Children with ASD and Communication Regression: Examining Pre-Loss Skills and Later Language Outcomes Through the Preschool Years

Abstract

This study investigated receptive and expressive language outcomes in children with autism spectrum disorder (ASD) with and without a history of language/communication regression, employing three progressively less stringent definitions of regression. Data were derived from a large, longitudinal sample of children with ASD in which regression was assessed at approximately 30 months. Results indicated poorer receptive language and larger discrepancies between receptive and expressive language in the regression group than the group without regression at 44 months but not 66 months. Number of words used before loss predicted receptive language at 44 months. Overall, results suggest that a regression profile in ASD is associated with modest and transient impacts on language outcomes that are no longer discernable at school entry.

Apparent loss of previously attained developmental milestones, known as regression, is an early cause of concern for many parents of children with autism spectrum disorder (ASD). Given the high frequency of this phenomenon in ASD, a considerable body of literature has emerged over recent years examining regression and its consequences for development. Yet even as regression has become better understood in the context of early development in ASD, several important questions remain regarding the nature of these skill losses, pre-loss development, and the relationship between early skill loss and long-term outcomes - particularly in the domain of language and communication.

The main goal of this longitudinal investigation was to examine regression in language and preverbal communication in toddlers with ASD (30 months on average) in order to compare outcomes in receptive and expressive structural language (vocabulary/grammar) approximately 1 year and 3 years later. Additional aims of this study were to determine whether number of words before loss predicted language outcomes and to characterize language/communication skills prior to loss. As background, we will begin by reviewing the extant literature on regression in ASD including definitional criteria, language outcomes, and the limitations of previous work addressed by the current study.

Characterizing Regression in ASD

Regression can occur in many developmental domains, including social communication, motor ability, adaptive skill, language, or a combination thereof (Davidovitch, Glick, Holtzman, Tirosh, & Safir, 2000; Goldberg et al., 2003; Matson, Wilkins, & Fodstad, 2010). But among children with ASD who experience regression, language appears to be the most common domain in which losses occur (94%; Estabillo, Matson, & Cervantes, 2018; 74.3%; Matson et al., 2010). Studies involving clinical samples that have used retrospective parent measures with stricter criteria, such as that of the Autism Diagnostic Interview – Revised (ADI-R; Rutter, Le Couteur, & Lord, 2003), with a requirement of loss of 5 words used communicatively for 3 months for language regression, have found the prevalence of regression in children with ASD to be within a range of 20-50% (Barger, Campbell, & McDonough, 2013). When defined more broadly as a loss of any skill, prevalence increased to 63% (Thurm, Manwaring, Luckenbaugh, Lord, & Swedo, 2014). Prospective studies and those employing home video review methodologies – which defined regression as a reduction in the frequency of social communication behaviors across time beyond what is expected of neurotypical peers - reported much higher incidence of skill loss (69-88%; Ozonoff et al., 2018, 2010). Thus, certain researchers have suggested that some early skill loss may occur in almost all cases of ASD and represent one end of a more dimensional continuum of ASD symptom onset patterns (Boterberg, Charman, Marschik, Bölte, & Roeyers, 2019; Ozonoff et al., 2018; Szatmari et al., 2016; Thurm, Powell, Neul, Wagner, & Zwaigenbaum, 2018; Zwaigenbaum, Bryson, & Garon, 2013). Among previous studies of regression in the communication domain, definitional criteria has been limited to loss of spoken words in keeping with the ADI-R. As such, many children who may have experienced language or communication skill loss at levels below the ADI-R threshold, such as those who are delayed in achieving first words but do demonstrate preverbal communication skills, have been excluded from regression groups in previous studies (Pearson et al., 2018). Thus, it remains unknown to what extent children might also experience losses in preverbal communication skills such as cooing, babbling, and vocal imitating. One goal of the current study is to better characterize regression in the communication domain overall, including among preverbal skills.

Language Outcomes

Because language losses occur most frequently (Estabillo et al., 2018; Matson et al.,2010) and are perhaps the most observable and memorable domain for parents (Luyster et al., 2005), many studies have defined regression by specific language loss criteria. Yet, relatively few studies have examined language as an outcome measure. Among those that did, the evidence for group-level differences in language outcomes is contradictory, much like the body of literature on regression outcomes generally. The majority reported worse language outcomes in the regression group (Bernabei et al., 2007; Estabillo et al., 2018; Hansen et al., 2008; Kalb et al., 2010; Kobayashi & Murata, 1998; Norrelgen et al., 2015; Rogers & DiLalla, 1990). Some studies found no differences (Pickles et al., 2009; Tamanaha et al., 2014), and one reported better current levels of verbal communication in the regression group (Davidovitch et al., 2000). No regression group comparison studies to date have analyzed receptive language skills as an outcome measure, a gap addressed by the current study.

In addition to the limited focus on language as an outcome measure in general, very little is known about differences in language outcomes over time in children with ASD with and without history of regression. In ASD regression studies examining later language outcomes, the majority reported data from a single time point (Davidovitch et al., 2000; Estabillo et al., 2018; Hansen et al., 2008; Kalb et al., 2010; Pickles et al., 2009; Rogers & DiLalla, 1990; Tamanaha et al., 2014). Only three studies to date have reported language outcome data across multiple time points (Bernabei et al., 2007; Kobayashi & Murata, 1998; Norrelgen et al., 2015). Moreover, interpretation of findings in these studies is limited due to the nature of the language outcome measures that were used. Norrelgen et al. (2015) relied upon Vineland Adaptive Behavior Scales 2^nd Edition (VABS-II) parent interview data as a language outcome measure, while Bernabei et al. (2007) utilized a parent interview measure and rating scales developed in their clinic. Kobayashi and Murata (1998) conducted clinical evaluations to rate current levels of communication as “good,” “fair,” or “poor,” with an unspecified assessment tool. In defining language regression, each of the three studies relied on parent interview and/or medical records and a strict operational definition of regression, potentially obscuring dimensional aspects of skill loss and further limiting interpretation of language outcomes in the extant literature.

