Calm with horses? A systematic review of animal-assisted interventions for improving social functioning in children with autism

The aim of this systematic review was to evaluate the effect of animal-assisted interventions on social functioning in children with autism spectrum disorder, based on evidence from randomized control trials. Included studies were articles published in English, with school aged children from 4 to 18 years with autism spectrum disorder. Databases searched were MEDLINE, PsycINFO, EMBASE, Web of Science, CINAHL and Zoological Record. Data extraction from included studies included demographics and sample features, interventions and controls descriptions, outcome measures, study funding and descriptive statistics. Risk of bias was assessed, considering randomization, allocation concealment, blinding, attrition, selective reporting and other sources of bias. Studies were synthesized narratively based on the animal approach taken and the use of waitlist versus active controls. Nine studies were included reporting across eight trials. Studies overall reported improvements in social functioning following equine-assisted services, with preliminary evidence suggesting improvements are sustained in the short and medium term. Insufficient evidence was available to draw conclusions on the efficacy of other animal-assisted interventions. Future research should aim to address the limitations common to included designs. Lay abstract Children with autism typically experience difficulties interacting socially with others when compared to their non-autistic peers. Establishing how effective interventions are for improving social functioning is important to help inform what should be offered to children with autism. This study reviewed how effective interventions that involved interaction with a live animal, known as animal-assisted interventions, are in improving social functioning in children with autism. A systematic search of the evidence on this topic found nine studies, which were explored for the effectiveness of animal-assisted interventions and the quality of methods used. Overall, these studies showed improvements in social functioning following equine-assisted or therapeutic horse-riding interventions, with initial evidence showing improvements are sustained in the short and medium term. However, several issues were identified, which limit the strength of any conclusions that can be drawn from this evidence. For example, in many studies people assessing the children were aware that they received the intervention or were in a control group. There was also not enough evidence available to draw conclusions on the effectiveness of other animal-assisted interventions. Future research should address the limitations that were common in the designs of these studies and investigate the potential benefit of other animal populations, such as dogs and cats.

, with a higher proportion of parents (1.7%) reporting being told their child is on the autism spectrum by a health professional (Russell et al., 2014). Of those children receiving a diagnosis, there is a male to female ratio of 3:1 (Loomes et al., 2017). Children with autism also experience an increased likelihood of receiving other co-occurring diagnoses, most commonly attention deficit hyperactivity disorder, oppositional defiant disorder and anxiety (Brookman-Frazee et al., 2018;Simonoff et al., 2008). Rather than focussing on deficits and a diagnosis of 'disorder', many proponents within the autistic and research community favour a perspective of autism as reflective of neurodiversity (Baron-Cohen, 2017). Accordingly, calls to focus on improving quality of life and well-being in people with autism have been made in preference to treatments aiming to reduce autistic traits (den Houting, 2018). Nevertheless, for many children with autism, difficulties in interacting socially can present a range of immediate problems starting in education settings, such as experiences of exclusion (Pellicano et al., 2018) and bullying (Park et al., 2020).
A range of psychosocial interventions are currently recommended for use in children with ASD, aiming to increase joint attention, engagement and reciprocal communication (Lord et al., 2022; National Institute for Health and Care Excellence (NICE), 2013). However, existing interventions are not universally effective in children with autism (Jobin, 2020) and from the perspective of adults with autism, there is a greater willingness to take part in complementary interventions in the community over established socio-behavioural interventions such as Applied Behavioural Analysis (Benevides et al., 2020). One type of complementary intervention, acceptable to adults with autism and parents of children with autism (Benevides et al., 2020;London et al., 2020), is animalassisted interventions (AAIs). AAIs incorporate the presence of a live animal, most frequently horses or dogs and, more rarely, other animals such as dolphins or guinea pigs  and are offered in many countries as complementary support for children with autism, including in the United States and United Kingdom (Eaton-Stull et al., 2020;Malcolm et al., 2018). AAIs are prominent in the public sphere, in media such as The Horse Boy (Isaacson, 2009) and Calm with Horses (Rowland, 2020). Proposed mechanisms of AAIs for children include reduction of stress -contact with animals has been shown to reduce anxiety in children (Crossman et al., 2020), and tactile contact may alter stress hormones, increasing peak oxytocin and reducing cortisol (Handlin et al., 2011). Reduced cortisol responses in children following AAIs may allow for reduced hyperactivity  and provide an 'open' context for children to engage with therapists and their environment (Malcom et al., 2018). Animals may also provide a less complex social stimulus for children with autism (Martin & Farnum, 2002) as their behaviour may be more predictable, and less challenging, as for example, animals can demand less eye contact than typical human interactions (Malcolm et al., 2018).
Despite their potential benefit, the evidence base for AAIs is limited (O'Haire, 2017). Previous systematic reviews have considered the impact of AAIs using less strict criteria, including lower quality evidence such as case study designs (O'Haire, 2017;Trzmiel et al., 2019), summarizing a range of preliminary and in some cases anecdotal evidence indicating AAIs may be beneficial for social functioning in children with ASD. In contrast, a meta-analysis including only higher quality randomized control trials (RCTs) had excluded AAIs from inclusion due to a limited number of trials and risk of bias concerns (Sandbank et al., 2019). To progress the evidence base for this potentially beneficial intervention, it remains important to evaluate the existing highquality evidence. This systematic review addresses the gap in the literature by narratively synthesizing evidence on the effect of AAIs on social functioning in children with diagnoses of ASD based only on RCTs.

