Data and sample

This report is based on an analysis of data from an ongoing longitudinal study of families of autistic adolescents and adults [33]. The study began in 1998 with families of 406 adolescents and adults diagnosed with autism living in Massachusetts and Wisconsin, and to date it has extended over 22 years. All participating families initially met three inclusion criteria: (1) the family had a son or daughter with an autism diagnosis given by an educational or health professional, (2) the proband was age 10 or older, (3) a researcher-administered Autism Diagnostic Interview-Revised (ADI-R) [34] profile was consistent with the diagnosis. When the study began, almost all (94.6%) met criteria for a diagnosis of autistic disorder. The remaining 22 cases (5.4%) were determined to have ADI-R profiles consistent with a diagnosis of Asperger’s disorder or Pervasive Developmental Disorder-Not Otherwise Specified (PDD-NOS) (see [35]), diagnoses in use at that time. The present analysis used repeated measures that were collected over nine study waves from the primary caregivers, mostly (96%) mothers, via in-home interviews as well as self-administered questionnaires. Although some autistic individuals directly provided data at three of the waves of the study, in the interest of including data from those of all cognitive and communicative ability levels, we use parent-report data for this analysis.

The nine waves of data (here referred to as Time 1 through Time 9) spanned 22 years. Previous reports of analyses from this study reported on shorter spans of time, and were based on fewer rounds of data collection gathered earlier in the life course (e.g., [8, 19, 36]). Although all of the measures reported here were used in our previous reports, here we nearly double the duration of the study period and add three additional repeated measures (Times 6 through Time 9), extending most measures through 2022. Table 1 shows the timing of each wave of data collection along with average ages and age ranges of the probands at each wave. As shown in Table 1, on average, most waves of data collection were approximately 18 months apart except for Time 4 and 5 (about 44 months) and Time 8 and 9 (about 8 years). The average number of study waves for which participants contributed data was six, and more than a half (52.2%) participated in seven or more study waves. Although attrition remains a limitation of this study, those who were lost to attrition did not differ from those who remained in the study through Time 9 with respect to sex, ID status, and the outcome variables. Those who were lost to attrition were an older age at Time 1 (23.1 vs. 19.2 years, p < 0.001) than those who remained.

Table 1 Waves, dates of data collection, and ages of autistic adolescents and adults

The analytic sample consisted of all 406 autistic individuals. Their average age at Time 1 was 21.9 (SD = 9.4), ranging from 10 to 52. The majority were males (73.2%) and over two-thirds (69.7%) had a co-occurring diagnosis of ID. Nearly two-thirds (65.0%) lived with their mothers (and often other family members) at Time 1. During the study, 28 autistic individuals died. The data describing these individuals prior to death are included in the analysis. A larger number of mothers died during the study period (n = 73), resulting in change of reporters from the mother to another family member or in a few cases to a family friend. We conducted a sensitivity analysis including only those cases where the mother was the source of data. It revealed that all of the results presented below were fully replicated (see Supplemental Material), reflecting the selection of measures that were well-validated and more objective than subjective reports of life course patterns.

Although the great majority of the families of the autistic individuals were White non-Hispanic (92.6%), there was significant socioeconomic heterogeneity. Median annual household income at Time 1 was between $50,000 and $60,000. Notably, 11.5% earned less than $20,000 per year, when the US poverty line was $17,050 for a family of four (Federal Register, Vol. 65, No. 31, Tuesday, February 15, 2000). Fewer than half of the mothers had achieved a bachelor’s degree (45.1%), and fully one-quarter (26.8%) had no education beyond high school.

Measures

Bivariate correlations among the measures used in the present research, as well as their associations with age, are presented in Table 2. Although there were significant associations among the measures within a given domain, the associations across domains were generally non-significant to moderate.

Table 2 Bivariate Correlations of Study Variables at Time 1

Measures of autism symptoms included impairments in social reciprocity, impairments in verbal and non-verbal communication, and repetitive behaviors. Measures of behavioral functioning included independence in activities of daily living, maladaptive behaviors, and social participation. Measures of health included ratings of health, number of psychotropic medications prescribed for mental health symptoms, and number of non-psychotropic medications prescribed for physical health symptoms. All measures have been shown to be sensitive to change in prior research [19, 23, 36,37,38]. Not all measures were obtained at all nine timepoints, but all were collected at least six times over the study period. The specific times of data collection for each measure are indicated below.

