Abstract
There is growing interest in the role of the oxytocin system in social cognition and behavior. Peripheral oxytocin concentrations are regularly used to approximate central concentrations in psychiatric research. This methodological approach has obvious appeal given the invasiveness of cerebrospinal fluid collection. However, the validity of this approach and potential moderators of the association between central and peripheral levels are unclear. Thus, we conducted a pre-registered systematic search and meta-analysis of correlations between central and peripheral oxytocin concentrations. A search of databases yielded 17 eligible studies for effect size synthesis and moderator analysis, resulting in a total sample size of 516 participants and subjects. Overall, a positive association between central and peripheral oxytocin concentrations was revealed [r=0.29, 95% CI (0.15, 0.42), p<0.0001], along with a moderate-to-high level of heterogeneity across effect sizes [Q=88.14, p<0.0001], and no evidence of publication bias (p=0.45). This association was significantly moderated by experimental context [Qb(4), p=0.0016]. The strongest association was observed after intranasal oxytocin administration (r=0.67, p<.0001), a correlation that was significantly greater (p=.0002) than the equivalent association under baseline conditions (r=0.08, p=.31). These results support the use of peripheral levels of oxytocin as a marker of central levels, but only after exogenous oxytocin administration. Despite the popularity of using peripheral OT levels to approximate central levels during baseline conditions, this approach is not supported by the present results.
Introduction
Oxytocin is a nine amino acid neuropeptide that acts on the widely distributed G-protein coupled oxytocin receptor in humans and almost all other vertebrate species (1). Oxytocin is released both into the central nervous system (CNS) and peripheral circulation from neurosecretory cells in the paraventricular (PVN) and supraoptical (SON) nuclei of the hypothalamus, where most endogenous oxytocin is synthesized. Central and peripheral compartments of the oxytocin system are separated anatomically by the blood-brain barrier, that only in exceptional cases is appreciably permeated by oxytocin (2).
Through central action, oxytocin is critically involved in a range of social behaviors and social cognitive functions (3). Endogenous, or naturally produced, oxytocin appears to co-vary with social cognition at all levels of information processing in humans and other mammals, with similar effects after administration of exogenous oxytocin (4). Growing clinical interest (5) has focused on neurodevelopmental and psychiatric conditions characterized by social cognition and behavioral impairments, such as autism spectrum disorder (ASD) (6,7) and schizophrenia (8), with the hope to explore the potential of oxytocin as a biomarker of these conditions, better understand their potential etiological pathways, and ultimately to ameliorate the associated social-cognitive and behavioral symptoms.
Several methodological approaches have been adopted to the study of oxytocin involvement in normal and impaired social behavior and cognition. These include the measurement of psychological or neurobiological outcomes after administration of exogenous oxytocin, and the assessment of endogenous oxytocin concentration covariance with psychological phenotypes and psychiatric disorder status. While crucial to the latter, concentrations of oxytocin have been sampled within both of these research traditions. Although the social cognitive effects of oxytocin are attributed to central mechanisms, oxytocin concentrations have typically, but not universally, been sampled in peripheral fluids such as blood plasma, saliva, and urine (9). Consequentially, that peripheral oxytocin concentrations approximate central bioavailability of the neuropeptide has been a crucial assumption in research where peripheral oxytocin concentrations are correlated with psychological phenotypes or psychiatric disorder status.
Although some animal research indicates that central release from the hypothalamus and peripheral release via the posterior pituitary is coordinated (10-12), other research does not support this (13,14). Research is also mixed in humans, with some results consistent with related levels of central and peripheral endogenous oxytocin (15), while others report no significant associations (16). After exogenous oxytocin delivered via intranasal administration in humans, one study found a significant association between cerebrospinal fluid (CSF) and blood plasma concentrations of oxytocin (17), while another found no significant association (18). Using peripheral oxytocin concentrations to index central concentrations is clearly appealing, given the more invasive procedures required to collect centrally circulating fluids in humans. However, it is currently unclear whether and when peripheral oxytocin measures can be used to index CNS concentrations and central oxytocin bioavailability.