Eligibility criteria
Studies eligible for inclusion were RCTs comparing AAIs to active controls without animal involvement or waitlist controls. Eligible studies included child participants of school age (from 4 to 18 years) with a diagnosis of ASD according to Diagnostic and Statistical Manual of Mental Disorders (5th ed.; DSM-5) criteria for ASD (APA, 2013). This included participants with prior diagnoses of Asperger's or Pervasive Developmental Disorder Not Otherwise Specified (PDDNOS) as in the International Classification of Diseases (ICD-10; World Health Organization (WHO), 1993). Inclusion criteria required studies to report participant's social behaviour as an outcome, assessed either by self-report or an external rater (parent, teacher, caregiver or other professional assessment) for both pre-and post-intervention. Dissertations and conference abstracts were excluded, as well as any studies without a live animal, such as virtual or robot animal interventions.

Information sources
Searches were completed across six electronic databases on 28 October 2020; Ovid MEDLINE(R) (1946-present), APA PsycInfo (1806-present), Embase Classic+Embase (1947-present), Zoological Record (1978-2010, Web of Science (1900-present) andCINAHL(1960-present;via EBSCO databases). Search terms included variants of 'Autism' AND 'Animal Intervention' AND 'Social Interaction' AND 'Child' AND 'Randomised control trial', as shown in full in Appendix 1. When data were not available or more details about studies were needed, the corresponding author of each study was contacted, resulting in further data requests from Gabriels et al. (2015Gabriels et al. ( , 2018 and . An updated search of the literature was performed covering five databases between 28 October 2020 and 8 October 2021; Ovid Medline (R) ALL 1946 to 8 October 2021, Embase Classic+Embase 1947 to 8 October 2021, APA PsycInfo 1806 to October 2021.

Study selection
After removal of duplicates, remaining studies underwent abstract and title screening. Four researchers (E.B., J.H.S., N.S. and H.S.) each screened 252 abstracts, with any resulting disagreements discussed and resolved afterwards. Studies at this stage were removed if they had an adult sample, did not use live animals, included no ASD diagnosis, recorded no social outcome or were a previously missed duplicate. Remaining studies underwent fulltext screening, with four researchers (E.B., J.H.S., N.S. and H.S.) each screening 15 or 20 articles with 10 articles overlapping (so that 36% of articles were double screened). Studies with quasi-experimental designs, lacking sufficient evidence of randomization and dissertations or conference abstracts were excluded at this stage.

