The clinical relevance of subgroups of autistic adults: Stability and predictive value

Autism in adulthood is characterized by heterogeneity, complicating the provision of tailored support. In previous work, we aimed to capture this heterogeneity by determining subgroups of autistic adults that differed in clinical outcomes: cognitive failures, psychological difficulties, and quality of life (QoL). Two subgroups were identified: a “Feelings of Low Grip” subgroup characterized by experiencing a lower sense of mastery and a higher susceptibility to difficulties in daily life, and a “Feelings of High Grip” subgroup characterized by a higher sense of mastery and lower susceptibility to difficulties in daily life. The current pre‐registered study involves a longitudinal extension to determine (a) stability and (b) predictive value of the previously identified two subgroups. Subgroups were identified using community detection based on 14 self‐report measures related to demographic, psychological, and lifestyle characteristics in two samples (aged 31–86 years) that were analyzed separately: Sample 1 (NAutism = 80) measured 5 years after baseline and Sample 2 (NAutism = 241, NComparison = 211) measured 2 years after baseline. The stability over time was assessed based on (a) the number of subgroups, (b) subgroup profiles, and (c) subgroup membership. Predictive validity was assessed for cognitive failures, psychological difficulties, and QoL. Results indicated that autistic and non‐autistic adults formed distinct subgroups. Within both autism samples, the two previously identified autism subgroups were replicated at follow‐up. Subgroup profiles were similar for >50% of the variables at two‐year follow‐up, and 21% at five‐year follow‐up. Moreover, ≥76% remained in the same subgroup at two‐year follow‐up, and ≥ 57% after 5 years. Subgroup membership was predictive of external clinical outcomes up to 5 years. Thus, this study demonstrated the stability and predictive value of the autism subgroups, especially for the two‐year follow‐up. A further focus on their clinical utility might increase the aptness of support, and may provide more insight into the aging process when being autistic.


Abstract
Autism in adulthood is characterized by heterogeneity, complicating the provision of tailored support.In previous work, we aimed to capture this heterogeneity by determining subgroups of autistic adults that differed in clinical outcomes: cognitive failures, psychological difficulties, and quality of life (QoL).Two subgroups were identified: a "Feelings of Low Grip" subgroup characterized by experiencing a lower sense of mastery and a higher susceptibility to difficulties in daily life, and a "Feelings of High Grip" subgroup characterized by a higher sense of mastery and lower susceptibility to difficulties in daily life.The current pre-registered study involves a longitudinal extension to determine (a) stability and (b) predictive value of the previously identified two subgroups.Subgroups were identified using community detection based on 14 self-report measures related to demographic, psychological, and lifestyle characteristics in two samples (aged 31-86 years) that were analyzed separately: Sample 1 (N Autism = 80) measured 5 years after baseline and Sample 2 (N Autism = 241, N Comparison = 211) measured 2 years after baseline.The stability over time was assessed based on (a) the number of subgroups, (b) subgroup profiles, and (c) subgroup membership.Predictive validity was assessed for cognitive failures, psychological difficulties, and QoL.Results indicated that autistic and non-autistic adults formed distinct subgroups.Within both autism samples, the two previously identified autism subgroups were replicated at follow-up.Subgroup profiles were similar for >50% of the variables at two-year follow-up, and 21% at five-year follow-up.Moreover, ≥76% remained in the same subgroup at two-year follow-up, and ≥ 57% after 5 years.Subgroup membership was predictive of external clinical outcomes up to 5 years.Thus, this study demonstrated the stability and predictive value of the autism subgroups, especially for the two-year follow-up.A further focus on their clinical utility might increase the aptness of support, and may provide more insight into the aging process when being autistic.

Lay summary
Autism in adulthood is characterized by differences between autistic people, in terms of autism characteristics, experienced strengths, and difficulties.In our earlier work, we identified two subgroups that differed in their susceptibility to experienced difficulties.The current study involves a longitudinal extension of our

