The Effects of Third-Party Intervention on Prosocial Behavior: A Three-Level Meta-Analysis

SHEN Yinqi, CAI Yi, WU Jixia*
School of Education, Soochow University, Suzhou, 215123

Abstract
As a crucial mechanism for maintaining social norms, third-party intervention has garnered widespread attention from researchers for its potential to promote prosocial behavior. To systematically examine the magnitude of this prosocial effect and its moderating factors, this study employed a three-level meta-analytic approach, synthesizing 130 effect sizes from 40 empirical studies (totaling 10,289 participants). Main effect analyses revealed that third-party intervention exerts a moderately strong facilitative effect on prosocial behavior. Moderator analyses indicated that the intensity and probability of third-party intervention significantly influence its prosocial impact, with stronger and more frequent interventions yielding larger effect sizes. However, participant age, gender, intervention type, form, agent, cost, prosocial behavior measurement paradigm, and control group setting did not demonstrate significant moderating effects. These findings provide robust evidence for the positive impact of third-party intervention on prosocial behavior and clarify key boundary conditions, offering valuable insights for future theoretical and empirical research.

Keywords: third-party intervention, prosocial behavior, three-level meta-analysis, moderating effect

"The sage does not hoard; the more he gives to others, the more he possesses; the more he provides for others, the more he gains." This core prosocial value, deeply embedded in Chinese traditional culture and modern thought, represents a key driving force for maintaining and advancing healthy social relationships (Fehr et al., 2002; Wu et al., 2022). Prosocial behavior encompasses actions that benefit individuals, groups, organizations, or society at large (e.g., helping, donating, volunteering, cooperation), often at a personal cost to the actor (Bradley et al., 2018; Thielmann et al., 2020). While biological evolution and the internalization of prosocial cultural values have shaped human prosociality and intuition-based prosocial behavior (Zaki & Mitchell, 2013; Shi & Liu, 2019), these factors alone cannot fully explain large-scale cooperation and altruism, particularly interactions among non-kin or strangers. The development of social norms and the evolution of reciprocal mechanisms provide more comprehensive theoretical foundations for costly prosocial behavior (Glowacki & Lew-Levy, 2022; Rand & Nowak, 2013). Third-party intervention facilitates the enforcement and reinforcement of social norms and the realization of indirect reciprocity (Wu et al., 2022; Guo et al., 2024), serving as an important exogenous mechanism for resolving cooperation dilemmas and promoting prosocial behavior in social interactions (Qin & Wang, 2013). However, existing research presents inconsistent conclusions regarding the effectiveness of third-party intervention in promoting prosocial behavior (Mulder et al., 2006; Windrich et al., 2024; Chen et al., 2021), suggesting that its effects may be contingent on complex boundary conditions. Therefore, a meta-analytic integration of existing studies is necessary to systematically examine sources of heterogeneity and further clarify the relationship between third-party intervention and prosocial behavior and its influencing factors, thereby providing a more comprehensive perspective for future theoretical development and empirical exploration.

1.1.1 The Concept of Third-Party Intervention

Third-party intervention refers to actions taken by an uninvolved third party who, upon observing others' behavior that violates, conforms to, or exceeds social norm expectations, actively punishes, rewards, or compensates the involved parties (Guo et al., 2024; Putz et al., 2016). Specifically, third-party punishment (TPP) involves an impartial party punishing norm violators regarding fairness, cooperation, or other social standards (Chen et al., 2021; Chen & Chen, 2014), including monetary penalties, verbal condemnation, and spreading negative gossip (Festré & Garrouste, 2014; Cui et al., 2017; Chen & Ma, 2011). Third-party reward (TPR) involves an impartial party rewarding those who comply with norms or exceed normative expectations (Fiedler & Haruvy, 2017; Sutter et al., 2009), including monetary rewards, verbal praise, and spreading positive gossip (Charness et al., 2008; Feinberg et al., 2012). Third-party compensation (TPC) refers to an impartial party compensating victims of norm violations (Lotz et al., 2011), including monetary compensation, verbal apologies, and behavioral assistance (Nakashima et al., 2017; Guo et al., 2024).

1.1.2 The Effect of Third-Party Intervention on Prosocial Behavior

Most theoretical and empirical research supports a positive effect of third-party intervention on prosocial behavior. Deterrence theory emphasizes the advantage of third-party punishment in suppressing norm violations (Guo et al., 2024): when third-party punishers are present, individuals anticipate potential punishment and violation costs before acting, thereby reducing behavior that contravenes social norms. For example, this manifests as decreased free-riding or defection in social dilemmas and increased contributions to and cooperation within groups (Nakashima et al., 2017; Cui et al., 2017; Chen et al., 2021). Correspondingly, third-party rewards can produce commitment effects (Fiedler & Haruvy, 2017); when actors anticipate that rewards for prosocial behavior can offset behavioral costs, they become more willing to engage in costly prosocial actions (Sefton et al., 2007; Sutter et al., 2009).

Indirect reciprocity theory also supports the notion that indirect monetary benefits from third-party rewards or punishment enhance the attractiveness of prosocial behavior (Wu et al., 2022). This theory also provides a basis for the effectiveness of social forms of third-party rewards and punishment: individuals may be motivated to build a good reputation by exhibiting more prosocial behavior to obtain positive social evaluations, avoid the spread of negative gossip, and increase their likelihood of being selected as cooperation partners (Liu & Xin, 2011; Yuan et al., 2016), thereby gaining potential indirect benefits in future group interactions (Barclay et al., 2021; Roberts et al., 2021). Thus, the reputational costs associated with non-monetary third-party interventions also predispose individuals toward prosocial behavior.