Data collection
Data extraction was completed using an adapted Cochrane Collaboration data extraction form. All nine included studies were double extracted and checked by two reviewers, with the first five checked by E.B. and J.H.S. and the latter four included studies by N.S. and H.S..

Data items
From each study, the following information was extracted: sample demographics (including age, gender); sample features (verbal or non-verbal, diagnosis severity or description, intelligence quotient (IQ), prescribed medication); intervention and control description (components, staff involved in delivery, treatment timing, duration and frequency); outcome measures (relevant scale and subscales used, time points measured and reported, scale validity); study funding sources; and reported descriptive statistics with any associated p values. Where descriptive statistics were missing, authors were contacted via email requesting original data.

Risk of bias in individual studies
To assess risk of bias in included studies, the Cochrane 'Risk of Bias' assessment tool was used by considering the criteria guidelines for each risk with respect to each study, or in the case of multiple outcomes per study, each outcome. Considered risks included selection bias, performance bias, detection bias, attrition bias and any other bias.
For selection bias, evidence of random sequence generation to avoid bias in the allocation to intervention and control groups was evaluated, as well as the concealment of these allocations to researchers so that they could not be predicted and influence procedure. For performance bias, the blinding of participants and personnel to the conditions participants were assigned to was considered, where interventions compared only to a waitlist control were assumed to be incompatible with blinding of participants. For detection bias, the blinding of outcome assessment was considered separately for each outcome measure where multiple were reported by a single study. For attrition bias, the incompleteness of reported outcome data was evaluated, indicated by a significant proportion of missingness or evidence of missingness related to the intervention or outcomes Missing Not At Random (MNAR). For reporting bias, reporting of results selectively was assessed, such as reporting based on significance or to support a hypothesis. Any other evident sources of bias were also considered, including baseline imbalances in measures or relevant characteristics, undeclared or inappropriate influence of study funding sources and specific sources of bias related to the design.
Risk of bias forms were completed for all included studies across each risk described, using ratings of low, high or unclear risk of bias.

Synthesis of results
A narrative synthesis was used to bring together and summarize quantitative results across studies. Studies were described and analysed in terms of trial design, intervention content, outcome assessors, controls used and efficacy of results. No community involvement was incorporated into this process.

Study selection
Study selection produced nine studies eligible for inclusion involving eight trials. CINAHL, EMBASE, Medline, PsycInfo, Web of Science and Zoological Record databases were searched, producing a total of 359 studies. Two hundred and fifty-two studies remained after removal of duplicates. During abstract and title screening, 197 studies were removed for failing to meet inclusion criteria. Full texts for a total of 55 remaining studies were retrieved and screened, resulting in a further 46 studies excluded for failing to meet inclusion criteria. Complete inter-rater agreement was reached for articles, which were double screened. A final total of nine studies were selected for inclusion in narrative synthesis. Reasons for exclusion at each stage are detailed within the flow diagram of the study selection process ( Figure 1) and listed individually in Appendix 2.
An updated search was performed from October 2020 to October 2021. The previous procedure was repeated, with 42 studies produced, 17 remaining after deduplication, of which seven were removed at title and abstract screening. Of 10 full texts screened, three further studies were selected for inclusion.

