The sample included 122,345 individuals (51.4% female) from the United Kingdom, with a mean age of 43.7 (SD 16.5, range 9-86) years, who participated in the online Great British Intelligence Test (GBIT). The GBIT is a collaborative citizen science project with BBC two Horizon Program, designed to investigate cognition and well-being in the general population. The study was an open, web-based survey using convenience sampling, accessible to all UK residents. It was promoted across BBC media channels, both online and offline, to recruit participants. Interested individuals could access the study through links provided in mainstream media and news articles, which directed them to the study’s landing page on the Cognition website. The recruitment for the GBIT study commenced in December 2019, with the specific sample used in this paper recruited between May and June 2020. For the current study, we focused solely on individuals who provided complete responses for both measures. The compulsivity measure was introduced during a later phase of the GBIT. Among the 122,680 participants who took part after the compulsivity measure was included, only a small proportion of responses were missing, specifically, 335 for the compulsivity measure and 45 for PUI items. These incomplete responses were excluded from the data analysis.

Descriptive statistics were calculated for demographic variables, CHI-T subscales, and PUI items. Before conducting our formal analysis, we used the goldbricker function to check for significant topological overlap among nodes, which identifies pairs with highly similar connectivity patterns, indicative of duplicates [32]. Specifically, we looked for node pairs with a topological overlap exceeding 80%, which corresponds to less than 20% difference in their connectivity patterns (significant at P<.01). No topological overlap was identified (nor was it for the detailed item-level analysis in the supplement). We constructed a regularized partial correlation network, where the variables (ie, the 3 dimensions of compulsivity and individual PUI symptoms) are represented as nodes. The edges between nodes represent regularized partial correlations, controlling for all other variables within the network. The Graphical Least Absolute Shrinkage and Selection Operator algorithm was applied for regularization [33]. The regularization procedure forces weak or spurious partial correlations to become exactly zero, leading to a sparse network where only the most robust edges are retained [33]. Following recommendations from [33], we set the tuning parameter (γ) for regularization to 0.5, which balances the trade-off between model fit and sparsity of the network.

The Fruchterman-Reingold layout algorithm was used for network visualization [34]. Within each network, strongly related nodes were placed closer together, while weakly related nodes were placed further apart. The magnitudes of partial correlations (edge weights) are reflected as edge thickness, with thicker edges indicating stronger partial correlations between nodes. Blue edges represent positive partial correlations, and red edges represent negative partial correlations between nodes. The visualization procedures were conducted via the “qgraph” package in R (R Core Team) [34]. To enhance visual clarity, we adjusted the scaling of edge widths and color saturation using the cut argument. The improved visual representation of the network is available (Figure S5 in Multimedia Appendix 1).

We used the “networktools” package [32] to quantify to what extent each compulsivity dimension was related to the PUI symptom cluster and identify influential bridge nodes. Number of 2 clusters (subnetworks) were predefined for the network, namely, the compulsivity cluster (consisting of 3 compulsivity dimensions) and the PUI community (consisting of PUI items). Bridge expected influence (ie, the sum of the edge weights connecting a given node to all nodes in the other cluster; [25] was computed for each node. The compulsivity dimension with a high positive bridge expected influence is positively related to the PUI symptom cluster. Conversely, the compulsivity dimension with a negative bridge expected influence is negatively related to the PUI symptom cluster.

We also examined the predictability of nodes, assessing how well a specific node can be predicted by all other nodes within the network [35,36]. Since all the variables included are continuous, we selected R² as the measure of predictability for the estimated network. A node with higher predictability suggests that a larger proportion of its variance can be explained by other nodes in the network [35]. Further, we also computed the mean predictability across nodes within the network. This allows us to determine whether, on average, the variance observed in the nodes of the network can be explained by other nodes within the network [35,36]. A high mean predictability indicates that a network is largely determined by internal nodes within itself [35,36].