Social norm focus theory (Cialdini et al., 1991) provides another layer of support for the prosocial effects of third-party intervention: punishment and reward are themselves externalizations of social norms, and monetary reward-punishment mechanisms can activate individuals' perception of social norms, while verbal evaluation or gossip dissemination similarly increases norm salience (Eriksson et al., 2021; Shank et al., 2019; Chen, 2022). Third parties maintain norms by punishing or rewarding actors and compensating victims, which makes those intervened upon aware of group-endorsed or group-rejected behavioral standards, thereby enhancing norm compliance (Guo et al., 2024). In other words, third-party intervention can promote prosocial behavior by strengthening individuals' perception of both injunctive and descriptive norms. Moreover, this norm focus effect can spill over to bystanders of the intervention or extend to new interaction contexts (Guo et al., 2024; Chen et al., 2021). Additionally, repeated third-party interventions may lead to reinforced learning of social norms, thereby improving prosocial behavior in future situations (Fiedler & Haruvy, 2017). For instance, third-party reward or punishment feedback in multi-round interactions leads to continuously increasing cooperation rates among participants (Hou et al., 2019; Nakashima et al., 2017). Based on this analysis, this study hypothesizes that third-party intervention has a significant positive effect on prosocial behavior.

1.2 Moderators of the Relationship Between Third-Party Intervention and Prosocial Behavior

To further clarify the boundary conditions of third-party intervention effects, this study examined multiple moderating variables identified in existing theoretical and empirical research as potential sources of effect heterogeneity, thereby providing a more comprehensive understanding of the conditions and mechanisms through which third-party intervention influences prosocial behavior.

1.2.1 Age

The prosocial effects of third-party intervention may differ across developmental stages, closely related to individuals' moral development levels and social cognitive abilities. Children in the norm-learning stage are particularly susceptible to third-party intervention. For example, 6-7-year-olds are in the formation and transformation stage of fairness concepts, and the normative signals conveyed by third-party punishment help deepen their understanding and internalization of social rules, thereby promoting prosocial behavior to a greater extent (Martin et al., 2021; Xiao, 2024). Although some studies suggest that children under 5 do not yet understand the indirect reputational benefits of gossip (Hill & Pillow, 2006), evidence indicates that 4-year-olds, when facing third-party gossip threats, already engage in reputation management by exhibiting more prosocial behavior, similar to adults (Shinohara et al., 2021). Adolescents are particularly sensitive to social evaluation, with more salient reputation motives, making them especially susceptible to social forms of third-party intervention (Cui et al., 2017). In summary, age may be a potential moderator of the relationship between third-party intervention and prosocial behavior.

1.2.2 Gender

Women typically exhibit higher sensitivity to rewards and punishments (Blackwell, 2000), greater sensitivity to social evaluation (Vanderhasselt et al., 2018), and higher risk aversion (Agnew et al., 2008), and are more responsive to monetary and social feedback (Ding et al., 2017). Therefore, when facing potential third-party punishment, women may be more inclined to engage in prosocial behavior to avoid negative consequences (Pablo & Stefania, 2009). Additionally, women tend to focus more on the norm enforcement function of interventions, whereas men are more likely to weigh costs and benefits from a profit perspective (Burnham, 2018; Qian et al., 2023), with differences also existing in their tendencies toward norm enforcement and maintenance (Boschini et al., 2011; Mieth et al., 2017). Overall, gender may influence individuals' responses to third-party intervention, thereby moderating its effect on prosocial behavior, which warrants investigation in this study.

1.2.3 Type of Third-Party Intervention

Based on target recipients, third-party interventions can be categorized into three types: third-party punishment, third-party reward, and third-party compensation (Gummerum et al., 2016; Sutter et al., 2009; Guo et al., 2024). These target violators, norm compliers or prosocial actors, and victims of violations, respectively. This study included these three intervention types as well as pairwise combinations. Existing research has not reached consensus regarding the relative effectiveness of different intervention types. Early meta-analytic results indicated that punishment and reward have statistically equivalent positive effects on enhancing cooperation (Balliet et al., 2011). However, some studies suggest that rewards promote greater contributions to public goods than punishment (Heine & Strobel, 2020; Rand et al., 2009), while others find third-party rewards less effective than punishment in suppressing free-riding (Fiedler & Haruvy, 2017; Zhang, 2019). Similarly, negative gossip has shown stronger effects than positive gossip in promoting cooperation (Wang, 2018). Regarding third-party compensation, although some theorists argue it has less deterrent power than punishment (Chavez & Bicchieri, 2013), empirical evidence suggests it is equally effective as punishment in promoting prosocial behavior (Guo et al., 2024; Wang, 2021). Furthermore, researchers and practitioners generally agree that combinations of intervention types outperform single-type interventions (Chen et al., 2015; Hou et al., 2019; Liu, 2018). In summary, intervention type may moderate the effect of third-party intervention on prosocial behavior.