Study characteristics
Of all nine included studies, eight reported unique RCTs, with one reporting a 6-month follow-up  to a previous trial . Seven out of the eight trials assessed the impact of an equine-assisted intervention, referred to as either therapeutic horse riding (THR; Bass et al., 2009;Gabriels et al., 2015; or equine-assisted therapy/activity (EAT/EAA; Borgi et al., 2016;Coman et al., 2018;. As recommended by Wood et al. (2021), the term equine-assisted services (EASs) will be used herein to describe different intervention approaches utilizing horses. One trial assessed the impact of a reading programme with the presence of dogs . Although a higher proportion of non-equine-based AAIs were identified in a previous systematic review , many of these studies used singlesubject or within participants designs and were, therefore, excluded from this review. In the trial using a dog-based Autism 26 (6) intervention , the lowest sample size of nine participants was reported, while sample sizes in the remaining equine-based studies ranged from 16 to 116. Three studies used Diagnostic and Statistical Manual of Mental Disorders (4th ed., text rev.; DSM-IV-TR; APA, 2000) criteria for ASD diagnosis Borgi et al., 2016;Coman et al., 2018) and one used the more recent DSM-5 (APA, 2013;. Remaining studies used cut-off scores on the Autism Diagnostic Observation Schedule (ADOS/ADOS-2; Gabriels et al., 2015Gabriels et al., , 2018Lord et al., 2012; or the Childhood Autism Rating Scale (CARS; Park & Kim, 2016;, with the exception of the study by , which reported participant's diagnosis of autism but not the diagnostic assessment used.
Although given different names, no differences between THR and EAT/EEA interventions were evident, and all EASs incorporated skills mounting and riding horses. EASs predominantly included a form of warm-up or preparation Coman et al., 2018;Gabriels et al., 2015; and skills caring for the horse Coman et al., 2018;Gabriels et al., 2015;. Some studies included additional components, such as mounted games , drawing activities  or specific time for 'touch stimulation' .
Equine-based studies predominantly used a waitlist control group, with the exceptions of a barn activity (BA) control without horse interaction in two studies ) and a dance group control within a crossover design in one study . In the study by , dog-assisted reading was compared to reading in the absence of a dog. Full study characteristics are reported in Table 1.

Risk of bias within studies
Included studies were assessed using the Cochrane Risk of Bias tool and assigned either 'low', 'high' or 'unclear' risk of bias for each risk. Unclear risk of bias was assigned where studies did not describe sufficient details, such as the randomization method, allocation concealment or blinding. Only one study was judged to not have any high risk of bias, although risk of bias was unclear for four of the risks for this study. All remaining studies had a mixture of low, high and unclear risks of bias. Detection bias was a consistent issue across studies, with no studies at low risk of bias for adequately blinding outcome assessment, often due to assessment by parents or teachers inevitably aware of group assignment. Judgements for risk of bias across each risk across all nine studies are shown in Figure 2.

