Network perspective

Psychological research taking a network perspective was first used as a theoretical model in cognitive developmental research [1, 2]. It was later developed into a psychometric model, extensively applied in clinical psychology [3]. This perspective has recently also been adopted in the study of attitudes. The Causal Attitude Network (CAN) model [4] conceptualizes attitudes as networks, consisting of evaluative reactions (nodes) and interactions between them (edges, i.e., links between the nodes). Edges represent either excitatory or inhibitory relations with varying weights (i.e., strength of relations varies between evaluative reactions). These nodes, consisting of the cognitive, affective and behavioral elements of attitudes, together with the edges connecting them, form a network. To exemplify, Fig 1 shows a hypothetical and simplified network concerning attitudes towards hand hygiene during pandemics. Dalege, Borsboom [5] show how the CAN model explains individual attitudes and their dynamics. In Dalege and van der Maas [6] this model is used to explain the differences between implicit and explicit attitude measures.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Part of hypothetical and simplified attitude network.

The network concerns the attitude towards hand hygiene, consisting of a behavioral element (‘Washing hands frequently’), a cognitive evaluation (‘Hand hygiene is effective to prevent spread’) and three affective evaluations (‘Fear of disease’ and ‘I dislike [parched hands resulting from hand hygiene / hand moisturizer]’). Strength of the relations is indicated by edge width. In this example, disliking parched hands has a stronger relation to washing hands frequently than perceiving hand hygiene as effective to prevent the spread, indicating that the former consideration has a stronger association with the behavioral element than the latter (which is indicated in the network by different edge width). Furthermore, although disliking hand moisturizer is indirectly associated with the behavioral element of washing hands through the disliking parched hands node, this affective evaluation is (in this fictitious case) not directly related to washing hands.

https://doi.org/10.1371/journal.pone.0276439.g001

Networks can be calculated with (e.g., survey) data. Nodes represent the measured psychological constructs, which can consist of single items or a combination of items (e.g., average on multiple items). Edges between nodes in psychological networks cannot be directly observed and are therefore parameters that are estimated from data [7]. In Gaussian networks estimated with continuous and ordinal data, edges represent partial correlations between nodes, where the correlation between two nodes is computed conditional on all other nodes in the network [e.g., see 8, 9]. Attitude network models can also be statistically estimated with empirical data [10]. Empirical applications of the CAN model are predominantly based on cross-sectional data [e.g., 11–13]. This enables computing undirected networks: networks in which edges represent associations and do not provide information on directions of these relations. Calculating networks with directed edges that represent predictive effects is also possible but requires a repeated measures design.

In addition to graphically representing the interplay of variables, network analysis also computes valuable properties of networks such as overall connectivity (i.e., average strength of connections between nodes) and centrality of nodes within the network (with central nodes having more and/or stronger connections to other nodes). The most commonly used centrality measures are Strength (calculated by summing absolute edge weights), Closeness (calculating distances between nodes based on the shortest path length) and Betweenness (calculated based on how often it lies on the shortest path between nodes) of nodes [14]. Of these centrality metrics, Closeness and Betweenness are considered least suitable for psychological networks [15], and can be problematic because they treat associations between nodes as distances [14].

In the context of attitudes, connectivity of an attitude network determines its stability and resistance to change [16], while centrality informs about a node’s connectedness with other nodes in the attitude network. Such network properties are also considered informative for designing effective interventions. Theoretically, changing a central (i.e., highly connected) node is likely to have a more profound effect on a network than changing a peripheral (i.e., less connected) node, given the central nodes’ relatively high connectiveness to other nodes [17]. This is however not undisputed in undirected networks for reasons such as missing latent common causes and problems regarding the specification of boundaries of networks [see 15, 18, 19 for an overview]. Also, high connectivity of a node can result from different scenario’s, namely a) that the node highly affects other nodes, b) that the node is highly affected by other nodes, or c) a combination of the first two directions of effects. Results of interventions depend on these directions of effects: intervening on a node that highly affects other nodes is likely to have a profound effect on the network, whereas changing a node that is highly affected by other nodes is unlikely to (durably) affect the network. Nevertheless, some studies have demonstrated the value of a network perspective in studying interventions [20–22].

