Analytical Robustness and Competing Interpretations in Violent Video Game Research: A Response to Teng and Bushman's (2026) Reanalysis of Lacko et al. (2024).

For any dataset, multiple plausible analytic approaches usually exist, and reasonable choices can lead to meaningfully different conclusions (Wagenmakers et al.2022). Given this analytic flexibility, thorough robustness checks are essential (Aczel et al.2026). Teng and Bushman's (“TB”;2026) response to our article (Lacko et al.2024) provides an opportunity to further examine the robustness and interpretation of our findings. Because TB's reanalysis reflects one set of analytical choices, we test whether their interpretation holds across alternative specifications and operational decisions. All robustness checks are available on OSF (https://osf.io/q5wv6/).

Reaction to TB's Criticism

First, we address concerns that were either unsupported or reflected matters of preference and therefore did not threaten the validity of the original findings. TB raised issues regarding our interpretation of the General Aggression Model (GAM), missing information, methodological decisions, missing data, measurement of aggression, causal inference, time constraints, and overparameterization. Their arguments, our responses, and the supporting robustness checks are summarized in Table1.

Table: Summary of unsupported TB's concerns.

Next, we address arguments that we consider relevant and potentially threatening to the validity of the original results. TB argued that the positive between‐person association between affective empathy and aggression constitutes a serious issue. This concern would be warranted if the association were robust and substantial. However, meta‐analyses (cited also by TB) indicate that the relationship is weak and heterogeneous (r≈ −0.10; Ritchie et al.2022; Vachon et al.2014), with empathy accounting for only ~1% of the variance in aggression (Vachon et al.2014). Moreover, similar positive associations have appeared in recent studies (Kahhale et al.2024; Palumbo and Latzman2021). Importantly, prior studies reporting negative associations did not distinguish within‐ and between‐person variance; their findings are not directly comparable to our RI‐CLPM results. For robustness, we conducted additional analyses: testing strict measurement invariance, estimating models with only aggression and affective empathy to rule out suppressor effects, and re‐estimating the model without affective empathy (OSF tab “AMES”). None indicated problems with the empathy measure.

TB further argued that our operationalization of VVG using the maximum score across up to three games is inappropriate. Although reasonable, alternative approaches have clear drawbacks. Because we did not measure time spent per game, weighted or frequency‐based scores would confound violent content with overall gaming frequency. Mean or sum scores introduce similar inconsistencies. For example, a participant playing three games with scores 1, 2, and 2 would obtain a highersumthan someone playing a single game scored 4, despite lower overall violent exposure. A participant playing two highly violent games (5 and 5) and one low‐violence game (1) would have a lowermeanthan someone playing only one game scored 4. TB's suggestion that prosocial gaming offsets violence is also speculative, as low violence does not imply higher prosocial content, and vice versa. Given that GAM assumes that mere exposure to violent content can increase aggression, the maximum score remains the most defensible operationalization. Nevertheless, we re‐estimated models using alternative operationalizations (max, sum, mean, and mean × time). Some inconsistencies emerged because these indicators are correlated but not identical (r> 0.75). Most differences occurred in the T3 → T4 cross‐lagged paths and varied across operationalizations. Importantly, although our findings are not fully robust across VVG operationalizations, none consistently supported any GAM interpretation (desensitization or selection), nor challenged our conclusion that overall evidence for GAM is limited (OSF tab “VVG”).

TB also argue that longitudinal analyses capture participants “mid‐stream” in developmental processes, so cumulative VVG effects manifest primarily as between‐person differences, limiting the value of within‐person estimates. While theoretically interesting, this claim lacks empirical support. Longitudinal studies capable of testing cumulative within‐person effects beyond our approach are lacking, and evidence for GAM's cumulative mechanism remains limited (Drummond et al.2020; Ferguson and Dyck2012). Addressing this fully would require intensive burst designs and methods for modeling random slopes. Conceptually, within‐person change aligns with GAM processes, as TB acknowledge (Teng et al.2019,2022). GAM treats VVG as a situational input influencing proximal affective and cognitive processes, which should manifest as within‐person fluctuations.

