Influence of User Personality Traits and Attitudes on Interactions With Social Robots: Systematic Review

We conducted a systematic review of the literature on user personality traits and attitudes as predictors of human-robot interaction. We registered our review protocol in the PROSPERO database [CRD42021233557].

We considered English-language peer-reviewed journal articles and conference proceedings, with no restrictions based on date. To be included, we required that studies contain a quantitative measure of personality traits or attitude, and report outcomes (either quantitative or qualitative) related to an interaction between a person and a social robot. To reduce the variability of the interaction outcome measures and to improve our ability to compare between the studies, our sample only included research that involved direct human-robot interaction; studies were excluded if they involved the participant observing a third party interacting with a robot, or viewing photos or video. Included studies could feature participants of any age or diagnostic group.

The literature search was conducted on November 14, 2020 using the following electronic databases: MEDLINE®, IEEE (including ACM), EMBASE, PsycINFO, and Web of Science. Search strategy was developed in consultation with a research librarian. Exact search terms varied slightly across databases, but included: (robot*), AND (interaction*) OR (relationship*), AND (personality) OR (attitude*). Full details are available in Supplementary Text S1.

A data extraction template was developed and piloted by K.K and J.A.D. Following piloting and discussion, the following data were extracted from each paper: bibliographic information (publication type, year of publication, etc.), country of study, relevant research aims, study design, human-robot interaction protocol, robot used and its characteristics (humanoid/non-humanoid, speaking/non-speaking, etc.), mode of robot control, predictive and outcome measures, participant characteristics, and findings.

To assess risk of bias, we used a modified version of an existing assessment tool which has been previously used to evaluate the literature on attitudes and anxiety directed at social robotics (Naneva et al., 2020). The quality measure assessed each paper in two broad domains 1) Study validity and 2) Outcome measure quality. The assessment includes three questions to evaluate study validity focusing on existence of alternative explanations, sampling bias and representativeness of the sample. These dimensions should be measurable during review of the paper and are important for gauging the quality of the reported results. We chose to evaluate outcome measure quality because it is vital for determining whether or not there was an effect of personality traits or attitudes towards robots on the HRI. Thus, for outcome measures we have asked questions regarding validity, test-retest reliability and internal consistency. Additionally, we have added the size of the sample to the quality criteria. The full instrument used for quality assessment, including the definitions user for “Low” and “High” assessment, is included in Supplementary Text S2.

The sample obtained in this study was not suitable for meta-analysis due to the heterogeneity outcome measures, both qualitative and quantitative, and robot-interaction protocols. As a result, based on guidance from the Joanna Briggs Institute (JBI) on mixed methods systematic reviews, we took a convergent integrated approach to analysing the data (Stern et al., 2020). This approach allows for combining qualitative and quantitative data to perform a synthesis. As recommended by JBI, we transformed all quantitative results from reviewed studies into qualitative statements. Once all data was in qualitative form, we performed meta-aggregation of all findings. In this process, the statements describing results from all included studies were categorized. The resulting categories were then combined to form synthesized findings (Pearson et al., 2011). This process was collaborative, with two authors (K.K and J.A.D.) reaching consensus through discussion.

An initial pool of 5267 references was obtained (Figure 1). These materials were managed using Covidence, an online software platform (Veritas Health Innovation, n.d.). First, duplicate results were removed. Then, authors K.K and J.A.D. each screened the titles and abstracts of all works against the inclusion/exclusion criteria. Disagreements were resolved through discussion. The two authors then independently examined the full text of the remaining potentially eligible studies, with further discussion of any discrepancies. A final set of fifty-six eligible studies were identified and subjected to extraction.

The characteristics of studies included in this review are summarized in Table 1. The number of participants in the studies ranged from 3 to 164 (mean = 46).

Full summary of included studies is shown in Table 2. Number of publications by year is depicted in Figure 2.

After data extraction and qualitization of the findings, we obtained 102 individual findings that related to our guiding questions (see introduction).

The findings were subsequently grouped into 31 categories and formed 11 synthesized findings. Two findings remained ungrouped. Supplemental Table S1 contains all the findings, categories and synthesized findings.

