectories, but rather averages cases at each timepoint)

Other

• True trajectory results NOT reported; did not at a minimum report the slope or direction of change of a trajectory outcome

• Non-English, -French or -Spanish publication

• Non-journal article publication (dissertations, grey literature)

We used a set of sub-questions to determine the fields to include in the data extraction form (developed in Microsoft Word), as described previously (Gentles et al., 2021). Data from included articles were extracted individually by one of the four reviewers and verified by the review coordinator. Deviations from the published protocol (Gentles et al., 2021) in terms of what was extracted are described here. While we initially expected trajectory results to be presented with reference to child age, this was not the case for all studies; we therefore extracted information about whether trajectory assessments and results were age-referenced or not. Trajectory analysis type was categorized as either traditional, GCM or group-based approaches, or the combination of both GCM and group-based approaches. Cohort research group names were listed only for cohort groups with multiple trajectory study publications (see Supplemental Material: Appendix 1). The final extraction form (Supplemental Material: Appendix 2) includes definitions for the most data fields that were extracted. After extraction, data were transferred to a Microsoft Excel database for aggregation and analysis.

Community involvement

Following the (optional) methodological recommendation of Arksey and O’Malley (2005), we engaged two stakeholders in this review – an adult autistic self-advocate and a parent of a child on the autism spectrum. Both provided feedback face-to-face (at an autism research conference) on aspects of the extraction form during planning stages; and verbally (in video conference meetings) shared their stakeholder-perspective interpretations of the findings, which we have incorporated into the discussion. Stakeholders were recruited by invitation, selected for their interest and expertise in partnering on autism research, from different regions of Canada. Ethical approval for this engagement process was considered unnecessary by our research ethics board. Terms of agreement for these non-authorship roles were mutually agreed on early in the project. Both stakeholders were compensated financially.

Results

Figure 1 shows the results of the searches and screening. We included 104 articles corresponding to 103 studies – two articles (Baghdadli et al., 2018, 2019) were considered the same study for purposes of this review strictly for presenting the same trajectory-related findings (they would otherwise be considered separate studies based on distinct non-trajectory findings). All except three articles were retrievable with MEDLINE: one (Lord & Luyster, 2006) retrieved with EMBASE and two (Magiati et al., 2011; Manti et al., 2011) retrieved with PsycINFO – a finding that may be relevant to inform planning of potential future rapid reviews in this area or for those who can only access MEDLINE (e.g. through PubMed).

Figure 1.

PRISMA flowchart showing identification and screening of eligible publications for inclusion.

Key characteristics of included studies are displayed in Appendix 1 (Supplemental Material), organized hierarchically by country where the study’s cohort originated, cohort research group name for groups with multiple trajectory study publications and individual study ID (author, year). Most studies originated in the United States (n = 69), followed by Canada (n = 13), the United Kingdom (n = 8), the Netherlands (n = 4), Australia (n = 3), France (n = 2) and four countries with one cohort (Israel, Italy, Sweden and Switzerland). We identified six consistently-named cohort research groups with multiple published trajectory studies from the United States, the United Kingdom (with two cohorts), Canada, France and Australia.

Autism-specific sample sizes ranged from 10 to 6975, with 32% categorized as small (n = 10–40), 48% moderate-to-large (n = 41–200) and 20% very large (n > 200). For 50 studies that included comparisons to non-autistic participants, overall sample sizes (i.e. including participants without autism) ranged from 17 to 8653.

Methodological characteristics

Most studies (91.3%) reported methods for ascertaining autism diagnoses. Diagnostic ascertainment was reported to be facilitated by the Autism Diagnostic Observation Schedule (ADOS; Lord et al., 1989; Lord & Rutter, 2012) in 75.7% of studies, the Autism Diagnostic Interview-Revised (ADI-R; Rutter et al., 2005) in 57.3% of studies, other standardized instruments in 18.4% of studies or clinical judgement in 82.5% of studies. Multiple diagnostic instruments were used in 57.3% of studies, and a gold standard approach (ADOS, ADI-R and clinical judgement) was used in 46.6% of studies. All studies were prospective except eight studies which involved retrospective analyses of cohort data. The sample source was community-based (recruitment through community-based settings, such as primary care doctors’ offices, preschools and public advertisements) for 36.9% of studies, clinical (recruitment from clinical settings, such as diagnostic or autism treatment centres) for 36.9% of studies and population-based (cases from population or surveillance databases or epidemiological samples) for 9.7% of studies; no sample source information was given for 16.5% of studies.