Contributions of the Current Study

The present study addresses several gaps in the extant literature on regression in ASD. First, while prior studies have attempted to include subthreshold word losses in their regression groups (e.g., Goin-Kochel et al., 2014), they did not report language outcomes. No group comparison studies to date have included loss in preverbal communication skills such as cooing, babbling, and vocal imitation in their regression criteria. As such, it remains unknown whether children who have experienced word loss below the ADI-R threshold demonstrate differences in later language outcomes, or whether children who experience loss in preverbal communication skills develop differences in later language outcomes compared to children who experience word loss or those with no communication loss. A crucial contribution of the present study is the inclusion of a broader spectrum of skills within the communication domain (including preverbal communication skills and word loss of any amount) within our analyses. These skill losses that fall below the ADI-R the present study addresses the concerns of some researchers (Boterberg et al., 2019; Ozonoff et al., 2018; Ozonoff & Iosif, 2019; Szatmari et al., 2016; Thurm et al., 2018; Williams, Brignell, Prior, Bartak, & Roberts, 2015) that strict language regression criteria may under-identify or over-simplify what may be a more dimensional onset pattern in ASD. By including preverbal communication skills within our criteria, we are able to analyze a more representative group of children with losses in the communication domain. While we will continue to use dichotomous operational definitions and categorical variables in our analyses consistent with previous literature and with the ADI-R coding scheme, the present study will begin to address the need for a more dimensional approach through multiple levels of analyses using incrementally less strict loss criteria. In an additional attempt to address this need, this study will examine the ADI-R Toddler Module item 37 (‘words used prior to loss’) as a continuous predictor variable to answer our second research question. Finally, in order to address the call to better define pre-loss levels of communicative function (i.e. Boterberg et al., 2019) and include data on individual ADI-R loss items (Pearson et al., 2018), this study will include descriptive statistics detailing the level of communication achieved prior to onset of regression in the ASD population as defined by individual ADI-R loss items.

Second, this study will add needed refinement to the body of literature examining language outcomes by measuring both receptive and expressive language outcomes and employing a developmentally-appropriate measure of regression. Previous retrospective regression studies employed the ADI-R, which was designed for use in individuals with a cognitive age of at least 24 months (Rutter, Le Couteur, & Lord, 2003) which would exclude many preschool-age children with ASD. While newer algorithms have improved the diagnostic validity of the ADI-R for children aged 12-47 months, many previous regression studies were published prior to the release of the new algorithms (Kim & Lord, 2012) when the instrument had lower sensitivity in children under 3 years (Ventola et al., 2006). In the present study we employed the Toddler Module of the ADI-R. The Toddler ADI-R is a standardized parent report measure including many of the same items as the standard ADI-R, but the Toddler Module was developed specifically for research purposes with children under age 4 and includes 32 additional questions regarding early development and onset of ASD symptoms not present in the standard version (Kim, Thurm, Shumway, & Lord, 2013).

Third, all children were between 23-39 months of age when the ADI-R Toddler Module was administered, minimizing the telescoping effects to which prior language outcome studies were susceptible (Kobayashi & Murata, 1998; Norrelgen et al., 2015). Telescoping refers to the increasing inaccuracy with which caregivers report the timing of children’s early milestones over time (Ayhan & Işiksal, 2004; Hus, Taylor & Lord, 2011; Lord, Shulman, & DiLavore, 2004). Previous language regression studies administered the ADI-R years after the regression occurred, increasing the risk of telescoping effects (Ozonoff, et al., 2018). Because parents were interviewed closely following the average age that regression occurs (24 months; Backes, Zanon, & Bosa, 2017; Bradley, Boan, Cohen, Charles, & Carpenter, 2016), the present study reports a more accurate measure of children’s regression than that of previous studies collected at later ages.

Finally, while three studies have examined language outcomes following regression across multiple time points, (Bernabei et al., 2007; Kobayashi & Murata, 1998; Norrelgen et al., 2015), this is the only study to date to present language outcome data from multiple time points using a standardized assessment specifically designed to measure children’s language skills. As a result, this study provides a more interpretable measure of language outcomes with an assessment familiar to many clinicians who work with preschool children with ASD.

To address the gaps identified in the literature to this point, the current study poses three primary research questions:

1. Do children with ASD and a history of language/communication regression demonstrate different language outcomes across the preschool years from children with ASD without a history of language/communication regression?

2. Among children with ASD with a history of language/communication regression, does the number of words used before loss predict language outcomes across the preschool years?

3. What levels of language skill do children with ASD and history of language regression achieve prior to regression onset?

Methods

Participants

The data for this study were initially collected as part of a larger longitudinal project examining language outcomes in 129 preschool children with ASD, approved by the Education and Behavioral/Social Sciences Institutional Review Board at the University of Wisconsin-Madison (e.g., Davidson & Ellis Weismer, 2014; Ellis Weismer & Kover, 2015; Ray-Subramanian & Ellis Weismer, 2012; Venker et al., 2014). Participants were recruited through the community, local clinics, and early intervention programs and parents provided written consent prior to study participation. Participants were screened via parent report prior to study enrollment. Exclusionary criteria included speaking any language(s) besides English in the home and/or history of chromosomal abnormalities, cerebral palsy, prematurity, multiple birth, or seizure disorder (see Table 2 for sample demographic information and participant characteristics at visit 1). At the time of recruitment, participants either had an existing ASD diagnosis or had been flagged by parents or professionals as demonstrating behaviors consistent with ASD. All participants had ASD diagnoses confirmed by our study team prior to data collection. In order to be included in the sample, each participant had to meet DSM IV-TR (American Psychiatric Association, 2000) diagnostic criteria for ASD at each visit (which was the current diagnostic standard at the initiation of the longitudinal study). ASD diagnoses were determined by an experienced, licensed psychologist based on comprehensive evaluations that included the ADI-R Toddler Module, the Autism Diagnostic Observation Schedule or Autism Diagnostic Observation Schedule Toddler Module (ADOS or ADOS-T; Lord, Rutter, DiLavore, & Risi, 2002; Luyster et al., 2009) and best estimate clinical diagnosis. Best estimate diagnosis is determined by applying expert clinical opinion to the current autism diagnostic classification system, using behavioral observations during assessment and interaction with participants, parent report, and assessment results. The race/ethnicity composition of the full sample was as follows: 86% White, 2% Black, 3% Hispanic, 2% Native American, and 8% Other. See Table 2 for participant demographics and clinical characteristics within the entire sample and each regression group. Three participants were excluded from analyses for this study due to missing data at visit 1, resulting in a sample of 126. Due to attrition between visits, language outcome data were available for 111 participants at visit 2 and 98 participants at visit 4.