The present systematic review and meta-analysis synthesized studies in which central and peripheral measures of oxytocin were simultaneously sampled into a summary effect size. The strength of the summary effect size is indicative of the plausibility of peripheral oxytocin as an index for central oxytocin concentrations. As eligible studies were likely to vary in a range of contextual specifications, several potential moderator variables were considered, including experimental paradigm, oxytocin sampling location, subject species, biochemical analysis methods, year of publication, and study quality. Such differences between contexts may contribute to variance in the correlations between central and peripheral oxytocin. Thus, it is possible that peripheral oxytocin can index central oxytocin concentrations in some contexts, but not others. Together, the purpose of this study was to examine whether, and under which circumstances, peripheral oxytocin is a correlate of central oxytocin concentrations.
Materials and Methods
The systematic search and meta-analysis was conducted in accordance with the PRISMA guidelines (19) (Supplementary material I) and recent recommendations for conducting correlational meta-analyses (20). Prior to the execution of the systematic search and meta-analysis, the protocol for this systematic review and meta-analysis was published (21) and pre-registered on the PROSPERO registry (CRD42015027864).
Systematic literature search and inclusion of eligible studies
A systematic literature search was performed in two iterations to retrieve studies in which oxytocin had been simultaneously sampled in fluids or tissues located in central (e.g., local extracellular fluid or CSF) or peripheral (e.g. blood plasma or saliva) regions of the body. In the first iteration, a search was performed, using Ovid, in Embase and Medline with the following combination of terms: (oxytocin) AND (concentration* OR level*) AND (plasma OR blood OR saliva* OR urin*) AND (central OR csf OR “cerebrospinal fluid”). The following constraints were applied to limit search results: the result should be (i) a full-text article or a conference abstract, (ii) written in English, that was (iii) published after 1971, when biochemical analysis of oxytocin content using enzyme immunoassay was made commercially available. The search was conducted on August 2, 2016, and resulted in a total of 572 studies. Out of these, 111 were relevant. A second iteration was performed in which citing articles and reference lists of included studies were examined for remaining relevant studies (Fig. 1). After retrieval, relevant studies were screened for inclusion based on the criterion that effect sizes for the correlation between central and peripheral concentrations of oxytocin must be obtainable. While 121 of the studies retrieved in the systematic search were relevant, only 17 of these satisfied this criterion.
Data extraction and management
Effect sizes and sample sizes were extracted from eligible studies. For some articles, effect sizes were stated explicitly, or directly obtainable through tables of individual values. In other articles, individual values were represented in graphs such as scatterplots, in which case a web plot digitizer (22) was used for conversion of plots into numerical values. Since some articles contained both a scatterplot and a directly stated effect size, this plot digitizer was validated through comparing effect sizes provided by authors with plot digitizer outputs, revealing almost perfect precision (Supplementary material II). Some articles did not provide relevant effect sizes, individual values in tables, or scatterplots. Since 15 years is a common time frame for the retention of clinical data, authors of such articles published from 2001 were contacted and asked to provide effect sizes. Articles lacking this information that were published before 2001 (n = 69), and studies performed by authors that were not able to respond to the data request (n = 35), were not included in the meta-analysis. Data were extracted from all eligible studies using a custom data extraction form (see Supplementary material III).
