Introduction: Cancer combination therapies are widely used, but tolerability may limit their application. Although ~95% of approved combination therapies show additive or less-than-additive effects on Progression-Free Survival (Hwangbo et al. (2023)), it remains unclear whether severe AEs follow similar patterns. Since grade 3-4 AEs are consistently reported in clinical trials and critical determinants of clinical trial outcomes, we assessed whether grade 3-4 combination AE incidence follows independent action of individual drugs or by a model of maximum-monotherapy-incidence (MMI), in which the expected combination incidence equals the greater monotherapy value. Methods: We systematically identified 29 phase III trials of combination therapy that included matching monotherapy arms and extracted all grade 3-4 AEs reported (A vs. B vs. A+B). Combination AE incidences were calculated under the null hypothesis of independent action: P(A+B) = 1 - (1-P(A))(1-P(B)) (Palmer et al. (2017)). Deviations from independent action were quantified using Wald z-scores with Benjamini-Hochberg adjustment. We also compared the number of AEs better explained by independent action versus MMI model, and whether monotherapy AE incidence affected the results, and performed subgroup analyses by cancer type, drug class, and AE category. Results: Across all grade 3-4 AEs, 6% (12/198) largely exceeded expected incidence, 23% (45/198) fell below it, and 71% (141/198) remained within a 0.5-2.0 fold range of expectation. Deviation from the null, measured as mean squared error (MSE), varied markedly by drug class: immunotherapy combinations showed the least MSE (0.0003), solid cancers without immunotherapy were intermediate (0.0028), and hematologic cancers deviated most (0.0121). Independent action explained more AEs (130) than the MMI model (68), and accuracy did not depend on monotherapy AE incidence. Many hematologic AEs were below independence, consistent with shared mechanisms such as bone marrow suppression, whereas immunotherapy AEs aligned with independence, suggesting distinct pathways to toxicity. Conclusions: The incidences of serious adverse events from combination therapies are predominantly explained by independent action, similar to prior findings on efficacy of combination therapies. The proportion of AEs with ‘more than expected' (6%) or ‘less than expected' (22%) incidence are interestingly similar to rates of ‘more than expected' efficacy (5%) and ‘less than expected' efficacy (27%) (Hwangbo et al. (2017)). These results indicate that toxicity, like efficacy, is driven mainly by independent drug action and inter-patient heterogeneity, with synergistic or antagonistic toxicities being uncommon. The predictability of combination AEs supports integrating toxicity modeling into trial design to improve combination selection and escalation or de-escalation strategies.