Synthesis of results
Efficacy within equine-based approaches. Seven of eight original studies evaluated EASs, with four of these assessing social outcomes with the Social Responsiveness Scale (SRS; Constantino et al., 2003). All of these reported significant improvements in SRS total scores, but results varied across different SRS subscales, with significant improvements in social motivation Coman et al., 2018), social communication Gabriels et al., 2015;, social cognition Gabriels et al., 2015), social awareness  and autistic mannerisms  all reported. Coman et al. (2018) reported significant improvement on four of five subscales of the SRS in a sample of 50. However, by contrast, the largest powered study of 116 participants by Gabriels et al. (2015) only reported significant improvements in social cognition and social communication subscales.  aimed to replicate the intervention previously evaluated in the study by Gabriels et al. (2015); however, the subscales of the SRS, which significantly improved were inconsistent between these studies. There is, therefore, limited evidence to suggest that variance in subscale improvement was related to heterogeneity in intervention delivery. The remaining three studies reported significant improvements in the Vineland Adaptive Behaviour Scale (VABS; Sparrow et al., 1984) socialization subscale , social participation (Souza-Santos et al., 2018) and Social Communication Questionnaire (Avcil et al., 2015) communication subscale . Although as previously described, there was some variation in the content of EAS, with some interventions including additional activities , the mechanisms proposed to be beneficial within the literature (such as tactile contact with animals, relaxation with animals and skills learning) were incorporated into all approaches through riding and horsemanship activities with horses.
Although Gabriels et al. (2018) reported on a 6-month follow-up to a previous trial , as SRS descriptive statistics were not reported for follow-up, the authors were contacted requesting data. Results showed that across SRS subscales, which significantly improved in the study by Gabriels et al. (2015), SRS communication and SRS cognition remained over twice the standard error below mean scores post treatment, while social awareness scores increased above post-treatment mean (Gabriels & Pan,Personal communication,7 December 2020). Coman et al. (2018) also collected follow-up 8 weeks post intervention, retaining 50% (25/50) of the sample and reporting sustained improvements in SRS total, social cognition, social communication and autistic mannerisms.
In an updated search from 2020 to 2021, two further studies evaluating equine-based approaches were Efficacy in non-equine-based approaches. As only one intervention assessed the impact of a dog-based intervention, comparisons cannot be drawn between intervention components. In this intervention,  randomized nine children to either a reading with dogs group or reading without dogs group, where children were instructed to read the same book on a weekly basis. Physical contact with the dogs was not allowed; potential mechanisms of change instead involved reading and talking to the dogs, which was suggested to be beneficial by providing a non-judgmental environment to practice reading in, with emotional support from the dogs actively listening. Sessions were conducted in the presence of a psychologist; otherwise, the intervention included no other targeted mechanisms or skills. Two female dogs were selected for their suitability for the intervention, based on their cooperation with children, reduced anxiety and aggression. The intervention partly aimed to improve reading abilities; however, in terms of social outcome, no significant improvements on the VABS socialization were reported in the reading with dogs' group . Notably, this intervention focussed on improving reading skills with social communication as a secondary outcome, whereas previous interventions used in case studies delivered dog-assisted interventions programmes focused on social skills (Silva et al., 2011). Results across all animal approaches are reported in full in Table 2. An updated search also identified another study taking a non-equine-based approach by Hernández-Espeso et al. (2021) in which dolphin-assisted therapy (DAT) was delivered to 48 children with ASD, involving structured games and activities in water equivalent to those with horses in equine-assisted services. Significant improvements in VABS 2 socialization were reported in the DAT group; however, these improvements were not significantly different to those found in an active therapy without dolphins control.
Efficacy in studies using active versus waitlist controls. Of eight included studies, four utilized waitlist controls Borgi et al., 2016;Coman et al., 2018; and four used active controls . Bass et al. (2009) delivered a 12-week EAS programme to 36 children, diagnosed with mild-to-severe ASD and Asperger's, resulting in improved social motivation on the SRS. Coman et al. (2018) also delivered an EAS intervention for a period of 12 weeks in a sample of 50, predominantly male children with autism. Again, they reported improvements in social functioning on the SRS, with some sustained changes in SRS total, social cognition, social communication and autistic mannerisms at 8-weeks follow-up. Borgi et al. (2016) delivered EAS to 28 boys over 25 weeks, reporting improved social functioning on the VABS. All three of these studies were limited by high risk of performance bias, as blinding was not possible due to use of waitlist controls.  successfully blinded personnel but not participants; however, in this context, children are not expected to have expectations of intervention effects and are, therefore, of less concern as a source of risk of bias. Gabriels et al. (2018) reported the effects of a 10-week EAS in the largest sample of 116 children, in comparison to a barnyard activity control. Results demonstrated significant improvements in social functioning measured by SRS total score, as well as social cognition, social communication and social awareness subscales, which were sustained at a 6-month follow-up in 64 of these participants in social cognition and social communication . As  replicated this procedure in a smaller sample of 16 children aged 6-16 years, they utilized the same control, where participants interacted with a life-sized stuffed horse in a barn to learn horsemanship skills without any live horse interaction.  reported improvements in social functioning, but in this case only in SRS total, SRS awareness and SRS communication.  utilized a crossover design, in contrast to the parallel designs used in all other included studies. In this study, the efficacy of an EAS was evaluated in comparison to a dance group control, as well as a combined equine and dance control over a 12-week period delivered to 45 children. Results demonstrated improved social participation as measured by the WHO Disability Assessment Scale (Huang et al., 2017) after receiving EASs in comparison to the dance group control.
Finally,  evaluated the impact of a non-equine-based approach, comparing the impact of a dog-assisted reading programme to a programme of reading without a dog over 10 weeks, in a sample of nine children. Results from this study demonstrated non-significant improvement in social skills in either group on the VABS. Although this meant that three out of four studies using active controls reported significant effects in comparison to four out of four studies using waitlist controls, it is difficult to draw any conclusions on this basis as the latter study was the only one to not evaluate an equine-based intervention. Studies using active controls nevertheless reduced the chance of reporting overinflated outcome effects, by controlling for the possibility of benefits to social functioning by engaging in activities within an intervention rather than remaining on a waitlist. Risk of performance bias was low in some of these studies using active controls , as blinding of participants and personnel was more feasible as a result of using active controls.
Efficacy in studies using parent, teacher or caregiver versus clinician reports. Of the included studies, the majority collected outcomes using either parent Borgi et al., 2016), caregiver  or teacher assessment . The assessors collecting outcomes were unclear in  as well as  who may have used a clinician assessor. Only one study unambiguously reported use of a clinician evaluator .
As Coman et al. (2018) collected both parent and teacher report, only teacher report was extracted, assuming parents may be less impartial and more susceptible to bias than teacher reports (Jones et al., 2017). Nevertheless, Coman reported statistically significant (p < 0.05) reliability coefficients between parent and teacher reports for each aspect of the SRS, except for the autistic mannerism's subscale. Despite this agreement between raters, of more concern is the extremely limited number of studies using independent evaluators to assess outcomes. As the one clear exception also reported no significant improvements in social outcomes , there is limited evidence to exclude the possibility that reported results are influenced by bias in outcome assessors. However, as this study was also the only study to evaluate a dog-assisted intervention, no comparisons based on outcome assessors can be made between EASs.
Efficacy in studies with low risk of bias. None of the included studies were at low risk of bias consistently across all risk of bias judgements. Although two studies  received no high risk of bias judgements, the number of unclear risks for these studies renders any focus on these studies inappropriate, as risk of bias that is less apparent is not necessarily any less likely to be high.