One example of applying the CAN model to interventions is by Wicker, Nohlen [22], in the context of sustainable consumer behaviors. They determined consumers’ attitude network about plastic and included their willingness to pay for sustainable alternatives as a node. It was found that a persuasive message aimed at a central node that was most strongly connected to this willingness to pay was an effective means to change actual behavior. This underlines that interventions tailored to attitude networks can be effective in changing behavior [22]. This study specifically focused on behavior, and research into how interventions impact attitude networks as a whole seems to be mostly lacking. Two notable exceptions were conducted in the mental health domain. Bekhuis, Schoevers [20] and Blanken, Van Der Zweerde [21] demonstrated how a network approach can be used to gain insight in treatment effects aimed at specific symptoms of depression and insomnia. By including the treatment in the network as a node they provide insight into direct and indirect treatment effects.

Thus, although decades of research has investigated how attitudes are affected by interventions [23–25], research into attitude networks and change is sparse. Such research could however improve our understanding of (indirect) intervention effects. In addition to the effects of interventions on the node the intervention was aimed at, there can be ‘downstream’ effects (on other nodes) in the attitude network. This is important, since certain structures can either reinforce or attenuate effects of interventions, depending on connections between nodes. For instance, a triangular motif (three interconnected nodes) of negative relations between nodes can possibly undermine intervention effects: an intervention aimed at increasing node A can lead to decreasing node B, which can lead to increasing node C, which can lead to decreasing node A, thus undermining the intervention’s effect. This depends not only on the type of connections (i.e., positive or negative), but also on the number and weights of connections with other nodes (i.e., centrality).

Finally, while networks and their properties are empirically determined, the question what does and what does not fall within the scope of a network is to a large extent dependent on what is included in the analysis (as in any statistical analyses). The CAN model focuses on elements that make up individual attitudes, but attitudes do not exist within a vacuum and are related to each other and to other psychological variables. Accordingly, the scope of attitude networks has recently been broadened by including attitudes as well as a wider range of social psychological constructs such as trust and social norms. This broad attitude networks approach provides insight into the interplay of attitudes and their elements within a system of other relevant variables [12]. In the present research we also focus on such broad networks, encompassing different attitudes as well as other psychological variables. With this we aim to do justice to the inherent complexity of health behavior in the context of a pandemic, in this case the COVID-19 pandemic.

The present research

We aim to contribute to designing effective interventions in complex psychological systems by providing empirical insight into how network structures explain intervention effects. In this study, we examine how broad attitude networks respond to tailored interventions aimed at specific nodes selected based on network properties (i.e., connectiveness with other nodes). We investigate whether these interventions a) affect the node the intervention was aimed at (i.e., target node), b) affect nodes other than target nodes, and if so, whether these effects can be explained by connections with target nodes (i.e., downstream effects). It is hypothesized that interventions impact target nodes and that node centrality predicts whether interventions have downstream effects on other nodes as well. We expected that changing more central nodes affected broad attitude networks more than changing peripheral nodes, because central nodes are more connected to other nodes in the network. Furthermore, we expect that downstream effects of interventions are induced through the nodes that interventions were aimed at, thus that target nodes explain effects of interventions beyond the target node. Note that the wording of affecting nodes is used for brevity and refers to affecting scores on the item(s) forming that node. States and changes of nodes were only measured at the behavioral level (i.e., observed responses on survey items). Our research can therefore only speak to that level of explanation and is not on the level of mental processes. The current study thus goes beyond what earlier intervention studies have done [e.g., 22], by formally comparing intervention effects on nodes with relatively high and low centrality, and testing mediation effects.

To examine the above, we designed a longitudinal study in which participants responded to a survey on COVID-19 related variables five times over a period of approximately 10 weeks (April 23^th–June 30^th 2020). This study design also enables examination of the temporal dynamics of the broad COVID-19 attitude network. This is however not the focus of the current paper and will be addressed in the discussion section. The COVID-19 pandemic was deemed a suitable empirical setting to conduct this study, given the pandemic’s complex and unprecedented nature, and the importance of (preventive) behavior. The psychological variables included in the survey are based on an extensive literature review of health behavior during a pandemic by Bish and Michie [26]. Their proposed frameworks, with determinants depending on types of protective behaviors, are covered by several prominent generic models within psychology contributing to explaining behavior, such as the Theory of Planned Behavior [TPB; 27] and Health Belief Model [HBM; 28], but also include other social psychological factors relevant in the context of pandemics. These factors, for instance trust, perceived knowledge, health complaints, well-being and individual differences, were also identified in more recent literature on psychological determinants of compliance during pandemics (see S1.1 in S1 File for more information and literature). Importantly, this research does not aim for an exhaustive combination or comprehensive network of variables relevant for compliance, since such a set of variables is arguably extremely large. Instead, it aims to demonstrate how networks with variables relevant for compliance during pandemics that extend beyond attitudes, respond to tailored interventions. We present the current research in two parts: the first part presents the case study of this research (i.e., broad COVID-19 attitude network), and the second part presents results of intervening in the broad COVID-19 attitude network.

