The effects of oxytocin administration on social and routinized behaviors in autism: A preregistered systematic review and meta-analysis

Oxytocin administration has demonstrated considerable promise for providing individualized support for autistic people. However, studies evaluating the effects of oxytocin administration on autistic characteristics have yielded inconsistent results. This systematic review and meta-analysis investigates the effect of oxytocin administration on social and routinized behaviors in autism using recently developed methods to accurately assess potential impact of effect size dependency and publication bias. Our frequentist meta-analysis yielded a significant summary estimate for the effect of oxytocin administration on social outcomes in autism ( d = 0.22, p < 0.001). The summary effect size estimate for routinized behavior outcomes was not statistically significant ( d = 0.14, p = 0.22), with follow up tests indicating that the effect size estimate was not either statistically equivalent ( Z = -1.06, p = 0.2), assuming a smallest effect size of interest of 0.25. Frequentist and Bayesian assessments for publication bias, as well as results from Robust Bayesian meta-analysis of oxytocin effects on social outcomes in autism, indicated that the summary effect size might be inflated due to publication bias. Future studies should aim to reduce bias by preregistering analysis plans and to increase precision with larger sample sizes.


Introduction
Oxytocin is a neuropeptide that is primarily known for its involvement in childbirth, milk let down, and bonding between parents and infants 1 .Recent research has provided insight on its role in social cognition and behavior, which includes the modulation of the perception of social information 2,3 , an increased attention to the eye region of faces 4 , a reduced aversion to fearful stimuli 5 , and the promotion of in-group cooperation 6,7 , among other mechanisms.However, some of these early reports of oxytocin's positive effects on social cognition and behavior have failed to replicate [9][10][11] .
Still, oxytocin's observed effects have led to increased interest in exploring its potential as an individualized support aid for autism 12 .Autism is a neurodevelopmental difference * characterized by a range of challenges associated with social interaction and communication, focused interests, and routinized/repetitive behaviors.Globally, approximately 1% of the population have an autism diagnosis, with a male-to-female ratio of about 4:1 13 .The autism spectrum includes individuals with diverse characteristics and required levels of support, varying from those requiring extensive support to those needing minimal or no support at all.For autistic individuals seeking assistance, there is a need for additional research into personalized support options.
Autism is typically identified early in life and has a strong genetic basis 14 .However, there is still much uncertainty regarding the neurobiological mechanisms underlying the diverse differences in autism.Recent evidence suggests that variations in the functioning of the oxytocin system in the central nervous system may be linked to challenges associated with social behavior and routinized behaviors, which are both core features of autism.For example, genes associated with the oxytocin system, including the oxytocin receptor gene (OXTR) and CD38 (a protein that regulates oxytocin release), have been linked to autism [15][16][17][18] .Variations in oxytocin signaling pathway genes can affect the affinity between oxytocin and its receptors, the distribution of receptors, and peripheral oxytocin levels.Moreover, some studies have suggested that individuals with autism have lower levels of circulating oxytocin compared to non-autistic individuals.
Numerous studies have investigated the cognitive and behavioral effects of oxytocin in autism.The first study that examined the effect of oxytocin administration on core autism features was conducted by Hollander and colleagues 21 in 2003, in which they reported that, compared to placebo, intravenous oxytocin administration decreased routinized behaviors.Subsequent trials further explored the effects of oxytocin on social cognitive and behavioral outcomes, generating significant interest in its potential as an individualized support tool in populations that experience difficulties with social interaction.Additionally, several functional neuroimaging studies reported changes in brain activation patterns in individuals with autism following oxytocin administration, particularly in brain regions that support the processing of social information, including the amygdala, medial prefrontal cortex, anterior insula, as well as reward-processing regions such as the nucleus accumbens.
However, mixed results, especially from studies investigating chronic oxytocin administration over extended administration periods, underscore the need for further validation of oxytocin effects in autism, particularly regarding long-term administration effects.Mixed results could be due to individual factors that can influence the effects of oxytocin administration, such as age, sex, and diagnostic heterogeneity, and etiological heterogeniety.Differences in study design may also contribute to these mixed results, such as the dosage and duration of the oxytocin administration schedule 33,34 , sample size 35 , and choice of outcome measure.Another significant concern is publication bias and lack of rigorous pre-registration, which can impact the outcomes of evidence synthesis.While several previous meta-analyses have not reported evidence of publication bias (e.g., 36,37 ), these studies have used tools that evaluate small study bias, which can include publication bias but also capture other types of bias associated with small studies.Furthermore, if any presence of publication bias exists, the summary effect size estimates have not been adjusted for the effect size inflation associated with publication bias.Recently, tools have emerged that can specifically test for the presence of publication bias and provide publication bias-adjusted effect sizes 38,39 .A more accurate synthesis of current evidence via publication bias-adjusted effect sizes will provide a more precise estimation of evidence to date and aid the planning of future investigations.
The objective of this systematic review and meta-analysis is to examine the impact of oxytocin administration compared to placebo on social behavior and routinized behaviors in individuals with autism by synthesizing data from available studies.While previous meta-analyses have provided valuable insights into the current state of oxytocin research in autism 36, , the recent availability of new meta-analysis tools allows the reevaluation of existing evidence using methods that more accurately evaluate effect size dependency and publication bias.This manuscript adheres to the PRISMA reporting guidelines 41 (Supplementary Table 1) and is pre-registered on PROSPERO J o u r n a l P r e -p r o o f (ID: CRD42022308149) and via a meta-analysis protocol 42 to minimize potential bias by outlining a priori analysis plans.