1.2.4 Form of Third-Party Intervention

Third-party intervention forms can be broadly categorized as monetary or social (Asulin et al., 2024; Liu et al., 2010; Chen & Chen, 2014). Monetary interventions typically involve third parties increasing or decreasing others' monetary payoffs (Chen et al., 2021), whereas social interventions involve criticism/praise or spreading positive/negative gossip (Festré & Garrouste, 2014; Cui et al., 2017). Some research suggests that, compared to monetary interventions' potential dual effect of enhancing external motivation while undermining internal motivation, social interventions place individuals in a moral context, activating internal and external attributions for cooperative behavior, thereby producing stronger and more sustained prosocial effects (Asulin et al., 2024; Liu et al., 2010). However, other studies indicate that praise and criticism are less effective than monetary rewards and punishments in promoting public goods contributions (Zhu, 2009). Thus, different intervention forms may also lead to differential effects on prosocial behavior.

1.2.5 Agent of Third-Party Intervention

Third-party interventions can be implemented by humans or computer systems. Computer-executed interventions operate according to pre-programmed rules, providing feedback based on actors' responses (Xiao & Houser, 2011; Liu et al., 2010), whereas human third parties can freely decide whether to intervene based on observed behavior (Nakashima et al., 2017; Sutter et al., 2009), a process that typically involves subjective intentions. As a spontaneous norm maintenance behavior (Guo et al., 2024), human-executed third-party intervention may be more effective at enhancing individuals' perception of descriptive norms, thereby reducing violations to a greater extent (Chen et al., 2021; Guo et al., 2024). For example, Zhu (2023) manipulated the intervention agent and found that human-imposed punishment significantly affected individuals' fairness perception and emotional experiences, subsequently guiding their prosocial behavior, whereas computer-administered punishment had relatively weaker effects. However, in terms of intervention accuracy, computer systems may deliver more precise interventions and feedback, potentially producing stronger deterrent effects. Thus, the intervention agent may be an important moderating variable in the relationship between third-party intervention and prosocial behavior.

1.2.6 Probability, Intensity, and Cost of Third-Party Intervention

Third-party intervention probability refers to the likelihood that involved parties will be punished, rewarded, or compensated. This variable is typically preset in experiments through two manipulation methods: (1) pre-setting the probability that third parties will intervene on specific behavioral consequences (Halevy & Halali, 2015; Xiao & Houser, 2011); or (2) manipulating the percentage of individuals in a group who receive intervention (Chen et al., 2015). Intervention intensity reflects the actual loss or gain imposed on the target (Charness et al., 2008; Guo et al., 2024). Given the difficulty of quantifying social intervention intensity, this study only included monetary intervention intensity in the analysis. To control for differences in original payoff structures across paradigms and enhance cross-study comparability, we operationally defined intervention intensity as the ratio of the monetary gain/loss from a single minimum-level intervention to the maximum possible gain from a single violation.

Deterrence theory posits that punishment probability and intensity jointly determine deterrent power (Becker, 1968). Theoretically, higher probability and greater intensity of third-party punishment produce stronger deterrent effects, a view supported by some empirical research (Windrich et al., 2024; Chen et al., 2015). However, other studies find that mild third-party punishment is sufficient to effectively promote prosocial behavior, with effects even surpassing high-intensity punishment (Kamei, 2020; Chen et al., 2021; Guo et al., 2024). The relationship between punishment probability and prosocial behavior may not be linear but rather follow an inverted U-shaped curve, where moderate probability is most effective (Qin & Wang, 2013). Other intervention types may also be moderated by intensity and probability. For example, low-probability, low-intensity rewards may be insufficient to create deterrent effects or build reciprocal mechanisms, thus failing to significantly enhance prosocial behavior (Almenberg et al., 2011; Zhang, 2019). High-intensity compensation may be more effective than moderate or low-intensity compensation in conveying and maintaining fairness norms, thereby more strongly eliciting prosocial behavior (Wang, 2021).

Intervention cost can be categorized as costly or costless: the former involves third parties paying monetary or time costs when implementing punishment, reward, or compensation, while the latter involves no resource expenditure (Chen & Chen, 2014). In the included literature, cost manipulation was based on monetary forms—whether third parties needed to pay money or tokens when intervening. Intervention cost largely determines intervention intensity and probability (Fiedler & Haruvy, 2017; Chen & Bo, 2016). Individuals often expect costless interventions to occur more frequently than costly ones, theoretically giving costless interventions stronger deterrent and norm-transmission effects (Guo et al., 2024). However, inconsistent findings suggest that costly interventions are more likely to be perceived as legitimate and altruistic (Raihani & Bshary, 2015), thus promoting prosocial behavior more effectively than costless interventions (Balliet et al., 2011; Kuwabara & Yu, 2017). In summary, intervention intensity, probability, and cost may all serve as potential moderators.

1.2.7 Prosocial Behavior Measurement Paradigm

Prosocial behavior measurement paradigms typically include single-interaction and repeated-interaction tasks. The former requires participants to make a one-time prosocial decision, while the latter allows individuals to make continuous behavioral choices across multiple rounds, potentially receiving third-party intervention feedback each round to learn about behavioral consequences. Research indicates that the repetitiveness of third-party interventions helps strengthen social norm learning, thereby more effectively promoting prosocial behavior (Martin et al., 2021; Xiao, 2024). Whether monetary rewards/punishments, social evaluations, or compensation for others, all can serve as feedback mechanisms for norm internalization (Raihani et al., 2012; Guo et al., 2024), with multi-round intervention feedback being more conducive to forming stable prosocial behavior patterns (Zheng et al., 2024). Therefore, compared to single-interaction contexts, individuals may exhibit higher levels of prosocial behavior in repeated-interaction contexts with ongoing third-party intervention or feedback mechanisms (Bradley et al., 2018).