Table 1 lists the terms authors used to describe the trajectory-specific analytic methods employed in their study, which may be useful for future literature review search strategies, and to facilitate understanding of terms likely to be encountered in primary study reports. Each specific method is classified into one of the three fundamental categories of approaches introduced above (see background): group-based, GCM and traditional approaches. We have included SAS (statistical software) function names since these were the descriptors used by authors in some instances and may therefore also be useful as search terms.

Table 1.

Terms used within study reports to describe trajectory-specific statistical methods grouped under the major categories of trajectory approaches to which they belong.

Term used	Studies using the corresponding method
Group-based approaches – GMM
• GMM	Anderson et al. (2007), Farmer et al. (2018), Henry et al. (2018), Tomaszewski et al. (2019) and Frazier et al. (2021)
Group-based approaches – LCGM
• Generalized linear latent and mixed model (GLLAMM)	Gotham et al. (2012), Lord and Rutter (2012) and S. H. Kim et al. (2018)
• Group-based approach	Richler et al. (2010)
• Group-based trajectory modelling/analysis (GBTM)	Ozonoff et al. (2018), McCauley et al. (2020), Peverill et al. (2019), Clarke et al. (2021) and Baribeau et al. (2021)
• Group-based latent trajectory modelling	Fountain et al. (2012)
• LCGM/LCGA	Ozonoff et al. (2018), Gotham et al. (2012), Pickles et al. (2014), Venker et al. (2014), Visser et al. (2017), Baghdadli et al. (2019), Vaillancourt et al. (2017) and Landa et al. (2012)
• Latent growth curves	Lord and Rutter (2012)
• Multivariate cluster analysis	Georgiades et al. (2022)
• Semiparametric (and) group-based methods/approach	Visser et al. (2017), Szatmari et al. (2015) and Vaillancourt et al. (2017)
• SAS PROC TRAJ	Anderson et al. (2007), Hus Bal et al. (2015), Lord and Luyster et al. (2006), Richler et al. (2010), Fountain et al. (2012), S. H. Kim et al. (2018), Baghdadli et al. (2012, 2018, 2019) and Baribeau et al. (2021)
GCM approaches – MLM
• Functional data analysis (FDA)	Jones and Klin (2013)
• General linear mixed-effects regression	Bradshaw et al. (2019)
• Generalized additive mixed-effects models (with semiparametric smoothing splines)	Travers et al. (2017)
• Generalized linear mixed models (GLMM)	Gotham et al. (2015), McDonald et al. (2020), Gulsrud et al. (2014) and Harrop et al. (2021)
• Generalized (linear) mixed-effects models	Gangi et al. (2020), Gotham et al. (2015) and Miller et al. (2020)
• Growth curve analysis (GCA)	Anderson et al. (2007, 2009, 2011), Lord et al. (2015), Tiura et al. (2017) and Szatmari et al. (2009)
• GCM	Yoder et al. (2015), Woynaroski et al. (2016), Wagner et al. (2019) and Frazier et al. (2021)
• Hierarchical generalized linear modelling (HGLM)	Ashley et al. (2020)
• Hierarchical GCM	Midouhas et al. (2013)
• Hierarchical linear modelling (HLM)	Wei et al. (2015), Martin et al. (2013), Rosen and Lerner (2016), Tiura et al. (2017), Campbell et al. (2017), Choi et al. (2018), Ekas et al. (2017), Ellis Weismer and Kover (2015), Green and Carter (2014), Kover et al. (2016), Leezenbaum and Iverson (2019), Munson et al. (2008), Toth et al. (2006), West et al. (2020), Yoder et al. (2006), Flouri et al. (2015) and Szatmari et al. (2009)
• Individual growth curve (ICG) analysis/modelling	Tek et al. (2014), Bopp et al. (2009), Bopp and Mirenda (2011)
• Linear mixed (effects) models	Bal et al. (2020), Spurling Jeste et al. (2014), MacDuffie et al. (2020), Saul and Norbury (2020), Nystrom et al. (2019), Bos et al. (2018) and Li et al. (2020)
• Longitudinal growth modelling	Tomaszewski et al. (2020)
• Mixed-effects growth curves	Fusaroli et al. (2019)
• Mixed-effects (linear) models	Meyer et al. (2018)
• Mixed-level modelling	Yoder et al. (2015) and Di Rezze et al. (2019)
• Mixed models with repeated measures (MMRM)	Estes et al. (2015)
• MLMs	Cain et al. (2019), McCormick et al. (2016), Tiura et al. (2017), Siller and Sigman (2008) and Greenlee et al. (2021)
• Ordinal logistic random effects models	Hutman et al. (2010)
• SAS PROC MIXED	Anderson et al. (2007), McCormick et al. (2016), Anderson et al. (2009), Lord et al. 2015), Meyer et al. (2018), Siller and Sigman (2008) and Bopp et al. (2009), Bopp and Mirenda (2011)
• SAS PROC GLIMMIX	Hutman et al. (2010)
GCM approaches – LCM
• Latent GCM (LCM)	Bedford et al. (2016), Grimm et al. (2017), Bennett et al. (2014), Stringer et al. (2020) and Tesfaye et al. (2021)
Traditional approaches
• Generalized estimating equations (GEE)	Ozonoff et al. (2010), Cohen and Flory (2019), Landa and Garrett-Mayer (2006), Landa et al. (2013), Brignell et al. (2019) and May et al. (2021)
• Mixed ANOVA (age/time factor, no missing data)	Rutherford et al. (2015)
• Repeated-measures ANOVA (no missing data)	Manti et al. (2011), Bernabei et al. (2007), Ben-Itzchak et al. (2014), Flanagan et al. (2015) and Clark et al. (2017)