Participants and design

This study was approved by the Ethics Review Board of the University of Amsterdam (2020-SP-12194) and not preregistered due to its explorative nature. Dutch participants were recruited via a research panel from Ipsos. Regarding sample size, the aim was to collect as large a sample as possible to ensure sufficient power to find between-subject effects in the last measurement, after which we checked the stability of the estimated network. We aimed for, and far exceeded, a minimum of 500 participants because this is the highest advised number of participants for a moderately sized network with either continuous or binary data [9, 29], and these types of data were combined in the current study. These sample sizes are advised in order to obtain accurate network estimation, indicating that the estimated network is an accurate representation of the true underlying network [7]. The initial sample was representative of the Dutch population in terms of gender, age and country region. These participants were invited for subsequent measures and no new participants were added in subsequent measures. Each measurement contained two attention checks to enhance data quality, and participants who failed both attention checks within one measurement were excluded from thereon (see Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Demographic and intervention specifics.

https://doi.org/10.1371/journal.pone.0276439.t001

The broad COVID-19 attitude network was based on all participants that completed the first survey (valid N = 6,093). Research into effects of interventions on the network was based on two samples: the first sample included respondents that participated in the third wave in which we presented the first intervention, and the second sample included respondents that participated in the fifth wave in which we presented the second intervention. Respondents that participated in the first intervention (third wave) were also eligible for inclusion in the second intervention (fifth wave). Therefore, respondents that were included in the second intervention were also included in the first intervention. Table 1 provides sample and descriptive information for all samples, including waves in between.

In waves 3 and 5, interventions were designed for two nodes, with two experimental conditions per node (low / high). In the low (high) experimental conditions, interventions were aimed at decreasing (increasing) scores on nodes. Each wave also contained a control condition, resulting in a total of five intervention conditions for both wave 3 and 5. Interventions in wave 3 were aimed at the nodes Trust and Social Norm. Wave 5 contained interventions aimed at the nodes Measures Support and Economic Consequences. Again, these nodes were selected based on network properties, and therefore not necessary directly related to the behavioral element (i.e., compliance) in the broad COVID-19 attitude network. More specifically, target nodes were selected based on different node strength (i.e., central and peripheral node) in the first wave and feasibility of designing a targeted intervention (see results section for more information on selecting nodes). As mentioned, node strength is considered the most suitable centrality measure for psychological networks. This research thus focuses on node strength as a centrality measure. Node strength represents the conditional association between a node and other nodes in the network and is meant to aid interpretation of networks. It is calculated by summing absolute weights of edges a node has with connected nodes. High node strength thus represents the number and strength of a node’s relations but does not inform us about directions of relations: edges can represent effects from or to nodes, or be bidirectional.

The interventions contained a manipulation check consisting of a multiple-choice question following the intervention to assess whether participants read provided information (e.g., Social Norm conditions: ‘According to the information above, has the level of compliance with the corona measures increased, decreased or remained the same?’). Table 1 reports the number and percentage of participants answering this question correctly, thus passing the manipulation check, resulting in the intervention subsamples.

Measures

Data was collected with an online survey in Dutch. The first step in composing the survey was to identify relevant constructs in the literature (see S1.1 in S1 File). As mentioned, our aim was not to include an exhaustive set of variables related to behavior during pandemics, nor to evaluate a specific (set of) model(s). Instead, we identified important constructs that extend beyond attitudes to include in the network to broaden the scope of the network. Subsequently, a survey was developed with items based on these constructs. An overview of the survey items is provided in S1 File (S1.2). After data collection in wave 1, we constructed psychological variables as nodes by either a predetermined combination of items or based on components in the data as identified through Principal Axis Factoring (PAF; a data reduction technique). Constructs that were surveyed with a single or two items were predetermined nodes (see Table 2 for the number of items per node). Other predetermined nodes were validated scales (see Table 2 –the nodes with references in the footnotes were the adopted validated scales). The remaining nodes were constructed based on the results of the PAF. S1 File (S2.2) provides node-specific descriptions of the approach to combine items into nodes, including PAF results.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Nodes (psychological variables) based on items in the survey, including item examples and answer scales.