Search Procedure
The meta-analysis has been reported using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.A set of relevant search terms was carefully chosen in collaboration with an academic librarian to identify the key topics and terminology related to oxytocin administration studies in autism.The search terms used were "oxytocin" AND "intranasal" OR "intravenous" AND "autism".These terms were used to conduct a comprehensive literature search across three databases: Embase, PsycInfo, and Web of Science, focusing on titles, abstracts, and keywords (Supplementary Table 3).The search was completed in February 2023 with no specific dateof-publication limitation was applied.In addition to the database search, studies were also identified through previous meta-analyses 36,37,40,50,51 and by examining the reference lists of included articles.Furthermore, relevant databases were explored for related articles and articles that cited the included studies to identify additional relevant research.

Inclusion and Exclusion Criteria
Studies were selected to be included in the present meta-analysis if the following inclusion criteria were met: 1) Included the administration of oxytocin to participants.
2) Included a sample of individuals with autism.
3) Provided at least one measure of social functioning/cognition or routinized behavior.4) Provided effect sizes or sufficient statistical information for calculating effects sizes and their variances.
5) The study was double-blind, randomized, and placebo controlled.
6) The study was published in English.
The included studies had to contain statistical information for computing effect sizes for oxytocin administration effects on social functioning and/or routinized behavior in autism.If this information was not readily available in the article, we attempted to contact the authors.If the authors did respond and we did not obtain the required statistics to calculate effect sizes or if the effect sizes were not J o u r n a l P r e -p r o o f available in a previously reported meta-analysis, the study would be excluded, even though it met all other inclusion criteria.Results from peer-reviewed articles, preprints and clinical trial registries were eligible for inclusion.Studies were excluded if they meet the following exclusion criteria: 1) The study was a meta-analysis, review, case-study, animal study, or qualitative study.
2) The study did not involve intranasal or intravenous oxytocin administration.

Data extraction and management
Studies collected through the database searches were assessed for initial eligibility independently by first author K.A. based on screening of titles and abstracts.If studies appeared to fulfill eligibility, or if this was unclear, the full article was evaluated for study inclusion.Author K.A. selected studies that would be evaluated for study inclusion.Each of the studies were given an identification number, and 20% of these studies were randomly selected, using a custom R script for inclusion evaluation by author A.M.S to confirm reliability.Any disagreements were adjudicated by author D.S.Q.If there were more than 20% disagreement in this sample, author A.M.S evaluated the full text of all possible studies, with author D.S.Q.adjudicating any disagreements.When the final list of potentially eligible studies was identified, author K.A. extracted the data required to calculate effect sizes and their variances, as well as information to conduct moderator analysis.
We extracted data necessary to calculate standardized mean differences (SMD) between oxytocin and placebo conditions from each eligible study.Mean scores from each condition, standard deviations and sample sizes were extracted, in addition to study characteristics such as publication year, participant demographics, study design, outcome variable of interest, administration method, peer review status and administration type (for full list of moderators, see below).If means and standard deviations were not readily available in the articles, alternative methods for calculating SMD were used, including calculation of Cohen's d from t-statistics, F-statistics, or conversion of descriptive statistics from figures using WebPlotDigitizer.The precision of this method was validated on other eligible studies that included both bar plot and mean and standard errors (see Supplementary Text 1).
If alternative methods for calculating Cohen's d were not feasible, we attempted to contact the study authors.If we received no response, the article was excluded from the analysis.20% of calculated effect sizes were then randomly selected for verification by author B.G.All data is available on a dedicated Open Science Framework page (https://osf.io/xf9kr/).All decisions on inclusion and exclusion were documented to ensure an open, transparent, and systematic selection process.
J o u r n a l P r e -p r o o f