INTRODUCTION
Autism in adulthood is characterized by heterogeneity at multiple levels ranging from biology to behavior (Masi et al., 2017).At the behavioral level there are differences in the presentation, extent of autism characteristics, and experienced strengths and challenges (Happé et al., 2006;Masi et al., 2017).This heterogeneity complicates provision of tailored support and the search for prognoses.Hence, autistic adults often do not receive the specific support they need and wish (Hwang et al., 2017).Moreover, many autistic peoples live with the uncertainty of what to expect as they reach old age (Finch et al., 2022).Therefore, gaining insight into this behavioral heterogeneity is needed to improve support and care for autistic individuals throughout adulthood.
One promising solution to this heterogeneity is to determine subgroups within the autism spectrum with predictive power for the adulthood developmental trajectory.Many studies have focused on the identification of subgroups (for a review, see Agelink van Rentergem et al., 2021), but only a few focused primarily on adulthood (Elwin et al., 2017;Gonthier et al., 2016;Lewis et al., 2008;Lombardo et al., 2016;Radhoe et al., 2023;Ring et al., 2008) of which even fewer included multiple clinically relevant variables or domains (Bishop-Fitzpatrick et al., 2016;LaBianca et al., 2018;Radhoe et al., 2023).Including such variables (like well-being, autonomy, or social satisfaction) is of importance when the goal of subgrouping is not primarily to understand the observed heterogeneity, but also to implement this knowledge into clinical practice.
A crucial additional step in subgrouping research involves the validation of the identified subgroups.If the aim is to utilize the subgroups in practice, it is essential to determine their validity.However, it is remarkable how little attention has been devoted to validation in the subgrouping studies within the autism research field so far (see for review Agelink van Rentergem et al., 2021).Specifically, of the subgrouping studies in autistic adulthood, only two studies performed an independent replication (i.e., identified subgroups in an additional, independent sample; Lombardo et al., 2016;Radhoe et al., 2023).Moreover, none of the studies included longitudinal data to (a) test predictions over time, or (b) assess the temporal stability of the subgroups.This lack of validation leaves the question whether the identified subgroups are genuinely robust, which would limit the (clinical) applicability of the identified subgroups.
In the autism literature thus far, there have been two subgrouping studies-apart from our own-focusing on clinically relevant measures in adulthood, while also adopting some form of subgroup validation.Bishop-Fitzpatrick et al. (2016) identified three subgroups using normative outcomes and objective quality of life (QoL).They assessed the external validity by showing that the subgroups differed on related external measures (i.e., measures not used in the subgrouping analysis), such as employment status and independent living.LaBianca et al. (2018) aimed to assess need for healthcare across the autism and/or ADHD spectrum.They identified five subgroups and demonstrated the external validity by showing subgroup differences on genetic risk factors.The difference between these findings already shows how the study goal impacts the design choices (in terms of sample and included variables), and which subgroup validation approaches may be most suitable.Thus, the studies focusing on clinically relevant subgroups in autistic adults are (a) still limited in numbers, and (b) difficult to bring together as their goals, approaches, and results are diverse.
Moreover, there have not been any subgrouping studies focusing on clinical predictions throughout autistic adulthood.Even when we do not consider the subgroups and focus on autism in adulthood in general, it becomes clear how much is still unknown about the developmental process of autistic adults (Tse et al., 2022;Wise, 2020).This is both surprising and alarming, as the evidence shows that autistic adults have a poorer overall QoL compared with the general population (Ayres et al., 2018) and increased rates of all major psychiatric conditions such as depression and anxiety (Hand et al., 2020;Nylander et al., 2018).The few longitudinal studies that have followed autistic people throughout adulthood show that there is marked heterogeneity in living arrangements, employment, and medical, and psychiatric co-occurring conditions (see for review Wise, 2020).Therefore, most autistic adults do not know what to expect as they grow older.This highlights the need for (a) knowledge on the developmental trajectory in adulthood, while (b) considering individual differences between autistic adults.
In our previous work, we have focused on subgroup identification in autistic adults using clinically relevant variables: psychological, demographic, and lifestyle characteristics (Radhoe et al., 2023).Our goal was to detect subgroups that might have an impact on clinical practice.Two subgroups were identified that differed in susceptibility to experienced difficulties.As suggested by a group of older/autistic adults, the subgroups were labeled as "Feelings of High Grip" (HighGr) and "Feelings of Low Grip" (LowGr).We showed that (a) the subgroups can be replicated in an independent sample, and (b) demonstrated the external validity as the subgroups differed on clinically meaningful variables.Specifically, the Feelings of Low Grip subgroup showed the most vulnerable profile on the cluster variables, and was associated with the lowest QoL, most psychological difficulties, and most cognitive failures.Thus, two autism subgroups were identified of which the validity has been demonstrated in multiple ways.However, while of value in itself, it is yet unknown whether these subgroups are informative for the future prospects of autistic adults.
In this study, we follow up on our earlier work by assessing the prognostic utility of the previously identified subgroups in a longitudinal extension.First, we assess whether the identified subgroups are stable as people age.Second, we determine whether the subgroups can be used to predict clinically relevant outcomes (i.e., QoL, psychological difficulties, and cognitive failures) over time.

METHODS
This study is the longitudinal follow up (data collection 2020-2022) of our previous cross-sectional study (data collection 2015-2020; Radhoe et al., 2023).Therefore, the same exclusion criteria and materials were used.Concise information on each of these elements is provided below, and a detailed description is provided in the crosssectional study and protocol paper (Geurts et al., 2021).