https://doi.org/10.1371/journal.pone.0276439.t002

Table 2 presents the resulting nodes, including examples of survey items for those constructs and their answer scales. Nodes that were based on multiple items consisted of mean scores of items relevant for that node, except for Risk Perception (i.e., the product of likelihood and severity of an infection) and Mental Well-being (i.e., sum score). A detailed description of each node in the network, its interpretation and the scale reliability as observed in the current study is provided in S1 File (S2.3).

Procedure

Participants that subscribed to Ipsos’ research panel received an invitation via e-mail to participate in our study. Only participants that finished the survey received an invitation for subsequent waves. They received compensation in the form of points that can be spent at web shops. Participants were informed about participation and provided written consent. They were also asked to commit to the longitudinal research design beforehand.

Participants were randomly assigned to an intervention condition by the software in which the questionnaire was programmed (i.e., Qualtrics). The intervention was presented at the beginning of the survey within a specific wave. Each experimental condition consisted of two headings of online news articles and additional information in text, followed by a question that served as a manipulation check. Additional information consisted of (fictional) preliminary results of the current study supporting the presented news articles and possible explanations for these results. The control condition consisted of one news article and a manipulation check. The interventions, including references to the news articles, can be found in S1 File (S1.3). Fig 2 shows an example of an intervention (high social norm condition; original in Dutch), which argued that the vast majority of Dutch people follow and support behavioral guidelines [i.e., descriptive and injunctive norm; 38]. The control condition in wave 3 and wave 5 covered news on topics that were expected to minimally affect nodes: increases in Netflix-subscribers in wave 3 and increases in sale of used cars in wave 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Example of intervention for ‘Social Norm high’ condition.

News articles are replaced due to copyright. Sources of original news articles are provided in S1 File (S1.3).

https://doi.org/10.1371/journal.pone.0276439.g002

Analysis

The broad COVID-19 attitude network was estimated using mgm [Mixed Graphical Models to combine continuous and binary variables; 39] with k = 2 (all pairwise interactions). An edge was included in the network if any of the two possible directions between edges were selected with 10-fold cross-validation (lambdaSel = CV [cross validation] and lambdaFolds = 10, both default settings of mgm). We have opted for k-fold CV because it emphasizes prediction more than with the EBIC, but often they give similar results. Both methods have also been investigated and have good properties. The k-fold CV is sometimes more conservative (i.e., allows fewer edges) than the EBIC, especially for smaller sample sizes. Besides the values specified in the R-script (see S2.4 in S1 File), the default values were used. Readers are referred to Haslbeck and Waldorp [39] for in-depth information about this method. Communities (i.e., groups of highly interconnected nodes) were identified with a walktrap algorithm (see S3.1.2 in S1 File).

One-way ANOVA and Kruskal-Wallis tests were conducted to examine differences between intervention conditions (i.e., low, high and control condition). The Kolmogorov-Smirnov value was significant for all variables (p < .001). Based on the QQ-plots, the variables Health Risk, Economic Consequences, Negative Affect, Worries Virus and Measures Ease were treated as normally distributed and therefore analyzed with a one-way ANOVA test. The rest of the variables were analyzed with the non-parametric Kruskal-Wallis test.

Given the many effects of the interventions, we adjusted the alpha level to p < .01 to reduce the type I error probability (see S3.2.2 in S1 File for significance values of each node per intervention). This adjusted significance value is determined based on the fact that we hypothesized a number of effects based on edges in the network, so Bonferroni correction for multiple comparisons based on all 27 nodes would increase type II errors, and tailored significance levels per intervention would decrease readability. Note that this adjusted significance level does not apply to post hoc tests given their included correction for multiple comparisons. For ANOVA, in case of equal variances, post hoc analysis was conducted with Tukey, and in case of unequal variances, Welch statistics are reported and post hoc analysis was conducted with Games-Howell. For Kruskal-Wallis, post hoc analysis was conducted with Dunn-Bonferroni post hoc method, and significance values were adjusted with a Bonferroni correction for multiple comparisons. Mean ranks are reported for significant differences between conditions with similar medians.