Summary of evidence
Overall, across a small number of studies, this systematic review found some evidence of the efficacy of EASs in improving social functioning in children with autism, but insufficient evidence of the benefits of AAIs more broadly. Most included studies evaluated the efficacy of EASs, with all reporting significant improvements across varied measures of social functioning, but some inconsistencies in changes in subscales of the SRS across those reporting this outcome. In two studies reporting follow-up outcomes, improvements in social communication and social cognitions remained significant at 8 weeks and 6 months post intervention. Included interventions were similar to those in earlier reviews; between 8 and 12 weeks in duration and involving an approximate average of 10 h contact for participants (O'Haire, 2017). All nine primary studies within the present review utilized RCT designs; however, multiple study limitations were prevalent -risks of bias were identified, namely that 66% of studies were at high risk of detection bias and 44% of studies were at high risk of performance and reporting bias. Given these limitations, caution should remain in drawing strong conclusions from this evidence and further trials should aim to minimize these sources of bias. Included studies also provided limited evidence for any mechanisms of change underlying a beneficial effect of AAIs on social functioning. One proposed mechanism of change is that AAIs function as calming stimuli reducing stress responses (O'Haire, 2017), which can be a source of difficulty in social interactions in children with autism (Corbett et al., 2010).  measured salivary cortisol before and after children received EAS or a barnyard activity control. Although changes in post-session cortisol over the 10-week period did not occur, there were significant pre-to post-session reductions in cortisol in the EAS group. These changes were associated with improvements in irritability and hyperactivity, although no equivalent analysis was performed for social outcomes. While this provides some evidence of the role of AAIs in reducing stress hormones, whether this is associated with an improved ability to develop social skills remains uncertain.
In the one included study evaluating the impact of a dog-assisted intervention, no tactile contact was allowed between children and the dogs, which may have removed the benefit of stress reduction in AAIs (Handlin et al., 2011). This was the only included study, which reported no significant improvements in children's social functioning following the intervention ; however, there should be caution in comparing dog-and equine-assisted interventions and further evidence is required to draw conclusions on the efficacy of dogassisted approaches. Rather than acting primarily as a reading programme , other dogassisted interventions within the literature instead aim to improve social skills in children with autism and allow tactile contact as a possible beneficial mechanism (Silva et al., 2011) and, therefore, might produce a different effect.
An update to the literature search produced three further studies, two of which provided results consistent with previous trials demonstrating improvements in socialization in children with ASD receiving equine-assisted services Zhao et al., 2021). These studies were, however, limited by similar issues identified in previous trials, such as a lack of blinding in outcome assessment. The remaining study by Hernández-Espeso et al. (2021) evaluated the efficacy of a dolphin-assisted intervention and reported significant improvements, which did not differ significantly from an active control. This demonstrates the importance of trials using active controls for animal-assisted interventions, especially in the case of 'exotic' animal interventions where costs are likely to be significantly higher than equivalent interventions without animals.