Measures of autism symptoms

Autism symptoms were assessed using the Autism Diagnostic Interview-Revised (ADI-R) [34] at Time 1 through Time 6. The ADI-R is a standardized diagnostic interview administered to a parent or primary caregiver and used to diagnose autism based on a specified subset of 37 items that constitute a validated algorithm. Our study administered these 37 items at Time 1 to confirm diagnostic status. At each subsequent point of data collection, we administered the 33 items from the core diagnostic algorithm that are applicable to adolescents and adults (4 of the 37 items are specific to childhood). Ratings of current functioning were made at each point of data collection by interviewers who had participated in an approved ADI-R training program. We created four ADI-R sub-scales using 32 of the 33 items based on consultation with one of the instrument’s designers (C. Lord). This grouping of items is based on the clustering of items established by the ADI-R scoring protocol [34], our prior work using this instrument [39], and analysis of the factor structure of the instrument [40]. The sub-scales were impairments in social reciprocity, impairments in verbal communication, impairments in non-verbal communication, and repetitive behavior. A code of 0 signifies the absence of a given symptom, while codes of 1 and 2 indicate impairments characteristic of autism. Some items also used codes of 3, but these are routinely recoded as 2 s [34]. Algorithm items were summed to create the four domain scores. ADI-R items reflecting current levels of impairments in verbal communication were assessed for those individuals who were able to communicate verbally using at least three-word phrases on a daily basis (ADI-R item 30, n = 318), 19 of whom shifted from being classified as non-verbal to using three-word phrases on a daily basis during the study period. The ADI-R has demonstrated good test–retest reliability and validity in past research [34, 41]. In this sample, the internal consistency coefficients (Cronbach’s α) at Time 1 were 0.84, 0.71 and 0.53 for social reciprocity, impairments in non-verbal communication, and repetitive behavior, respectively.

Measures of behavioral functioning Independence in Activities of Daily Living (ADL)

Independence in activities of daily living was measured longitudinally at seven times of data collection (Time 1, Time 4—Time 9) using the Waisman Activities of Daily Living Scale (W-ADL) [37]. Mothers rated the level of independence of their son or daughter with regards to 17 activities of daily living, measuring performance of personal hygiene (e.g., washing/bathing, grooming, toileting), housekeeping (e.g., home repairs, laundry), meal preparation (e.g., preparing simple food, drinking from a cup, washing dishes), and financial management (banking and managing daily finances) on a scale of 0 to 2 (0 = does not do at all, 1 = does it with help, 2 = does independently). Item scores were summed into a total score with higher scores signifying greater independence in daily living skills. For the present sample, scores ranged from 2 through 34. Past research has shown that the W-ADL is strongly correlated (r = 0.82) with the Daily Living scale within the Vineland Screener [8, 37]. The internal consistency (Cronbach’s α) of the W-ADL at Time 1 was 0.903. Criterion validity of the W-ADL for adults with ASD was previously established [37]. The items in the W-ADL span skills generally acquired in early childhood (e.g., drinking from a cup) through those acquired in adulthood (e.g., banking), suggesting good representation of independence in daily living skills across the life course.

Maladaptive Behavior

Maladaptive behavior was measured longitudinally at all nine times of data collection using the Behavior Problems subscale of the Scales of Independent Behaviors-Revised (SIB-R) [42]. The SIB-R measures behavior problems, grouped in three domains: internalized behaviors (hurtful to self, unusual or repetitive habits, withdrawal or inattentive behavior), externalized behaviors (hurtful to others, destructive to property, disruptive behavior), and asocial behaviors (socially offensive behavior, uncooperative behavior). If a given behavior problem was manifested during the past 6 months, then frequency (1 = less than once a month to 5 = 1 or more times/hour) and severity (1 = not serious to 5 = extremely serious) of the behavior were rated by mothers. Standardized algorithms [42] translate the frequency and severity ratings into a General Maladaptive Behavior Index, with higher scores indicating more severe behavior challenges. Reliability and validity have been established by Bruininks et al. [42]. The present analysis uses the General Maladaptive Behavior Index.

