Interplay of network structure and neighbour performance in user innovation

The study investigates how an individual's (ego's) innovation in user communities is affected jointly by two factors: the openness of the ego’s surrounding social network structure and the innovation performance of directly connected neighbours. Prior work separately shows that open-network structures (low constraint, bridging unconnected sub-groups) enhance innovation by providing diverse, nonredundant information, while neighbours with high innovation performance can provide valuable resources and experience. Network theory suggests these effects should interact: the network structure moderates the direction of neighbours’ influence, and neighbours’ performance adjusts its intensity. The authors hypothesise that in open networks, higher-performing neighbours exert stronger positive effects on an ego’s innovation, whereas in enclosed networks, higher-performing neighbours exert stronger negative effects, potentially via redundancy and echo-chamber dynamics. The paper tests this interplay using longitudinal data from Dell’s IdeaStorm community.

The literature highlights two drivers of individual innovation: network structure and alters’ (neighbours’) capabilities. Open structures with structural holes (bridging unconnected groups) provide diverse, nonredundant information beneficial for creativity and innovation (Burt, Perry-Smith & Shalley, Phelps et al.). Enclosed, dense networks tend to produce redundant information and echo chambers that can suppress novelty (Granovetter; Pentland; Newman & Dale; Prell et al.). High-performing or high-status neighbours can strongly influence egos by supplying more and higher-credibility resources (Grosser et al.; Shah et al.; Lin). However, prior studies typically analyze these factors separately. Network theory (Lin; Pentland) implies their interaction is critical: the network context shapes whether neighbours’ influence helps or harms. This study addresses that gap by empirically testing the moderating role of network openness on the effect of neighbours’ innovation performance.

Data source: Dell’s IdeaStorm platform (public, anonymous), capturing user interactions (votes and comments) and idea submissions with timestamps.

• User network: Directed, weighted network of 6,333 users built from vote/comment ties; a tie exists if either a vote or comment occurs; direction and frequency define tie direction and weight. Votes and comments were treated equivalently based on prior work and observed high synchronicity (1,000 of 1,057 commenters also voted on the same user).
• Idea outcomes: 326 users submitted 493 ideas deemed effective by Dell (implemented or partly implemented) from 2007–2018. Panel construction: Built an unbalanced daily user-level panel using all interactions up to each time point to compute contemporaneous network measures and cumulative counts of outcomes and activities prior to t. This yielded 2,213,782 user-day observations capturing antecedent network structure before each effective idea submission. Variables:
• Dependent variable: Ego’s innovation ability measured as the cumulative count of effective ideas submitted up to time t.
• Independent variables: • Network openness: Constraint (Burt-style). Lower constraint indicates more open structure (bridging), higher indicates enclosed structure; values conceptually range from 0 (fully open) to 1 (fully enclosed). • Neighbours’ innovation performance: Average number of effective ideas submitted by the ego’s neighbours (chosen to reduce multicollinearity with constraint; correlation r=0.08, p<0.01; total number had high negative correlation with constraint r=-0.43). • Interaction: Neighbours’ average effective ideas × Constraint.
• Controls: (1) Number of days observed; (2) Frequency of interactions sent by ego; (3) Number of neighbours; (4) Total number of ideas submitted (effective + ineffective). Individual fixed effects account for time-invariant unobserved user characteristics. Estimation: Panel Poisson fixed-effects model (Hausman tests favored fixed effects: without interaction χ²=53.861, p<0.01; with interaction χ²=18.738, p=0.009). Variables were standardized. Maximum likelihood estimation in R. Robustness checks included constructing panels using 1-week and 1-month moving windows (results consistent). Supplementary simulation (SI 5) further tested moderating effects at the network level, consistent with regression findings.

• Main panel Poisson models (N=2,213,782 user-days): • Without interaction: Neighbours’ average effective ideas negatively associated with ego’s effective idea submissions (coef = -0.21, SE=0.02, p<0.01); Constraint negative (coef = -1.35, SE=0.04, p<0.01). Controls: Frequency of interactions positive (0.05***), Days positive (0.01***), Number of neighbours n.s., Total ideas positive (0.02***). LogLik = -130,548.57. • With interaction: Neighbours’ average effective ideas positive (coef = 0.17, SE=0.04, p<0.01); Constraint negative (coef = -1.77, SE=0.06, p<0.01); Interaction (Neighbours’ avg × Constraint) negative and large (coef = -1.89, SE=0.17, p<0.01). Controls: Frequency 0.04***; Days 0.003 (n.s.); Neighbours 0.003 (n.s.); Total ideas 0.02***. LogLik = -130,460.79.
• Interpretation: The significant negative interaction indicates that as network constraint increases (more enclosed), the effect of neighbours’ high performance becomes increasingly negative; when constraint is low (open networks), neighbours’ higher performance has a stronger positive association with ego’s innovation. Visualization (marginal effects) shows positive slopes for low constraint and negative slopes for high constraint.
• Robustness: Results held when constructing panel data via 1-week and 1-month moving windows. Additional model including interactions between neighbours’ performance and all control variables still showed a significantly negative interaction with constraint (coef = -0.14, SE=0.07, p=0.05), ruling out that the main interaction is a by-product of other interactions. In that extended model, several interactions with controls were significant, suggesting ego attributes also condition neighbours’ effects.
• Overall, findings strongly support the hypothesis that network openness moderates the direction and neighbours’ performance adjusts the intensity of their influence on ego innovation.

The findings directly address the research question by demonstrating that the impact of neighbours’ innovation performance on an ego’s innovation depends critically on the ego’s network structure. Open networks (low constraint) enable access to diverse, nonredundant information from high-performing neighbours, amplifying positive effects on innovation. Enclosed networks (high constraint) foster redundancy and potential echo-chamber effects, making influence from high-performing neighbours detrimental to an ego’s innovation outcomes. This interplay reconciles mixed evidence from prior studies that examined network structure and alter performance separately, showing that considering only a single network feature can yield misleading conclusions. Beyond user innovation, similar moderating dynamics likely operate in other domains (e.g., online investment, organizational collaboration), underscoring the broader relevance of network context in shaping how influential alters affect individual performance.

This study fills a key gap by empirically establishing that the openness of an ego’s network moderates the effect of neighbours’ innovation performance on the ego’s innovation: high-performing neighbours help in open networks and hinder in enclosed ones. Using large-scale longitudinal data from IdeaStorm and fixed-effects panel Poisson models, the paper shows a strong negative interaction between constraint and neighbours’ performance, robust across model specifications and supported by simulation. Future research directions proposed include: (1) examining how ego experience (e.g., novice vs. experienced users) conditions these effects; (2) validating results in information-flow networks with explicit transmission links; (3) enriching analyses with datasets that include individual attributes beyond anonymous activity traces; and (4) distinguishing effects of different interaction types (votes vs. comments).

• Data limitations: IdeaStorm users are anonymous; individual-level demographic or trait information is unavailable. The study controls for time-invariant individual differences via fixed effects, but cannot directly test heterogeneity by user characteristics.
• Network operationalization: The network reflects preference/attention (votes/comments) rather than explicit information flow; the authors expect effects could be stronger in an information-flow network.
• Interaction types: Votes and comments were combined due to their essential similarity in this dataset; separate analysis of distinct interaction networks was not conducted.
• Generalizability: Findings are based on a single platform/community and may not generalize without validation in other contexts.
• Measurement timing: Dependent variable is cumulative; while panel construction uses antecedent networks, cumulative measures may have dynamics not fully captured by the model.