To estimate the accuracy and stability of the network model, we followed the steps described in [37]. First, we bootstrapped 95% CIs (with 10,000 bootstrap samples) of edge weights to determine the accuracy of the estimated edges within each network. Second, we estimated the stability of edges and bridge centrality using the case-dropping bootstrap procedure (with 10,000 bootstrap samples) and calculated the correlation stability coefficient (CS-coefficient) for edge weight and bridge expected influence. An acceptable CS-coefficient should be above 0.25, and a value of 0.5 or higher indicates optimal stability. Finally, in alignment with the tutorial paper [37], we examined whether the weights for each edge and the bridge centrality estimates for each node differ significantly (α=.05) from one another. This was achieved using bootstrapped difference tests for edge weights and bridge centrality indices. The “bootnet” package was used to perform the accuracy and stability tests [38].

We acknowledge that the item-level findings may add further nuance to the current results. To offer a comprehensive overview, we have included the item-level network analysis in the for readers interested in detailed item-level analysis (Figure S6 in Multimedia Appendix 1).

The study adhered to the ethical principles for medical research involving human subjects outlined in the Helsinki Declaration of 1975, as revised in 2008, and received approval from the institutional review board at Imperial College London (approval number 17IC4009). All participants were fully informed about the study details and provided their informed consent before participation. All data collected were anonymous. Participation was voluntary, and no financial incentives were offered.

This study is the first and largest to date to explore the relationships between different dimensions of compulsivity and individual PUI symptoms through network analysis. Extensive intercluster connections were found, indicating that compulsivity and PUI were strongly interrelated. We identified several strong mechanistic pathways, notably the edge between cognitive rigidity and coping and the edge between reward drive and functional impairment (ie, negative financial consequences). Finally, reward drive and cognitive rigidity were identified as the bridge nodes.

When elucidating interrelationships between compulsivity dimensions and PUI symptoms, we found that cognitive rigidity was associated with coping-motivated internet use. The results are in line with previous research [39-41], which highlights the role of rigid coping tendencies as a core mechanism underlying the relationship between compulsivity and addictive behaviors. Specifically, individuals high on cognitive rigidity may have a narrowed coping repertoire when experiencing negative emotions that, in turn, limits engagement in alternative coping strategies [42]. Several features of internet use (eg, easy accessibility and providing immediate stress relief) make it easy for individuals to use it as a coping strategy. However, such a coping strategy may be counterproductive in the long run (eg, due to excessive time spent online and avoiding dealing with daily challenges). For individuals with high cognitive rigidity, the inability to generate and adopt alternative coping responses when they need these most (ie, under stress) may result in their persistent engagement in coping-motivated internet use, even if this coping strategy proves to be ineffective and impairing.

Another strong pathway emerged between reward drive and functional impairment (ie, negative financial consequences and sleep loss), which may be explained by the counterproductive attentional capture mechanism characteristic of compulsivity [11,42]. Specifically, individuals with high reward drive may exhibit stronger attraction toward reward-related cues, finding it difficult to resist approaching these cues even when doing so may be counterproductive. In these individuals, such approach tendencies could be triggered by reward-related stimuli presented on the internet (eg, applications and websites for recreational online gaming, gambling, and cybersex), leading to internet use despite such use not being in line with an individual’s goals and resulting in negative consequences.

Interestingly, negative relationships were observed between perfectionism and PUI symptoms. The relationship between perfectionism and addictive behaviors tends to be inconsistent across studies, with some studies finding that perfectionism may actually protect individuals from addiction-related impairments [43] and problematic drinking [44]. This inconsistency may stem from variations in how perfectionism is measured. Perfectionism is considered as a multidimensional construct, encompassing types like self-oriented, other-oriented, and socially prescribed perfectionism [45]. In this study, the perfectionism dimension captures self-oriented perfectionism (ie, holding high standards and striving for perfection or “just right”). It has been posited that specific aspects of perfectionism (ie, self-oriented perfectionism) may be considered a form of hyperconscientiousness [44]. Such hyper-conscientiousness may protect individuals from particular addictive behaviors, which are considered impairing and not aligned with their goals (eg, being perfect in everything; [44]). Our results provide support for this hypothesis by showing that perfectionism was negatively associated with self-report items measuring functional impairment.