The summary of quality assessment is available in Table 3. The overall assessment is based on the total rankings a study has received. If a study scored High, Low or Unsure in at least half (4/8 or more) of assessment areas, the overall score was the same as the majority of the scores. If no clear majority of assessments could be established, the overall ranking of the study was rated as Unsure. There were 8 studies in the sample that received an overall High quality score, 35 studies received an overall Low score and for 13 studies we were unable to establish a quality assessment score with the information available in the publications.

The most frequently used measure of personality traits in the reviewed studies was the Big Five Personality Inventory e.g., (Gosling et al., 2003; Rammstedt & John, 2007) (k=17). One specific measure based on this model was NEO Personality Inventory-3 (Costa & Mccrea, 1992). The second most frequently used measure were the Eysenck Personality Inventory and Questionnaire (Eysenck, 1991; Eysenck & Eysenck, 1965) used in 5 studies. The detailed information on user trait measures is summarized in Table 4.

The Association Between User Traits, Robot Traits and HRI Assessment. Establishing the association between user traits and general interaction outcomes such as interaction length, preference for the robot, engagement with the robot and evaluation of the robot was the dominant theme in 17 extracted study findings (Agrigoroaie & Tapus, 2020; Aly & Tapus, 2013, 2016; Andrist et al., 2015; Celiktutan et al., 2019; Celiktutan & Gunes, 2015; Correia et al., 2019; Craenen et al., 2018; Cruz-Maya & Tapus, 2017; K. M. Lee et al., 2006; Park et al., 2012; So et al., 2008; Tapus et al., 2008).

For some users, interacting with robots matching their traits led to more positive human-robot interaction outcomes. Users tended to prefer robots that match their traits (Aly & Tapus, 2013, 2016; Andrist et al., 2015; Correia et al., 2019; Craenen et al., 2018; HeeSeon Park et al., 2012; So et al., 2008) have longer interactions with robots that have traits similar to theirs (Agrigoroaie & Tapus, 2020; Andrist et al., 2015; Tapus et al., 2008) and tended to be more engaged and perform better in interactions with robots similar to them (Andrist et al., 2015; Celiktutan et al., 2019; Celiktutan & Gunes, 2015). While these findings support the similarity principle, some of the reviewed studies also provided evidence to support the complementarity principle. In these studies, participants preferred the robots that exhibited traits different from their own, or enjoyed the interaction with these robots more than with those matching their traits (Craenen et al., 2018; Cruz-Maya & Tapus, 2017; K. M. Lee et al., 2006). These findings tended to be specific to particular user groups. For instance, in a study by Cruz-Maya and Tapus (Cruz-Maya & Tapus, 2017) extroverted female participants were found to prefer the distance of the introverted robot in an interaction (0.8m distance, as opposed to 0.6m distance of the extroverted robot).

User Traits Are Related to Their Impressions of Robots. Another focus of investigation was whether user traits are associated with perceptions or impressions of robots. Extroversion was the most frequently investigated human trait in this context. The reviewed studies found that extroverted participants tend to anthropomorphize the robots more (Park et al., 2012; Salem et al., 2015), report higher psychological closeness to robots (Salem et al., 2015) and perceive them as more friendly (Park et al., 2012). One study reported findings providing evidence against positive association between user extroversion and perceived anthropomorphism of the robot (Haring et al., 2015).

Other user traits such as emotional stability, attachment style and trait loneliness were related to perceived anthropomorphism of the robot (Dziergwa et al., 2018; Haring et al., 2015; Li et al., 2020; Salem et al., 2015). Two findings reported an association between user traits and perceived social presence of robots. Park et al. (Park et al., 2012) found that introverted participants perceived the robot as less socially present, while Rossi et al. (Rossi et al., 2020) found that participants with greater ability to recognize the mental states of others rated robot sociability lower. Other correlations included higher perceived positivity of the robot assigned by conscientious participants, higher perceived realism assigned by participants with higher neuroticism, higher perceived positivity and realism of the robot rated by participants with greater openness to experience (Celiktutan & Gunes, 2015). Additionally, participants higher in conscientiousness perceived the robot as more stressful and more expressive in certain conditions (Cruz-Maya & Tapus, 2016a).

User Traits Are Related to Robot Acceptance and Attitudes Towards Robots. User extroversion (Park et al., 2012), agreeableness (Celiktutan & Gunes, 2015) and tendency towards parasocial interaction (N. Lee et al., 2011) were found to be related to greater enjoyment of interaction with the robot and more positive attitudes towards robots. Meanwhile, trait loneliness and avoidant attachment style were associated with lower acceptance of robots (Dziergwa et al., 2018; Li et al., 2020). The higher participants scored in conscientiousness, the less they liked a robot that behaved socially (Looije et al., 2010).