Data quality, small study bias and publication bias
We calculated the statistical power of each included study to evaluate its evidential value.This was done using the 'metameta' R package.Small study bias, which can encompass publication bias but also include other sources of bias, was assessed using Egger's regression test 51 and visualized via a contour-enhanced funnel plot 53 .A sensitivity analysis for publication bias was implemented using the R package 'PublicationBias' 38 , which can account for smaller sample sizes and provide a publicationbias adjusted summary effect size, along with a 'worst-case' summary estimate based on meta-analysis of only non-affirmative studies.This frequentist analysis was complemented by a Robust Bayesian meta-analysis, which can evaluate the relative evidence for a "publication bias" model against a "no publication bias" model, along with a publication bias-adjusted summary effect size estimate 39 .The Cochrane Risk of Bias tool 44 was also used to assess risk of bias in six domains at the study level: selection bias, detection bias, performance bias, attrition bias and other bias (i.e., not covered by specified domains).A table describing the risk of bias across these domains and a text summary is provided.The conclusions of the meta-analysis may be tempered if the risk of bias is judged to be high for a large proportion of studies.The GRADE system was used to assess and report the strength of evidence 53 .

Statistical analyses
We used R version 4.3.1.(R Core Team, 2023) for the statistical analyses.All primary analyses were conducted using the following R packages: metafor (version 4.4), meta (version 6.5) 56 , and RoBMA (version 3.1) 39 .Random-effects models were fitted with a restricted maximum-likelihood estimator.
Measures of effect size and variance for each study were calculated depending on the available information.If effect sizes are described in included articles, these will be extracted.If not, effect sizes will be calculated as standardized mean differences (SMD) based on available descriptive statistics.
When both per protocol and full analysis set data were presented, the full analysis set data was used for analysis to increase sample size.
Dependent effect sizes (i.e., effect sizes extracted from the same study or sample) were expected, as oxytocin administration studies in autism often include and report more than one outcome measure for social and/or routinized behaviors.Furthermore, as the included studies likely adopt a wide range of different measures of social functioning and/or routinized behaviors, and given mixed findings on oxytocin's effects, we expected significant heterogeneity, which is indexed by calculating J o u r n a l P r e -p r o o f I 2 and τ.To account for potential dependencies as well as within-and between study heterogeneity, a three-level meta-analytical model was constructed.
The frequentist meta-analysis was complemented by a multilevel Robust Bayesian metaanalysis using the RoBMA R package, which can quantify the relative likelihood of presence of true effect, heterogeneity, publication bias through the calculation of a Bayes factor (BF) and compare random effects models assuming the presence of a hierarchical structure to models assuming absence of hierarchical structure.The following thresholds were used to interpret results from RoBMA: BF < 0.1 corresponds to strong evidence, 0.1 < BF < 0.33 corresponds to moderate evidence, 9 0.33 < BF < 1 corresponds to anecdotal evidence for the null hypothesis, whereas 1 < BF < 3 corresponds to anecdotal evidence, 3 < BF < 10 corresponds to moderate evidence, and BF > 10 corresponds to strong evidence for the alternative hypothesis 57,58 .We carried out two separate meta-analyses: one synthesizing the effects of oxytocin administration on social functioning and cognition and one synthesizing at effects of oxytocin administration on routinized behaviors.We also carried out sensitivity analysis to identify potential outliers in our data (leave-one-out analysis) as well as sensitivity analysis excluding effect sizes derived by methods that are sometimes associated with less precision (e.g., t-statistics, summary statistics extracted from plots), and sensitivity analysis for publication bias in meta-analysis.

Moderator analyses
Moderators were specified prior to data collection and analysis.The following variables were analyzed as potential moderators of oxytocin effects on social and routinized behavior outcomes: 1. Age of participants: As oxytocin's role in cognition and behavior likely shifts throughout the lifespan 59 , the effects of oxytocin administration on social and routinized behaviors in autism could be moderated by the age of the sample used.The mean age of the sample was extracted for analysis.