Social participation

Social participation was assessed longitudinally at all nine times of data collection. At each time point, mothers reported on the frequency with which their son or daughter spent time with friends or neighbors (0 = once a year or never, 1 = several times a year, 2 = once or twice a month, 3 = once a week, 4 = several times a week), an item drawn from the National Survey of Families and Households (www.ssc.wisc.edu/nsfh/).

Measures of health Health Ratings

Health ratings were obtained at all nine times of data collection. Mothers rated the health of their son or daughter (1 = poor, 2 = fair, 3 = good, 4 = excellent). Considerable previous research has provided evidence of the validity of such health ratings in predicting mortality [43, 44]. In prior analyses of data from the present study, this measure of health was found to significantly predict mortality over the course of two decades [45].

Number of Prescription Medications

As a separate and objective indicator of physical health, at each time point, mothers listed names of all prescription medications currently taken by their son or daughter along with dosage and reason for taking each medication [23]. Medications were separated into psychotropic and non-psychotropic categories. As shown in Table 2, self-rated health was not significantly associated with the number of psychotropic medications, whereas there was a significant negative association between self-rated health and the number of non-psychotropic medications (r = -0.115, p < 0.05).

Psychotropic medications were prescribed for mental health problems and included antipsychotics, antidepressants, anxiolytics and sedative-hypnotics, central nervous system (CNS) stimulants, antimanic medication, anticonvulsant medications that were prescribed to an individual with no comorbid diagnosis of epilepsy or seizures (usually for bipolar symptoms), and hypotensive medications that were prescribed to an individual with no comorbid diagnosis of hypertension.

Non-psychotropic medications were prescribed for physical health problems and included anticonvulsants (for seizures), antiparkinson medications prescribed for side effects of antipsychotic medications (i.e., not prescribed to a person diagnosed with Parkinson's disease), antiemetics, and medications for hypertension, thyroid, diabetes, respiration, hormones, ocular, gastrointestinal (GI), and other miscellaneous purposes. Excluded from our analysis were over-the-counter medications, such as analgesics, laxatives, vitamins, antifungal medication, antacids, and topicals.

Medications were coded and classified based on Physician’s Desk Reference Drug Guide for Mental Health Professionals [46]. These classifications were reviewed and verified by a university-based pharmacist with over 20 years of experience. Separate counts of the number of psychotropic and non-psychotropic prescription medications were used in this report.

Predictor variables

The main predictor variable in the present study was age at each study point. Age for each individual was calculated from date of birth to the date of that individual’s data collection at each time point. In addition, we included three other predictors: age at Time 1, sex, and ID status. Age at Time 1 was controlled to evaluate whether the trajectories differed depending on the age of the participant when the study began. This variable was included to provide insight into possible cohort effects in the age-related trajectories (for example, whether individuals who were adolescents at the start of the study showed a different pattern of change as compared to those who were in adulthood when the study began). Sex was coded as 0 = male, 1 = female. ID status (0 = no intellectual disability, 1 = intellectual disability) was determined using a variety of sources. Individuals with standard scores of 70 or below on the Wide Range Intelligence Test (WRIT) [47] and the Vineland Screener [48] were classified as having intellectual disability, consistent with diagnostic guidelines [49]. For individuals with scores above 70 on either measure or when either of the measures for the person was missing, clinical consensus among three psychologists was reached to determine their ID status based on a review of medical and psychological records.

Data analysis

We used an accelerated longitudinal design (ALD; also referred to as a cohort-sequential design or cross-sequential design) to estimate trajectories in autism symptoms, behavioral functioning, and health of autistic adolescents and adults. The ALD estimates a single long-term longitudinal trajectory by combining multiple short-term longitudinal trajectories of each individual covering different periods. This way, the ALD makes it possible to estimate a growth trajectory spanning wider age ranges than the duration of time covered by a longitudinal study [50]. The present study spanned 22 years, including individuals as young as 10 and as old as 52 when the study began. We are thus able to estimate age-related trajectories spanning approximately 60 years.