User Traits Are Related to User Behavior Towards Robots. Studies investigating the correlation between the extroversion personality dimension and behavior found that participants who were more extroverted tended to be more engaged in the interaction with the robot and try different types of interactions (Abe et al., 2017; Haring et al., 2013, 2015; Hwang & Lee, 2013; Ivaldi et al., 2017; Salam et al., 2017). Agreeable and conscientious participants tended to be more comfortable and engaged interacting with a robot (Salam et al., 2017; Takayama & Pantofaru, 2009), with the exception of children with high agreeableness who were less likely to maintain eye contact with a robot (Abe et al., 2017). High neuroticism was associated with lower engagement in interaction (Haring et al., 2015; Salam et al., 2017) and greater distance from the robot (Takayama & Pantofaru, 2009). Other traits associated with behaviors towards a robot include introversion correlated with higher head nod synchrony with a robot (Thepsoonthorn et al., 2018) and more monotonous touch (Hwang & Lee, 2013); internal locus of control correlated with greater disclosure during conversation with a robot (Nomura & Kanda, 2012); secure attachment style related to treating the robot like a human being and avoidant attachment style related to treating the robot more like a machine (Dziergwa et al., 2018). Technology competence and enthusiasm for technology influenced the money offers participants made to a robot during a game (Nitsch & Glassen, 2015).

User Traits Are Related to Performance/Outcomes of HRI. Since potential future uses of robots include administering assessments and supporting people during work tasks, the influence of individual differences on outcomes of interaction with a robot is of interest for studies included in this review. Introverts were found to perform better in robot-delivered tasks (Agrigoroaie & Tapus, 2020; Cruz-Maya & Tapus, 2016a). For robot administered tests, openness to experience and extroversion (Kanero et al., 2018; Rossi et al., 2018, 2020), as well as neuroticism (Cruz-Maya & Tapus, 2016b; Rossi et al., 2020)) were linked to higher scores. A study by Jeong et al. found that participants with high conscientiousness and low neuroticism were more ready to change after interacting with a robot (Jeong et al., 2020). Participants with higher conscientiousness were also more motivated by robot to perform a task, under certain conditions (Cruz-Maya & Tapus, 2016a).

Two dominant measures assessing users’ attitudes towards robots were the Negative Attitudes Towards Robots Scale (Nomura et al., 2006) (12 studies) and Robot Anxiety Scale (Nomura & Kanda, 2003) (7 studies). Other measures used in the studies included Godspeed Questionnaire (Bartneck et al., 2009), Technology Commitment Scale (Neyer et al., 2016), Attitudes Towards Robots Taking Social Roles (Xu, 2019), Unified Theory of Acceptance and Use of Technology Questionnaire (Venkatesh et al., 2003), Technology Affinity (Leichtmann & Nitsch, 2021), Implicit Anthropomorphism Scale (Sundar, 2004), Ezer Scale, Kozak Scale, Demoulin Scale and Cottrell Scale (Wullenkord et al., 2016). Some of these instruments were adapted for individual studies.

Attitudes Towards Robots Are Related to User Behavior and Performance in HRI. Evidence from the analyzed studies suggests that the attitudes that users hold towards robots predicted the allowable distance between the robot and the user. Generally, the more negative the attitudes were, the larger physical distance from robots was preferred (Nomura et al., 2007; Takayama & Pantofaru, 2009). Further, attitudes were reflected in user behavior towards robots. Participants with more negative attitudes were more likely to turn off the robot when it asked them not to (Spatola & Wudarczyk, 2021), looked at the robot’s face less (Ivaldi et al., 2017), were less engaged with the robot (Chen et al., 2020) and made emotional utterances towards the robot (Nomura et al., 2008). De Graaf and Allouch found that for women, negative attitudes towards robots explained some variance in how much they talked to a robot (de Graaf & Ben Allouch, 2013). Attitudes were also positively correlated with scores on test following robot-delivered instruction (Kanero et al., 2018) and with robot evaluation (Stafford et al., 2014).