Participants
In total, 592 (348 autistic and 244 non-autistic adults) were screened for inclusion.Autistic participants were recruited via mental health institutions in the Netherlands, and advertisements on client organization websites and social media.Participants in the non-autistic comparison group were recruited through advertisements on social media, and via the social network of researchers and research assistants involved in this study.
As in the cross-sectional study, we applied the following exclusion criteria to all participants: (1) intellectual disability (ID), (2) insufficient understanding of the Dutch language required to complete the self-report questionnaires, (3) age lower than 30 years.For the autism group, we only included adults who received a clinical DSM-III, DSM-IV or DSM-5 diagnosis of an autism spectrum disorder (ASD) (American Psychiatric Association, 1987, 2000, 2013).For the non-autistic comparison (COMP) group, we applied the following additional exclusion criteria: (1) history of more than one psychotic episode, (2) a present or past diagnosis of ASD or a total score higher than 32 on the autism spectrum quotient (Baron-Cohen et al., 2001), (3) a present or past diagnosis of AD(H)D or a score of six or higher on the Dutch translation of the ADHD DSM-IV Rating Scale (Kooij et al., 2005), (4) ASD diagnosis in close family members (i.e., parent(s), child(ren), sibling(s)), and (5) AD(H)D diagnosis in close family members.The exclusion criteria were checked at baseline and follow-up, based on data from self-report questionnaires.In total, 532 participants met inclusion criteria and had sufficient data to be included: 321 autistic and 211 non-autistic adults.Participant characteristics are described in Table 1, and participant numbers and reasons for exclusion are described in more detail in Figure 1a.Two subsets of participants were included, overlapping with those included in the cross-sectional study (Radhoe et al., 2023) where all baseline measures were analyzed, but not the follow-up measures.The subsets 1 differed in the time interval until follow-up: Sample 1 (N AUT = 80) with a time interval of 5 years until follow-up, and Sample 2 (N Total = 452, N AUT = 241, N COMP = 211) with a time interval of 2 years until follow-up.These subsets did not overlap in included participants.Please note that Sample 1 (with a smaller sample size, but a longer time interval until follow-up) was included to explore the stability of the autism subgroups over a longer time interval.Therefore, although data were collected from both autistic and non-autistic adults at five-year follow-up, only autistic adults were included in this specific sample.A schematic representation of the data collection timepoints and corresponding samples is depicted in Figure 1b.

Measures
For a more detailed description, including psychometric properties, of the measures see (Radhoe et al., 2023).Below each measure is described briefly.

Cluster variables
Autism characteristics were measured using the autism spectrum quotient (Baron-Cohen et al., 2001).Subscale scores for social skills, attention switching, attention to detail, communication, and imagination were included (10 items per subscale, range 0-10).Higher scores indicated more autism characteristics.
Educational level was classified using the Dutch Verhage scale (Verhage, 1964), consisting of seven categories (1 indicating less than 6 years of primary education, and 7 indicating a university degree). 1 The same subsets of participants were used in the cross-sectional study with different labels (Radhoe et al., 2023).Sample 1 was referred to as "original sample" (i.e., Cohort 2) and Sample 2 was referred to as "replication sample" (i.e., Cohort 3).
Mastery was measured with the Pearlin Mastery Scale (Pearlin et al., 1981).A sum score was included (7 items, range 7-35).A higher score reflected more feelings of being in control.
Worries/fears were assessed with a combination of the Worry Scale (Wisocki et al., 1986) and fear questionnaire (Marks & Mathews, 1979).A sum score was included, where a higher score indicated more worries (15 items, range 15-75).
Physical activity was measured with the International Physical Activity Questionnaire (IPAQ) (Craig et al., 2003).The included cluster variable represented the total number of minutes during which someone was physically active during the past 7 days.
Negative life events were measured with the list of threatening experiences (Brugha et al., 1985).A sum score was calculated that indicated the number of negative life events someone has experienced during the past year (12 items, range 0-12).
Emotional support was assessed using the Close Persons Questionnaire (Stansfeld & Marmot, 1992).A sum score was included, where higher scores indicated higher levels of received emotional support (12 items, range 12-60).
Positive and negative affect were assessed with the positive and negative affect schedule (Watson et al., 1988).Two subscale scores were included, for positive and negative affect (10 items per subscale, range 10-50).Higher scores indicated, respectively, more positive or negative feelings.

External validators
Cognitive failures were measured with the Cognitive Failures Questionnaire (Broadbent et al., 1982).A sum score was calculated, where higher scores indicated more cognitive failures (25 items, range 0-100).
Psychological difficulties were assessed with the symptom checklist-90 revised (SCL-90-R) (Derogatis, 1977).The total score (based on 90 items) and nine subscale scores were included: agoraphobia, anxiety, depression, somatization, cognitive performance deficits, interpersonal sensitivity, hostility, sleep difficulties, and items not included in any specific factor.Higher scores represented more psychological difficulties.
QoL was measured with the World Health Organization Quality of Life Questionnaire-BREF (WHOQoL-BREF) (THE WHOQOL GROUP, 1998).Scores on four subscales were included: physical health, psychological, social relationships, and environment.Higher scores represented a higher QoL.
T A B L E 1 Participant characteristics for two separate samples at baseline and follow-up after 5 and 2 years.

