Prescriptive Survival Analysis

Prescriptive Survival Analysis#

This type of analytics answer the question “what should be done?”

Note

This chapter is out of the scope of this tutorial. However, for interested readers some pointer to the relevant topics are provided for further reading.

Introduction to Causal Inference (CI)#

There are two schools of thought in Causal Inference.

  1. Potential Outcome

  2. Structural Causal Models

See also

The Causal Inference school of thoughts are extensively covered in below talks

Applying CI tools to Survival Analysis#

  1. Potential Outcome:

    • Calculate the propensity score given relevant variables

    • Correcting for unbalanced confounding using the Inverse Propensity Weights to re-weight Kaplan-Meier (KM) estimator.

  2. Structural Causal Models: Given the desired variable for effect estimation

    • Identify the causal graph

    • Fit a Cox proportional hazard with only confounders (avoid including mediator & Collider)

    • Ensure the cox proportional hazard assumption is satisfied

    • Average Treatment Effect (ATE) is the coefficient of the desired variable i.e., treatment

Note

Checking cox proportional hazard assumption is not important for predictive analytics, while the situation is different in prescriptive analytics.

  • In prediction the objective is to maximize an accuracy metric, and not to learn about how individual feature contribute to the model making that prediction.

  • In prescriptive analytics, the focus is on inference & correlation and to understand the influence of individual variables on the survival duration & event.

Checking & Ensuring Proportional Hazard Assumption#

Under preparation

Contents