The bridge centrality analysis identified reward drive and cognitive rigidity as bridge nodes that positively relate to the PUI symptom cluster, whereas perfectionism negatively relates to it. The roles of reward drive and cognitive rigidity in PUI are in line with previous studies that have demonstrated attentional approach tendencies toward reward cues and rigidity of such attentional responses are linked to obsessive-compulsive and addictive behaviors, including PUI [11,41]. Meanwhile, by showing that perfectionism is negatively related to the PUI symptom cluster, our findings extend previous assumptions on the potential protective role of perfectionism (underpinned by conscientiousness) in problematic alcohol use [44] to suggest that such a protective effect may be transdiagnostic across substance use and behavioral addictions.

Current findings addressed several key research priorities identified by experts [1,2]. First, our findings help elucidate the relations between compulsivity dimensions and PUI symptoms, which addressed the need for clarifying the role of compulsivity in PUI [1,2]. The distinct connectivity between traits and symptoms may suggest that interventions should be personalized to address individual symptom profiles. For example, interventions might focus on reducing reward-driven tendencies rather than targeting perfectionism in individuals whose problematic usage is characterized by financial consequences, to achieve better intervention outcomes. Second, by demonstrating the role of reward drive and cognitive rigidity, our results may inform efforts to identify at-risk individuals at the earliest stages of PUI and provide early intervention [2]. Specifically, given that compulsivity traits can exist before the emergence of PUI symptoms, reward drive and cognitive rigidity traits should, in future work, be evaluated in terms of their potential value in identifying individuals who are most at risk of developing PUI. In turn, this may inform our screening processes.

Previous simulation research on the role of bridge nodes has demonstrated that “focusing treatment on the detected nodes is effective in preventing the contagion of comorbid disorders” ([25], P. 365). Empirically, if we can establish that the identified bridge nodes, namely reward drive and cognitive rigidity, exhibit temporal precedence, these findings could mark a significant step toward intervention target selection.

Finally, node predictability may shed several practical insights. For instance, using the internet to cope with loneliness or boredom was identified as one of the nodes with the highest predictability. The predictability allows an understanding of the clinical relevance of the identified mechanism. For instance, we found that a significant proportion of variance in using the internet to cope with loneliness or boredom may be explained by its neighboring nodes (reward-drive and cognitive rigidity). This may indicate that intervention on reward drive and cognitive rigidity may have a higher success potential in reducing coping-motivated internet use. Further, the mean predictability of the network suggests that although compulsivity dimensions may play significant roles in PUI, other factors may also be involved in driving PUI and should be considered when planning interventions.

Based on our findings, targeting reward drive and cognitive rigidity (the bridge nodes) via digital personality change interventions [46] may be expected to improve treatment outcomes. Specifically, digital personality change interventions incorporate evidence-based techniques that foster and sustain personality change (eg, tailored implementation intentions (if-then plans), psychoeducation, behavioral activation, self-reflection, resource activation, and individual progress feedback), alongside a chatbot that facilitates coaching dialogues to simulate a personal coaching process [47]. A large-scale randomized controlled trial (n=1523) has shown that a digital personality change intervention was associated with significant changes in big 5 personality traits, versus waiting list control, with effects persisting beyond the intervention [47]. The promising findings suggest that such interventions have promise in potentially improving compulsivity traits such as reward drive and cognitive rigidity. It should be noted that the aforementioned study did not directly measure compulsivity, but rather traits associated with it (eg, conscientiousness), as classic personality inventories typically do not assess compulsivity. Future studies may consider directly examining the extent to which the intervention can effectively reduce dimensions of compulsivity.