The trajectories were estimated for each variable separately in order to evaluate their potentially unique age-related functions. As the present study analyzed data spanning the longest available period during adolescence and adulthood, our goal was to determine for each indicator whether there were age-related increases or decreases, whether the change was linear or curvilinear, and whether the trajectories differed between those who have ID and those who have average or above-average intellectual functioning. Subsequently, the slopes of the trajectories were evaluated to determine if they differed depending on age when the study began, with the goal of identifying possible cohort effects.

To assess linear and quadratic trajectories of each dependent variable, we estimated the mixed-effects growth curve models with polynomial functions of age. The Level-1 equation of the quadratic growth model is:

$$_ = _ + _(_-10) + _(_ - 10)}^ + _,$$

where Yit is a dependent variable for person i at time t, t = 1,…, Ti, where Ti is the number of observations for person i; ait represents age of person i at time t; π0i represents the level of a measure at age 10 (intercept) for person i; π1i represents the instantaneous linear change at age 10 for person i; π2i represents the acceleration in each growth trajectory for person i; eit is the random within-person error for person i at time t. For linear growth models, π2i is set as “0” and π1i represents the linear growth rate for person i at time t. For Level-2 equations, the person-level covariates – age at Time 1, sex, and ID status – were added to predict the baseline (intercept) differences of measures, with age variables (linear and quadratic terms) of person i specified as random. Then, to evaluate effects of ID status on these trajectories, cross-level interaction terms between ID status and age were added to the models.

For each measure, we estimated four models, each building on the previous one. Model 1 estimated the linear age effect and also included the age of the autistic individual when the study began, sex, and co-occurring ID status. For Model 2, to assess whether age was best estimated as a linear or curvilinear function, we added a term of age-squared to the variables in Model 1. In Models 3 and 4, to assess whether the age-related linear or curvilinear trajectories differed by ID status, we additionally included interaction terms (age X ID in Model 3 and age-squared X ID in Model 4). Following the approach of Joiner, Bergeman, and Wang [51], the best model was selected based on log-likelihood tests for nested models (Model 1 versus Models 2, 3, and 4; Model 2 versus Model 4; Model 3 versus Model 4), and Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) values for non-nested models (Models 2 versus Model 3). We report the results of the best model. Once the best model was selected, we tested the significance of random effects components – random intercept variance and random slope variance – using likelihood-ratio tests. Note that the random effects components of each best model showed that there were significant variations in initial levels of all outcomes among individuals with ASD even after the baseline age and the ID status were controlled. Similarly, there were significant between-individual differences in the age-related trajectories for all outcomes except for impairments in non-verbal communication, as we report below. Following these main analyses, we also conducted an exploratory analysis with each best fitting model to probe for potential cohort effects.

All analyses were conducted using Stata version 17.0 [52]. The level of significance was set at equal to or less than 0.05. The examination of distributions of the variables yielded no evidence of skewed distribution, with skewness ranging from -0.61 to 1.26. For the two count variables (the number of prescription psychometric and non-psychometric medications), additional analyses using random intercept Poisson regressions were also performed. Since the results were similar to those from the growth curve models treating the outcomes as continuous variables, we reported the results based on the continuous variable analyses.

We included all individuals in the study sample in the analysis, regardless of the number of data points they contributed. Those with just one data point contributed to the estimation of the intercepts and the effect of age at the start of the study on the intercept, but they did not contribute to estimates of the trajectories. With Maximum Likelihood (ML) estimation, the greater the number of data points, the more influential the case was to the estimation of the age-related trajectories. By fitting our models in a multilevel framework (as opposed to repeated measures ANOVA, for example), we were able to retain all observed data and accommodate missing observations under the same missing at random (MAR) assumptions as are adopted when using full information maximum likelihood (FIML) estimation in other modeling contexts. For descriptive purposes, the individual trajectories of each participant are portrayed in figures that were generated by the equation in the best model for each indicator of autism symptoms, behavioral functioning, and health.