Overall, 26 studies used a group-based approach (i.e. GMM or LCGM; including studies combined with other approaches), while 77 studies used other analytic approaches (i.e. GCM or traditional approach only). Figure 2 illustrates historical trends of numbers of studies using group-based and non-group-based approaches.

Figure 2.

Numbers of studies either employing group-based analytic approaches (includes eight studies that used both a group-based and non-group-based approach; solid bars) or not employing a group-based approach (i.e. a non-group-based or traditional approach only; open bars), by publication year (to May 2021).

We had hypothesized that studies employing group-based approaches would less often feature comparisons to children not on the autism spectrum (i.e. by focusing on within-autism heterogeneity) and would have higher sample sizes, compared to studies employing non-group-based approaches (GCM- or traditional-type) (Gentles et al., 2021). The type of analytic approach employed varied significantly by whether studies included comparisons to children not on the autism spectrum (χ²(2) = 7.016, p = 0.030) – only 17% (3/18) of studies employing group-based approaches featured comparisons to children not on the autism spectrum, compared to 50% (38/76) of studies not employing group-based approaches (GCM- or traditional-type) and 56% (5/9) mixed-approach studies. On average, autism-specific sample size for studies employing group-based approaches (M = 599, SD = 134) was significantly higher than for studies employing non-group-based approaches (M = 96, SD = 108; t(92) = 2.77, p = 0.007).

Timing of outcome measurement

For some studies, authors reported the timing (schedule) of trajectory outcome assessment (i.e. in the methods) differently from how they represented the timing of outcome measurements in the trajectory results. In either case (assessment schedule or trajectory results), the timing could be age-referenced (months since birth) or assessment timepoint-referenced (e.g. reporting timepoints in terms of time after baseline assessment – but not by age). Studies also varied in the consistency of the number and timing of assessments among study participants, generally because some analytic methods were capable of handling variation in these measurement parameters.

Both the reported assessment schedule and the reported trajectory results were age-referenced for 57 (55.3%) studies. The assessment schedule was assessment timepoint-referenced, while the reported trajectory results were age-referenced for 25 (24.3%) studies. Neither the assessment schedule nor the reported trajectory results were age-referenced for 21 (20.4%) studies. Thus, there were 82 (79.6%) studies for which the ages of trajectories studied for individual outcome domains could be summarized (see below, and Figures 2 to 12).

Figure 12.

Age intervals and assessment timepoints of published trajectories of language in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

BPVS-2: British Picture Vocabulary Scales, Second Edition; CASL: Comprehensive Assessment of Spoken Language core tests; EOWPVT: Expressive One-Word Picture Vocabulary Test; EVT-1: Expressive Vocabulary Test – First Edition; MSEL: Mullen Scales of Early Learning, EL: Expressive language domain, RL: Receptive language domain; PLS-4: Preschool Language Scale-4, AC: Auditory Comprehension, EC: Expressive Communication; PPVT-3: Peabody Picture Vocabulary Test – Third Edition; VABS-2: Vineland Adaptive Behaviour Scales, EL: Expressive language, RL: Receptive language; WJ-III-ACH-LWI: Woodcock Johnson – Third Edition Tests of Achievement: Letter Word Identification subtest.

