PBC: Time-Dependent Accuracy

• Background Reading
  • Original "Mayo Model" paper ----- Dickson1989_MayoPBCOriginalArticle.pdf

• Data
  • pbc-data.txt --- Primary Billiary Cirrhosis (PBC) survival data
  • pbcseq R documentation --- Note that the pbc-data-seq.txt file has been modified for this tutorial

• Required R Packages
  • survival -- survival analysis (Cox regression, Kaplan-Meier)
  • survivalROC -- methods for cumulative/dynamic time-dependent ROC

• Required R Functions
  • WeightedMeanRank.zip -- Functions needed for Weighted Mean Rank Estimator

• Exercises
  • [1] Input the data.
  • [2] Use Cox regression to obtain a risk score that uses log(bili), log(protime), edema, albumin, age. The below exercises will be conducted on this 5-covariate risk score.

  • Cumulative Cases / Dynamic Controls
    Using survivalROC,
    [3] Estimate AUCs for ROC curves corresponding to cases defined cumulatively from
    • t = 0 to 1 year
    • t = 0 to 2 years
    • t = 0 to 5 years
    
    [4] Plot the above three ROC curves.

  • Sequential Cumulative Cases / Dynamic Controls
    [5] Estimate the AUCs for cases defined cumulatively over 1-year windows starting from
    • t = 0
    • t = 2 years
    • t = 5 years

  • Incident Cases / Dynamic Controls.
    Using the Weighted Mean Rank method,
    [6] Plot the time-varying AUC of the risk score.
    [7] Plot the time-varying true positive fraction (or ROC value) for a fixed false positive fraction of 10%.
    [8] Estimate the AUCs at times t = 6 months, 2 years, 5 years.
    [9] Estimate the C-index to summarize the time-varying performance represented in [6] above.

  • Example R Files
    • run_me_tutorial1.r

  • Example Results
    • [4] pbc-survivalROC-plot.ps
    • [6] pbc-weightedMeanRank-plot-AUC.pdf
    • [7] pbc-weightedMeanRank-plot-ROC.pdf