Limitations
Despite our focus on RCTs, improvements to the rigour of research methods used could still be made, such as clearer reporting of randomization methods used. While random allocation to groups is preferable to non-randomized designs, many included studies used waitlist rather than active controls as comparison groups Borgi et al., 2016;Coman et al., 2018;. Waitlist controls may inflate reported effect sizes (Michopoulos et al., 2021) and active controls may provide an opportunity to reduce risk of bias by better enabling blinding of participants to their group allocation. High risks of bias were a persistent issue across most studies, with consistent issues with detection bias. Many studies failed to adequately blind outcome assessment, largely due to the use of parent-or carer-recorded outcome measures, which is a notable limitation within the literature on autism interventions for children (Jones et al., 2017). Efforts to provide blinded assessment of outcomes in RCTs are arguably the most essential design improvement for future RCTs to make in this area. As no restrictions on sample size were included, some studies may also have been underpowered to detect any significant effects, such as a sample of only nine children in the study by . In terms of the review itself, as it was not preregistered, this introduces the potential for bias resulting from any changes made to the method. All procedures were kept the same throughout the trial with the exception of GRADE ratings for the overall body of evidence, which were removed from the discussion.
There are also a series of practical limits to the results reported across included studies. Scaling up EASs could present practical challenges, as for example, in the largest scale study, Gabriels et al. (2015) delivered an EAS in sessions of two to four participants at a time. As the intervention required trained staff, volunteers and animals, the resource constraints of a riding centre could limit the expansion of EAS to larger scales. In the study by , children with uncontrolled seizures were unable to participate due to risk of danger during horseriding. As there are higher rates of epilepsy in people with autism than the general population (Spence & Schneider, 2009), risk of seizures may exclude a significant portion of children with autism from participation in AAIs. Generalizability of AAIs is also limited, as subgroups of children with autism were excluded from many studies, such as children with intellectual disability Gabriels et al., 2015Gabriels et al., , 2018. Of the remaining studies, only  reported the mean IQ of the sample. While some studies included only verbal children with autism , improvements in social functioning in mixed samples of both verbal and non-verbal children with autism have been demonstrated Coman et al., 2018).
Although the present review was limited to a narrative synthesis and not a meta-analysis, it acts as a stop gap in evaluating the efficacy of AAIs for social functioning in children with autism as the quality of available evidence improves. In subsequent years, further RCTs, which build upon the limitations highlighted in the present review ought to be reviewed and synthesized in a meta-analysis to estimate the size of effects on social communication and provide guidance for the most effective intervention.

Conclusion
This review reported on evidence from nine RCTs, many of which were published in recent years and have not been included in previous systematic reviews (O'Haire et al., , 2017Trzmiel et al., 2019). We found evidence to support the efficacy of the most prominent form of AAI -EASs -in improving social functioning in children with autism. A small amount of evidence supported the continuation of benefits in social functioning at short-(8-week) and medium-term (6-month) follow-ups. Insufficient evidence was available to conclude on the efficacy of other AAIs such as those including dogs. Similarly, no comparisons could be made between outcomes based on the measures used. Future studies should aim to address the limitations common to included designs.