Mediation analyses were conducted with PROCESS in R [40]. We conducted multi-categorical independent variable mediation analyses with the control group as the reference group [41]. Given this reference group, mediation analyses were conducted for interventions resulting in significant differences for scores on target nodes between control and experimental conditions.

Broad COVID-19 attitude network structure

The broad COVID-19 attitude network, obtained through nodewise regression, is shown in Fig 3a (left). Nodes represent measured psychological factors. In general terms, the right section of the network displays (psychological) health nodes. The left section is comprised of cognitive and behavioral attitude nodes and additional psychological nodes (e.g., social norm, perceived knowledge, trust and perceived control). Edges represent linear relations between two nodes after controlling for every other node in the network. These edges are associations, meaning that the direction of the relation is not determined. Interpretation guidelines can be found in the caption of Fig 3a. Edge weights are regression coefficients that represent the strength of a relation between two nodes after removing effects from all other nodes in the network. Edges with weights below the value of .10 are omitted from the figure to facilitate readability (see S3.1.3 in S1 File for network without threshold). The edges discussed in this section had sufficient estimation accuracy and their edge weights are reported in parentheses. A complete overview of the edge weights and their accuracy is provided in S1 File (S3.1). In order to provide a brief descriptive account of the broad COVID-19 attitude network, we focus on the behavioral (i.e., compliance) and health elements (see S3.1.1 in S1 File for a more detailed interpretation of the network).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Psychological broad COVID–19 attitude network.

a) Network obtained through nodewise regression, based on sample of participants that completed the first wave (N = 6,093). Nodes represent the measured psychological factors and edges represent the relations between nodes. For binary nodes (marked with ^), a positive relationship indicates that increasing the other node results in a higher probability for category one of the binary node (i.e., Gender 1 = Female; Education 1 = Higher; Smoking 1 = Yes; Illness 1 = Yes). Positive edges (blue) represent excitatory relations and negative edges (red) represent inhibitory relations. The strength of the relation is indicated by the edge weight (visualized by edge width). Edges with edge weights below the value of .10 are omitted to facilitate readability. Colored groups represent communities (i.e., clusters with higher interconnectedness) consisting of nodes being more connected to each other than to other nodes in the network; b) Centrality measure ‘Strength’ for each node in the broad COVID–19 attitude network. This measure represents the conditional association between a node with other nodes in the network and is calculated by the sum of absolute edge weights of relations a specific node has with connected nodes.

https://doi.org/10.1371/journal.pone.0276439.g003

Nodes in the broad COVID-19 attitude network that showed the strongest relation with Compliance, were Gender (.28), Measures Support (.24) and Self-efficacy (.13). This indicates that compliance was positively associated with being female, support for behavioral measures and perceived self-efficacy. Edges between Compliance and Gender, and Compliance and Measures Support, were of comparable weight, as indicated by edge weight and formally tested with an edge difference test (see S3.1.6 in S1 File). Furthermore, these edges were both significantly stronger than the edge between Compliance and Self-efficacy. Regarding nodes on individual differences, Compliance was positively related to Consideration of Future Consequences and Coping (both .07).

The relatively important nodes that were directly associated with Mental Well-being were Coping (.22; significantly strongest positive edge), Resilience (.16; significantly second strongest positive edge) and Health General (.12), but also Loneliness (-.19; significantly strongest negative edge) and Depressive Complaints (-.15). This implies that higher mental well-being during the COVID-19 pandemic was related to a positive coping strategy, more resilience, and better perceived general health, and with less experienced loneliness and depressive complaints. Two pairs of these nodes connected to Mental Well-being were also directly connected among each other, namely Coping and Resilience (.33), and Depressive Complaints and Loneliness (.36). Such a triangle is a particular motif that lends support to the idea that there is a reinforcing structure.

The broad COVID-19 attitude network also showed that Vaccination Intention was higher among men (-.20) and positively associated with Trust (.17) and Age (.12). Another salient detail was that the node Negative Affect, positioned at the center of the network, had eight relations to other nodes in the network: the significantly strongest, positive edges of comparable weight with Involvement (.23), Anxiety Complaints (.22), Worries Virus (.21) and Worries Measures (.20), followed by the significantly weaker edges of comparable weight with Gender (.11) and Compassion (.10), and negative edges with Measures Ease (-.13) and Resilience (-.13).

