Penalized Composite Quasi-Likelihood for Ultrahigh-Dimensional Variable Selection
Citations
892 citations
265 citations
Cites methods from "Penalized Composite Quasi-Likelihoo..."
...Moreover, we show that the proposed method is much more efficient than the least-squares-based method for many non-normal errors and that it only loses a small amount of efficiency for normal errors....
[...]
228 citations
199 citations
152 citations
References
5,231 citations
4,600 citations
4,382 citations
3,652 citations
"Penalized Composite Quasi-Likelihoo..." refers background or methods or result in this paper
...2 Penalized composite quantile regression Weighted composite quantile regression (WCQR) was first studied by Koenker (1984) in a classical statistical inference setting. Zou and Yuan (2008) used equally weighted composite quantile regression (EWCQR) for penalized model selection with p large but fixed....
[...]
...2 Penalized composite quantile regression Weighted composite quantile regression (WCQR) was first studied by Koenker (1984) in a classical statistical inference setting....
[...]
...This term is identical to the situation that was dealt with by Portnoy (1985). Using his result, the second conclusion of theorem 2 follows....
[...]
...Particular focus will be given to the oracle property of Fan and Li (2001), but we shall strengthen it and prove that estimator (5) is an oracle estimator with overwhelming probability. Fan and Lv (2010) were among the first to discuss the oracle properties with nonpolynomial dimensionality by using the full likelihood function in generalized linear models with a class of folded concave penalties....
[...]
...This method is particularly computationally efficient in ultrahigh dimensional problems and here we retained the top 100 SNPs with the largest F -statistics. In the second step, we applied to the screened data the penalized L2- and L1-regression, L1–L + 2 , L1–L2, EWCQR, WCQR + and WCQR with local linear approximation of the SCAD penalty. All the four composite quantile regressions used quantiles at .10%, . . . , 90%/. The lasso was used as the initial estimator and the tuning parameter in both the lasso and the SCAD penalty was chosen by fivefold cross-validation. In all three populations, the L1–L2- and L1–L + 2 -regressions reduced to L2regression. This is not unexpected owing to the gene expression normalization procedure. In addition, WCQR reduced to WCQR+. The selected SNPs, their coefficients and distances from the TSS are summarized in Tables 5–7. In the Asian population (Table 5), the five methods are reasonably consistent in not only variables selection but also coefficient estimation (in terms of signs and order of magnitude). WCQR uses the weights .0:19, 0:11, 0:02, 0, 0:12, 0:09, 0:18, 0:19, 0:10/. There are four SNPs which were chosen by all five methods. Two of them, rs2832159 and rs2245431, up-regulate gene expression, whereas rs9981984 and rs16981663 down-regulate gene expression. EWCQR selects the largest set of SNPs, whereas L1-regression selects the smallest set. In the CEPH population (Table 6), all five methods consistently selected the same seven SNPs with only EWCQR choosing two additional SNPs. WCQR uses the weight .0:19, 0:21, 0, 0:04, 0:03, 0:07, 0:1, 0:21, 0:15/. The coefficient estimations were also highly consistent. Deutsch et al. (2005) performed a similar cis-EQTL mapping for gene CCT8 using the same CEPH data as here....
[...]
3,539 citations