Statistical analysis
Statistical analysis was performed with R statistical software version 3.2.4. (R Core Team, 2016), using the MAc (23), metafor (24), and multcomp (25) R packages. The dataset and script to perform the analyses are available at http://osf/aj55y/
Prior to meta-analytic synthesis, raw effect sizes were transformed to Fischer’s z for variance stabilization (26). Raw effect sizes given as Spearman’s ρ were first transformed to Pearson’s r according to Gilpin (27), and then transformed to Fischer’s z for meta-analysis. For studies reporting several effect sizes, or reporting one effect size based on repeated measures, within-study variance was estimated using the procedure described in supplementary material IV. A random effects model (28), where between-studies variance (τ2) was estimated using a restricted maximum likelihood method was used in the synthesis of individual effect sizes into a summary effect size. Outlier diagnostics were also performed to identify potential effect size outliers (24). Point estimates were converted back to Pearson’s r for interpretive purposes. The observed variance between studies may be due to heterogeneity (variance in the true effect sizes between studies) and within-study variance. Q, the significance of Q, and I2 were computed in order to examine variance and heterogeneity among effect sizes of included studies. I2 values of ~25%, ~50%, and ~75% were interpreted as low, moderate, and high, respectively (29).
Potential moderator variables were defined a priori (21). Some of the levels for moderator variables were also defined a priori, such as the levels baseline condition (lack of experimental intervention) and intranasal administration for the experimental paradigm moderator. Other levels of moderator variables were adjusted from preplanned analyses post hoc based on the specific characteristics of included studies (for details, see Supplementary material V). A random effects model with separate estimates of between studies variance was applied for all categorical moderator variables, yielding summary weighted mean effects and the significance of subgroup effects, which were calculated for each subgroup. Although mammals share essential oxytocin system characteristics, such as production of oxytocin in the hypothalamus, peripheral and central release of oxytocin from hypothalamus, and a blood brain barrier that inhibits diffusion of oxytocin between the CNS and systemic circulation (1), the between-species differences (21) necessitated an additional analysis to examine the role of species in the different effects observed between experimental paradigms. When there were more than two subgroups, pairwise comparisons were performed between all moderator categories with Holm-adjusted p-values to control the family-wise error rate. Meta-regression models were fitted to account for heterogeneity of continuous moderator variables.
Data quality measures
Small study bias, which includes both publication bias and study quality bias (30)(31), was assessed by visually inspecting a funnel plot and performing Egger’s regression test (24). A significant test (p < .05) is indicative of small study bias. A contour enhanced funnel plot, which superimposes key areas of statistical significance (p = .1, p = .05, p = .01), was constructed to specifically assess for risk of publication bias (32). An over-representation of effect sizes in the key areas of significance is indicative of publication bias risk. Since the decision to report a specific effect size, in contrast to the decision to publish a study, is not directly dependent on sample size, the regression test for funnel plot asymmetry does not rule out the possibility that there could be a bias in the type of evidence that is reported in published studies. To examine whether this was a source of bias in the set of included studies, the included studies that explicitly stated effect sizes were compared to the studies where effect sizes were obtained by other means, such as data scraping or author request.
There may also be problems with validity of the data that are internal to included studies. A custom risk of bias tool (Supplementary Material VI) was used (by ME and AMM) to systematically assess within-study risk of bias in included studies. This tool was developed by adapting the tool used in another meta-analysis (33) to the context of oxytocin research.
Results
17 studies yielding 32 effect sizes were included in the meta-analysis (Table 1; Fig. 1). The total number of participants/subjects across studies was 516. Among these, 257 were human, 248 were rodents, 7 were sheep, and 4 subjects were non-human primates.
Association between central and peripheral concentrations of oxytocin
There was a positive correlation between central and peripheral concentrations of oxytocin [r = 0.29, 95% CI (0.15, 0.42), p < 0.0001; Fig 2]. Egger’s regression test revealed no evidence of publication bias (p = .45; Fig 3A). An inspection of the contour enhanced funnel plot did not reveal an over-representation of effect sizes in the significance contours (Fig. 3B). Furthermore, a meta-regression revealed that risk of bias did not influence effect sizes (p= 0.22; Fig 3C). Influence diagnostics identified one potential outlier (17). A sensitivity analysis, which involved re-analysis without the identified outlier, revealed a similar summary effect size as the original analysis that was also statistically significant [r = 0.24, 95% CI (0.12, 0.36), p = 0.0002]. As the sensitivity analysis suggested that this single effect size only had a modest effect on the overall meta-analysis, it was retained for the remainder of the analyses. In the total sample of included studies, there was a moderate-to-high level of heterogeneity [Q = 88.14, p < .0001, I2 = 63.5% (37.8%, 76.8%)]. Accordingly, moderator analyses were performed to identify sources of heterogeneity.