Positive Attitudes Towards Robots Positively Impact User Evaluation And Acceptance. Negative attitudes towards robots at baseline were negatively correlated with robot evaluation following an interaction (Stafford et al., 2010, 2014). Participants who were more comfortable with robots taking social roles had greater intention to use robots in the future (Xu, 2019). Similarly, the acceptance subscale of technology commitment scale was a good predictor of using a robot during an experiment (Bernotat & Eyssel, 2017). The more negative attitudes participants held towards robots, the less safe they felt during an interaction (Takayama & Pantofaru, 2009).

Interaction with the Robot is Related to Change in User Attitudes. There are mixed findings regarding the attitude change following an interaction with a social robot. Björling et al. and Stafford et al. reported an improvement in attitudes following an interaction (Bjorling et al., 2020; Stafford et al., 2010), and de Graaf and Allouch found no change (de Graaf & Ben Allouch, 2013). It is also suggested that negative emotions towards robots prior to interaction may predict more negative attitudes towards robots following the encounter, especially if the participant is asked to touch the robot (Wullenkord et al., 2016). Thus touch may be a factor in predicting user attitudes towards robots following an interaction. For instance, Haring et al. found that length of touch predicted higher anthropomorphism scores assigned to the robot while comfort with casual touch did not predict attitudes towards robots (Haring et al., 2015).

Robot Interaction Leads to Increased Anxiety. Five of the studies included in this review showed consistent evidence that participants tend to experience higher anxiety towards robots following the interaction (de Graaf & Ben Allouch, 2013; Nomura et al., 2007, 2008; Nomura & Kawakami, 2011; Rossi et al., 2020). Further, anxiety before the interaction explained how much male participants talked to the robot (de Graaf & Ben Allouch, 2013) and was associated with making negative disclosures to the robot (Nomura & Kawakami, 2011). It is unclear whether the increase in anxiety is attributable to neuroticism.

This goal of this systematic review was to characterize how individual differences and attitudes towards robots influence human-robot interaction. The distribution of included studies by year shows a steadily increasing interest in this topic, as social robots become more prevalent and tailoring the interaction to an individual increases in importance. Our novel application of the mixed-methods meta-synthesis in the field of social robotics revealed and described key relationships in HRI and illustrate the variety, complexity and interrelatedness of individual traits and interaction outcomes investigated in social HRI field.

We have assessed the quality of evidence, based on 1) Study validity and 2) Outcome measure quality. The assessment included three questions to evaluate study validity focusing on existence of alternative explanations, sampling bias and representativeness of the sample; four questions regarding outcome measure quality, and sample size. Overall, the majority of studies included in this review were assessed to have low quality of evidence, which is not an uncommon finding in social robot research (Kabacińska et al., 2020) given the fast-moving pace of the field and the complexity of factors in the design of an HRI study. Studies tend to have small sample sizes and usually include only one brief interaction with the robot and a single assessment, which increases the risk of bias. It is important to note that not all included studies focused on our topic of interest as the main research question, which limits the usefulness of quality evaluation in these instances. Further, 33 of the studies were conference proceedings. This inherently shorter format results in less detailed reporting and contributes to lower quality scores, which may not reflect the quality of research conducted and is the standard for reporting much of HRI research. Therefore, the quality assessment results should be interpreted with these contextualizing factors in mind. Suggestions on how the quality of the studies could be improved were discussed in Kabacińska et al. (2020) and are relevant to the present study. Publications which were rated as high quality were well-controlled, had larger sample sizes and included thorough descriptions of study designs (e.g., Rossi et al., 2018).

The CASA framework posits that human-computer interactions can be modelled on human-human interactions, and as such it is not surprising that many of the included studies investigated the complementarity and similarity attraction hypotheses, which are also investigated in human social interactions. While the majority of findings reported in this review support the similarity hypothesis, as is the case for human studies, some studies provided evidence to the contrary. Therefore, it is unclear whether or not there is support for CASA in the HRI space. To better establish which hypothesis holds, there is a need to develop a coherent framework that would allow for efficient synthesis of studies across the field of HRI (Robert, 2018). Social robot-specific interaction framework would allow for development of models for social HRI that could generate testable hypotheses. Better-established outcome measures and more detailed reporting would allow for more quantitative investigation into which hypothesis is supported by direct HRI. Future quantitative studies evaluating these two hypotheses would be valuable in modelling and designing successful HRI. Further, there is a need to investigate other potential factors that could influence users’ preferences such as demographic variables (Esteban et al., 2022). The studies analyzed in this review also suggest that attitudes towards robots may be correlated with interaction outcomes and robot evaluation (Bernotat & Eyssel, 2017; Stafford et al., 2010, 2014; Xu, 2019) and that personality traits are associated with user attitudes towards robots (Celiktutan & Gunes, 2015; N. Lee et al., 2011) which adds to the limitations of modeling interaction outcomes based on personality traits alone.

