|
Activity Number:
|
286
- Missing Data Methods
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 11, 2021 : 1:30 PM to 3:20 PM
|
|
Sponsor:
|
Biometrics Section
|
|
Abstract #318749
|
|
|
Title:
|
Comparison of Variable Selection Methods Based on LASSO After Multiple Imputation of Missing Covariates in Survival Data
|
|
Author(s):
|
Qian Yang* and Susan Halabi and Bin Luo
|
|
Companies:
|
Duke University and Duke University and Duke University
|
|
Keywords:
|
multiple imputation;
proportional hazards model;
missing data;
penalized method
|
|
Abstract:
|
Multiple imputation (MI) is a standard approach to handling missing data. MI is more efficient and avoids the potential risk of bias in complete cases studies. The implementation of penalized methods in multiple imputed data can be complicated due to the inconsistent variable selection when it is applied on each imputed data set. The motivation was to develop a predictive model of overall survival, and to find an optimal variable selection method when there are missing baseline covariates. We conducted a simulation study with a penalized proportional hazards model in survival data to assess the performance of several types of variable selection methods: variable selection on each MI dataset, and take average of variables selected by models in all/any/>50% datasets; and variable selection on stacked MI dataset with/without weighted regression. We used least absolute shrinkage and selection operator (LASSO) and adaptive LASSO (aLASSO) penalty functions with different choices of tuning parameters. The stacked method with weight provides a simple and less computationally demanding approach. We applied all variable selection methods to a real dataset of 1050 prostate cancer patients.
|
Authors who are presenting talks have a * after their name.