Table 2.

Descriptive statistics of demographic information and clinical characteristics across entire sample and broken down by participants who met loss criteria for 1) Definite Word Loss, 2) Any Word Loss, and 3) Any Communication Skill Loss

Variable	Entire Sample N = 126	Definite Word Loss n = 35	Any Word Loss n = 58	Any Communication Skill Loss n = 61
Gender
Male	110 (87%)	30 (86%)	49 (84%)	51 (84%)
Female	16 (13%)	5 (14%)	9 (16%)	10 (16%)
Race/ethnicity
White	108 (86%)	27 (77%)	61 (90%)	51 (84%)
Black	2 (2%)	1 (3%)	1 (1%)	1 (2%)
Hispanic	4 (3%)	3 (9%)	4 (7%)	4 (7%)
Asian	0 (0%)	0 (0%)	0 (0%)	0 (0%)
Native American	2 (2%)	1 (3%)	2 (3%)	1 (2%)
Other	10 (8%)	3 (9%)	4 (7%)	4 (7%)
Maternal Education
≤12 years	42 (33%)	17 (49%)	25 (43%)	25 (41%)
13-16 years	67 (53%)	14 (40%)	25 (43%)	29 (48%)
17-20 years	17 (14%)	4 (11%)	8 (14%)	7 (11%)
	M (SD)	M (SD)	M (SD)	M (SD)
Age	30.80 (4.08)	31.24 (3.85)	31.23 (3.78)	31.03 (3.84)
ADOS CSS	7.59 (1.92)	7.24 (2.15)	7.39 (2.09)	7.52 (2.13)
NVIQ	76.73 (14.48)	80.16 (16.45)	77.80 (15.19)	77.58 (15.01)
Speech-Language Therapy
Hours per Month	3.35 (1.65)	3.46 (1.75)	3.47 (1.83)	3.45 (1.81)
Cumulative Months	11.24 (6.17)	12.35 (6.00)	10.57 (6.13)	10.90 (6.06)

Note: ADOS CSS = Autism Diagnostic Observation Schedule Calibrated Severity Score (N=125), NVIQ = Mullen Scales of Early Learning Nonverbal Ratio IQ (N=108), Speech-Language Therapy received in Early Intervention program as reported at visit 2, Hours per Month N = 37, Cumulative Months N = 71

Procedure

Participants were administered a battery of autism diagnostic measures, cognitive, and language assessments at each of 3-4 visits, a subset of which were included in analyses for the present study. Each of the assessment measures was the most current version available at the time of data collection. Visits were scheduled approximately 12 months apart and occurred at the following mean chronological ages (in months): 30.8 (SD=4.1, range=23-39; visit 1), 44.1 (SD=4.1, range=37-53; visit 2), 56.8 (SD=4.7, range=49-66; visit 3), and 66.5 (SD=4.9, range=57-79; visit 4). The current paper reports data on regression from visit 1 and language outcome data from visits 2 and 4.

Measures

The toddler version of the ADI-R was administered by a trained psychologist to caregivers of each participant at visit 1. Like the standard ADI-R, the ADI-R Toddler Module is a standardized, semi-structured, clinician-administered diagnostic caregiver interview and is considered a gold standard diagnostic measure for ASD (Rogers, Young, Cook, Giolzetti, & Ozonoff, 2010). Individual loss items from the ADI-R were used to operationally define regression in three ways (see Table 1) and to describe level of communication achieved prior to regression onset in order to answer the second and third research questions. To answer the second research question, ADI-R Toddler Module item 37 (“number of words used before loss”), coded numerically, was utilized as the predictor variable. For the third research question, descriptive data was calculated using both ADI-R toddler module item 37 and level of communication achieved prior to loss, which was derived from the ADI-R toddler module item 36 (“level of communication before loss,”) which is coded with the following criteria: 0=daily, spontaneous, and meaningful speech used communicatively, with at least 3 different words used at some point before change, 1=occasional and/or fewer than 3 words used spontaneously and communicatively (alone or in combination with imitative abilities), 2=produced speech or sounds upon request (may or may not have also spontaneously imitated), 3=spontaneous imitations of vocalization (without ever having any completely spontaneous speech) with no elicited imitation or spontaneous communicative speech, 8=no change or loss, or 9=not known or not asked.

Table 1.

Schema for operationally defining language regression using ADI-R Toddler Module specific loss questions

Level of regression	Criteria
Definite Word Loss (per ADI-R criteria)	Definite loss of 3 or more words; defined by score of 2 on ADI-R Toddler Module Q34	Strict ↓ Broad
Any Word Loss	Word loss of any amount; defined as score of 2 or 1 on Q34 and/or Q42 and score of <0 on Q37
Any Communication Skill Loss	Definite loss of any communication skill; defined as meeting criteria for definite word loss or any word loss or score of 2 on any of the following preverbal communication skill loss items: Q38, Q39, Q40, Q41

The Mullen Scales of Early Learning (MSEL; Mullen, 1995) was administered to participants at visit 1 by a trained psychologist. The MSEL is a standardized developmental assessment for children ages birth-68 months with five scales measuring specific developmental domains including fine motor, gross motor, visual reception, expressive language, and receptive language. The scales measuring nonverbal skills (fine motor and visual reception) were used to calculate nonverbal ratio IQ scores for each participant. The nonverbal ratio IQ scores were derived by comparing nonverbal mental age equivalent to chronological age and were utilized in the present study as a proxy for nonverbal cognitive ability (Farmer, Golden, & Thurm, 2016). MSEL data was available for 108 participants in the sample.

