Original citation: J Clin Invest. 2010;123(9):3740–3750. doi:10.1172/JCI68509.
Citation for this erratum: J Clin Invest. 2013;123(12):5410. doi:10.1172/JCI74035.
Some expressions and notations related to Equations 1 and 2 were presented incorrectly. The correct text and equations are below.
The coefficients (β) in Cox’s regression model are estimated by maximizing the partial likelihood function subject to a constraint on the L1-norm of the coefficients. The lasso estimator (β̂) maximizes the objective function given below:
(Equation 1)
Here l(β) is the log partial likelihood in the Cox model; for the exact form of this function, see ref. 41. The tuning parameter, λ in Equation 1, was chosen by 10-fold cross-validation. For the implementation, we used the R package “glmnet” (39).
PROVAR was defined for each of the 222 TCGA samples as the sum of the estimated coefficients multiplied by protein expression levels, as shown below. Here i represents patients (i = 1, ..., 222), j represents proteins with nonzero coefficients (j = 1, ..., m), β̂j is the lasso coefficient of the jth protein marker, and Xij is the expression level of the jth protein for the ith patient.
(Equation 2)
The JCI regrets the error.