Bonferroni Correction Is Not Applicable to the Problem of Determining Dose Volume Predictors of Clinical Outcomes. an Example of Alternative Statistical Approaches.
Abstract
Purpose
Dose Volume Histogram (DVH) indices remain a backbone of treatment planning in radiation therapy. Many outcomes studies treat DVH indices as independent candidate predictors and retain only indices that meet the statistical significance cutoff of p=0.05. Such a procedure raises multiple comparisons concerns (Bonferroni correction). We aim to show that Bonferroni Correction does not apply to this problem because of strong correlations between indices. Systematic scanning of the index space is more appropriate, combined with statistical methods that account for correlations.
Methods
We used two patient outcomes studies: (1) grade >= 2 acute rectal toxicity in a cohort of 79 patients treated with conventionally fractionated radiation therapy (RT) for prostate cancer, and (2) cardiac toxicity in a cohort of 134 locally advanced, Non-small Cell Lung Cancer (LA-NSCLC) patients treated with conventionally fractionated radiation therapy, with overall survival (OS) as a clinical endpoint. Both cohorts were treated at Mayo Clinic Arizona. All patients had their treatment plans in Eclipse TPS (Varian, Inc). ESAPI interface was used to extract arrays of V%_D indices computed at 1Gy and 5Gy intervals. The V%_D index is the percentage of OAR volume subjected to dose “D” or greater. We first fit a family of NTCP models, each using one DVH index as a potential predictor. Subsequently, we fit generalized linear model with Fused Lasso constraint to the same data.
Results
p-values associated with NTCP models are not randomly distributed, but show systematic structure suggesting predictive and transition regions, with the transition region dominated by correlations with indices in the predictive region. The GLM model finds distinct dose thresholds separating predictive and transition regions.
Conclusion
Bonferroni correction is not applicable to this problem because DVH indices are not independent predictor candidates. Systematic scanning of index space, combined with advanced statistical methods, needs to be used instead.