# Predictive analysis of discrete survival time

I have a standard time to event data set, with both time-independent and time-varying covariates.

I assume the time to event is a discrete random variable, and construct the extended data set to estimate a logistic model (following this link: http://data.princeton.edu/wws509/notes/c7s6.html).

My questions are:

1. What is the independence assumption for the observations in the extended dataset?

2. My goal is the predict if $T_{i,t}$ will be an event for individual $i$ at time point $t$ with the trained logistic regression above. How should I split the extended dataset into training and testing set? Split by individuals or split by individual-time observations?

That individuals are independent. There are dependence between each of your rows in the extended data within individual as each outcome at $t$ is conditional on having survived at time $t - 1$.
I would split by individuals due to the former mentioned dependence. You are observing failure times $T_i$ and not binary indicators $y_{i1},\dots,y_{iT_i}$ in your initial non-extended dataset.