TB's claim that cumulative processes appear mainly as between‐person variance is also problematic. Between‐person variance reflects individuals' standing during the measurement window, not earlier developmental histories. If cumulative desensitization effects accrue prior to the study, older participants should differ systematically from younger participants in VVG‐related associations. In other words, age‐related differences in desensitization would be expected. However, neither our analyses nor TB's reanalysis provided evidence of such patterns. Both analyses identified only selection effects. Moreover, in the original analysis, we found no evidence of trait‐level change (trends or growth) when examining grand means over time. We also tested mediation for VVG → empathy → aggression to capture potential cumulative effects across waves. None of the indirect effects were statistically significant (OSF tab “Mediation”).

Limitations and Inconsistencies of TB's Response

TB's reanalysis appears to rely on a subset of estimated models, leaving the full scope of results unclear. Although multiple models were estimated, only selected results were interpreted. In contrast, we interpret results across all estimated models, including inconsistencies, to provide a complete and robust interpretation. TB also evaluated time constraints using fit indices rather than the recommended chi‐square difference test (Mulder and Hamaker2021). Even by their criteria, Model 4 (full constraints over time) should be preferred over Model 3 (partial constraints over time). Notably, in Model 4, the within‐person VVG → empathy effect (the only effect supporting GAM at the within‐person level) becomes non‐significant. This effect also disappears when T1 aggression is included as a time‐invariant covariate, suggesting it may be spurious. Re‐estimating the models using alternative VVG operationalizations confirmed this pattern: although some operationalizations produced negative within‐person effects on aggression (contrary to GAM), none yielded a significant negative effect of VVG on empathy (OSF tab “Inconsistencies”).

Further issues arise from TB's VVG computation. We identified systematic discrepancies between the means reported by TB and the correct values derived from the raw data. For example, participant 210 played games scored 2, 4, and 3 at Wave 1; the correct mean is 3, yet TB's dataset lists 4. Across participants, TB's index correlates onlyr≈ 0.88 with the correct mean, leaving ~22% of variance unexplained (OSF tab “VVG”). Because TB did not provide the processing syntax, these discrepancies cannot be verified, and their derivation is unclear. This lack of detail in reported procedures is critical given the central role of TB's operationalization in their analyses. As shown in Table2, relative to the original max‐score approach, TB's VVG operationalization based on unexplained data modification yields the highest proportion of non‐robust effects (i.e., deviated the most), compared to all other plausible alternatives (including in the correctly computed mean×time specification; OSF tab “Inconsistencies”). TB also interpreted effect sizes inconsistently: catharsis effects contradicting GAM were labeled small even when they exceeded the magnitude of desensitization effects considered meaningful when supporting GAM. This asymmetric application of effect size thresholds, depending on theoretical alignment, complicates interpretation and undermines the reliability of their conclusions.

Table: Proportions of non‐robust effects using different VVG operationalization under time‐constrained models.

Broader Implications for GAM in Future Research

This highlights a broader problem for GAM. TB repeatedly refer to meaningful effect sizes and practical significance, yet GAM provides no criteria for what is practically significant. The theory also does not specify cumulative mechanisms or how age and exposure duration shape effects. In our 2‐year study, with 6‐month intervals starting at age 11, we expect such patterns to emerge if they exist. Instead, TB argue that long‐term effects occurred before the study, implying no cumulative change should be detectable. They further suggest that negative within‐person effects of VVG on aggression align with catharsis theory. However, catharsis theory is incompatible with GAM and has been explicitly rejected by GAM proponents (Bushman et al.1999; Bushman2002). Such findings should therefore be interpreted as contradictory to GAM. Consequently, if null, positive, and negative effects can all be framed as evidence in support of the model, a serious epistemological problem emerges: the theory becomes unfalsifiable and thus unscientific (Devilly et al.2023; Finkel2014).

