Abstract:

We use pairwise fusion penalty regularization method， based on Tobit regression model， to perform subgroup analysis on left-censored data with heterogeneity， simultaneously estimating regression parameters and identifying subgroups. By introducing a set of new parameters， the original optimization problem is transformed into a multivariate optimization problem with equality constraints only that can be solved by alternating direction method of multipliers. Moreover， the multivariate function related to the loss in each iteration is transformed into a group of quadratic surrogate functions of single variable by generalized coordinate descent algorithm. We prove that the proposed algorithm is convergent， and establish the large sample properties of the obtained parameter estimators. Simulation studies and real data analysis show that the proposed method has good performance.

Key words: left-censored data, Tobit model, subgroup identification, pairwise fusion penalty