Procedure
For the exact procedure and included measures we refer to the cross-sectional study and published protocol (Geurts et al., 2021;Radhoe et al., 2023).Briefly, after written informed consent was received, participants filled out questionnaires, either online or on paper.Each participant spent around 2 h to complete the questionnaires.A subset of participants was also interviewed (e.g., ADOS-2, N autism = 97), and tested using the shortened WAIS-IV (Wechsler, 2012;

Missing data on cluster variables
We distinguished between item-level missingness and instrument-level missingness.At the item level, we imputed a maximum of 10% of missing data per participant for each questionnaire (Bennett, 2001).The manner of imputation was dependent on the instrument: for autism characteristics, mastery, sensory sensitivity, worries/fears, positive and negative affect, and emotional support; a maximum 10% of missing values was recoded to the median of the participant's responses on the specific measurement instrument.For negative life events and physical activity, a maximum of 10% of missingness was recoded to zero, which implied either the absence of a negative life event/specific physical activity.No missing values were imputed for the education variable.
For the instrument-level missingness, we only included participants with a maximum of one missing value out of the total of 14 cluster variables.Thus, participants with more than one missing value on the cluster variables were excluded from the analyses.

Community detection: analysis
The goal of a community detection analysis is to identify communities (/subgroups), which are locally dense connected subgraphs in a larger network.To perform this analysis, the scores on cluster variables were first transformed to z-scores, to ensure that different measurement scales did not affect results.Next, a Pearson correlation matrix was created, that included person-to-person correlations between scores of all participant pairs in the sample.These correlations represented the similarity between the scores of two participants: The higher the correlation, the more similar the scoring patterns on the cluster variables.This correlation matrix was used as input for the community detection analyses, and the spinglass algorithm (with γ = 0) was used to identify the communities (Reichardt & Bornholdt, 2006).
The community detection analysis was performed in three different steps in two independent samples.First, the analysis was performed using Sample 2 (two-year follow-up) including autistic and non-autistic participants.Second, the analysis was performed including only the autistic adults from Sample 2. Third, the analysis was repeated for Sample 1 (five-year follow-up) including only autistic participants.

Stability of autism subgroups over time
We used three criteria to determine the similarity of the subgroups identified at follow-up to those at baseline.To conclude that the subgroups are stable over time, at least Criteria 1 and 2 or Criteria 1 and 3 had to be met.
Criterion 1: The community detection analysis again results in two major autism subgroups at follow-up (i.e., each subgroup should include more than 25% of the sample and should be the largest subgroups identified at follow-up).The 25% cut-off per subgroup was selected to ensure the subgrouping solution remained true for the majority of the sample at follow-up.This involved comparing the number and size of subgroups at both timepoints.
Criterion 2: Scores on the individual cluster variables per subgroup are similar between both timepoints (i.e., baseline and follow-up).Bayesian t-tests with a standard/flat prior were used for each of the cluster variables.As we hypothesize that the subgroups are stable over time (i.e., similar at follow-up), the likelihood of the data fitting under the null hypothesis (H 0 : cluster variable scores are similar over time) is contrasted with the likelihood of the data fitting under the alternative hypothesis (H 1 : cluster variable scores are not similar over time) (Wagenmakers, 2007).The resulting Bayes factor (BF 01 ) quantifies the evidence in support of the null hypothesis as compared with the alternative hypothesis (Jarosz & Wiley, 2014;Wagenmakers & Lee, 2014).For example, a BF 01 of two indicates that the data are two times more likely to occur under the null hypothesis than under the alternative hypothesis.
Criterion 3: The same participants cluster together in one subgroup at follow-up as they did at baseline.A Bayesian contingency table test (with standard/flat prior) was used to test similarity between subgroup membership at baseline and follow-up.Moreover, two measures were calculated that indicate similarity between subgroup membership at two timepoints (i.e., baseline and two-year follow-up, or baseline and five-year follow-up): (1) the Rand index (RI; Rand, 1971) and (2) the Hubert-Arabie Adjusted RI (ARI HA ;Hubert & Arabie, 1985).Values above 0.90 for both measures represent excellent subgroup recovery (i.e., similarity between subgroups identified at two timepoints), whereas values below 0.65 represent poor recovery (Steinley, 2004).
Next to the pre-registered analyses, a complimentary analysis was performed to determine the stability over time of the mean scores on the cluster variables for the total autism group (see Supporting information S4).

