Determinants of recurrent-event model performance in TOPCAT-Americas
B Szabo-Soderberg, T Pol, L H LundAbstract
Background
Recurrent-event methods capture total disease burden but heterogeneity, treatment discontinuation and other factors limit their performance in randomized trials.
Methods
In TOPCAT-Americas, we evaluated: 1) how the choice between negative binomial regression (NBR) and Lin-Wei-Yang-Ying (LWYY) models affect treatment effect estimates on heart failure hospitalizations and cardiovascular death, 2) potential sources of power loss, such as treatment discontinuation, temporal clustering, covariate adjustment and competing risks, 3) the effect of risk marker adjustment on the statistical power of these models.
Results
Among 1765 patients, 1078 events of heart failure hospitalizations and cardiovascular death occurred (49% were recurrent events). Spironolactone consistently improved outcomes. Accounting for repeated events increased treatment effect estimates both for unadjusted NBR (IRR 0.80 [95% CI 0.65–0.98], p=0.0335 for the first event and 0.73 [0.59–0.91], p=0.0054 for all events) and for unadjusted LWYY models (HR 0.84 [0.71–0.99], p=0.0431 for first event and 0.80 [0.66–0.97], p=0.0262 for all events). Adjustment for risk markers improved the performance of both NBR (IRR 0.69 [0.56-0.84], p=0.0003 for all events) and LWYY models (HR 0.77 [0.63-0.93], p=0.0067). Accounting for competing risk of death in LWYY models further reduced statistical uncertainty (HR 0.82 [0.72-0.92], p=0.0009). Treatment discontinuations and temporal clustering of events had limited effect on the treatment effect estimates.
Compared to unadjusted models, adjustment for baseline risk markers increased statistical power for both NBR (from 57% to 67%) and LWYY models (from 77% to 94%).
Conclusion
Accounting for repeated events increase the estimated treatment effect for both NBR and LWYY models. Adjustment for baseline covariates efficiently increases statistical power by reducing heterogeneity. Considering competing risks may further improve model performance.