Evaluating a candidate biomarker or developing a predictive model score for event-time outcomes is frequently an important clinical goal. However, model development and assessment may be complicated in the presence of competing risks. The time-dependent incident/dynamic cause-specific (CS) area under the ROC curve (AUC) proposed by Saha and Heagerty (2010, Biometrics 66, 999-1011) is an appealing semi-parametric measure to capture the predictive performance of a biomarker and incorporate competing risks. We propose a local and a global non-parametric estimator for the time-dependent CS AUC from censored survival data in the presence of competing risks. The first proposed estimator is a local average of time-dependent CS AUCs. The second estimator is based on modelling the CS AUC as a function of t through fractional polynomials. We investigated the performance of the proposed estimators through both simulation and real-life data analysis.