Participants’ language skills were assessed at each visit through administration of the Preschool Language Scale – Fourth Edition (PLS-4; Zimmerman, Steiner, & Pond, 2002) by a certified speech-language pathologist. The PLS-4 is a standardized language assessment with two subscales: Auditory Comprehension (AC), a measure of receptive language skill, and Expressive Communication (EC), a measure of expressive language skill. In order to avoid the floor effects associated with the use of standard scores in assessment of language skills in young children with ASD (Volden et al., 2011), PLS-4 raw scores were utilized in analyses. PLS-4 AC and EC raw scores from visits 2 and 4 comprised language outcome data in the present study.

Defining Regression

The standard ADI-R requires a reported loss of 5 or more meaningful words not including ‘mama’ or ‘dada,’ used for at least 3 months then lost for at least 3 months (Rutter, Le Couteur, & Lord, 2003) in order to code a child as having experienced language regression. With the precedent set by studies employing ADI-R criteria, regression has long been conceptualized within a dichotomous framework. However, researchers have criticized this operationalization as failing to capture what may be a more continuous, dimensional symptom onset pattern (Jones et al., 2014). In order to begin to address this criticism while also aligning with prior literature, the present study employed three progressively less strict dichotomous operational definitions of regression within the language domain (see Table 1 for definition schematic). All analyses were repeated for each of these three sets of criteria.

For the first definition, participants were classified as having experienced a regression if they were coded as 2 (“definite loss of 3 or more words (not including ‘mama’ or ‘dada’”) for at least 1 month”) on item 34 (“loss of language skills”) on the ADI-R Toddler Module. These participants were determined to meet criteria for Definite Word Loss (n = 35). While strict, we chose to include this first set of criteria because it is among the most commonly used in previous studies (Backes et al., 2017). By including this set of criteria in our analyses, we sought to make our findings directly comparable to similar studies in the extant literature.

Next, in order to capture subthreshold language losses, all participants in the sample were reclassified according to less strict word loss criteria. We defined subthreshold language loss as spoken word loss below the defined criteria set by the ADI-R. That is, a loss of fewer than 3 words. Participants were determined to have experienced Any Word Loss (n = 58) if they were coded as 2 (“definite loss of 3 or more words (not including ‘mama’ or ‘dada’”) or 1 (“probable loss of specified skill, including language (<3 words or not clear loss) or communication skills”) on ADI-R Toddler Module item 34 (“loss of language skills”) and/or if they were coded as 2 (“definite loss of specified skill”) or 1 (“probable loss of specified skill”) on item 42 (“loss of spontaneous use of at least 3 meaningful words”) and were coded as <0 on item 37 (“number of words used before loss”). This group included all participants previously defined as having Definite Word Loss in addition to those who experienced word loss below the standard ADI-R criteria.

Finally, all participants in the sample were reclassified according to the broadest loss criteria in order to account for preverbal skill losses in the communication domain that may occur in nonverbal children, such as cooing, babbling, vocal imitation, and non-word vocalizations. Participants were considered to have experienced definite Loss of Any Communication Skill (n = 61) if they met criteria for Definite Word Loss or Any Word Loss or were coded as 2 (“definite loss of specified skill”) on any of the following ADI-R Toddler Module items related to pre-verbal communication skills: 38 (“loss of cooing”), 39 (“loss of vocalizations”), 40 (“loss of speech-like babbling”), 41 (“loss of vocal imitation”). This group included all participants previously defined as having Definite Word Loss, and/or Any Word Loss in addition to those who experienced loss of a preverbal communication skill.

Results

No significant differences were found between loss and no loss groups in participant demographics (χ² range from .11-7.47; ps ranged from .08-.74) or clinical characteristics (ts range from −1.57-1.27; ps ranged from .12-.68) at visit 1 for Definite Word Loss, Any Word Loss, or Any Communication Skill Loss regression criteria (descriptive data are summarized in Table 2).

In order to answer the first research question, analyses of covariance (ANCOVA) models were fit comparing PLS-4 language outcomes from visit 2 and visit 4 between the loss and no loss groups for each set of regression criteria, with child age and MSEL nonverbal ratio IQ (NVIQ) scores from visit 1 included as covariates. Although there were not statistically significant group differences in NVIQ (p=.12-.57), the groups were not well matched on this variable (see Kover & Atwood, 2013). Further, since we were using raw scores it seemed appropriate to adjust for age differences. Use of these covariates that have been shown to be related to language abilities would adjust for subtle differences that may have affected differences in language outcomes across the loss and no loss groups. To investigate collinearity among our predictors, we conducted Variance Inflation Factor (VIF) analyses. Across all models, VIF values (range = 1.01-2.01) were well below the widely accepted cutoff of <10 (Salmerón Gómez, García Pérez, López Martín, & García, 2016). Across all analyses, both child age and NVIQ accounted for a significant portion of the variance in language outcomes (ps<.001) for each of the three groupings. For the first models, the strictest regression criteria (Definite Word Loss) were employed. Results revealed a significant main effect of group (F(1, 91) = 4.93, p = .03) on visit 2 PLS-4 Auditory Comprehension (AC) raw score, with significantly higher scores in the no loss group (EM = 29.63, SE = 1.01) than the loss group (EM = 25.30, SE = 1.66). There were no significant group effects found for any other language outcome measure at visit 2 or visit 4 using Definite Word Loss criteria (statistics reported in Table 3). We then reclassified the sample according to Any Word Loss regression criteria to define the loss and no loss groups. ANCOVAs revealed a significant effect of group on visit 2 PLS-4 AC raw scores (F(1, 91) = 4.49,p = .037) with the no loss group (EM = 30.07, SE = 1.15) outperforming the loss group (EM = 26.38, SE = 1.30). Analyses did not reveal any significant group effects on any other language outcome measure at visit 2 or visit 4 using Any Word Loss criteria (statistics are reported in Table 3). Analyses were repeated once more after reclassifying the sample according to the broadest criteria, Any Communication Skill Loss. ANCOVAs revealed a significant effect of group on visit 2 PLS-4 AC raw scores (F(1, 91) = 6.61, p = .01) with higher scores in the no loss group (EM = 30.44, SE = 1.15) than the loss group (EM = 26.03, SE = 1.27). No significant group effects were found on any other language outcome measure at visit 2 or visit 4 (statistics reported in Table 3).