Considering the transdiagnostic characteristics of mechanisms discussed above (ie, counterproductive attentional capture, rigid coping styles, and conscientiousness), such interventions may potentially help reduce other compulsive behaviors (eg, problematic drinking, binge-eating, and obsessive-compulsive behaviors) that frequently co-occur with PUI [48]. These suggestions should be tested in future research studies.

Several limitations should be noted when interpreting current findings. First, PUI symptoms in the current study were measured using items drawn from expert consensus, which may raise concerns regarding the psychometric properties of these items. This was the case for pragmatic reasons as the data were collected as part of a citizen science study [31] that needed to have relatively brief questions that were also acceptable to other stakeholders. The advantage of this approach is that it enables a very large sample size to be collected. Second, the findings were analyzed and discussed with the theoretical assumption that personality traits act as predisposing factors and predict PUI (instead of the other way around; [49-52]. However, due to the cross-sectional nature of the study, we cannot disregard the possibility that PUI may contribute to compulsive traits. It will be important for future research to test the temporal precedence of compulsivity traits and PUI symptoms with longitudinal network analysis. Third, the majority of the sample self-identified as White or Caucasian and resided in a Western country (ie, the United Kingdom), potentially limiting the generalizability of these findings to individuals from other ethnic or cultural backgrounds. Fourth, the current findings were based on data from the general population, which may not be generalizable to individuals with severe levels of PUI. It is important for future research to replicate current findings among individuals with severe levels of PUI.

In theory, targeting bridge nodes could limit the transmission of activation between different psychological symptoms (or conditions), potentially preventing the onset of comorbid psychopathologies and thus making them valuable targets for intervention [25]. However, empirical evidence on the predictive validity of bridge nodes is inconsistent. For instance, findings from 6 randomized control trials suggest that bridge nodes may predict treatment outcomes in patients with depressive disorders [53]. In addition, studies indicate that bridge nodes can statistically predict the persistence of psychopathology, such as poorer psychosocial functioning and lower BMI in adolescents with anorexia nervosa [54], as well as the future onset of anxiety disorders [55]. Conversely, some evidence suggests that bridge nodes might only partially predict treatment outcomes (ie, predict weight restoration, but not full remission, in patients with anorexia nervosa [56]) or yield nonsignificant results in terms of relationships to outcomes [57]. Given these mixed results, future longitudinal and intervention research is essential to ascertain the predictive validity of bridge nodes before they can be confidently used as treatment targets.

Most research in the fields of psychology and psychiatry that uses network-based approaches has relied on cross-sectional, between-subject designs. However, there is an increasing shift toward using longitudinal designs, including panel data and intensive time-series data [58-60]. These methods enable researchers to differentiate within-subject and between-subject effects and to elucidate the temporal dynamics among psychological constructs. A recent study using graphical vector autoregressive network models to analyze panel data revealed that coping-motivated substance use could predict posttraumatic stress disorder symptoms over time [59]. In addition, when combined with intensive time-series data, graphical vector autoregressive network models can elucidate the interrelationships among psychological symptoms and the impact of risk factors on these symptoms at an individual level, offering insights into precision medicine (eg, by enabling personalized diagnostic profiling and treatment planning; [58,61]). To improve the clinical relevance of current findings, future studies should use longitudinal designs to further explore the network structure and bridge nodes identified in this study.

The current study addresses evidence gaps in the literature by investigating the relations between different dimensions of self-reported compulsivity and PUI symptoms. using network analysis, we identified distinct connections between dimensions of compulsivity and PUI symptoms. Specifically, we found that reward-drive and cognitive rigidity were differently associated with PUI symptom clusters, which may reflect distinct mechanisms (ie, counterproductive attentional capture and rigid coping response). Further, we identified “reward drive” and “cognitive rigidity” as bridge nodes within the network. These findings represent an initial exploration into the role of transdiagnostic compulsivity in developing future interventions for PUI. Examining the direction of the mechanistic pathways identified, whether compulsivity drives PUI or vice versa, through longitudinal studies represents a promising next step for applying network-informed treatment strategies in clinical practice for PUI.