We acknowledge GAM's historical significance and its potential contemporary relevance, and we note that a single study cannot disprove the model, though our findings may limit its generalizability across cultural and contextual settings that may shape how violent media exposure relates to aggression. This makes cross‐context testing especially important for identifying potential boundary conditions. However, for GAM to remain scientifically useful, greater theoretical specificity is needed. Many scholars argue that addressing the broader theory crisis requires a shift toward formal modeling, which enhances rigor, transparency, and predictive specificity (Eronen and Bringmann2021; Oberauer and Lewandowsky2019; van Dongen et al.2025). Such a shift would require articulating precise hypotheses, including expected effect sizes and causal pathways, thereby rendering GAM empirically testable and falsifiable. Although formal modeling demands more rigorous groundwork than traditional post hoc interpretations, it is essential for strengthening the field's predictive power and theoretical depth.

Ethics Statement

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

References

1. Aczel, B. , B. Szaszi, H. T. Clelland, et al. 2026. “Investigating the Analytical Robustness of the Social and Behavioural Sciences. ”Nature652, no. 8108: 135–142. . doi.org/10.1038/s41586-025-09844-9
2. Bushman, B. J. 2002. “Does Venting Anger Feed or Extinguish the Flame? Catharsis, Rumination, Distraction, Anger and Aggressive Responding. ”Personality and Social Psychology Bulletin28, no. 6: 724–731. . doi.org/10.1177/0146167202289002
3. Bushman, B. J. , R. F. Baumeister, andA. D. Stack. 1999. “Catharsis, Aggression, and Persuasive Influence: Self‐Fulfilling or Self‐Defeating Prophecies?”Journal of Personality and Social Psychology76, no. 3: 367–376. . doi.org/10.1037/0022-3514.76.3.367
4. Devilly, G. J. , A. Drummond, J. D. Sauer, A. Copenhaver, J. Kneer, andC. J. Ferguson. 2023. “Directional Is the New Null? A Comment on Bushman and Anderson (2021). ”Psychology of Popular Media12, no. 3: 364–372. . doi.org/10.1037/ppm0000447
5. Drummond, A. , J. D. Sauer, andC. J. Ferguson. 2020. “Do Longitudinal Studies Support Long‐Term Relationships Between Aggressive Game Play and Youth Aggressive Behaviour? A Meta‐Analytic Examination. ”Royal Society Open Science7, no. 7: 200373. . doi.org/10.1098/rsos.200373
6. Eronen, M. I. , andL. F. Bringmann. 2021. “The Theory Crisis in Psychology: How to Move Forward. ”Perspectives on Psychological Science16, no. 4: 779–788. . doi.org/10.1177/1745691620970586
7. Ferguson, C. J. , andD. Dyck. 2012. “Paradigm Change in Aggression Research: The Time Has Come to Retire the General Aggression Model. ”Aggression and Violent Behavior17, no. 3: 220–228. . doi.org/10.1016/j.avb.2012.02.007
8. Finkel, E. J. 2014. “The I3 Model. ” InAdvances in Experimental Social Psychology, 1–104. . doi.org/10.1016/b978-0-12-800052-6.00001-9
9. Kahhale, I. , J. L. Hanson, A. Raine, andA. L. Byrd. 2024. “Associations Between Subtypes of Empathy and Aggression in High‐Risk Adolescents. ”Journal of Psychopathology and Behavioral Assessment46, no. 1: 62–75. . doi.org/10.1007/s10862-023-10112-1
