(002) A Quantitative Bias Analysis to Assess the Impact of Tumor Selection on Linear Growth in Observational Research: a Case Study in Renal Cancer.

Background: For observational cohort studies of cancer, differential selection of patients with faster or slower growing tumors can mischaracterize the linear growth rate (LGR). One mechanism by which this can occur is with censoring of faster growing tumors due to surgical treatment, which can potentially underestimate the LGR.

Objectives: Conduct a quantitative bias analysis (QBA) to estimate the impact of differential selection of renal cell carcinoma (RCC) tumors on LGR estimation.

Methods: The index date was the earliest radiology report during the study window (2004-2020) with measurable disease (≥1 cm tumor diameter) for patients with von Hippel-Lindau disease-associated RCC. Based on radiology review by an independent committee, up to five target tumors at the index date were identified and measured during follow-up. LGR estimates were calculated for tumors using a hierarchical linear model. For the QBA, tumor measurements were classified into four mutually exclusive groups defined by key variables related to potential selection bias (tumor diameter during follow-up and follow-up time). Selection probabilities were derived by 1) estimating the probability of target tumor selection based on the distribution of new tumors during follow-up to assess differential selection into the cohort, and 2) prespecifying hypothetical probabilities of selection (i.e., 0.9 to 0.5) for larger tumors at the end of follow-up with shorter follow-up duration to assess differential loss due to surgery during follow-up. In both approaches, Monte Carlo sampling was used to draw values for the selection probabilities over 1,000 iterations which were used to weight target tumors as 1/selection probability in the LGR analysis. Bias-corrected LGRs and simulation intervals (SI) are presented.

Results: The cohort included 244 VHL RCC patients with 601 target tumors with 2139 tumor measurements. 60% of tumors underwent surgery in follow-up. The LGR was 1.6 mm/year (IQR: 0.7, 2.4) among target tumors, and higher among new tumors and tumors with renal surgery. The corrected LGR (95% SI) using selection probabilities based on the distribution of new tumors was 1.7 mm/year (1.3, 2.0). Estimates based on hypothetical levels of selection were similar.

Conclusions: For tumors systematically measured and recorded, this analysis found little impact on the LGR indicating minimal selection bias based on differential selection of tumor size or degree of follow-up. The analysis does not address bias potentially related to development and validation of the LGR model or other potential forms of bias.