Statistical methods to analyze treatment strategies from randomized and non-randomized studies in oncology

This thesis focuses on statistical methods to analyze data from different types of sources: data from randomized trials, from registries, and from single-arm studies combined with registry data. The methods were applied to studies in different areas of oncology: early-stage breast cancer, diffuse large B-cell lymphoma (DLBCL), and advanced-stage ovarian cancer. The objective was to develop and validate treatment strategies, and provide predictions for individual patients, with the aim to support clinicians in shared decision making and advance personalized care.
Key results include:
- Development of a model for early-stage breast cancer patients treated with breast-conserving therapy in the randomized Boost vs No Boost trial. A novel variable-selection procedure was used to reduce the effect of over-fitting and make the model easier to use in practice. Its application may aid in individual-patient decision making by predicting the 10-year local recurrence-free probability.
- Expansion of the analysis of the HOVON-130 DLBCL trial by a cohort of patients who received standard R-CHOP treatment. By employing three statistical methods that reduce treatment-selection bias (multi-variable regression, one-to-one matching, and inverse probability of treatment weighting), it was demonstrated that lenalidomide as an add-on treatment improves overall and progression-free survival for DLBCL patients with a MYC rearrangement.
- Validation of a treatment-selection strategy for choosing between primary surgery and neoadjuvant chemotherapy for patients with advanced epithelial ovarian cancer. This validation involved combining multiple imputation with inverse probability of treatment weighting. The results may contribute to adoption of the strategy on a broader scale, facilitating the decision-making process when selecting a treatment for women with advanced stage ovarian cancer.