10. Lacko, D. , H. Machackova, andD. Smahel. 2024. “Does Violence in Video Games Impact Aggression and Empathy? A Longitudinal Study of Czech Adolescents to Differentiate Within‐ and Between‐Person Effects. ”Computers in Human Behavior159, no. 108341: 108341. . doi.org/10.1016/j.chb.2024.108341
11. Mulder, J. D. , andE. L. Hamaker. 2021. “Three Extensions of the Random Intercept Cross‐Lagged Panel Model. ”Structural Equation Modeling: A Multidisciplinary Journal28, no. 4: 638–648. . doi.org/10.1080/10705511.2020.1784738
12. Oberauer, K. , andS. Lewandowsky. 2019. “Addressing the Theory Crisis in Psychology. ”Psychonomic Bulletin & Review26, no. 5: 1596–1618. . doi.org/10.3758/s13423-019-01645-2
13. Orth, U. , D. A. Clark, M. B. Donnellan, andR. W. Robins. 2021. “Testing Prospective Effects in Longitudinal Research: Comparing Seven Competing Cross‐Lagged Models. ”Journal of Personality and Social Psychology120, no. 4: 1013–1034. . doi.org/10.1037/pspp0000358
14. Palumbo, I. M. , andR. D. Latzman. 2021. “Parsing Associations Between Dimensions of Empathy and Reactive and Proactive Aggression. ”Journal of Personality Disorders35, no. Supple C: 56–74. . doi.org/10.1521/pedi_2021_35_522
15. Ritchie, M. B. , R. W. J. Neufeld, M. Yoon, A. Li, andD. G. V. Mitchell. 2022. “Predicting Youth Aggression With Empathy and Callous Unemotional Traits: A Meta‐Analytic Review. ”Clinical Psychology Review98: 102186. . doi.org/10.1016/j.cpr.2022.102186
16. Teng, Z. , andB. J. Bushman. 2026. “Best Practice Procedures in Longitudinal Research of Violent Media Effects: Reanalysis of the Lacko et al. (2024) Study as a Case Study. ”Aggressive Behavior52, no. 1: e70054. . doi.org/10.1002/ab.70054
17. Teng, Z. , Q. Nie, C. Guo, Q. Zhang, Y. Liu, andB. J. Bushman. 2019. “A Longitudinal Study of Link Between Exposure to Violent Video Games and Aggression in Chinese Adolescents: The Mediating Role of Moral Disengagement. ”Developmental Psychology55, no. 1: 184–195. . doi.org/10.1037/dev0000624
18. Teng, Z. , C. Yang, M. Stomski, Q. Nie, andC. Guo. 2022. “Violent Video Game Exposure and Bullying in Early Adolescence: A Longitudinal Study Examining Moderation of Trait Aggressiveness and Moral Identity. ”Psychology of Violence12, no. 3: 149–159. . doi.org/10.1037/vio0000424
19. Usami, S. 2021. “On the Differences Between General Cross‐Lagged Panel Model and Random‐Intercept Cross‐Lagged Panel Model: Interpretation of Cross‐Lagged Parameters and Model Choice. ”Structural Equation Modeling: A Multidisciplinary Journal28, no. 3: 331–344. . doi.org/10.1080/10705511.2020.1821690
20. Vachon, D. D. , D. R. Lynam, andJ. A. Johnson. 2014. “The (Non)Relation Between Empathy and Aggression: Surprising Results From a Meta‐Analysis. ”Psychological Bulletin140, no. 3: 751–773. . doi.org/10.1037/a0035236
21. van Dongen, N. , R. van Bork, A. Finnemann, et al. 2025. “Productive Explanation: A Framework for Evaluating Explanations in Psychological Science. ”Psychological Review132, no. 2: 311–329. . doi.org/10.1037/rev0000479
22. Wagenmakers, E. J. , A. Sarafoglou, andB. Aczel. 2022. “One Statistical Analysis Must Not Rule Them All. ”Nature605, no. 7910: 423–425. . doi.org/10.1038/d41586-022-01332-8

---

Republished from the open web under CC-BY. Authors: Lacko D, Machackova H, Smahel D. Read the original.