Interquantile shrinkage and variable selection for longitudinal data in regression models (Q6961897)

For longitudinal data with potential correlations, this paper considers quantile regression models which assume the effects of the covariates vary at different quantile levels. A generalized estimating equation (GEE) approach is proposed to jointly estimate these effects across a finite number of quantile levels. The working correlation structure in the proposed approach incorporates both within subject correlation and cross-correlation among quantiles. In practice, the effects of covariates may remain constant within a specific quantile range while varying across other quantiles. An adaptive fused penalization approach is proposed to detect commonality by shrinking the differences in the effects between adjacent quantiles toward zero. An adaptive LASSO penalty is also proposed to eliminate irrelevant covariates and induce sparsity within the model. An alternating method of multipliers (ADMM) algorithm based on smoothed GEE function is developed to estimate the unknown parameters in the model. Model selection consistency and asymptotic normality of the estimators are established under some regularity conditions. Simulation studies under both non-sparsity and sparsity cases are performed to evaluate the accuracy of the estimates and the ability of the proposed procedure to identify the model structure. The proposed methods are also applied to analyze the longitudinal wage data from the National Longitudinal Surveys (NLS) conducted by the US Department of Labor. While the methods proposed in this paper provide valid estimates for the magnitude of the covariate effects, estimation for the variances of these estimates is not discussed in this paper. Therefore, full assessment of covariate effects by the proposed method may be difficult in some applications since it would include hypothesis testing and confidence interval estimation which require variance estimates.