External and predictive validation
For external and predictive validation, we compared the autism subgroups identified on variables not included in the community detection analysis, that is, cognitive failures, psychological difficulties, and QoL.The subgroups were considered meaningful if they differed on these external measures.For external validation, the independent variable was subgroup membership established at follow-up, with the external variables (i.e., dependent variables) also measured at follow-up.For predictive validation, subgroup membership established at baseline was used as the independent variable, whereas the external variables (i.e., dependent variables) were measured at follow-up.An Analysis of Variance (ANOVA) was used to assess whether the subgroups differed in the total amount of experienced cognitive failures.Ten ANOVA's were used to test for subgroup differences in psychological difficulties (i.e., total score and scores non nice subscales).Four ANOVA's were used to assess differences in QoL between the subgroups (i.e., for four subscales).

Community involvement
For this study, and the overarching project, we collaborated with a group of older/autistic adults.We met at least three times a year (either online or in person), and discussed relevant matters including recruitment strategies, questionnaires, and interpretation of analysis results.All members were paid for their contribution.

RESULTS
The attrition analyses are described in detail in the Supporting information S1.Overall, there was more dropout in the comparison group than in the autism group.Between the autism subgroups, there were differences in total AQ score: higher scores for those who dropped out in the "Feelings of High Grip" (HighGr) subgroup compared with the "Feelings of Low Grip" (LowGr) subgroup.

Two-year follow-up: Autistic and non-autistic adults again form separate groups
Two subgroups were identified that corresponded to an autistic or non-autistic subgroup.The first subgroup (N = 227) mostly included autistic adults (96%), and the second subgroup (N = 225) mostly included non-autistic adults (89%).The profiles of the two subgroups are depicted in Figure S1.To obtain more insight into the heterogeneity within the autism group, the community detection analysis was repeated for just the autism group.

Two-year follow-up: Replication of autism subgroups over time
The community detection analysis on data from the autism group again resulted in three subgroups.The profile of the first subgroup (N = 109, 45%) resembled that of the "Feelings of High Grip" (HighGr) subgroup that was identified at baseline (Radhoe et al., 2023).This subgroup was again characterized by a higher score on variables in the social domain, a higher mastery, lower level of worries, more positive affect and less negative affect.The second subgroup (N = 122, 51%) was similar to the "Feelings of Low Grip" (LowGr) subgroup that was previously identified.This subgroup was again characterized by lower scores on the social domain, lower mastery, more worries, less emotional support, less positive affect, and more negative affect.The third subgroup (N = 10, 4%) did not seem similar to the "Rest" subgroup that was previously identified.As this subgroup only included 10 people, we did not consider this a valid separate subgroup, and did not include this subgroup in further analysis.Thus, the HighGr and LowGr subgroups were replicated at follow-up after an interval of 2 years (i.e., Criterion 1 was met).Subgroup profiles are depicted in Figure 2a.Descriptives and raw cluster variable scores of the subgroups can be found in the Supporting information S3.
Regarding the stability of subgroup profiles, Bayesian analyses indicated that most scores on the individual cluster variables were similar at follow-up (Figure 3 and Table S6; see S4 for results of the total autism group).For the HighGr subgroup, there was moderate evidence (BF 01 >3) for 11 out of 14 cluster variables (78%) that the data were in favor of the null hypothesis (i.e., the cluster variable scores were similar over time).For social skills, communication, and worries, there was only anecdotal evidence in favor of the null hypothesis.For the LowGr subgroup, most Bayes factors also provided moderate evidence (BF 01 >3) in favor of the null hypothesis, except for attention switching (anecdotal evidence H0), attention to detail (anecdotal evidence H0), and positive affect (anecdotal evidence H1).As the scores on at least 11 cluster variables (out of 14, i.e., 78%) were similar between the two timepoints, we conclude that the subgroup profiles were similar at follow-up (i.e., Criterion 2 was met).

Stability of memberships: Most autistic adults remain in the same subgroup after 2 years
In both subgroups, the majority of autistic adults retained their subgroup membership at follow-up (Figure 2b).Specifically, from the HighGr subgroup 87 adults (76%) remained in this subgroup after 2 years, whereas 23 (20%) switched to the LowGr subgroup.From the LowGr subgroup 98 adults (81%) remained in this subgroup, whereas 17 (14%) switched to the HighGr subgroup at follow-up.In both subgroups, the number of participants that were not analyzed at follow-up (due to missing data or drop-out) were similar: 10 from the HighGr subgroup and 9 from the LowGr subgroup.A Bayesian test of association produced a Bayes factor >100 (decisive evidence), in favor of the alternative hypothesis (i.e., there was an association between subgroup membership at both timepoints).Therefore, Criterion 3 was met.Nonetheless, as there were some changes in subgroup membership over time, traditional measures typically used to assess convergence on the same sample showed that recovery was not 100% (RI = 0.71, ARI HA = 0.41).
Although not preregistered, we explored changes in cluster variable scores for those autistic adults whose subgroup membership was not stable over time (see Supporting information S7).Switches from the HighGr subgroup to the LowGr subgroup associated with a decrease in mastery, emotional support, positive affect, and an increase in negative affect.Switches from the LowGr subgroup to the HighGr subgroup were associated with an increase in mastery, emotional support, physical activity, positive affect, negative life events, and a decrease in negative affect.Therefore, switches in subgroup membership seem associated with changes on a variety of factors (i.e., change in profile) rather than an individual cluster variable.