Sex of participants: As oxytocin's effects on certain social cognitive processes and behaviors
appear to be sexually-dimorphic, sex was assessed as a potential moderator (percentage of males).
3. Year of publication: Earlier studies might exhibit differences in effect sizes due to improvements in methodology or publication bias.
4. Pre-registration in a recognized registry 5. Whether the effect size was associated with primary or secondary outcome J o u r n a l P r e -p r o o f 6. Oxytocin dosage: while 24 IU is the most commonly administered dose, some studies use higher or lower dosages.Considering the dose-dependent effects of oxytocin, oxytocin dosage was be assessed as a potential moderator.
7. Administration type: single administration vs. chronic administration 8. Administration duration 9. Outcome measure type: There are various approaches for collecting outcome measures that may influence effect sizes.Outcome measures will be categorized as task-based (e.g., Reading the Mind in the Eyes task), self-reported questionnaire, caregiver-reported questionnaire, or observation evaluation by expert/clinician.10.Level of required support: Oxytocin might be more beneficial for those with autism that require more support than others.If available, a common measure of support requirements will be collected from studies, such as IQ.
11. Study design: Within vs. between-participants design

Oxytocin administration
All studies administered oxytocin intranasally, except for one study 21 , which used intravenous infusion.Oxytocin dosage (International Units; IU) varied between studies, ranging from 3IU to J o u r n a l P r e -p r o o f 48IU, the most common dosage being 24IU (48%).12 studies investigated the effect of a single administration and 14 trials assessed oxytocin administration effects after repeated administrations over longer periods of time.The duration of the chronic administration trials ranged from 4 to 168 days.

Study participants
All studies included a sample of participants with an autism diagnosis.The mean age of study samples ranged from 5 years to 38 years.All study samples had a mean IQ above 70.Most included studies only recruited male participants (58%), while the percentage of males in trials that also recruited female participants ranged from 79% to 93% (see Supplementary Table 2 for overview of characteristics of included studies).

Risk of Bias
Risk of bias for every study included in the meta-analysis was rated using the Risk of Bias 2.0 tool (RoB-2) 43 , which evaluates the risk of bias across five domains: 1) the randomization process, 2) deviations from the intended interventions, 3) missing outcome data, 4) measurement of the outcome, and 5) selection of the reported result.If study articles provided sufficient information on randomization and blinding procedures, made pre-registered protocols and statistical analysis plans available, and accounted for potential missing data, this lowered a study's risk of bias.The lack of such information would contribute to higher risk of bias.Although preregistration has become more common in randomized controlled trials, pre-specified analysis plans are still not always available in clinical trial registries.No studies had indication of high risk of bias, 65 % of studies had at least one concern, and 35 % of studies had low risk across all categories.See Supplementary Figure 1 for an overview of results from the risk of bias assessment.

Meta-analysis
As stated in our pre-registered analysis plan 42 , we anticipated that several eligible studies would report more than one outcome measure on the effect of oxytocin on social cognition, social behavior or routinized behaviors.To address effect size dependency, which would violate meta-analysis assumptions, we conducted a three-level multilevel meta-analysis fitted with a restricted maximumlikelihood estimator 45 for both the social outcomes and the routinized behavior outcomes.This was complemented by a Robust Bayesian meta-analysis 39 , which provides a comparison of random J o u r n a l P r e -p r o o f effects models assuming the presence of a hierarchical structure vs. non-hierarchical structure, and a publication bias-adjusted summary effect size estimate.

Social outcomes
The meta-analysis on the effect of oxytocin administration on social outcomes in autism was conducted with 57 outcomes from 25 studies.The multilevel meta-analysis yielded a summary effect size of d = 0.22 [0.12, 0.32], p = <.0001 (Fig. 2).A sensitivity analysis for publication bias in metaanalyses yielded a publication bias-adjusted summary effect size of d = 0.15 [95% CI (0.07, 0.23), p < 0.001], and a worst-case estimate summary effect of d = 0.13 [95% CI (0.05, 0.21), p < 0.001], based on a publication bias selection ratio of 4.84 38 .However, a robust Bayesian meta-analysis revealed moderate evidence against the effect, BF10 = 0.177, with mean model-averaged estimate of d = 0.004, 95% CI [-0.13, 0.21].Comparing random effects models assuming the presence of a hierarchical structure to models assuming absence of hierarchical structure found anecdotal evidence in favor of clustering effect sizes within studies (BF  = 2.34).A power calculation for individual studies included in the meta-analysis on oxytocin effects on social outcomes in autism using the 'metameta' tool 46 found that the median statistical power to detect the observed effect size (d = 0.22) is 10.4%.Assuming that the observed effect size is the true effect size, the statistical power of included studies ranged from 7% to 41% (Fig. 3).