Impact of moderators on effect size
A moderator analysis revealed that part of the heterogeneity in the model was due to the type of experimental paradigm [Qb(4) = 17.44, p = 0.002; Fig 4A]. Across experimental paradigms, positive associations were observed for the intranasal oxytocin (IN-OT) condition (r = .67, p < .0001, k = 4) and after stress interventions (r = .49, p = 0.0005, k = 5; Supplementary material VII). In contrast, no association was observed in the baseline condition (r = 0.08, p = .31, k = 15). The subgroup effects for the peripheral oxytocin administration category (r = 0.29, p = .28, k = 3), as well as for the ‘other’ category (r = 0.30, p = .09, k = 5) were not significant. The results for the baseline condition were similar (r = 0.10, p=0.26, k = 13) when applying a strict rather than inclusive extension of ‘baseline’. A comparison of all possible pairwise comparisons with Holm corrected p-values revealed that the IN-OT point estimate was significantly greater than the baseline point estimate (p = .002; Fig. 4A). While there were no other significant pairwise comparisons, the increased stress point estimate compared to baseline point estimate was on the border of statistical significance (p = .089). When constrained to human studies, results for the levels of the experimental paradigm moderator were reproduced [Qb(2) = 7.56, p = 0.02], with no significant correlation in the baseline condition [r = 0.05, 95% CI (-0.19, 0.29), p = .58, k = 7, I2=0%], and a significant correlation in the intranasal condition [r = 0.71, 95% CI (0.34, 0.89), p = 0.013, k = 2, I2 = 93%].
Analysis of the effect of central sampling location on effect sizes was on the border of significance [Qb(3) = 6.71, p=0.08; Fig. 4B], suggesting that specific brain sampling location differences may contribute to observed heterogeneity. Across levels of the central sampling location moderator, subgroup effects for hypothalamus (r = 0.42, p = 0.0003, k = 10), central amygdala (r = 0.52, p = 0.034, k = 3), and hippocampus (r = 0.53, p = 0.023, k = 3; Supplementary material VII) were significant. The subgroup effect for samples taken from CSF (r = 0.14, p = 0.14, k = 16) was not significant. Pairwise comparisons did not reveal any significant difference between any of these subgroups.
The moderator analysis for species was not significant [Qb(3) = 1.93, p = 0.59; Fig. 4C], suggesting that species diversity might not contribute to heterogeneity among effect sizes. Across levels of the species moderator, only the subgroup effect for rodents was significant (r = 0.36, p = .0004, k = 19; Supplementary material VII). The point estimates for human (r = 0.22, p = 0.081, k = 10), sheep (r=0.19, p = .54, k = 2), and non-human primate (r = -0.34, p = .53, k = 1) subgroups were not significant. There was no significant difference between any of the levels of this moderator variable. An exploratory mixed-effect meta-regression model was fitted to assess whether the influence of species type (human vs. rodent) on the correlation between central and peripheral levels varied between experimental paradigms (intranasal oxytocin vs. baseline vs. other). This model revealed no evidence for a significant interaction (Qb(2) = 0.43, p =.81).