1.2.8 Control Group Setting

When examining the effect of third-party intervention on prosocial behavior, experimental studies typically employ two types of control groups: no-third-party control groups and third-party observer groups with no intervention power. Compared to third-party intervention groups where third parties can punish, reward, or compensate, control groups receive no additional experimental manipulation, while observer groups involve a third party present to observe the interaction but unable to intervene in any way. Theoretically, these two control groups provide different comparison baselines. When contrasted with no-third-party groups, intervention effects may include both "observer effects" and "intervention effects" (Xiao, 2024). When contrasted with third-party observer groups, intervention effects primarily reflect mechanisms beyond observer effects, such as monetary gains/losses (Sutter et al., 2009), higher reputational costs (Fehr & Sutter, 2016), and norm reinforcement (Chen et al., 2021). Therefore, effect sizes measured against the former may be higher than those against the latter, suggesting that control group setting may moderate the magnitude of third-party intervention effects.

2 Method

To ensure systematicity and reproducibility, this study followed the PRISMA2020 guidelines (Page et al., 2021) and pre-registered the protocol on the Open Science Framework (OSF) before literature screening and data analysis (Registration ID: 10.17605/OSF.IO/E9BTG).

2.1 Literature Search

We conducted comprehensive searches of both Chinese and English literature. Chinese literature was searched in CNKI (China National Knowledge Infrastructure), China Doctoral and Master's Dissertations Full-Text Database, Wanfang Database, and VIP Database. English literature was searched in PubMed, Web of Science, Elsevier, EBSCO, ProQuest, and Google Scholar.

The search procedure was as follows: (1) Search format: "third-party intervention search terms" AND "prosocial behavior search terms" (specific terms shown in Table 1 [TABLE:1]); (2) Search fields: Title and abstract; (3) Final search date: March 2025. A total of 9,324 articles were retrieved. After initial screening to remove duplicates, 5,987 articles were included in the literature pool for management. (4) Data import: Literature data were imported into Zotero for organization and screening.

2.2 Literature Screening

Inclusion criteria were: (1) Chinese or English language only; (2) Quantitative empirical studies only, excluding reviews, meta-analyses, theoretical models, and qualitative research; (3) Studies must include a third-party intervention group (i.e., a third party with no direct interest in the interaction who can influence others' outcomes through punishment, reward, or compensation) and at least one control group (no-third-party condition or third-party observer without intervention power); (4) Prosocial behavior must be measured in potential norm compliers/violators affected by third-party intervention, excluding studies measuring "violation victims (second parties)" or "intervention implementers (third parties)"; (5) Prosocial behavior indicators include but are not limited to: prosocial behavior in Dictator Game, Public Goods Game, Prisoner's Dilemma, Investment Game, and prosocial behavioral intentions in social contexts; (6) Studies must report sample size, mean, and standard deviation sufficient for calculating Hedges' g, or other statistics convertible to g (e.g., t, χ², F values). Studies reporting only prosocial behavior under third-party intervention conditions or with incomplete data were excluded; (7) For articles containing multiple independent samples, each sample was coded separately; (8) Duplicate publications were excluded, and when duplicate data existed, the article with more complete information was selected. The screening process is illustrated in Figure 1 [FIGURE:1].

2.3 Literature Coding

Two independent coders coded the included articles using a coding manual. Basic information included: (1) Publication status: published journal article (J), dissertation (D), conference paper (C); (2) Mean age and age group: children (Ch), adolescents (Te), adults (Ad); (3) Female proportion: continuous variable ranging from 0-100%; (4) Study ID: author + year; (5) Experiment number: original experiment or study number in the source article.

Third-party intervention characteristics were coded as: (1) Intervention type: third-party punishment (TP), third-party reward (TR), third-party compensation (TC), TP+TR, TP+TC, TR+TC; (2) Intervention form: monetary intervention (MI), social intervention (SI), MI+SI; (3) Intervention agent: human (H), computer/system (S); (4) Intervention probability: continuous variable (0-100%) representing either the probability of intervention on specific behavioral consequences or the percentage of individuals receiving intervention; (5) Intervention intensity: continuous variable q calculated as q = x/y, where x = monetary change from a single minimum-level intervention and y = maximum possible gain from a single violation (see Appendix B for calculation details across paradigms). Intensity ranged 0 < q ≤ 3; (6) Intervention cost: costly (YES) or costless (NO).

Prosocial behavior characteristics were coded as: (1) Measurement paradigm: single interaction (SG) or repeated interaction (RG); (2) Measurement tool: Dictator Game (DG), Public Goods Game (PGG), Prisoner's Dilemma (PD), Investment Game (IG), or other tools (OT). Control group settings were coded as: no-third-party control group (CC) or third-party observer group (OC).

