Handling missing covariates in observational studies: An illustration with the assessment of prognostic factors of survival outcomes in sarcoma in irradiated fields (SIF).

Abstract

11558 Background: Missing covariates are common in observational research. They can result in biased estimations and loss of power to detect associations. Limited data regarding prognostic factors of survival outcomes of SIF are available. We assessed prognostic factors of overall (OS), progression-free (PFS) and metastatic- progression-free (MPFS) survivals in SIF using 3 methods to account for missing covariates. Methods: We relied on the NETSARC database of the French Sarcoma Group: Cox models for OS/PFS and competitive hazard survival models for MPFS. Prognostic factors investigated were age, sex, tumor characteristics (histology, size, depth, grade), metastasis, surgery, surgical resection, expertise of the surgeon (NETSARC/other), pre-surgical imaging (Y/N) and neoadjuvant treatment. To account for missing covariates, we used multiple imputation (MI) by fully conditional specification (Bartlett 2015). The observed data are used to estimate the missing covariate and subsequently replace (impute) the missing value by that estimate. With the “missing-data modality” alternative approach, a category “missing” was created (e.g. yes/no/missing). With the “complete-case” alternative approach, analysis was restricted to patients with all covariates available. Results: Among the 504 patients, 169 had all covariates available (33%). Rate of missing data was greater than 20% for imaging, neoadjuvant treatment, and surgical resection. In the complete-case analysis (N = 169), factors associated with OS included R1/R2 (vs R0) surgical resection (p = 0.03). In opposite, MI (N = 504) revealed an association with metastasis (p = 0.03), and a time-varying effect for surgical resection (p < 0.001), with a risk increasing over time for R1/R2 vs R0). For PFS and MI, associated factors included higher grade and ungradable tumors (p = 0.002) and R1/R2 vs R0 resection (p < 0.001), not found with the complete-case analysis. For MPFS and MI, factors included metastasis (p = 0.03), higher grade tumor (p = 0.007) and R1/R2 resection (p = 0.001). Grade and resection were not found in the complete-case analysis. The missing-data modality led to slight differences with MI, including significant association due to missing covariate (e.g. tumors with size missing associated with shorter PFS). Conclusions: We identified prognostic factors of survival outcomes for SIF. The complete-case analysis led to reduced statistical power and population was non-representative of the full sample, introducing bias. In non-randomized studies, as the outcome may be related to variables with missing values, the missing-data modality method will typically result in biased estimates. Appropriate statistical methods for missing covariates, such as MI, should be considered in particular for observational studies.