Caren van Roekel

188 Chapter 7 Statistical Analyses Patient demographics and treatment characteristics were summarized using descriptive analyses. The strength of association between CTCAE toxicity grade and parenchymal-absorbed dose was assessed using linear regression models with CTCAE grade in categories as the dependent continuous variable and parenchymal absorbed dose as the independent continuous variable. For clinical significance, CTCAE grading of any clinical and laboratory toxicity was also dichotomized in the following categories: grade 0/I/II versus grade III/IV/V and analyzed using logistic regression with Firth’s correction for small sample bias (17). The association between relative change in laboratory parameters (represented as Δ laboratory parameter) and healthy liver tissue dose was analyzed using simple linear regression models with percentage change as the dependent continuous variable and parenchymal absorbed dose as the independent continuous variable, after log-transformation of the dependent variable to fulfill model assumptions. All toxicity analyses were also adjusted for response to therapy (binary coded as response/non-response), previous treatment (defined as number of prior systemic treatment lines, categorical variable) and tumor load (defined as percentage involvement of the liver by tumors, continuous variable) as possible confounders, which were identified by making directed acyclic graphs. The relationship between tumor-absorbed dose and response was analyzed using a linear mixed-effects regression model with tumor-absorbed dose as the dependent variable. This type of analysis was chosen to account for correlation of tumors within patients. To fulfill model assumptions, dose was log-transformed. Nested models were compared using Akaike’s Information Criterion (AIC). The dose-effect relationship was best explained using a random intercept per patient without random slopes. A geometric mean of the tumor- absorbed dose per patient per response category was estimated because the anti-log of the arithmetic mean of log-transformed values is the geometric mean. A trend test was also done with response as a continuous variable in the model, to test the presence of an ordered relationship across response categories. By including them as co-variables, analyses were adjusted for the following possible confounders: previous treatment (coded as factor with the following categories: yes/no previous treatment with anti-VEGF medication) and tumor load (continuous). An ROC analysis, according for clustered data,