Determining which social rules govern interactions between humans and robots is crucial for better HRI design which will improve interaction quality and robot acceptance. Although individual personality trait differences surely play a role in how people interact with robots, there may be more interaction-specific ways to model behavior, that are less intertwined with other possible variables. One theory that enables modelling of interpersonal interaction is Affect Control Theory (Heise, 2007). Rather than focusing on relationship formation as similarity and complementarity attraction hypotheses do, this theory allows for modelling individual situational interactions which is more appropriate in the context of usually single, brief interactions with robots in the studies. Hoey, Schröder, and Alhothali (Hoey et al., 2013) used the theory to develop BayesAct, which models a two-way interaction for artificial intelligence applications and allows for behavior prediction. This model is being used in development of artificial agents that are emotionally aligned with the users (Ghafurian et al., 2020).

One of the reasons that prevents us from conducting a traditional meta-analysis of the studies in the sample and thus getting more unambiguous answers is the inconsistency of outcome measures between the studies. The HRI outcomes are often not clearly defined. The exception to this observation is proxemics studies in which allowable distance between the robot and the user is measured between the conditions. Among the studies that used quantitative outcome measures, many chose to design custom, ad-hoc questionnaires to investigate HRI outcomes. To allow for meaningful comparison between studies, there is a need to employ better-defined and more consistent outcome measures. Existing initiatives are working towards more standardized HRI outcome measures. For instance, National Institute of Standards and Technology Performance of Human-Robot Interaction project is focused on developing tests, metrics and databases that will allow for consistent performance measurement, establishment of key performance indicators and thus increase the quality of existing data in the field of HRI (Bagchi et al., 2020; Performance of Human-Robot Interaction, n.d.). Continuous work towards creating standard HRI measures is crucial for further development of not only social robotics, but any human-facing robotic technology.

When measuring independent variables, researchers use various forms of Big Five personality trait inventory (Eysenck & Eysenck, 1965) to establish personality traits. To avoid using proprietary measures and achieve greater consistency, resources such as the International Personality Item Pool should be used (Goldberg et al., 2006). It is an open-source collaboration that collects scales assessing personality traits. For measuring attitudes towards robots the Negative Attitudes Towards Robots Scale (Nomura et al., 2007) is being consistently used.

In this review we found evidence that users’ personality traits are correlated with their impressions of robots and robot evaluations (e.g., Park et al., 2012; Salem et al., 2015), however, similar findings were also reported when looking at the association between attitudes towards robots and robot evaluation and acceptance (e.g., Xu, 2019). To disentangle the attitude and personality trait findings, it is crucial to conduct HRI studies in the future that control for these two factors.

The consistent results regarding increased anxiety towards robots after interacting with a robot are contradictory to another finding that attitudes towards robots tend to improve after an interaction. It is unclear what contributes to this discrepancy and further research is required to explain this phenomenon. For instance, there is a well-established link between trait anxiety and neuroticism (Knowles & Olatunji, 2020), however the included studies did not investigate whether the increased anxiety following the HRI was linked to trait neuroticism. Based on a study of fifty seven adults, Broadbent et al. found that participants who held more human-like mental representations of robots had greater blood pressure increases when interacting with a robot, compared to participants who held less human-like representations (Broadbent et al., 2011). Therefore, the anxiety difference from pre-to post-interaction may be dependent on user expectations, but also on the type of robot that is being used. The large majority (39/56) of studies in this review used humanoid robots, which could contribute to the increase in anxiety. Researchers should carefully select robots for intended applications. For instance, when using social robots to support mental health (Kabacińska et al., 2020), especially in vulnerable populations like children, using non-humanoid robots could be more suitable to minimize possible increase in anxiety. Similarly, some relationships between personality traits and attitudes (for instance, trait loneliness and avoidant attachment style) may be especially relevant to certain social robot use settings and populations of interest such as older adult care, where robots are designed as companions to mitigate loneliness and provide companionship (Berridge et al., 2023; Dosso et al., 2022, 2023). These context -dependent considerations highlight the critical imperative of involving end-users in social robot development (Martin et al., 2024; Robillard & Kabacińska, 2020). Ultimately, as suggested by Naneva et al., more data is needed to determine whether and how robot design moderates users’ anxiety and attitudes towards robots (Naneva et al., 2020).