Subgroups useful for prediction over time
Statistics regarding the external and predictive validation can be found in Table 2.For external validation, subgroup differences were found on external variables NAff = negative affect.NLife = negative life events.NIA = not included in analysis (due to drop-out or missing data).Higher z-scores represent higher scores on Edu, Soc, AttD, AttS, Com, Imag, Mas, Sup, Phys, and pAff.Higher z-scores represent better scores on Sens, Wor, Naff, and NLife (i.e., less sensitivity, less worrying, less negative affect, fewer negative life events).Shaded area represents 95%-confidence interval.measured at the same occasion (i.e., Sample 2; two-year follow-up).The LowGr subgroup reported (a) more cognitive failures, (b) more psychological difficulties, and (c) a lower QoL on all subscales, compared with the HighGr subgroup.
For predictive validation, subgroup differences were found on external variables measured at follow-up (i.e., Sample 2, subgroups identified at baseline; external variables measured at two-year follow-up).Results indicated that the subgroups identified at baseline, differed on external outcomes measured at follow-up.Specifically, being a member of the LowGr subgroup at baseline was predictive of (a) more cognitive failures, (b) more psychological difficulties, and (c) a lower QoL when compared with the HighGr subgroup.Thus, subgroup membership at baseline was predictive of future external outcomes.

Autism subgroups largely stable at five-year follow-up
At five-year follow-up (N = 80), we again identified three subgroups of which two profiles were highly similar to those of the HighGr and LowGr subgroups identified at baseline.Subgroup profiles are depicted in Figure 2c.Descriptives and raw cluster variable scores of the subgroups can be found in the Supporting information S3.The HighGr (N = 30, 38%) and LowGr (N = 35, 44%) subgroups included most of the autistic adults (i.e., Criterion 1 was met), whereas the third "Rest" subgroup (N = 15, 18%) included a minority.
Bayesian analyses indicated that scores on most cluster variables pointed towards similarity over time, showing that Criterion 2 was met (Supporting information S6; see S4 for results of the total autism group).For the HighGr subgroup, there was evidence in favor of the null hypothesis (i.e., similar scores on cluster variables over time) for most cluster variables, although it did not meet the threshold for "moderate evidence".For variables related to autism characteristics, there was evidence in favor of the alternative hypothesis: a decrease in difficulties with social skills, and imagination, and an increase in attention switching.For the LowGr subgroup, there was either anecdotal or moderate evidence in favor of the null hypothesis for eight cluster variables.There was moderate evidence in favor of the alternative hypothesis for mastery (increase), and a decrease in worries, and negative life events.
For the stability of subgroup membership, a Bayesian test of association produced a Bayes factor of 50.8 (very strong evidence) in favor of the alternative hypothesis (i.e., there was an association between subgroup membership at baseline and five-year follow-up: Criterion 3 was met).Specifically, from the HighGr subgroup 21 (57%) autistic adults remained in this subgroup, whereas 12 (32%) switched to the LowGr subgroup.From the LowGr subgroup 17 (61%) adults remained in this subgroup, whereas 5 (18%) switched to the HighGr subgroup.Twelve participants from the HighGr subgroup, and 20 participants from the LowGr subgroup, were not analyzed at follow-up (due to missing data or drop-out).Detailed percentages regarding subgroup membership  a Subgroups identified at two-year follow-up based on Sample 2; external variables also measured at two-year follow-up.
b Subgroups identified at baseline based on Sample 2; external variables measured at two-year follow-up.
stability can be found in Figure 2d.These switches in subgroup membership were also reflected in lower values on other measures (RI = 0.57, ARI HA = 0.13).Thus, after 5 years (a) the same number of subgroups was identified, (b) scores on less than 50% of the cluster variables were similar (with BF 01 >3) according to Bayesian analyses, (c) stability of subgroup membership over time was at least 57%.
Five-year follow-up: Subgroups still predictive of clinical external outcomes Regarding predictive validation, the results were similar to the two-year follow-up.Subgroup membership established at baseline was predictive of external outcomes measured after 5 years.Specifically, being a member of the LowGr subgroup as compared witrh the HighGr subgroup was associated with more cognitive failures and psychological difficulties, and a lower QoL at follow-up.Detailed statistics of the predictive validation can be found in the Supporting information S8.