Table 3.

ANCOVA model results comparing language outcomes, with main effect of regression group per 1) Definite Word Loss, 2) Any Word Loss, and 3) Any Communication Skill Loss criteria, with child age and nonverbal IQ (NVIQ) as covariates in all 3 models

Visit	Outcome Measure	Variable	df	F	P	η P 2
Definite Word Loss
2	PLS-4 AC
		Loss	1,91	4.93	.029*	.05
		Age	1,91	34.85	.000***	.28
		NVIQ	1,91	71.24	.000***	.44
	PLS-4 EC
		Loss	1,91	.59	.445	.01
		Age	1,91	21.48	.000***	.19
		NVIQ	1,91	60.37	.000***	.40
4	PLS-4 AC
		Loss	1,82	1.03	.313	.01
		Age	1,82	15.73	.000***	.16
		NVIQ	1,82	68.77	.000***	.46
	PLS-4 EC
		Loss	1,82	.40	.529	.01
		Age	1,82	15.27	.000***	.16
		NVIQ	1,82	58.64	.000***	.42
Any Word Loss
2	PLS-4 AC
		Loss	1,91	4.49	.037*	.05
		Age	1,91	35.97	.000***	.28
		NVIQ	1,91	69.19	.000***	.43
	PLS-4 EC
		Loss	1,91	.78	.379	.01
		Age	1,91	21.92	.000***	.19
		NVIQ	1,91	60.70	.000***	.40
4	PLS-4 AC
		Loss	1,82	.34	.561	.00
		Age	1,82	16.07	.000***	.16
		NVIQ	1,82	67.72	.000***	.45
	PLS-4 EC
		Loss	1,82	1.33	.253	.02
		Age	1,82	16.38	.000***	.17
		NVIQ	1,82	60.30	.000***	.42
Any Communication Skill Loss
2	PLS-4 AC
		Loss	1,91	6.61	.012*	.07
		Age	1,91	36.49	.000***	.29
		NVIQ	1,91	70.94	.000***	.44
	PLS-4 EC
		Loss	1,91	1.63	.206	.02
		Age	1,91	22.08	.000***	.20
		NVIQ	1,91	61.60	.000***	.40
4	PLS-4 AC
		Loss	1,82	.75	.390	.01
		Age	1,82	16.13	.000***	.16
		NVIQ	1,82	68.29	.000***	.45
	PLS-4 EC
		Loss	1,82	.67	.416	.01
		Age	1,82	15.69	.000***	.16
		NVIQ	1,82	59.48	.000***	.42

To further examine our first research question, a secondary analysis was completed using only the first definition of regression (Definite Word Loss). An ANCOVA was conducted comparing the difference between PLS-4 AC and PLS-4 EC raw scores between the Definite Word Loss group and no loss group while controlling for child age and nonverbal IQ. Results indicated a significant effect of group on comprehension-production difference scores at visit 2 (F(1,90) = 7.70,p = .007, ${\eta }_p^2=.08$η_ρ²= .08), with a larger discrepancy in the Definite Word Loss group (EM= −6.44, SE = 1.07) than the no loss group (EM= −3.00, SE = .64). By visit 4, there was no longer a significant group difference in expressive-receptive language profile (F(l,82) = .41,p = .53, ${\eta }_p^2=.005$η_ρ² = .005).

To address the second research question, linear regression models were fit predicting PLS-4 raw scores from visits 2 and 4 from number of words used before loss including child age and MSEL nonverbal IQ scores as covariates. Only participants who met criteria for Definite Word Loss, Any Word Loss, and Any Communication Skill Loss were included in these analyses. As with the first research question, a separate model was fit for each of the three sets of criteria. Among participants who met criteria for Definite Word Loss, results indicated that number of words used before loss was not predictive of any language outcome measure at visit 2 or visit 4, although NVIQ did significantly predict PLS-4 AC and EC raw scores at visit 2 and 4 (ps<.01-.001). Child age was also found to significantly predict PLS-4 AC raw scores at visit 2 (p=.02) and 4 (p=.05) for the Definite Word Loss group. Number of words used before loss was not found to be predictive of any language outcome measure at visit 2 or visit 4 among participants who met criteria for Any Word Loss. NVIQ significantly predicted PLS-4 AC and EC raw scores at visit 2 and visit 4 (ps<.001) and child age significantly predicted PLS-4 AC raw scores at visit 2 (p=.003) and visit 4 (p=.028) for the Any Word Loss group. Among participants who met the broadest loss criteria, Any Communication Skill Loss, number of words used before loss was found to be a significant predictor of visit 2 AC raw scores (F(3, 39) = 18.24,p < .001, adj. R² = .55, t = 2.16,p = .037) and visit 2 EC raw scores (F(3, 39) = 11.77,p < .001, adj. R² = .44, t = 2.08, p = .045). However, by visit 4, number of words used before loss was no longer predictive of PLS-4 AC or EC raw scores (statistics are reported in table 4). NVIQ was found to significantly predict PLS-4 AC and EC raw scores at visit 2 and visit 4 (ps<.001) and child age was found to significantly predict PLS-4 AC raw scores at visit 2 (p=.001) and visit 4 (p=.022) for the Any Communication Skill Loss group.

Table 4.