Coding principles: (1) Each independent sample was coded once, with multiple effect sizes within a study coded separately; (2) When individual group sample sizes were not reported, we followed Quarmley et al. (2022) by dividing the total sample size by the number of groups; (3) When multiple variables were measured, each was coded separately. Inter-coder reliability was high: ICCs for continuous variables ranged 0.95-1.00, and Kappas for categorical variables ranged 0.86-1.00. Discrepancies were resolved through discussion and consultation with the corresponding author.

2.4 Quality Assessment

We assessed each included study using the Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies. Each item was rated as Yes (1 point), No, Cannot Determine (CD), Not Reported (NR), or Not Applicable (NA). Quality ratings were: good (>7 points), fair (5-7 points), or poor (<5 points) (Lin et al., 2025). Two independent coders conducted the assessment (Kappa = 0.93). Discrepancies were resolved through item-by-item comparison, discussion, and verification against original articles.

2.5 Effect Size Calculation

We used Hedges' g as the effect size measure. Most studies provided means, standard deviations, and sample sizes for calculating g, while a few required conversion from F, χ², t, β, or p values (Harrer et al., 2021). Effect size benchmarks were: small (0.20), medium (0.50), and large (0.80) (Cohen, 1992).

2.6 Model Selection

This study included multiple effect sizes from the same original study, violating traditional meta-analysis assumptions of independence. Three-level meta-analysis accounts for effect size dependency by explaining three variance sources: sampling variance (Level 1), within-study variance (Level 2), and between-study variance (Level 3) (Cheung, 2014; Van den Noortgate et al., 2013), thereby addressing non-independence while preserving information integrity and increasing statistical power (Cheung, 2019). Therefore, we employed a three-level random-effects model.

2.7 Heterogeneity and Moderator Testing

We used Q-tests for overall heterogeneity assessment and one-tailed log likelihood ratio tests to evaluate the significance of within-study and between-study variance (Assink & Wibbelink, 2016; Cheung, 2014; Gao et al., 2024). If heterogeneity was present, moderator analyses were conducted. Moderators included: (1) Continuous: female proportion, intervention intensity, intervention probability; (2) Categorical: age group, intervention type, intervention form, intervention agent, intervention cost, prosocial behavior measurement paradigm, control group setting. Categorical moderators required at least 5 effect sizes per level (Card, 2016).

2.8 Publication Bias and Sensitivity Analysis

We included published journal articles and unpublished dissertations/conference papers to control for publication bias. Funnel plots, Egger's regression, and trim-and-fill methods were used to assess bias. Symmetrical funnel plots indicate minimal bias (Rothstein et al., 2005), and non-significant Egger's regression suggests small bias (Rodgers & Pustejovsky, 2021). If asymmetry was detected or Egger's test was significant, trim-and-fill analysis assessed the impact—if the adjusted effect size did not change substantially, results were considered robust to publication bias (Duval & Tweedie, 2000).

Sensitivity analyses used outlier and influence diagnostics. Outliers were identified using studentized deleted residuals (SDR), with |SDR| > 1.96 indicating substantial deviation from predicted mean effects; outlier proportion should not exceed 1/10 of total effect sizes (Viechtbauer, 2010). Influence was assessed using DFBETAS, measuring standard deviation changes in correlations after excluding each effect size; DFBETAS > 1 indicates significant influence (Viechtbauer & Cheung, 2010).

2.9 Data Processing

Analyses were conducted in R 4.4.3 using the metafor (Viechtbauer, 2010) and esc (Lüdecke, 2019) packages, following tutorials by Assink and Wibbelink (2016) and Harrer et al. (2021).

3 Results

3.1 Study Characteristics

The final sample comprised 40 articles published between 2006-2024, including 57 independent samples, 130 effect sizes, and 10,289 participants. Effect sizes per study ranged from 1 to 10, with details for each moderator shown in Table 2 [TABLE:2]. Quality assessment scores ranged 5-11, with 27 rated as good and 13 as fair. Overall, the included literature was of good quality; basic information is provided in the Appendix.

3.2 Main Effect and Heterogeneity

Three-level meta-analysis revealed a significant effect of third-party intervention on prosocial behavior, g = 0.73 (p < 0.001, 95% CI [0.57, 0.88]). Q-tests indicated significant heterogeneity (Q(129) = 995.30, p < 0.001). One-tailed log likelihood ratio tests showed significant within-study variance (Level 2) (σ² = 0.19, p < 0.001, I² = 50.14%) and between-study variance (Level 3) (σ² = 0.15, p < 0.001, I² = 39.82%). Following Higgins et al. (2003), within-study heterogeneity was high (I² > 50%) and between-study heterogeneity was moderate (I² > 25%), justifying moderator analyses.

3.3 Moderator Effects

We examined age group (adults/adolescents/children), female proportion, intervention type (punishment/reward/compensation/punishment+reward), intervention form (monetary/social), intervention agent (human/computer), intervention probability, intervention intensity, intervention cost (costly/costless), measurement paradigm (single/repeated interaction), and control group setting (no third party/third-party observer). For continuous moderators, three-level meta-regression tested linear relationships; for categorical moderators, dummy coding tested between-level differences.

Results showed non-significant moderating effects for age group (F(2,115) = 0.47, p = 0.627), female proportion (F(1,96) = 0.41, p = 0.525), intervention type (F(3,126) = 2.09, p = 0.106), intervention form (F(1,127) = 0.86, p = 0.355), intervention agent (F(1,128) = 2.04, p = 0.156), intervention cost (F(1,127) = 0.44, p = 0.510), measurement paradigm (F(1,128) = 0.97, p = 0.327), and control group setting (F(1,128) = 1.20, p = 0.275).