DISCUSSION
The goal of this study was to determine the prognostic utility of two previously identified autism subgroups.
This study shows that (a) autistic adults and non-autistic adults formed separate subgroups, (b) the LowGr and HighGr autism subgroups were stable up to two to 5 years after baseline, (c) the subgroups can be used to make clinical predictions over time.When we do not focus on the subgroups, findings from the overall autism group show that most characteristics remain similar with age, some difficulties decrease (i.e., social interaction, communication, and worries), whereas none seem to increase with age.We conclude that the autism subgroups are stable over time based on three pre-registered criteria.First, the community detection analysis again resulted in two major subgroups at two-and five-year follow-up.Second, subgroup profiles (i.e., the average scores on the cluster variables per subgroup) were similar over time on at least half of the cluster variables.Specifically, from baseline to two-year follow-up, scores on 11 out of 14 cluster variables were similar according to Bayesian analyses.From baseline to five-year follow-up, the evidence did not meet the threshold for "moderate evidence", as average scores on only up to four out of 14 cluster variables were similar.The evidence still pointed in the same direction as for the two-year follow-up.Third, subgroup membership was stable from baseline to two-and five-year follow-up according to Bayesian tests of association.This implies that most autistic adults remained in the same subgroup over time.
Although most autistic people retained their subgroup membership over time, this study also shows that switches between subgroups were possible.After 2 years, 20% switched from the HighGr to the LowGr subgroup, whereas 14% switched from the LowGr to the HighGr subgroup.These percentages were even higher after 5 years: 32% switched from the HighGr to the LowGr subgroup, and 18% switched from the LowGr to the HighGr subgroup.In the current study, modifiable factors (e.g., physical activity, social skills) were included intentionally, to ensure that changes over time were a possibility.Thus, although the majority of autistic people retained their subgroup membership over time, changes to a subgroup with a different outcomes may occur.
Besides being stable over time, the autism subgroups showed potential utility for clinical practice.Subgroup membership at baseline was predictive of clinically relevant external outcomes (i.e., cognitive failures, psychological difficulties, and QoL) measured after two to 5 years.Membership of the LowGr subgroup, that was associated with the most vulnerable profile on the cluster variables, was predictive of more cognitive failures, more psychological difficulties and a lower QoL.This was the case even when these outcomes were measured after 5 years.By considering someone's prognosis based on current subgroup membership, we can focus on intervening on associated vulnerabilities to prevent more cognitive failures, more psychological difficulties, and a lower QoL later in life.This study shows which variables-next to autism characteristics-may distinguish these subgroups, namely mastery, worries, emotional support, and affect.As these variables are modifiable in varying degrees, they may be most fruitful for support.Therefore, it may be valuable for future studies to further investigate the potential of these factors for clinical practice.To this end, clinical institutions are recommended to include validated self-report instruments measuring these variables as part of their routine set of outcome measures.Another avenue for future research may be to assess whether the current findings can also be found using objective measures, such as health records.This could further demonstrate the validity of the subgroups.Moreover, the current findings also show which variables do not seem to distinguish the subgroups: education, physical activity, and negative life events.As there was ample variability on these variables, we can cautiously hypothesize that these variables may be less sensitive to differentiating between either diagnosis or autism subgroups.
Although the validity and replicability of the HighGr and LowGr subgroups were demonstrated in several ways, a "Rest" subgroup was also identified at both follow-up occasions that was not considered valid.At the two-year follow-up, 4% of autistic adults were assigned to this rest subgroup, and at the five-year follow-up this percentage was 18%.While the proportion of participants in the rest subgroup became higher after 5 years, we should stress that this subgroup did not replicate (a) in independent samples (i.e., Sample 1 and 2), and (b) over time (see Radhoe et al., 2023 for comparability with the findings at baseline).As the Rest subgroup can't be regarded as a valid subgroup, we should be cautious not to interpret the profile of this subgroup.
This study has several strengths.First, a sample of over 300 autistic adults was included, which was large compared with what is commonly reported in the autism subgrouping literature (Agelink van Rentergem et al., 2021).This sample was followed over a period of two to 5 years, providing valuable longitudinal knowledge on aging with autism, even when we do not focus on the subgroup level.When considering the total group of autistic adults, results indicated that (a) most factors seemed stable over time, (b) some difficulties decreased with age (i.e., worries, and difficulties with social skills and communication), whereas (c) none increased with age.Second, this study included a non-autistic comparison group to assess whether the observed heterogeneity in autism in distinct from variation in non-autistic adults.Third, the validity of the identified subgroups was demonstrated in multiple ways, substantiating the idea that the subgrouping solution was reliable.Fourth, the analysis plan was pre-registered.
There are also some limitations that need to be considered when interpreting the study findings.First, the representativeness of the study sample is restricted to autistic adults with an average to above average intelligence.Moreover, most autistic adults received their autism diagnosis in adulthood.Thus, the findings may not generalize to autistic adults with a below average intelligence, or those who received their diagnosis in childhood.Second, the dropout at follow-up may have influenced the results.Compared with the group of autistic adults that was included at follow-up, the dropout group included relatively more men.In the nonautistic comparison group, those who dropped out had an average lower age and IQ score compared with those included at follow-up.Third, it is important to note that the findings of the current study should be interpreted in light of the COVID-19 pandemic that occurred when the follow-up data were collected.As shown by a recent review, the pandemic had a significant impact on the lives of autistic adults (Scheeren et al., 2023).While there were large differences between autistic adults in the positive and negative effects of the pandemic, an overall decrease in wellbeing during the pandemic was reported.This potential influence on the current study could explain the finding that there were relatively more people switching from the HighGr subgroup to the LowGr subgroup at follow-up than vice versa.This finding applied to both autism samples (i.e., Sample 1 and 2) and suggests that switches were more likely to occur to the subgroup with more experienced difficulties (i.e., LowGr), which could be expected during a pandemic.Nonetheless, the body of research on the effects of the COVID-19 pandemic on those with autism is still limited and findings may be inconsistent across studies.Some might argue that there are limitations to inclusion of participants related to the presence of an autism diagnosis and lack of ID, as these were based on selfreport.However, this is unlikely to reflect a true limitation.Regarding autism diagnosis, in the Netherlands, this diagnosis is usually made by a multidisciplinary team rather than an individual clinician.Moreover, previous research has shown that self-reported diagnoses correlate highly with diagnoses that were independently verified (Daniels et al., 2012;Lee et al., 2010).Furthermore, we administered the AQ to all autistic participants.Two trained researchers also administered the ADOS-2 (Module 4) to a subsample of our autistic participants.The autistic participants in this subsample who met the cut-off for an ASD classification according to the ADOS-2 (N = 79) did not differ on the AQ from the participants to whom we did not administer the ADOS-2.Therefore, we argue that the self-reported autism diagnoses are reliable and valid.Regarding lack of ID, a part of the sample was also tested using the WAIS-IV.There were no exclusions based on the IQ-score resulting from the WAIS-IV, which underscores our confidence that a lack of self-reported ID was reliable in our study.
Finally, it should be acknowledged that longitudinal studies, while having many benefits over cross-sectional studies, also involve certain challenges.For example, changing instruments across waves of data collection are problematic, as it decreases comparability across waves.Moreover, it is often not feasible to include additional instruments in longitudinal studies due to the extra time and effort this would require from study participants.Therefore, in the process of developing new instruments, it is important that abbreviated versions are developed and validated to reduce time of participation, especially for use in longitudinal studies.
This study highlights the stability of two previously identified subgroups of autistic adults in terms of profiles and memberships, and demonstrates their predictive value for clinical outcomes measured up to 5 years in time.Although subgroup membership seems generally stable over time, switches to a subgroup with a different (perhaps more favorable) outcomes were possible.Therefore, even for those autistic adults who are susceptible to more day-to-day difficulties, aging does not inevitably lead to a less favorable outcome.Further considering these autism subgroups and focusing on their potential for clinical practice could be valuable to improve the lives of people on the autism spectrum.
subtests Vocabulary and Matrix Reasoning).The WAIS-IV was administered to 313 participants at baseline (N autism = 165, N comparison = 148), 170 at two-year follow-up (N autism = 80, N comparison = 90), and 46 at five-year follow-up.The ADOS-2 was only administered in person, and the WAIS-IV was administered in person or online.This study was approved by the local ethical review board of the department of Psychology at the University of Amsterdam (2018-BC-9285).