Outcome domains followed

There was substantial diversity across studies in the types of trajectory outcome domains assessed (see Supplemental Material: Appendix 1). The most commonly followed outcome domains for which age-referenced trajectory results were reported are summarized in age plots in Figures 3 to 12. In describing each outcome domains below, we state the number of latent subgroups defined by studies that used a group-based approach, showing where the number of subgroups defined across such studies was inconsistent. For the outcome domain adaptive behaviour functioning, 10 studies reported age-referenced results, 6 of which featured consistent timepoint assessments (definable in number and ages) and 4 of which featured inconsistent timepoint assessments across a range (Figure 3). The trajectory age ranges followed spanned 6–24 months (Estes et al., 2015) to 24 months–10 years (Meyer et al., 2018). The measures (outcome measurement instruments) used to quantify this adaptive behaviour functioning included different iterations of the Vineland Adaptive Behaviour Scales (VABS) composite score – the Vineland Social Maturity Scale (Doll, 1953), VABS-1 (Sparrow et al., 1984) and VABS-2 (Sparrow et al., 2005). Of the two (age-referenced) studies that used a group-based analytic approach, two latent subgroups were defined in one study (Farmer et al., 2018), and three latent subgroups were defined in the other (Szatmari et al., 2015). In a third group-based analysis study, whose trajectory results were not age-referenced, and which used the VABS-2 composite, four latent subgroups were defined (Ben-Itzchak et al., 2014).

Figure 3.

Age intervals and assessment timepoints of published trajectories of adaptive behaviour functioning in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

VABS: Vineland Adaptive Behaviour Scales.

*Study with group-based analysis – number of latent subgroups listed.

Figure 4.

Age intervals and assessment timepoints of published trajectories of social functioning in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

CDER Soc: Client Development Evaluation Report, rescaled sum of five social functioning items; VABS Soc: Vineland Adaptive Behaviour Scales (1 or 2), Socialization domain.

*Study with group-based analysis – number of latent subgroups listed.

Figure 5.

Age intervals and assessment timepoints of published trajectories of communication in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

CDER Com: Client Development Evaluation Report, rescaled sum of three communication items; VABS-1 or -2: Vineland Adaptive Behaviour Scales – First or Second Edition, Com: Communication domain.

*Study with group-based analysis – number of latent subgroups listed.

Figure 6.

Age intervals and assessment timepoints of published trajectories of daily living skills in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

VABS-1 or -2: Vineland Adaptive Behaviour Scales – First or Second Edition, DLS: Daily living skills.

*Study with group-based analysis – number of latent subgroups listed.

Figure 7.

Age intervals and assessment timepoints of published trajectories of autism symptom severity in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

ABC: Autism Behaviour Checklist; ADI-R: Autism Diagnostic Interview-Revised; ADOS: Autism Diagnostic Observation Schedule; CSS: Calibrated Severity Score; PDDBI: Pervasive Developmental Disorder Behaviour Inventory; RRB: Repetitive and Restricted Behaviour; SA: Social Affect; SEQ: Social Emotional Questionnaire; SRS-1: Social Responsiveness Scale, First Edition.

*Study with group-based analysis – number of latent subgroups listed.

Figure 8.

Age intervals and assessment timepoints of published trajectories of restricted and repetitive behaviours in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

ADI-R: Autism Diagnostic Interview-Revised; ADOS: Autism Diagnostic Observation Schedule; IS: Insistence on Sameness factor (three items of ADI-R); RBS-R: Repetitive Behaviour Scale-Revised; RSM: Repetitive Sensory Motor factor (four items of ADI-R).

*Study with group-based analysis – number of latent subgroups listed.

Figure 9.

Age intervals and assessment timepoints of published trajectories of internalizing problems in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

ABCL: Adult Behaviour Checklist; ASEBA: Achenbach System of Empirically Based Assessment; CBCL: Child Behaviour Checklist; CDI: Children’s Depression Inventory; CSI: Child Symptom Inventory; DBC: Developmental Behaviour Checklist; EAQ: Emotion Awareness Questionnaire; ECI-4: Early Childhood Inventory-4; SCL: Somatic Complaints List; SDQ: Strengths and Difficulties Questionnaire; W/RQC: Worry/Rumination Questionnaire for Children.

*Study with group-based analysis – number of latent subgroups listed.

Figure 10.

Age intervals and assessment timepoints of published trajectories of externalizing problems in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

ABCL: Adult Behaviour Checklist; CBCL: Child Behaviour Checklist; CSI: Child Symptom Inventory; RBS-R: Repetitive Behaviour Scale-Revised; SDQ: Strengths and Difficulties Questionnaire.

Figure 11.