Results of multiple linear regression models predicting language outcomes from number of words used before loss, nonverbal IQ (NVIQ), and child age, among participants who met 1) Definite Word Loss, 2) Any Word Loss, and 3) Any Communication Skill Loss criteria

Loss Criteria	Visit	Outcome Measure	Variable	t	P	β	F	df	P	adj. R 2
Definite Word Loss
	2	PLS-4 AC
			Overall Model				13.00	3,22	.000***	.59
			Loss	1.23	.234	.17
			Age	2.57	.017*	.35
			NVIQ	6.19	.000***	.89
		PLS-4 EC
			Overall Model				5.21	3,21	.008**	.35
			Loss	.89	.383	.15
			Age	1.51	.147	.27
			NVIQ	3.95	.001**	.73
	4	PLS-4 AC
			Overall Model				7.37	3,19	.002**	.47
			Loss	1.67	.112	.27
			Age	2.15	.045*	.35
			NVIQ	4.59	.000***	.78
		PLS-4 EC
			Overall Model				5.65	3,19	.006**	.39
			Loss	1.26	.222	.22
			Age	1.80	.087	.31
			NVIQ	4.05	.001**	.73
Any Word Loss
	2	PLS-4 AC
			Overall Model				14.55	3,38	.000***	.50
			Loss	1.83	.075	.21
			Age	3.17	.003**	.40
			NVIQ	6.58	.000***	.84
		PLS-4 EC
			Overall Model				9.76	3,37	.000***	.40
			Loss	1.81	.079	.23
			Age	1.54	.131	.22
			NVIQ	5.25	.000***	.75
	4	PLS-4 AC
			Overall Model				9.30	3,32	.000***	.42
			Loss	1.81	.080	.24
			Age	2.31	.028*	.33
			NVIQ	5.22	.000***	.75
		PLS-4 EC
			Overall Model				7.65	3,32	.001**	.36
			Loss	1.69	.102	.23
			Age	1.74	.091	.26
			NVIQ	4.73	.000***	.71
Any Communication Skill Loss
	2	PLS-4 AC
			Overall Model				18.24	3,39	.000***	.55
			Loss	2.16	.037*	.23
			Age	3.44	.001**	.40
			NVIQ	7.33	.000***	.85
		PLS-4 EC
			Overall Model				11.77	3,39	.000***	.44
			Loss	2.08	.045*	.25
			Age	1.77	.084	.23
			NVIQ	5.78	.000***	.77
	4	PLS-4 AC
			Overall Model				9.86	3,32	.000***	.43
			Loss	1.91	.065	.25
			Age	2.41	.022*	.33
			NVIQ	5.33	.000***	.75
		PLS-4 EC
			Overall Model			8.02	3,32	.000***	.38
			Loss	1.59	.122	.22
			Age	1.71	.098	.25
			NVIQ	4.85	.000***	.72

Note: PLS-4 AC = Preschool Language Scales – 4 Auditory Comprehension raw scores, PLS-4 EC = Preschool Language Scales −4 Expressive Communication raw scores, NVIQ = Mullen Scales of Early Learning Nonverbal Ratio IQ Score, * p<.05, ** p<.01, *** p<.001

Finally, in order to answer the third research question, descriptive statistics related to level of communication achieved prior to regression were calculated for loss groups according to each of the three sets of regression criteria. ADI-R toddler module item 37 (“number of words used before loss”) and ADI-R toddler module item 36 (“level of communication before loss”) were used to define communication skills achieved prior to the onset of regression. Among participants who met criteria for Definite Word Loss, 77% were reported to have achieved daily, spontaneous, and meaningful speech used communicatively, with at least 3 different words used at some point before change, 17.1% were reported to have achieved occasional and/or fewer than 3 words used spontaneously and communicatively (alone or in combination with imitative abilities), and 5.7% were reported to have produced speech or sounds upon request (may or may not have also spontaneously imitated). When participants were reclassified per Any Word Loss criteria, those figures were 50%, 45%, and 5%, respectively. Among those who met criteria for Any Communication Skill Loss, they were 48%, 44%, and 8%, respectively. Participants who met criteria for Definite Word Loss used an average of 13.00 (SD=15.74) words prior to the onset of regression. When reclassified according to Any Word Loss criteria, participants were reported to have used an average of 10.14 (SD=15.80) words before regression onset. Participants who met criteria for Any Communication Skill Loss were reported to have used an average of 9.56 (SD=15.60) words prior to regression onset.

Discussion

In the present study, we aimed to determine whether toddlers with ASD and a history of regression in the communication domain developed expressive and receptive language skills differently from those without a history of regression throughout the preschool years. We also sought to establish the role of number of words used before loss in the preschool language outcomes of children with ASD and a history of regression, and to describe pre-loss levels of communication achieved via ADI-R Toddler Module individual loss items.

Previous ASD regression group comparison studies have demonstrated equivocal evidence for differences in skill outcomes between children with ASD with and without a history of regression. These mixed findings were due (at least in part) to a lack of longitudinal data, strict dichotomous criteria for defining regression, reliance on parent report which is subject to significant telescoping effects in retrospective studies, and bias in clinical samples and data collection procedures in prospective studies. By minimizing telescoping effects, moving toward broader, more dimensional regression criteria, measuring receptive as well as expressive language outcomes, and analyzing language outcome data across multiple time points in the preschool years, the present study was able to extend and add clarity to the current body of literature on ASD regression and language outcomes. Most previous language outcome studies demonstrated relatively poorer outcomes in children who experienced regression (Bernabei et al., 2007; Norrelgen et al., 2015; but see Davidovitch et al., 2000). Consistent with those findings, our results did indicate significantly poorer outcomes in the regression group in some areas. However, unlike previous studies, those outcomes were detected only in receptive language (as measured by PLS-4 AC raw scores) one year after initial assessment among participants with a history of regression compared to those without, regardless of how broadly regression was defined. Because no previous ASD regression study has examined group differences in receptive language outcomes, this is a particularly notable finding. Group effects were small-medium (η_ρ²=.05-.07) and no longer reached significance by the second time point three years after the initial assessment (when participants’ mean ages were between 57-79 months), indicating that any disadvantage in receptive language development faced by children with ASD who experience regression is not only subtle, but also short-lived. By the end of the preschool years, children with and without a history of regression as defined by any of the three loss group criteria demonstrated similar performance in both the AC and EC subscales of the PLS-4. There was no evidence of group differences in later expressive language skills in the present study at either time point, in contrast to some previous studies (Bernabei et al., 2007; Davidovitch et al., 2000; Estabillo et al., 2018a; Hansen et al., 2008; Kalb et al., 2010; Kobayashi & Murata, 1998; Norrelgen et al., 2015; Rogers & DiLalla, 1990) but consistent with others (Pickles et al., 2009; Tamanaha et al., 2014). Additionally, a secondary analysis indicated a significantly larger discrepancy between comprehension and production in the Definite Word Loss group than the no loss group, suggesting that children with a history of regression may drive the discrepant expressive-receptive language profile at visit 2 found in the same overall sample by Davidson and Ellis Weismer (2017). As the discrepancy between PLS-4 AC and EC scores disappeared by visit 4 in the overall sample in Davidson & Ellis Weismer (2017), so too did the group effects on PLS-4 AC-EC raw score differences between Definite Word Loss and no loss groups in the present study.