F
I G U R E 1 (a) Flow diagram of participant numbers.(b) Schematic representation of data collection timepoints per sample.AUT, autism; COMP, comparison.

F
I G U R E 2 (a) Profiles of the three autism subgroups based on Sample 2 at two-year follow-up.(b) Stability of subgroup membership from baseline to two-year follow-up based on Sample 2. (c) Profiles of the three autism subgroups based on Sample 1 at five-year follow-up.(d) Stability of subgroup membership from baseline to five-year follow-up based on Sample 1. HighGr = Feelings of High Grip.LowGr = Feelings of Low Grip.Edu = education, Soc = social skills.AttS = attention switching.AttD = attention to detail.Com = communication.Imag = imagination.Sens = sensory sensitivity.Mas = mastery.Wor = worry.Sup = emotional support.Phys = physical activity.PAff = positive affect.

F
I G U R E 3 Bayes Factors (with standard/flat prior) indicating stability of cluster variable scores per subgroup over a two-year interval.Bayes Factors under 1 indicate evidence for a difference over time, a value of 1 represents no evidence, and values above 1 indicate evidence for stability of cluster variable scores.HighGr = Feelings of High Grip, LowGr = Feelings of Low Grip, Edu = education, Soc = aocial skills, AttS = attention switching, AttD = attention to detail, Com = communication, Imag = imagination, Sens = sensory sensitivity, Mas = mastery, Wor = worries/fears, Sup = emotional support, Phys = physical activity, PAff = positive affect, NAff = negative affect, NLife = negative life events.T A B L E 2 External and predictive validation measures for the two autism subgroups based on Sample 2