Age intervals and assessment timepoints of published trajectories of cognitive functioning in children with an autism diagnosis.

Solid line: Study with consistent assessment timepoints. Dotted line: Study with inconsistent assessment timepoints. Measures used are listed in parentheses.

ABAS-2: Adaptive Behaviour Assessment System – Second Edition; BSID: Bayley Scales of Infant Development; MP: Merrill-Palmer; MSEL: Mullen Scales of Early Learning, DQ: Developmental Quotient, ELC: Early Learning Composite, NVIQ: Non-verbal IQ, VIQ: Verbal IQ, VR: Visual reception; WASI-2: Wechsler Abbreviated Scale of Intelligence – Second Edition, FSIQ: Full-scale IQ, PRI: Perceptual Reasoning Index; VCI: Verbal Comprehension Index; VPT: Various psychometric tests; WISC: Wechsler Intelligence Scale for Children; WPPSI: Wechsler Preschool and Primary Scale of Intelligence.

For social functioning trajectories (Figure 4), 12 studies reported age-referenced results, 3 of which used a group-based analytic approach – 2 of these 3 had two latent subgroups (Baghdadli et al., 2012, 2018, 2019), while the third study had six subgroups (Fountain et al., 2012). In one group-based analytic study for which trajectory results were not age-referenced, two latent subgroups were defined (measure: VABS-2 Socialization domain) (Tomaszewski et al., 2019).

For communication trajectories (Figure 5), 14 studies reported age-referenced results, 4 of which used a group-based analytic approach – 2 of these 4 had two latent subgroups (Baghdadli et al., 2012, 2018, 2019), while a third study had six subgroups (Fountain et al., 2012), and a fourth study had seven subgroups (Pickles et al., 2014). In one group-based analytic study for which trajectory results were not age-referenced, two latent subgroups were defined (measure: VABS-2 Communication domain) (Tomaszewski et al., 2019).

For daily living skills trajectories (Figure 6), 11 studies reported age-referenced results, 4 of which used a group-based analytic approach – all with two latent subgroups (Baghdadli et al., 2018, 2019; Clarke et al., 2021; Hus Bal et al., 2015; Tomaszewski et al., 2020).

For trajectories of autism symptom severity, 13 studies reported age-referenced results (Figure 7). The most common symptom severity measure was the ADOS Calibrated Severity Score (ADOS CSS) (Gotham et al., 2009), used in seven studies, while six alternative measures were used in the remaining studies. However, 6 of the 13 age-referenced studies used a group-based analytic approach, with two latent subgroups defined in 2 studies (Georgiades et al., 2022; Szatmari et al., 2015) and four latent subgroups defined in 4 studies (Gotham et al., 2012; S. H. Kim et al., 2018; Lord et al., 2012; Venker et al., 2014). In two additional group-based analysis studies for which trajectory results were not age-referenced, two latent subgroups were defined in one study, which used the ADOS CSS (Ben-Itzchak et al., 2014), and five latent subgroups were defined in the other study, which used the ADOS total algorithm score (Visser et al., 2017).

Only four studies reported age-referenced results for trajectories of restricted and repetitive behaviours (RRBs; Figure 8), two of which used a group-based analytic approach – one with two latent subgroups (Richler et al., 2010) and one with four subgroups (Lord et al., 2012); in one additional group-based analytic study for which trajectory results were not age-referenced, five latent subgroups were defined (measure: ADOS Restricted and Repetitive Behaviour (RRB) score) (Visser et al., 2017).

For four well-studied outcome domains, there were no age-referenced studies that used a group-based analytic approach. For trajectories of internalizing problems (Figure 9) and of externalizing problems (Figure 10), eight and nine studies, respectively, reported age-referenced results. For trajectories of cognitive functioning, 16 studies reported age-referenced results (Figure 11). For language trajectories, another 16 studies reported age-referenced results (Figure 12).

Discussion

This review provides the first comprehensive map of the available published trajectory research on children with an autism diagnosis of which we are aware. As the review demonstrates, use of longitudinal trajectory methodology has rapidly increased since 2006 and a significant knowledgebase now exists regarding the development of children on the autism spectrum across numerous outcome domains. Over these years, a wide variety of methodological approaches have been used; and many different outcome domains have been studied. Not all study results have been age-referenced, however, rendering them less comparable to the majority of age-referenced findings. While the age coverage (up to 18 years) for the 10 most frequently studied outcome domains is relatively good, some understudied age gaps remain. Geographically, four anglophone-dominant countries and six non-anglophone countries are represented; however, no included studies originated from low- and middle-income countries.