Intervention intensity showed a significant moderating effect, F(1,95) = 4.27, p = 0.042. Stronger intervention intensity was associated with stronger prosocial effects (b = 0.28, 95% CI = [0.01, 0.55], p = 0.042). Intervention probability showed a marginally significant effect, F(1,51) = 3.97, p = 0.052, with higher probability associated with stronger effects (b = 0.01, 95% CI [–0.0001, 0.02], p = 0.052). Results are detailed in Table 2.

Reference groups for categorical moderators were: adult for age, punishment for intervention type, monetary for intervention form, human for agent, costly for cost, single interaction for measurement paradigm, and no-third-party for control group.

3.4 Sensitivity Analysis

Twelve effect sizes were identified as outliers (|SDR| > 1.96) (IDs: 13, 27, 28, 29, 30, 40, 47, 79, 87, 92, 104, 119), not exceeding 1/10 of total effect sizes. The DFBETAS plot (Figure 2 [FIGURE:2]) showed no standardized coefficient changes > 1 after sequentially removing each effect size (Cook & Weisberg, 1982; Viechtbauer & Cheung, 2010). Thus, outliers did not substantially affect results, indicating robust findings.

3.5 Publication Bias Assessment

The funnel plot (Figure 3 [FIGURE:3]) showed effect sizes distributed relatively evenly around the central effect, though some points fell in the lower right region. Egger's test was significant, t = 3.49, p < 0.001, intercept = 1.92, 95% CI = [0.84, 3.00], suggesting some publication bias. Trim-and-fill analysis added 30 effect sizes (k = 160, original k = 130), yielding g = 0.52, 95% CI = [0.38, 0.66], t = 7.20, p < 0.001. After removing the 12 outliers identified in sensitivity analysis, 22 effect sizes were added (k = 140, original k = 130), yielding g = 0.55, 95% CI = [0.46, 0.65], t = 11.36, p < 0.001. Rosenthal's fail-safe N = 5,766, exceeding 5k + 10 (k = 130). Therefore, publication bias had minimal impact on results.

4 Discussion

This three-level meta-analysis of 40 studies confirms that third-party intervention has a moderately strong positive effect on prosocial behavior. The relationship is moderated by intervention intensity and probability but not by age, gender, intervention type, form, agent, cost, measurement paradigm, or control group setting. Overall, third-party intervention demonstrates robust prosocial effects.

4.1 The Effect of Third-Party Intervention on Prosocial Behavior

The facilitative effect can be explained by deterrence theory, indirect reciprocity theory, and social norm focus theory. From a rational actor perspective, norm violations typically occur when individuals seek greater benefits (Becker, 1968). Deterrence theory similarly posits that when individuals realize third-party punishment costs may exceed violation benefits, violations are avoided (Akers, 1990). From an indirect reciprocity perspective, while prosocial actors cannot benefit directly from current interactions, they may gain indirect benefits through third-party rewards (Wu et al., 2022). This theory also emphasizes the importance of reputational benefits beyond monetary gains, suggesting prosocial behavior helps obtain positive evaluations and build favorable reputations, increasing future cooperation opportunities and achieving indirect reciprocity (Roberts et al., 2021). Wang's (2018) empirical research validated this, showing third-party gossip increases reputation concern and cooperation.

Social norms guide behavior and are prerequisites for prosocial behavior, explaining why humans act prosocially at personal cost beyond rational self-interest (Gross & Vostroknutov, 2022). Social norm focus theory emphasizes third-party intervention's role in activating and strengthening social norms (Cialdini et al., 1991; Chen et al., 2015). Punishing violators, rewarding compliers, and compensating victims all reaffirm and maintain social norms (Guo et al., 2024). Guo et al. (2024) and Chen et al. (2021) confirmed that social norm perception mediates the relationship between third-party intervention and prosocial behavior, demonstrating that third-party intervention effectively focuses individuals' attention on social norms, significantly inhibiting violations and increasing prosociality.

In summary, deterrence theory primarily explains third-party punishment's prosocial effects, indirect reciprocity theory emphasizes monetary and reputational incentives, and social norm focus theory supports both reward/punishment effects and provides a theoretical basis for compensation effectiveness. These three theories offer complementary explanations from the dimensions of violation costs, indirect benefits, and norm reinforcement.

4.2 Moderator Effects

Moderator analyses revealed significant effects of intervention intensity and marginally significant effects of probability. Specifically, stronger and more probable interventions produce stronger prosocial effects. Consistent with deterrence theory, higher intensity and probability mean greater costs and risks for violations, reducing violation rates (Becker, 1968; Chen et al., 2015). From an indirect reciprocity perspective, greater potential indirect benefits attract more prosocial behavior to offset costs (Wu et al., 2022). High-intensity external incentives also increase mutual cooperation expectations (Balliet et al., 2011; Liu et al., 2010), raising cooperation levels (Lergetporer et al., 2014; Chen & Chen, 2014). Social norm focus theory also explains this: higher intensity and probability increase norm salience, strengthening prosocial tendencies, as confirmed in previous research (Wang, 2021; Chen et al., 2015).