Heterogeneity among studies in the meta-analysis of social outcomes.
A test for heterogeneity indicated statistically significant variance among the social cognition and behavior outcomes [Q (df =56) = 80.12, p = 0.02)].The calculation of multilevel I 2 -statistics revealed that 5.6% of the total variance could be attributed to between-study variance, and 26.4% of the total variance could be explained by heterogeneity within studies.Robust Bayesian meta-analysis revealed weak evidence in favor of heterogeneity (BFrf =2.34), with a mean model-averaged estimate of τ = 0.1 [95% CI (0, 0.24)].Subgroup analysis was conducted to investigate the potential source of heterogeneity in accordance with our protocol 42 .Twelve moderator variables were specified in our preregistered protocol, but one was excluded from subgroup analysis (i.e., caregiver randomization allocation belief) due to lack of information in eligible studies.Subsequently, we fitted a randomeffects meta-regression model separately for 1) mean age of participants, 2) percentage of male participants, 3) year of publication, 4) primary vs. secondary outcome, 5) oxytocin dosage, 6) single vs. chronic administration, 7) administration duration, 8) outcome measure type (self-report, J o u r n a l P r e -p r o o f caregiver-report, cognitive task, observation/expert assessment, eye-fixation to social information), 9) level of required support (i.e., IQ), 10) study-design, and 11) preregistration status.See Table 1 for all moderator analysis results.

Publication bias and small study bias.
Small study bias was assessed using Egger's regression test and visual inspection of funnel plot asymmetry.Visual inspection of the funnel plot indicated the potential presence of asymmetry (Fig. 3).The results of Egger's regression test were consistent with funnel plot asymmetry (z = 2.62, p = 0.009).These tests for small study bias were complemented by an assessment for publication bias using RoBMA 38 , which found strong evidence in favor of presence of publication bias (BFpb = 47.28).This contrasted with the frequentist sensitivity analysis reported above, which yielded a publication bias-adjusted summary effect size of d = 0.15, and a 'worst-case' summary effect size of d = 0.13.

Sensitivity analysis.
We ran sensitivity analysis (leave-one-out) to identify potential outliers in our data.Influence diagnostics did not identify any significant outliers.Additionally, we ran sensitivity analysis excluding studies with effect sizes derived by alternative methods (data extracted from WebPlotDigitizer, t-statistics, F-statistics).The meta-analysis excluding effect sizes derived by alternative methods yielded a summary effect size of d = 0.19, [0.07, 0.3], p = 0.002, which was marginally smaller than the summary effect size from the original meta-analysis for social outcomes (d = 0.22 [0.12, 0.33], p = <.0001).Lastly, we ran a Bayesian sensitivity analysis, only including the effect size with the smallest standard error from each study.This sensitivity analysis found moderate evidence against the effect (BF10 = 0.21) with a mean model-averaged estimate of d = 0.004 [-.15, 0.17], which was the same model-averaged as the meta-analysis including all outcomes from each study (d = 0.004).

Routinized behavior outcomes
The meta-analyses on routinized behavior outcomes were conducted with 20 outcomes from 13 studies.The multilevel meta-analysis yielded a summary effect size estimate of d = 0.14 [-0.09, 0.36], p = 0.22 (Fig. 4).We followed this result up with an equivalence test for interpretation of statistically non-significant results.The results of the equivalence test indicated that we could not reject the null equivalence hypothesis with equivalence bounds of d= ±0.Assuming the observed effect size if the true effect size, the statistical power of included studies ranged from 6% to 11% (Fig. 5).

Heterogeneity among studies in the meta-analysis of routinized behavior outcomes. A test
for heterogeneity indicated significant variance among the routinized behavior outcomes [Q (df =19) = 34.31,p = 0.02)].I 2 -statistics revealed that approximately 50% of the total variance could be attributed to between-study variance, and less than 1% could be explained by heterogeneity within studies.The Robust Bayesian meta-analysis found strong evidence in favor of heterogeneity (BFrf =21.45), with mean model-averaged estimate τ = 0.24 [0.00, 0.47].Moderator analysis was conducted to investigate the source of heterogeneity.Separate random-effects meta-regression models were fitted for the same moderators as for the social outcomes.One study was excluded from the meta-regression model investigating dose as a moderator due to non-comparable dosage information 21 .Results from the moderator analysis are summarized in Table 2. Meta-regression analysis indicated that the duration of oxytocin administration was a statistically significant moderator of oxytocin's effects on for routinized behaviors (QM = 6.24, p = 0.02), with stronger effects associated with longer administration periods (Figure 6).