The biochemical analysis method moderator analysis was not significant [Qb(2) = 4.67, p=0.097; Fig. 4D], indicating that this moderator variable is not likely to contribute to heterogeneity among effect sizes. Across the levels of the biochemical analysis method moderator variable, subgroup effects for both RIA (r = 0.28, p = 0.0005, k = 24) and EIA (r=0.43, p=0.002, k = 6) were significant. There was no significant effect for LC/MS (r = -0.2, p = 0.43, k = 2) and no significant differences between the levels of this moderator variable. The peptide extraction moderator analysis was not significant [Qb(1) = 0.05, p = 0.82; Fig. 4E]. Both with (r = 0.31, p = 0.0001, k = 24) and without (r = 0.35, p = 0.028, k = 6) extraction subgroup effects were significant (Supplementary material VII). Finally, the year of publication did not significantly moderate the relationship between central and peripheral oxytocin concentrations [Q(1) = 2.19, p = .14; Fig 3D].
Discussion
The present systematic meta-analysis revealed a positive correlation between concentrations of oxytocin in blood plasma and oxytocin concentrations in the CNS. However, the association was moderate and showed a high degree of heterogeneity, suggesting that the observed association might not be present across all contexts. Experimental paradigm was the moderator variable most likely to account for this heterogeneity. After IN-OT, as well as after an experimental stressor, there was a positive correlation between central and peripheral oxytocin concentrations. However, in the baseline condition, there was no evidence of correlation, neither for the entire sample of subjects, nor for any of the species analyzed separately. Notably, there was a statistically significant difference between the summary statistic for IN-OT studies and baseline studies. Given the lack of evidence for a correlation between peripheral and central oxytocin levels in the baseline condition, the data suggest blood plasma may not efficiently index central oxytocin concentrations under baseline conditions. Furthermore, this result provides additional indirect evidence for the effectiveness of the blood-brain barrier in restricting oxytocin diffusion between systemic circulation and the CNS (2), as well supporting the hypothesis that under baseline conditions, hypothalamic oxytocin release into blood and into the CNS is uncoordinated (13).
There is a substantial body of research attempting to link peripheral oxytocin concentrations with psychological phenotypes or psychiatric disorder status. Since the social-cognitive effects of oxytocin have been assumed to arise from oxytocin action in the CNS, the assumption that peripheral and central oxytocin concentrations correlate in a baseline condition was crucial in the interpretation of the results from these two approaches (e.g. 34,35). This assumption is called into question by the present data. These results may have two possible, mutually exclusive, implications for the interpretation of studies within these research traditions: either the apparent social cognitive effects are type I errors produced by chance, or the demonstrated covariance between social cognition and endogenous oxytocin in systemic circulation arise from some phenomenon unrelated to central oxytocin levels. The former potential interpretation is consistent with the evidence of publication bias that has surfaced in the field of psychological and psychiatric oxytocin research (9,36,37). The latter interpretation points to a potential peripheral mechanism for the observed social cognitive correlates of basal peripheral oxytocin concentrations. One potential causal mechanism is oxytocin action on peripheral tissues that provide afferent feedback to the CNS (1).
In contrast to what was discovered under baseline conditions, this metaanalysis revealed a positive correlation between central and peripheral oxytocin after intranasal administration of oxytocin. Almost every study examining the effects of exogenous oxytocin on social cognition and behavior in normal and clinical populations have made use of the intranasal delivery route (e.g. 38-41). The motivation behind administering oxytocin intranasally is to obtain non-invasive delivery of oxytocin into the brain. Although vasopressin, which is structurally similar to oxytocin, has been shown to enter the CSF after intranasal administration (42), it is not entirely clear where intranasally administered oxytocin travels, or whether it actually reaches brain areas containing oxytocin receptors such as the hypothalamus or the amygdalae (43). However, recent work in humans comparing intranasal and intravenous oxytocin administration indicates that despite comparable peripheral oxytocin concentrations after both administration routes, social cognitive (44) and neural effects (45) were only observed after intranasal administration. Together, these results are consistent with a direct nose-to-brain transport of intranasally administered oxytocin via olfactory and trigeminal nerve fibers.