In order to address the need for a more dimensional understanding of language skill losses and gains in the ASD regression literature, the second research question sought to determine whether the number of words participants were reported to have used before the onset of regression was a continuous predictor of PLS-4 language outcomes. Results indicated that number of words used before loss significantly and positively predicted both expressive (PLS-4 EC) and receptive (PLS-4 AC) language outcomes one year after initial assessment when controlling for child age and nonverbal IQ, but only when regression was defined using the broadest criteria, Any Communication Skill Loss. Moreover, number of words used before regression onset was no longer predictive of any language outcome measure by the end of the preschool years regardless of which loss criteria were applied, indicating that any advantage in language skill that may exist as a result of a larger expressive vocabulary in the first two years of life disappears by age 5-6.

It should be noted that several researchers have used the term ‘recovery’ to describe the pattern of skill gain following regression in ASD (Boterberg et al., 2019; Goldberg et al., 2003; Pearson, Charman, Happé, Bolton, & McEwen, 2018; Pickles et al., 2009; Tamanaha et al., 2014; Thomas, Knowland, & Karmiloff-Smith, 2011). While it is the case that quantitative language measures (such as the PLS-4 used in this study) may indicate scores that return to and then exceed pre-loss levels, we were not able to specify whether specific skills lost during regression were recovered, or whether other, related skills simply improved to an extent that any losses were ‘washed out’. By washed out we mean that it is possible that gains in new language skills could offset the weaknesses in other language domains. Moreover, because we examined outcomes at the group level, it would not be appropriate to interpret a lack of differences between groups at age 5-6 as ‘recovery’ among children with regression, as we were not able to account for individual participant trajectories which may be more variable. Our findings also align with recent literature suggesting that regression is a common, dimensional symptom onset pattern in ASD (Boterberg et al., 2019; Ozonoff et al., 2018; Ozonoff & Iosif, 2019; Szatmari et al., 2016; Thurm et al., 2018; Williams, Brignell, Prior, Bartak, & Roberts, 2015). As such, it may be more appropriate to describe the skill losses and gains among children with regression and their later outcomes in the same terms as other early developmental patterns in ASD.

Finally, by examining individual language loss items from the ADI-R Toddler Module, this study was able to address the need to elucidate levels of communication skill achieved by children with ASD and regression prior to regression onset, an area previously under-explored in the extant literature as noted by several researchers (e.g., Boterberg et al., 2019; Pearson et al., 2018). Many parents insist that their child was developing typically prior to the onset of regression (Luyster et al., 2005; Rogers & DiLalla, 1990). However, several studies have reported atypical or delayed development before losses occurred in at least a subset of children with ASD and regression (Goldberg et al., 2003; Kurita, 1985; Luyster et al., 2005; Ozonoff, Williams, & Landa, 2005; Richler et al., 2006). Kurita (1985) reported that 93.8% of children were only able to use single words prior to loss and that the majority used “few or several words” infrequently, while 7.2% were reported to have used “many” words with a maximum of 30. In the present study, parent report on ADI-R loss items suggested relatively larger expressive vocabularies and more frequent communication prior to regression onset than that of Kurita (1985). Depending on the definitional criteria used, 48-77% of children with a history of regression in our sample were reported to have used “daily, spontaneous, and meaningful speech used communicatively, with at least 3 different words used at some point before change” with an average expressive vocabulary of 9.56-13.00 words (SD=15.60-15.74) and a maximum of 75 words. These findings indicate a wide range of pre-loss expressive language skills in the children with a history of regression, consistent with previous findings that suggest delayed or atypical early language development in some children with a history of language regression (Kurita, 1985; Luyster et al., 2005; Ozonoff et al., 2005; Pickles et al., 2009).

Limitations and Future Directions

While this study was able to address several important gaps in the ASD regression literature, the results are qualified by many of the same methodological limitations faced by previous studies that employed retrospective parent-report designs. Although the ADI-R remains a “gold standard” measure for ASD diagnosis (Rogers et al., 2010), the skill loss questions and coding scheme impose a strict, dichotomous operational definition for language regression that may fail to capture the full spectrum of losses (Pearson et al., 2018). It should also be noted that ADI-R skill loss questions fail to account for other developmental and environmental changes that might contribute to a change in language use, such as a change in home circumstances that lead to a reduced need to use certain words, or loss of words that may be obscured by rapid acquisition of new words in the same period. Many previous studies using the ADI-R were susceptible to forward telescoping effects which may have decreased the accuracy of parent report (Hus et al., 2011; Ozonoff, et al., 2018). However, participants were all between 23-39 months old when the ADI-R Toddler Module was administered in the present study, minimizing the time delay between regression onset and parent interview and reducing the potential effects of forward telescoping. Moreover, by operationally defining language regression using multiple levels of criteria less strict than that of the ADI-R, we were able to capture a broader spectrum of skill loss experienced by children with ASD in the communication domain. While a small group effect was found in receptive language outcomes one year later regardless of which regression criteria were used, we were only able to detect a relationship between pre-loss expressive vocabulary size and later expressive and receptive language outcomes among children with a history of regression when the broadest criteria were applied. This would suggest that future studies should also consider including broader operational criteria including preverbal communication skills in order to capture the full spectrum of skill losses and gains that may occur in children with ASD. Additionally, while this study was able to provide some needed insight into receptive language outcomes following language regression, the ADI-R does not probe for losses in receptive language. As such, there remains a need to examine skill loss in that domain which is beyond the scope of this study (Boterberg et al., 2019).