Publication bias and small study bias.
Small study bias was assessed using Egger's regression test and visual inspection of funnel plot asymmetry.Visual inspection of funnel plot and results of the test and did not support funnel plot asymmetry (z = -1.01,p = 0.31) (Figure 5).The test for small study bias was complemented by an assessment for publication bias using RoBMA 38 , which found weak evidence against the presence of publication bias (BFpb = 0.56).
J o u r n a l P r e -p r o o f 3.6.3.Sensitivity analysis.We ran sensitivity analyses (leave-one-out) to identify any potential outliers.Influence diagnostics did not identify any significant outliers.We also ran sensitivity analysis excluding one study where the effect size was derived by alternative methods (i.e., Fstatistic 47 ).After excluding this study, the meta-analysis yielded a summary effect size of d = 0.14, [-0.1, 0.38], p = 0.22, which was the same effect size estimate as the original meta-analysis for routinized behavior outcomes.Lastly, we ran a Bayesian sensitivity analysis, only including the effect size with the smallest standard error from each study.This sensitivity analysis found weak evidence against the effect (BF10 = 0.34) with mean model-averaged estimate d = 0.04 [0.00, 0.26], a slightly larger estimate than the results from the Bayesian meta-analysis including all outcomes from each study (d = 0.03).

Discussion
Here we performed a multilevel random effects meta-analysis with studies investigating the effects of oxytocin administration on social and routinized behavior outcomes in autism using recently developed methods that account for publication bias and that can evaluate evidence for null models.
For social outcomes, the analysis was conducted with 57 effect sizes derived from 25 studies, and for routinized behavior outcomes with 20 effect sizes derived from 13 studies.While our frequentist approach yielded a statistically significant bias-attenuated effect of oxytocin on social outcomes in autism, this was not consistent with our Bayesian analysis, which suggested moderate evidence against an effect.Our bias-attenuated frequentist analysis revealed no statistically significant effect of oxytocin on routinized behavior, which was consistent with the results of Bayesian analysis that revealed weak evidence against an effect.Follow-up equivalence testing indicated that effects as large as d = ± 0.25 could not be rejected.The promising, yet inconsistent, support for oxytocin's effects on social outcomes and inconclusive support for its effects on routinized behaviors indicate that additional data is required to draw more definitive conclusions.Indeed, power analysis for individual studies indicated that most of the included studies in both meta-analyses were statistically underpowered, with power ranging from 4% to 47% to determine the presence of an effect of d = 0.22 for social outcomes, and from 6 to 12% power to detect the presence of an effect of d = 0.16 for routinized behavior outcomes.In other words, the current evidence base of studies evaluating J o u r n a l P r e -p r o o f oxytocin's effects in autism is not suited to reliably detect or reject a wide range of small-to-medium effect sizes.
We found evidence of effect size heterogeneity among studies in evaluating social outcomes in both our frequentist and Bayesian analyses.A meta-regression model further investigating potential sources of heterogeneity found that accounting for oxytocin dosage significantly reduced the amount of residual heterogeneity, highlighting the importance of determining the most efficacious oxytocin dose 48 .However, interpreting the nature of this dose-effect relationship is challenging due to the many interacting influences that play into the relationship, such as whether a study used single vs. chronic administration, or administration duration.At present, sufficient data is not available for reliably fitting models to explore moderator interactions.We also found evidence of heterogeneity among studies in evaluating routinized behavior outcomes in both our frequentist and Bayesian analyses.Meta regression investigating the source of heterogeneity found that accounting for oxytocin administration duration reduced residual heterogeneity for routinized behavior outcomes.Our finding of administration duration as a significant moderator could indicate that chronic administration trials with longer administration durations may produce stronger oxytocin effects on routinized behaviors, compared to single administration trials, or chronic administration trials over shorter administration periods.This interpretation is still speculative, however, as the meta-analysis of oxytocin effects on routinized behavior outcomes includes a relatively small number of studies.
