467 Graft Versus Leukemia Effect Is Overestimated in Studies with High Rates of Graft Versus Host Disease Related Mortality: A Simulation Study

Track: Poster Abstracts
Saturday, February 14, 2015, 6:45 PM-7:45 PM
Grand Hall CD (Manchester Grand Hyatt)
Murtadha Al-Khabori, MD, MSc, FRCPC , Hematology, Sultan Qaboos University Hospital, Muscat, Oman
Mohammed Al Huneini, MD , Hematology, Sultan Qaboos University Hospital, Muscat, Oman
Khalil Al Farsi, MD, FRCPC , Hematology, Sultan Qaboos University Hospital, Muscat, Oman
Salam Al Kindi, MD , Hematology, Sultan Qaboos University Hospital, Muscat, Oman
David Dennison , Hematology, Sultan Qaboos University Hospital, Muscat, Oman
Abdulhakeem Al Rawas, MD , Child Health, Sultan Qaboos University Hospital, Muscat, Oman
Yaser Wali, MD , Child Health, Sultan Qaboos University Hospital, Muscat, Oman
Presentation recording not available for download or distribution as requested by the presenting author.

Graft Versus Leukemia Effect is Overestimated in Studies with High Rates of Graft Versus Host Disease Related Mortality: A Simulation Study

Introduction

Graft versus leukemia (GvL) effect decreases the risk of relapse. The presence of graft versus host disease (GvHD) was previously shown to correlate with the GvL effect. GvHD-related mortality may spuriously lead to this conclusion when regression models are used without adjusting for competing risk of death. Herein, we planned to assess the performance of Fine and Gray  proportional hazard model, a commonly used competing risk model in survival analysis, in a simulation study using different GvHD-related mortality rates and across different simulated GvL effects.

Methods

We simulated datasets of 500 patients with maximum follow up of 5 years using Weibull distribution with 30% risk for GvHD and relapse. The relapse status was assigned using different simulated impact of GvL effect (simulated Hazard Ratio [HR]: 0.8 to 0.2) and at different GvHD-related mortality (HR: 1.2 to 4.8). We then estimated the HRs from different datasets using the Fine and Gray model. Then, we plotted the estimated HRs against the different GvHD-related mortality at different simulated GvL effects. We used simulation scenarios that are close to the expected actual rates in real datasets. We used STATA 13 for all statistical analyses.

Results

Fine and Gray proportional hazard competing risk model overestimated the impact of GvL effect with progressively increasing bias as GvHD-related mortality increases. The median bias in the estimated HR ranged from 0.09-0.34. The maximum median bias was estimated in the dataset with the minimal simulated GvL effect (HR of 0.8) and the minimal median bias was estimated in the dataset with the maximum simulated GvL effect (HR of 0.2). The bias overall increased with increasing GvHD-related mortality across different simulated GvL effect (Figure 1).

Conclusion

Fine and Gray proportional hazard competing risk model overestimated the impact of GvL effect with progressively increasing bias as GvHD-related mortality increases. We caution its use in datasets with high GvHD-related mortality.

 

Figure 1: Estimated Impact of GvHD on Relapse

Disclosures:
Nothing To Disclose
See more of: Poster Session 2: GVH/GVL
See more of: Poster Abstracts