Study results should also be qualified by our intentional decision not to correct for multiple comparisons. While correction procedures are widely practiced in the literature to avoid type I error, they simultaneously decrease statistical power and increase the risk of type II error. In weighing the risks of type I and type II error with the understanding that the equivocal and conflicting findings in the previous literature suggested that any group effects would likely be small, we determined that there was a particularly high risk of type II error in our study. By avoiding correction for multiple comparisons, we were better able to increase the sensitivity of our analyses to detect any group differences that may exist. Most importantly, we considered the practical implications of committing either a type I or type II error in the context of our particular research questions. We felt that taking a more conservative approach and potentially failing to detect an effect of regression on language outcomes would be more detrimental to families and children with ASD, who as a result of type II error may then miss out on critical early intervention aimed specifically at language/communication, than the possibility of reporting an effect that does not exist but may lead to additional empirical scrutiny and clinical vigilance. It is also worth noting that we were not able to detect significant differences in language outcomes at visit 4 despite using a more lenient analytical approach, allowing us to conclude with a fair degree of certitude that regression patterns were not indicative of poorer outcomes at school entry.

Further, the effects of regression group on receptive language outcomes which we were able to detect in response to the first research question were small-medium in magnitude. While this study had the advantage of utilizing a standardized tool specifically designed to assess language skills for our outcome measures, child accuracy on a behavioral assessment remains a relatively crude index of language ability. It is possible that additional group differences may emerge with more sensitive, fine-grained measures of on-line language processing such as eye-tracking methodology.

In addition to the need for more precise language outcome measures and the development of a more dimensional operational definition for regression, longer-term longitudinal follow-up is recommended to determine whether the lack of group differences on language outcome measures between children with ASD with and without a history of regression remains stable throughout childhood and adolescence. While previous studies have measured language and other developmental outcomes at these later ages, none were longitudinal in nature and as such were subject to significant telescoping effects by relying on parents to report events that occurred several years prior in their child’s development. Future studies should also examine quantitative and qualitative data on language intervention received by children with and without regression. We chose not to include language intervention data in these analyses as parent report is an insufficiently precise measure of the impact of early intervention given natural differences in access, intensity, and its interaction with child characteristics such as autism severity. However, descriptive statistics regarding language intervention frequency and duration can be found in Table 2 for the entire sample and for each loss group. It is conceivable that parents of children who appear to lose language skills may be more likely to seek referrals for speech-language evaluation or may be recommended for more intensive intervention, potentially altering the developmental trajectory of children with ASD who have experienced regression. It will be critical for future studies to prospectively track the developmental progress in language, social and cognitive domains as well as early intervention received in order to better understand individual developmental trajectories and begin to parse child and environmental characteristics that lead to the most optimal language outcomes among those who experienced a regression. Pursuing this approach to studies of regression in ASD will also inform clinical best practice for this highly variable population.

Conclusions

While results indicate that differences in receptive language and receptive-expressive language profiles may exist in preschool-age children with ASD with and without a history of language/communication regression, those differences are small in magnitude and relatively transient, disappearing by the end of the preschool years. Among preschool children with ASD who have experienced language regression, pre-loss expressive vocabulary size is predictive of expressive and receptive language outcomes, but also only in the short term, and only when regression is defined using broad criteria. Moreover, parent report suggests that there is considerable variability in the pre-loss language skills of children with ASD who experience regression. These findings support the assertion made by several researchers (Ozonoff & Iosif, 2019; Thurm et al., 2018; Williams et al., 2015) that regression is likely a dimensional ASD symptom onset pattern that may not be fully captured within a dichotomous categorical framework. Study results do not support the theory that regression as a symptom onset pattern develops into a distinct language phenotype within ASD, although longer-term longitudinal follow-up is needed to confirm this assertion. While loss of early developmental communication milestones is often a cause of alarm for parents of children with ASD, these results suggest that language regression does not have a negative impact on language skills at school entry and does not appear to warrant a different approach to speech-language intervention from that of children with ASD with other symptom onset patterns. Based on these findings, parents should be reassured that communication skill loss is a common pattern of early development in ASD and does not appear to be detrimental to language outcomes beyond a relatively short-term (1 year) period. Clinicians are advised to proceed with speech-language intervention individualized to the needs and strengths of each autistic child’s unique language profile regardless of their history of regression.

Table 5.

Descriptive statistics for levels of communication achieved prior to regression onset, among participants who met the following sets of criteria: 1) Definite Word Loss, 2) Any Word Loss, and 3) Any Communication Skill Loss

Measure	Definite Word Loss n = 35		Any Word Loss n = 58		Any Communication Skill Loss n = 61
Level of communication achieved prior to loss
0= daily, spontaneous, and meaningful speech used communicatively, with at least 3 different words used at some point before change	27 (77%)		29 (50%)		29 (48%)
1= occasional and/or fewer than 3 words used spontaneously and communicatively (alone or in combination with imitative abilities)	6 (17.1%)		26 (45%)		27 (44%)
2=produced speech or sounds upon request (may or may not have also spontaneously imitated)	2 (5.7%)		3 (5%)		5 (8%)
3=spontaneous imitations of vocalization (without ever having any completely spontaneous speech) with no elicited imitation or spontaneous communicative speech	0 (0%)		0 (0%)		0 (0%)
8=no change or loss	0 (0%)		0 (0%)		0 (0%)
9=not known or not asked	0 (0%)		0 (0%)		0 (0%)
	M (SD)	Range	M (SD)	Range	M (SD)	Range
Number of words used before loss	13.00 (15.74)	(2-75)	10.14 (15.80)	(1-75)	9.56 (15.60)	(0-75)