We identified six major cohort research groups with multiple published trajectory studies. Consistent use of recognizable cohort names allowed easier cross-referencing of prior studies from these groups, something that could be useful to understand the influence of cohort-specific characteristics on serial reports of trajectory outcomes over years of observation. For some studies that included children from multiple different cohort sources, we observed that cross-referencing of cases from the same source was not possible. As a result, cohort information necessary for such cross-referencing remained inaccessible without contacting researchers.

Methodology of trajectory studies

The list of terms used to describe analytic methods employed in trajectory studies (Table 1) illustrates the wide range of available methods and terminology. This list may provide useful search terms to help systematically identify studies employing trajectory methods in future reviews. Of note, within the GCM-type approaches is the range of terms used for MLM analyses (e.g. MLM, linear mixed models, mixed-effect models and hierarchical linear models); these are generally interchangeable references to the modelling approach of analysing repeated data nested within a person without the assumptions of homogeneity-of-regression slopes that are required by traditional approaches, such as repeated-measures or multivariate ANOVA. A less used GCM-type approach was LCM, which falls within the structural equation modelling (SEM) framework and, like MLM, allows for the estimation of between-person differences in within-person change. Interestingly, while LCM provides a foundation for the group-based approaches for identifying unobserved subgroups (GMM and LCGM), more autism researchers have used MLM than LCM when examining longitudinal variability as associated with a priori covariates (e.g. autistic vs non-autistic, demographics). This may be because MLM is a direct extension of the familiar ANOVA-type approaches. In contrast, LCM might be preferable among those who are familiar with the SEM or latent variable framework. Although the two approaches are numerically identical and technically derive equivalent results, one approach may be more practically advantageous than the other in certain research contexts (e.g. complexity of data structures and research questions; see McNeish & Matta, 2018).

While GCM approaches focus on quantifying longitudinal trends as associated with a priori factors, group-based approaches have been gaining interest in the field for their power to parse out unobserved longitudinal heterogeneity – that is, chronogeneity, suggesting different subgroups of children exist that each follow different developmental paths (Georgiades et al., 2017). Confirming our protocol-specified hypothesis that group-based approaches have more often been used to parse out subgroups within purely autistic populations, studies employing such approaches featured comparisons to children not on the autism spectrum significantly less often than non-group-based approaches (GCM- or traditional-type).

Contrary to our additional hypothesis that group-based approaches may have increased in popularity compared to other methods over time, however, we saw no discernable trend in the use of this approach relative to non-group-based approaches since the first year of trajectory study publications in 2006 (Figure 2). This may be partially due to the general requirement of larger sample sizes for group-based approaches to ensure accurate estimation and sufficient statistical power for group-level analyses (S. Y. Kim, 2012). Indeed, sample sizes for studies employing group-based approaches were observed to be significantly larger than non-group-based studies. Among studies that used a group-based approach to empirically identify unobserved trajectory subgroups, LCGM (a parsimonious version of GMM assuming within-subgroup homogeneity) was the more commonly used method. This may be due to its less computationally intensive nature and to a research focus on between-subgroup variability, in contrast to the more complex GMM that allows for within-subgroup variability (Bauer & Reyes, 2010).

Selecting the right approach from the analytic options observed here requires familiarity with broad-ranging methods. Aside from carefully evaluating the data requirements for each approach, it is important to ensure appropriate fit with the research questions or underlying theory (Curran et al., 2010). For instance, the application of GMM approaches (i.e. subtyping) should be based on the hypothesis or previous evidence that the longitudinal outcome of interest can be better explained by more than one mean trajectory for all. For investigations of the contributions of specific factors (e.g. age of diagnosis, initial cognitive or language abilities) to various trajectories, GCM approaches are often suitable for testing such covariate effects and more flexible for exploring developmental complexity (e.g. developmental dynamics between multiple outcome domains, time-varying effects of changing factors (e.g. intervention elements) on trajectories). Figure 13 provides further clarity on how the different analytic approaches identified in this review may be appropriately selected for studying trajectories in different contexts.

Figure 13.

Trajectory analytic method selection flowchart.

Note. The GCM extensions may also be applied to group-based growth models.

