y**3
发帖数: 267 | 1 For time-to-event survival model,we may do time -sliced classification
models to predict event(yes or no). For the whole data set, the claims9 in
my data) are censored at different times. I can build different models using
data censored at 1 year, 2 year, 3 years, 4 years etc. For example, for the
2 year model, I use all the claims IDs of age >=2 years and censor them at
2 years. Then use the two years data to find the target and predictors. I
think that is statistically how we do it. Then this is the typical
classification models. No time component is involved.Then we cant predict
into future using each model.
But a different suggestion is using the 2 years data to find predictors,
but for the target, using the whole life of claim to find the event occured
or not.In this way, all the target values for all the time-sliced models
such as 1 year model 2-year modle, 3-years model etc are gonna be the same.
Is this statistically right?
Thanks so much! | y**3
发帖数: 267 | |
|