Intervening in the broad COVID-19 attitude network

This part presents whether interventions aimed at specific nodes in the broad COVID-19 attitude network a) affected the targeted node (i.e., manipulation check), b) affected other nodes, and if so, whether the network structure could explain these effects. The interventions, included in wave 3 and 5, were aimed to affect scores of nodes that varied in node strength (based on the results of the first wave). As mentioned, these nodes were selected based on network properties, and therefore not necessary directly related to compliance in the network. The networks of the different waves were highly stable: correlations between edge weights of the network based on the first wave and the two waves that included interventions were r = .96 (wave 1 and 3) and r = .93 (wave 1 and 5). Table 3 presents the descriptive statistics of nodes for the subsamples of each intervention condition in wave 3 and wave 5. These subsamples consisted of participants that both completed the survey up to and including that wave and passed the manipulation check (see Table 1 for the sample size of each intervention condition). Significant results of the interventions (i.e., differences between low, high and/or control conditions) are marked in Table 3 and depicted in Fig 4. A specification of these differences is provided in the text.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Visualization of significant effects of interventions on nodes in the broad COVID–19 attitude network.

See Fig 3a for node legend. The green arrow indicates a significant effect with the specific intervention included in the arrow. Note that the broad COVID–19 attitude network is based on participants from the first wave and not the wave in which interventions were included.

https://doi.org/10.1371/journal.pone.0276439.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Statistics of nodes that differed significantly (p < .01) between intervention conditions in wave 3 and 5.

Including one way ANOVA/Kruskal–Wallis results for comparing conditions (including control condition).

https://doi.org/10.1371/journal.pone.0276439.t003

Understanding intervention effects through broad attitude network structures

As mentioned, it was hypothesized that interventions aimed at target nodes (e.g., trust) would affect that target node, and that effects on nodes connected to target nodes (e.g., measures support for trust) are induced through target nodes (e.g., interventions change trust and therefore trust changes measures support). We thus tested whether target nodes served as mediators for downstream effects of interventions. Although mediation analysis cannot provide evidence of causal mediation effects, results of such an analysis suggesting that an effect is mediated is considered a possible (first) step to investigate causal structure.

Fig 5 depicts mediation models with nodes connected to the intervention’s target node as dependent variables. As described, the intervention aimed at the node Trust in the third wave significantly affected, in addition to the target node, the connected nodes Measures Support and Social Norm. There was a significant indirect effect of the trust intervention on the node Measures Support via the node Trust of being in the high intervention condition in reference to the control group (ab = 0.19, SE = 0.05, 95% CI [0.10, 0.27]). This did not apply to the low intervention condition in reference to the control group (ab = -0.08, SE = 0.05, 95% CI [-0.18, 0.01]). Similar effects were found for Social Norm: There was a significant indirect effect of the trust intervention on the node Social Norm via the node Trust of being in the high intervention condition in reference to the control group (ab = 0.10, SE = 0.02, 95% CI [0.05, 0.14]), but this did not apply to the low intervention condition in reference to the control group (ab = -0.04, SE = 0.02, 95% CI [-0.09, 0.00]). These results indicate that effects from the high Trust intervention on connected nodes Measures Support and Social Norm were not induced directly by the intervention but rather indirectly by effects of the intervention on Trust. Thus, the target node Trust partially mediated effects of the high trust intervention on connected nodes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Significant mediation models with nodes connected to the intervention’s target node as dependent variables.

Values of indirect effects are provided in the text. * indicates significant effects (for path a and c’ in reference to control group). Plots indicated that the homoscedastic assumption might be violated. Repeated analyses with transformed variables revealed comparable results.

https://doi.org/10.1371/journal.pone.0276439.g005

The intervention in the third wave aimed at the node Social Norm did not affect connected nodes, therefore no mediation analyses were conducted for this intervention. The interventions in the last wave (e.g., Measures Support and Economic Consequences) did not result in significant differences between the experimental and the control conditions (only between experimental conditions). Subsequently, mediation analyses were not conducted for interventions in the fifth wave, since multi-categorical mediation analyses uses the control group as reference group.

In summary, results showed that effects of the first set of interventions, on nodes connected to the node the intervention was aimed at, were partially mediated by the target node. This implies that downstream effects of these interventions can be partially explained by the variables the interventions were aimed at.