It was important to attend to the age-referencing of trajectory findings in this review. Approximately one fifth of studies did not report trajectories in terms of child age. Age-referencing of trajectories, however, provides a seemingly necessary basis for understanding the natural development of children on the autism spectrum. While reporting trajectories in terms of age may not be possible in some studies, such as those following interventions administered to children at different ages, their results are unlikely to contribute as meaningfully to understanding natural development. In addition, age-referencing allows a level of comparability necessary for potentially pooling or meta-analysis of trajectory results. Consequently, our summaries of the outcome domains focused predominantly on age-referenced studies. Finally, almost one third of studies had small autism-specific sample sizes (⩽ 40). We suggest a minimum sample size above 40 for GCM approaches, and potentially much larger numbers for group-based approaches, given the variation in trajectories among children on the autism spectrum.

Outcome domains followed

The summaries of the trajectory outcome domains that have been studied (Figures 3 to 11) illustrate the ages covered by existing research and identify age gaps where further studies may be warranted. In addition, the relevance of each outcome domain from stakeholder’s (i.e. autistic self-advocate and caregiver) perspectives is an essential consideration when deciding where future research efforts should be directed. Among studies following composite measures of adaptive behaviour functioning measured using the VABS (Figure 3), only two studies examined trajectories beyond age 7 years – ages that were noted by the stakeholder reviewers as characterized by important biological and social changes. While trajectory findings at these older ages may represent a potential knowledge gap, researchers have noted that following trajectories of individual subdomains of adaptive functioning on the VABS-2 (Social Functioning, Communication, Daily Living Skills, and Motor Functioning) may be preferable to using a composite measure given the moderate correlation of these subdomains among autistic people, which suggests they may capture divergent patterns in this population (Baghdadli et al., 2018; Yang et al., 2016). Summaries of the studies that followed the individual subdomains of social functioning (Figure 4), communication (Figure 5) and daily living skills (Figure 6) illustrate good age coverage. These adaptive function-related outcome domains were considered highly relevant from both stakeholders’ perspectives.

Among studies following autism symptom severity (Figure 7), only two examined trajectories beyond age 12 years, indicating a possible gap warranting future research. Similar to the VABS composite, measures of symptom severity generally represent multiple domains – for example, the ADOS CSS (Gotham et al., 2009) is further separable into two standardized measures, the Social Affect and RRB scores (Hus et al., 2014). Of the four studies following RRBs as a trajectory outcome domain (Figure 8), two (Lord et al., 2015; Richler et al., 2010) separated RRBs further into an Insistence on Sameness factor (IS) and Repetitive Sensory Motor factor (RSM); a third (MacDuffie et al., 2020) separated RRBs into Stereotyped/restricted, Self-injurious, Sameness/ritualistic/compulsive factors. The stakeholders agreed that symptom subdomains may be more relevant than a composite construct because some aspects of symptoms may not be problematic for individuals, as has been recognized previously (Bal et al., 2018; McConachie et al., 2015). Specifically, aspects of the RRB domain were highlighted as potentially non-problematic. From the autistic stakeholder’s perspective, while insistence on sameness was considered potentially impairing and therefore relevant, repetitive self-stimulating behaviour (whether motor or intellectual) was noted as useful to cope with real problems and therefore considered beneficial – consistent with broader published reports of autistic perspectives on stimming (Joyce et al., 2017; Kapp et al., 2019). From the caregiver stakeholder’s perspective, her child’s ‘perseverative interests’ led him to build expertise that resulted in studying math at university. These perspectives reflect research highlighting the potential value of preferred interests for autistic people to reduce anxiety and engage in vocational and other life pursuits (Bury et al., 2020; Patten Koenig & Hough William, 2017). Thus, a line of the trajectory research that separates and recognizes the varying problematic potential of the domains of RRBs, and symptom severity in general, is more likely to produce stakeholder-relevant findings.

For the outcomes of internalizing (Figure 9) and externalizing problems (Figure 10) – both relevant from stakeholder perspectives – there were no studies covering children 8- to 9-years old, except for one study of anxiety trajectories (Baribeau et al., 2021), potentially warranting future research covering these ages. For the trajectory outcomes of cognitive functioning (Figure 11) and language (Figure 12), there was generally good coverage of ages below 10 years, perhaps reflecting the relevance of these domains for understanding development at younger ages. Cognitive functioning was highlighted as potentially problematic from the autistic stakeholder’s perspective, who believed most IQ measures to be limited in their ability to reflect the true intellectual abilities of autistic people. The idea of language trajectories, meanwhile, was noted as potentially problematic from the caregiver stakeholder’s perspective because of the unhealthy pressure she understood caregivers to feel to achieve an externally fixed level or rate of language uptake that may be unreasonable or inappropriate for their child at that point in time. These perspectives are similar to ideas highlighted in the literature discussing how to improve on the concept of ‘optimal outcomes’ in autistic people (Bal et al., 2018; Georgiades & Kasari, 2018; Taylor, 2017), which have highlighted the problem of emphasizing inflexible outcomes that may be unreasonable for some to achieve and ignores individuals’ histories. Future trajectory research on cognitive functioning and language may benefit from further consideration of these ideas.