The reviewed studies used a variety of different robots, including research prototypes, which makes it difficult to ascertain to what extent the reported results are robot specific. For the purposes of reporting in the results sections, we categorized the robots into different types (i.e. humanoid, non-humanoid, animal-like, head only and other). Further, there are significant differences in how various robots of a certain type appear. For instance, humanoid robots can be more or less human-like, and can range from android robots, which are aiming at mimicking humans (Haring et al., 2013) to robots with some human-like features (Nitsch & Glassen, 2015). The more humanoid robots resemble humans, the more likely they are to be perceived as eerie or disturbing, as is described by the uncanny valley phenomenon (Zhang et al., 2020). Thus, since different robots are used across studies, it is unclear to what extent the human-likeness of the robot contributes to the results, especially when outcome measures include affective responses or attitudes towards robots. A 2021 study investigated whether this phenomenon could be reduced (Yam et al., 2021). Yam, Bigman and Gray found that this effect is mediated by individual’s tendency to dehumanize others but can also be reduced by explaining to participants that humanoid robots do not have mental states and cannot experience emotions. Including this simple procedure in the study design could mitigate undesirable uncanny valley effect in HRI studies. Further, more studies which compare multiple robots are needed to help isolate which robot abilities or features are contributing to HRI outcomes. It is also important to consider that different types of robots, especially robots at different stages of development, have different levels of ability to perform the intended actions. How well individual robots perform, can then affect the outcomes of research studies that are investigating social HRI. Thus, the wide range of robots used in included studies adds a layer of complexity to the synthesis of the findings. For instance, people may treat robots who are abstract differently from robots that attempt to mimic humans due to the negative influence of the Uncanny Valley phenomenon (Wang et al., 2015) and attempts at reducing this effect included de-humanizing of humanoid robots (Yam et al., 2021).

The main shortcoming of the present study is the lack of quantitative meta-analysis. While a meta-synthesis provides a thorough overview of the findings collected in this review and novel insights in the relationships that drive HRI, it does not allow for calculating effect sizes and confidence intervals, resulting in less precise, narrative synthesized findings. Additionally, findings included in this review were sometimes different from the main focus of the studies and as a result were reported in less detail.

Further, personality traits in the reviewed studies were largely assessed using the Big Five personality model, which is not fully applicable to all global contexts (Laajaj et al., 2019; Thalmayer et al., 2022). Amber Thalmayer and colleagues propose a new, culturally de-centered personality trait model that would allow for studying personality traits in HRI outside of the western, industrialized context (Thalmayer et al., 2024).

In this systematic review we have captured and evaluated studies investigating the relationship between personality traits, user attitudes and the outcomes of human-robot interaction.

Since the review combined quantitative and qualitative findings, we used a mixed methods approach to data analysis, which allowed for synthesizing diverse studies in the field. Despite low quality of evidence, we identified some emergent categories of findings, including extroversion being tied to better interaction outcomes. The categories and synthesized findings resulting from our research can serve as a guide for refining research designs in HRI and developing further hypotheses in areas where contradictory findings exist.

Contributed to conception and design: KK, KV, JAD, JMR

Contributed to acquisition of data: KK, KV, JAD

Contributed to analysis and interpretation of data: KK, JAD, KV, TJP, JMR

Drafted and/or revised the article: KK, JAD, TJP, JMR

Approved the submitted version for publication: KK, JAD, KV, TJP, JMR

Jill A. Dosso is an associate editor at Collabra: Psychology. She was not involved the review process of this manuscript.

The authors acknowledge funding from: AGE-WELL NCE, the New Frontiers in Research Fund, The BC Children’s Hospital Research Institute, the BC Children’s Hospital Foundation, and Michael Smith Foundation for Health Research.

All the data generated in this study is available in the Supplementary Material.