I tried to develop a cox regression model to predict the survival outcomes of the patients with a particular disease.
My data included X1, X2, X3, X and time, status. Let X1, X2, X3 be the independent variables and X be the covariate.
My cox model architecture is:
model = coxph(Surv(time, status) ~ X1 + X2 + X3, data = train_data)
From the established model, I create a "representative" predictor:
train_data$predictors = predict(model, newdata = train_data, type = 'lp')
The 'predictors' variable is continuous (or numerical).
I knew that an unobserved confounder (U) is obvious. I tried to use sensitivity analysis method proposed by Huang et al., 2019 ([PDF] Sensitivity analysis of treatment effect to unmeasured confounding in observational studies with survival and competing risks outcomes. | Semantic Scholar).
In R, we have survSens package to perform this method. Unfortunately, the treatment (in my case, the newly created 'predictors' variable) is binary.
I intended to use different cut-off of 'predictors' to transform it into a binary 'predictor' such as 25th, 50th, and 75th percentile of the linear 'predictors' variable.
Is it a correct way to perform this type of analysis?
Sorry if it's a clumsy question. I don't have a profound understanding of statistics.