In this meta-analysis, a positive association was also found between central and peripheral concentrations of oxytocin after experimental stress induction. Stress induction involved either separation from a mother (46), or a forced swim test (47). As the authors suggest (47), the hypothalamus-pituitary-adrenal (HPA) axis and related hormones such as corticosterone interact with the oxytocin system to regulate stress responses. Such interaction may occur through interneurons between magnocellular and parvocellular neurons in the PVN (48), from which oxytocin and corticotropin-releasing hormone are released. Furthermore, interaction may be mediated through corticosteroid effects on vasoconstriction and heart rate, which in turn could affect oxytocin release through baroreceptors and the vagal feedback system (1,43).
There are some limitations to the study worth mentioning. First, to estimate variances for effect sizes from repeated measures, dependent samples variance estimation was used to control for dependency between samples. Since exact dependencies between repeated samples were unknown, there is a chance that variances for effect sizes obtained in repeated measures designs were slightly overestimated or underestimated, relative to variances for effect sizes obtained in single sample designs. A differential variance estimation would favor one of the two study types with respect to the relative weight they were afforded in the main analysis. However, since there is no a priori reason to believe that study type should impact upon the estimated effect sizes, it is unlikely that this potential bias had any considerable effects on the results. Second, even if there was no evidence for publication bias, or for bias in report of effect sizes, there may be some bias in the subjects sampled for studies where CSF was collected. Across included studies, some of the human participants had medical conditions (15,17,18). Medical conditions are often associated with pain, and pain may influence oxytocin release: in one study, chemical pain stimulation increased oxytocin release within the brain, but not in plasma (49). If pain leads to uncoordinated release, then this may bias the results of this meta-analysis in a negative direction. However, the strongest correlation between central and peripheral concentrations of oxytocin among included studies – which was also identified as a potential outlier – was observed in a sample of headache patients (17). This may point to the opposite possibility that pain could bias the effect sizes of this meta-analysis in a positive direction. To ensure that this study did not inflate the effect size for the IN-OT moderator, a secondary analysis was performed with this study removed, yielding comparable results.
The collection of peripheral oxytocin measures to index central levels has obvious appeal given the difficulties surrounding central collection. However, research has yet to establish whether this is a valid experimental approach. The results of this meta-analysis indicate that there is a positive association between central and peripheral concentrations of oxytocin, but this association depends on experimental context. As there was evidence for a positive association between central and peripheral concentrations of oxytocin after intranasal oxytocin administration and experimental stress induction – both overall and when only including humans in the analysis – blood plasma could be used to approximate central levels under these circumstances. However, as there was no evidence for an association between central and peripheral oxytocin concentrations under baseline conditions, future studies on the role of basal oxytocin in cognition or social behavior should avoid using peripheral oxytocin measures to make inferences on central oxytocin concentrations.
Conflict of Interest
OAA, LTW, and DSQ are investigators in a project studying oxytocin’s effects after intranasal delivery partnered by OptiNose AS (Oslo, Norway) and funded by a User-driven Research based Innovation (BIA) grant (219483) from the Research Council of Norway (RCN). The RCN and partner contributed to funding this review (through salary support to DSQ and project support to OAA, LTW, and DSQ); however, they had no influence in the ideas contained in the manuscript and no role in the writing of the manuscript.
Acknowledgements
MV received salary support from the Research Council of Norway (RCN) via a grant for students in clinical psychology programmes. GAA is funded by the Cooperative Research Centre for Living with Autism (Autism CRC), established and supported under the Australian Government’s Cooperative Research Centres Program. The Research Council of Norway (RCN) and OptiNose AS contributed to funding this review through a BIA project grant (219483) via salary support to DSQ and project support to OAA, LTW, and DSQ. LTW is supported by the South-Eastern Norway Regional Health Authority (2014097). DSQ is supported by an Excellence Grant from the Novo Nordisk Foundation (NNF16OC0019856). We thank Hege Kristin Ringnes (University of Oslo Library) for providing guidance on our systematic search strategy.