Recorded trajectories that span broad age ranges and include assessment timepoints around ages known to be critical periods of transition for children have been demonstrated to be important to detect ‘turning points’, or changes in the rates of development that may be associated with commonly timed environmental or biological transitions (Clarke et al., 2021; Georgiades et al., 2022). Thus, not only are gaps in ages studied in the literature important, but the density of ages that have been assessed is also a crucial consideration for the ability to highlight turning points. For many outcome domains summarized in this review, there was often a higher density of assessments at younger compared to older child ages. Higher density of sampling at theoretically important transitions in adolescence may therefore be warranted in future trajectory studies.

Among studies using group-based trajectory approaches, there was inconsistency in the number of latent trajectory subgroups (clusters) defined for a given outcome domain. In the case of adaptive functioning, for example, three studies each identified different numbers of subgroups. This inconsistency, which reflects variations in methodological decisions (e.g. discarding subgroups with few cases; Gotham et al., 2012) and the underlying samples, highlights the challenges in identifying a commonly agreed-upon set of trajectory subgroups. In addition, the potential for varying membership in subgroups according to the outcomes used to define them (see Szatmari et al., 2015) suggests that it may be premature to assume that subgroups defined with this approach represent true phenotypes.

Strengths and limitations

The rigorous search of multiple databases increased confidence that publications within its scope were comprehensively captured. Some relevant studies may have been missed as we excluded non-published sources, such as dissertations, resulting in a potentially incomplete assessment of available evidence coverage across outcome domains and ages. Prior to the decision to exclude them, however, a partial scan revealed otherwise includable dissertations had later been published, suggesting that resource-intensive trajectory research rarely remains unpublished. Another potential limitation is the exclusion of predominantly adult-focused studies that may have also assessed outcome domains at one or more timepoints in the child age range (to 18 years). While justified on the basis of separate sets of outcome domains relevant to child versus adult populations, the possible exclusion of child age datapoints for child-relevant outcome domains would similarly yield an incomplete assessment of the available evidence. We suggest that determining the outcome domains common to both child and adult age ranges could be identified through an analogous adult-focussed scoping review – and that such a review should be a high priority.

While engagement of the two stakeholders may be considered a strength given that methodological guidance for scoping reviews frames this as optional (Anderson et al., 2008; Arksey & OʼMalley, 2005; O’Brien et al., 2016), the level of this engagement was nevertheless limited. It did not, for example, reach the level of co-creation as has been encouraged in more recent guidance (Pollock et al., 2022). Moreover, it included only two individuals, whose perspectives cannot be considered to represent all stakeholders’ views. Similarly, we did not explicitly account for clinician and autism researcher perspectives, who may have alternative legitimate views regarding outcome domain relevance. Nonetheless, we consider these stakeholders to have contributed valuable insights.

Conclusion

This scoping review of published trajectory studies of children on the autism spectrum has implications for primary trajectory research and utility for multiple audiences. First, the absence of trajectory studies in low- and middle-income countries is at odds with the global imperative to promote health research equity and outcomes (World Health Organization, 2012). We thus recommend a shift among researchers and funding bodies to prioritize such studies in future. Second, given stakeholder perspectives highlighting the variable relevance of different outcome domains, which have implications for producing research that is more relevant and better aligned with public priorities, we recommend those planning future trajectory studies consider the evolving literature on meaningful outcomes for caregivers and autistic people (e.g. McConachie et al., 2018) and engage stakeholders themselves to inform the design of their studies. Third, to address the evidence gaps we have highlighted, particularly for the outcomes of adaptive functioning, RRBs, internalizing and externalizing problems, we encourage researchers to consult the outcome domain summaries in this review and consider planning follow-up periods that cover the ages where more data are needed.

The outcome domain summaries, including the gaps in ages studied, provide a useful map of trajectory research available across clinically relevant outcome domains for clinicians, policymakers and others planning care or services; they may also serve the informational needs of families, known to proactively seek research findings (Gentles et al., 2019), by identifying the studies that have helped us understand ‘what to expect’ during development among children on the autism spectrum. Finally, this scoping review may help future reviewers to identify specific outcome domains for which sufficient evidence is available to conduct a synthesis of primary study results (e.g. a systematic review).