Regarding intervention type, third-party reward showed weaker effects than punishment, while other types did not differ significantly from punishment. Overall, punishment, reward, compensation, and punishment+reward combinations all significantly promoted prosocial behavior, but rewards were relatively less effective. This can be explained by prospect theory (Kahneman & Tversky, 1988), which posits that individuals experience unequal psychological weighting of equivalent gains and losses, showing greater loss sensitivity ("loss aversion"). Thus, the psychological deterrence from punishment's "loss" is stronger than reward's "gain," weakening its prosocial effect (Hou et al., 2019; Zhang, 2019). Additionally, from an information processing perspective, negative information is typically more persuasive and valuable for learning than positive information (Baumeister et al., 2001; Martinescu et al., 2014). As a negative social signal, third-party punishment may more effectively trigger understanding and processing of behavioral consequences, strengthening norm learning and compliance, thereby more effectively guiding prosocial behavior (Wang, 2018).

Age did not significantly moderate effects. Humans begin learning social norms from early social interactions (Bian et al., 2024), and both monetary and social interventions effectively promote norm compliance (Martin et al., 2021; Shinohara et al., 2021). Although children in the norm-learning stage may have greater room for prosocial improvement, third-party intervention effects did not vary significantly across age groups, demonstrating cross-age consistency in effectiveness.

Gender also showed no significant moderation, indicating consistent prosocial effects across genders. Despite potential differences in social evaluation and punishment sensitivity (Blackwell, 2000; Vanderhasselt et al., 2018), third-party intervention effectively promotes prosocial behavior in both men and women, as confirmed by Guo et al. (2024) and Chen et al. (2021).

Intervention form showed no significant moderation. Although monetary and social interventions may activate different motives, both material gain and social acceptance are fundamental behavioral motives (Chen & Chen, 2014), so any intervention form strengthens prosocial motivation. Some research suggests monetary interventions may crowd out intrinsic motivation, reducing prosocial behavior after withdrawal (Liu et al., 2010; Zhu, 2009). However, this study focused on behavior during intervention, where monetary and social interventions were equally effective. Combined monetary+social interventions could not be analyzed due to insufficient effect sizes (Card, 2016), leaving questions about combined forms for future research.

Intervention agent showed no significant moderation. Two possible explanations exist: First, participants may not psychologically distinguish between agents. Although computer interventions lack human intention, they remain rule-driven feedback mechanisms preset by humans (Liu et al., 2010; Cui et al., 2017). Participants may view computer agents as "rule enforcers," granting them normative roles equivalent to humans, limiting differential effects. Second, prosocial behavior may be more influenced by intervention outcomes than intentions. Most experiments did not manipulate intervention reasonableness or accuracy (Charness et al., 2008; Chen et al., 2021), so participants likely focused on consequences rather than inferring intentions, weakening agent effects.

Intervention cost showed no significant moderation. This may be because cost coding was researcher-based and differed from participants' actual perceptions (Lin et al., 2025). Within limited experimental timeframes, participants may not engage in complex reasoning about intervention processes, focusing instead on proximal outcomes rather than distal costs. This suggests intervention cost does not limit prosocial effects, though future research should examine whether explicit cost differences affect behavior.

Measurement paradigm showed no significant moderation, consistent with Halevy and Halali (2015). Both single and repeated interactions showed significant positive effects, with repeated interactions showing slightly but non-significantly stronger effects. This may be because single-intervention effects were already strong, and short-term repetition did not further enhance them. This demonstrates the robustness of intervention effects and suggests future longitudinal research should examine whether longer time spans produce stronger effects.

Control group setting showed no significant moderation. Both no-third-party and third-party observer controls yielded significant prosocial improvements, with non-significant differences between them. This may be because "observability" alone is insufficient to activate strong reputation mechanisms. Research indicates that observers effectively trigger reputation motives only when direct or indirect reciprocity is possible (Bradley et al., 2018). In third-party observer groups, the observer is an uninterested party without actual reciprocal relationships (Chen et al., 2021), limiting reputation concerns and making behavior similar to no-third-party conditions. This also demonstrates the robustness of intervention effects across different baselines.

5.1 Theoretical Contributions

First, this three-level meta-analysis reveals a moderately strong positive effect of third-party intervention on prosocial behavior, further validating its effectiveness as a norm maintenance mechanism. Second, we systematically examined multiple key moderators (age, gender, type, agent, form, probability, intensity, cost, measurement paradigm, control group), comprehensively exploring potential sources of effect heterogeneity. Moderator analyses confirmed overall robustness while identifying that higher intensity and probability enhance effects. These findings support deterrence theory, social norm focus theory, and indirect reciprocity theory, revealing their complementary explanatory power and applicable boundaries, while providing a systematic variable framework for future research. Additionally, comparing control groups (no third party vs. third-party observer) showed that active intervention significantly enhances prosocial behavior beyond passive observation, underscoring its norm maintenance function. Overall, this study advances understanding of third-party intervention's prosocial functions and mechanisms, offering theoretical and practical insights for enhancing public cooperation and social welfare.