While we did not identify any statistically significant moderators apart from oxytocin dosage affecting social outcomes, a more detailed examination of individual moderator factors from our meta-regressions indicates that oxytocin's impact on outcomes in autism varies in magnitude based on several factors.For instance, the meta regression model for primary vs. secondary social outcomes found that the summary effect for primary outcomes was statistically significant, however it was not significantly larger than the summary effect for secondary outcomes.The meta-regression model for single administration in social outcomes trials also revealed that the chronic administration summary effect size was statistically significant, however, it was not significantly larger than the summary effect size for the single administration effect size summary.Interestingly, the summary effect size for oxytocin effect on social outcomes for studies that were not preregistered was J o u r n a l P r e -p r o o f statistically significant, whereas the summary effect for studies that were preregistered was not, which may reflect the influence that preregistration has on effect size deflation.
One important criterion in meta-analyses is that effect size measures are comparable across studies 44 .In the present review we aimed to include a larger number of studies by including a wider range of outcome measures within both the social cognitive and behavioral domain and the routinized behavior domain.The heterogeneity of the outcome measures included in the present review could have had consequences for the comparability of oxytocin effects across studies.For instance, oxytocin's effects might differ between performance on cognitive tasks that assess processing of social stimuli and self-report questionnaires that assess social support needs in everyday life.Although outcome measure type was not a significant moderator of oxytocin effect on neither social nor routinized cognition or behaviors, the summary effect estimates for the separate outcome types from our moderator analysis differed between outcome types in both meta-analyses.
Still, as the included studies all aimed to assess oxytocin administration as a potential aid for individuals with autism, either within the social or routinized behavior domains, we believe their inclusion in the present meta-analysis was relevant but acknowledge that the variety of outcome measures included in our meta-analysis could have implications for the interpretation of results.This meta-analysis included studies published over the course of almost two decades.Over this period the diagnostic criteria for autism had changed, most notably from the sub-category system of DSM-IV, where autistic disorder, Asperger's disorder, pervasive developmental disorder not otherwise specified (PDD-NOS), Rett's disorder, and childhood disintegrative disorder were all classified as Pervasive Developmental Disorders (PDDs), to a single 'autism spectrum disorder' (ASD) category in DSM-V.Despite this change, the DSV-V criteria for ASD captures over 90% of people who meet criteria for PDD 60 .In addition, more recent work has adopted novel approaches, such as exploring the impact of repeated administration or investigating the effect of different dosages, and generally used larger sample sizes that can more reliably detect a wider range of effects sizes.But despite these changes over time, the reported effects of oxytocin on social and routinized behavior outcomes in autism do not seem to be moderated by the year of publication.
The initial enthusiasm for oxytocin as a potential individualized support for autism has been damped by subsequent mixed results, for which publication bias has been nominated as an important contributor.While the present meta-analysis represents an up-to-date synthesis of available evidence using gold standard synthesis tools that account for publication bias, our results altogether cannot provide sufficient evidence for either confidently supporting or rejecting oxytocin as an J o u r n a l P r e -p r o o f individualized support aid for autism at present.More studies with larger samples will be required to make such a determination.Larger sample sizes should also be coupled with other approaches that increase the precision of intranasal oxytocin research 31 , such as improving our understanding of underlying mechanisms and refining methodology (e.g., intranasal administration approaches).In addition, it is important to also establish critical auxiliary hypothesis 47 , such as the most efficacious oxytocin dose, and to determine individual and contextual factors 48 that influence the response to oxytocin administration.Given the heterogeneity of autism, which is associated with various physiological and genetic underpinnings that contribute to various mechanisms, it is important to determine for whom oxytocin may provide a useful support aid, given that individual differences are likely to influence response 49.There is some evidence to suggest that oxytocin receptor gene variants are associated with response to oxytocin administration in autistic individuals 72 , however, others have not found similar effects 80 .Considering that oxytocin signaling is influenced by variants across over a hundred genes, future research may benefit from moving beyond a single candidate gene approach 81 for better understanding oxytocin's role in wellbeing and the heterogeneity of response to oxytocin's effects in autism.As our data and analysis scripts are openly available, future metaanalytic work can easily build on the present research as more studies are reported.

Registration and Protocol
Pre-registration for this review is available on PROSPERO (ID: CRD42022308149).Protocol with detailed plan for data collection and analysis was published as a preprint on Open Science Framework (https://doi.org/10.17605/OSF.IO/XF9KR).

Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this preprint article.(presence of significant effect) = p < 0.05.Black diamond: robust independent point estimate within all studies; gray diamond: robust independent point estimate within only the non-affirmative studies.
J o u r n a l P r e -p r o o f

Declaration of interests
☒ The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
☐ The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

Highlights
• A meta-analysis was performed on the effects of oxytocin on au8sm features.
• Oxytocin's effect on rou8nized behavior outcomes was inconclusive.
J o u r n a l P r e -p r o o f 10 (Z = 0.339, p = 0.633), d = ± 0.20 (Z = -0.592,p = 0.277), or d = ± 0.25 (Z = -1.057,p= 0.204).Robust Bayesian meta-J o u r n a l P r e -p r o o f analysis found moderate evidence against the effect (BF10 = 0.299) with a mean model-averaged estimate of d = 0.03 [0.00, 0.25].Comparing random effects models assuming the presence of a hierarchical structure to models assuming absence of hierarchical structure found anecdotal evidence in favor of clustering effect sizes within studies (BF  = 21.45).The sensitivity analysis for publication bias in meta-analyses yielded a publication bias-adjusted summary effect size estimate of 0.08 [95% CI (-0.11, 0.28), p = 0.38], and a worst-case summary effect size estimate of 0.06 [95% CI (-0.14, 0.26), p = 0.53] based on a selection ratio of 4.84.Power calculation for individual studies included in the meta-analysis on oxytocin effects on routinized behavior outcomes in autism found that the median statistical power to detect the observed effect size (Cohen's d = 0.16) is 7%.

Figure 2 .
Figure 2. Forest plot for the meta-analysis of oxytocin effects on social outcomes in autism.A three-level multilevel model was conducted with 57 outcomes from 25 studies.The first three columns on the left describe the author, year, oxytocin dosage, and outcome measure, respectively.Effect sizes are represented by filled circles, the size of which indicate the weighting of the point estimate.95% confidence intervals are represented by thin lines.The filled diamond represents the summary estimate (with the left/right edges indicating 95% confidence interval limits).

Figure 3 .
Figure 3. Study-level statistical power and small-study bias for social behavior and cognition outcomes.(A) A firepower plot visualizing the median statistical power of studies included in the meta-analysis of oxytocin's effects on social outcomes assuming the observed effect size is the true effect size, along with a range of true effect sizes.(B) A Contour-enhanced funnel plot for meta-analysis on oxytocin effects on social outcomes.Black dots represent individual effect sizes, the vertical line represents the overall effect size, and contour-shaded areas represent p-value cutoffs, as indicated by the legend.(C) Significance funnel plot for meta-analysis on oxytocin effects on social outcomes.Effect sizes on the diagonal line report a p-value of 0.05.Effect sizes labeled as "affirmative" (presence of significant effect) = p < 0.05.Black diamond: robust independent point estimate within all studies; gray diamond: robust independent point estimate within only the non-affirmative studies.J o u r n a l P r e -p r o o f

Figure 4 .
Figure 4. Forest plot for the meta-analysis of oxytocin effects on routinized behavior outcomes in autism.A three-level multilevel model was conducted with 20 outcomes from 13 studies.The first three columns on the left describe the author, year, oxytocin dosage, and outcome measure, respectively.Effect sizes are represented by filled circles, the size of which indicate the weighting of the point estimate.95% confidence intervals are represented by thin lines.The filled diamond represents the summary estimate (with the left/right edges indicating 95% confidence interval limits).J o u r n a l P r e -p r o o f

Figure 5 .
Figure 5. Study-level statistical power and small-study bias for routinised behavior outcomes.(A) A firepower plot visualizing the median statistical power of studies included in the meta-analysis of oxytocin's effects on routinised behavior outcomes assuming the observed effect size is the true effect size, along with a range of true effect sizes.(B) A Contour-enhanced funnel plot for meta-analysis on oxytocin effects on routinised behavior outcomes.Black dots represent individual effect sizes, the vertical line represents the overall effect size, and contour-shaded areas represent pvalue cutoffs, as indicated by the legend.(C) Significance funnel plot for meta-analysis on oxytocin effects on routinised behavior outcomes.Effect sizes on the diagonal line report a p-value of 0.05.Effect sizes labeled as "affirmative"

Figure 6 :
Figure 6: Scatterplot with fitted regression line for the relationship between administration duration (total number of days) and effect size for routinized behavior outcomes.The grey region represents the 95% confidence interval of the linear regression line.