5.2 Limitations and Future Directions

Several limitations warrant attention. First, although we included 10 moderators, other potentially important factors could be explored, such as intervention fairness/legitimacy (Rand et al., 2009; Wu et al., 2022) and group relations/social distance between interveners and targets (Harris et al., 2012; Vollan, 2011). These have been underreported and warrant future theoretical and empirical attention. Second, some moderators had uneven effect size distributions (e.g., only 5 effect sizes for third-party compensation; adult samples dominated age groups), limiting representativeness and stability of results, which should be interpreted cautiously. Third, some moderator manipulations could be expanded. For example, because social intervention intensity is difficult to quantify, our intensity analysis only included monetary interventions. Future research could develop conversion methods or equivalence models between monetary and social interventions to enhance comparability and generalizability (Chen & Xu, 2020). Similarly, intervention cost only included monetary costs; whether non-material costs (e.g., time investment, retaliation risk) produce differential effects requires further investigation (Chen et al., 2020). Regarding intervention probability, most studies focused on high-probability conditions, with insufficient data for low-to-moderate probabilities, limiting examination of non-linear patterns (e.g., inverted U-shape). Future research should systematically manipulate probability across the full range. Fourth, data missingness in early literature could not be fully resolved. Despite contacting authors and checking supplements, some early studies lacked complete descriptive statistics, preventing inclusion of some effect sizes. Future research should improve reporting of key statistics to facilitate meta-analyses. Finally, although some studies examined mediation mechanisms, insufficient numbers precluded meta-analytic examination. Future research should expand exploration of mechanisms and psychological pathways to deepen theoretical development and intervention practice.

Conclusion

This three-level meta-analysis examined the relationship between third-party intervention and prosocial behavior, confirming a moderately strong positive effect. The relationship is moderated by intervention intensity and probability but not by age, gender, intervention type, form, agent, cost, measurement paradigm, or control group setting. Overall, third-party intervention demonstrates robust and stable prosocial effects.

Appendix B: Meta-Analysis Coding Manual

Basic Study Information

Variable Categories/Values Definition/Range Publication Status Published journal article (J), Dissertation (D), Conference paper (C) Whether the article is formally published Study ID Author name + year For 3+ authors, only first author's name Mean Age Arithmetic mean of all participants' ages Use reported mean; if only range reported, use midpoint Age Group Child (Ch), Adolescent (Te), Adult (Ad) As reported or classified by mean age: <12 years = child; 13-17 = adolescent; >18 = adult Female Proportion 0%-100% Percentage of female participants in final sample (after exclusions) Experiment Number Original experiment/study number in source article

Experimental Design

Variable Categories/Values Definition Design Type Between-subjects, Within-subjects Whether manipulation was between or within subjects

Independent Variable Characteristics

Variable Categories/Values Definition Third-Party Intervention Type TP, TR, TC, TP+TR, TP+TC, TR+TC TP: Punish violators (e.g., reduce payoff, negative gossip, criticism); TR: Reward compliance (e.g., increase payoff, positive gossip, praise); TC: Compensate victims (e.g., monetary compensation, apology); Combinations as specified Third-Party Intervention Form MI, SI, MI+SI MI: Monetary changes to payoff; SI: Non-material (verbal, gossip, facial feedback); MI+SI: Both forms Third-Party Intervention Agent Human (H), Computer (S) Whether implementer is human or computer program Third-Party Intervention Cost Costly (YES), Costless (NO) Whether third party pays cost to intervene Third-Party Intervention Probability 0% ≤ p ≤ 100% Probability of intervention on specific consequences OR percentage of group receiving intervention Third-Party Intervention Intensity q = x/y Ratio of minimum intervention monetary change (x) to maximum violation gain (y):
• Punishment: x = minimum punishment amount
• Reward: x = minimum reward amount
• Compensation: x = minimum compensation amount
• y = maximum gain from single violation (varies by paradigm)

Dependent Variable Characteristics

Variable Categories/Values Definition Prosocial Behavior Measurement Paradigm Single interaction (SG), Repeated interaction (RG) SG: Single-round decision; RG: Multi-round decisions (mean or sum) Prosocial Behavior Measurement Tool DG, PGG, PD, IG, Other (OT) DG: Dictator Game allocation; PGG: Public Goods contribution; PD: Cooperation choice; IG: Investment/return decision; OT: Other tools (specify)

Control Group

Variable Categories/Values Definition Control Group Setting No-third-party control (CC), Third-party observer (OC) CC: No third party, no manipulation; OC: Third party present but cannot intervene

Statistical Data

Variable Definition Original Data Page Page number where data appear Test Statistic t, F, χ², Pearson's r (within-subjects separately), standardized β, unstandardized β, Z, Kruskal-Wallis H, Mann-Whitney U, Wilcoxon signed-rank, Fisher's exact, Spearman's ρ, Wald χ², Kolmogorov-Smirnov, G², McNemar χ², Kendall's τ, Epps-Singleton Degrees of Freedom Numerator and denominator df Statistic Value Specific test statistic value Regression SE Standard error of regression coefficient p-value Probability of observing current or more extreme result under null 95% CI Upper and lower confidence limits Reported Effect Size Type (d, r, η², etc.) and value Experimental Group Sample Size Number of participants (if not reported, assume equal across conditions; for within-subjects, record participants not observations) Experimental Group Prosocial Proportion Proportion choosing prosocial option in binary-choice games Experimental Group Mean Mean prosocial behavior for >2 choice options Experimental Group SD/SE Standard deviation/error for >2 choice options Control Group Sample Size Same specifications as experimental group Control Group Prosocial Proportion Same specifications as experimental group Control Group Mean/SD/SE Same specifications as experimental group