15289-s-bos

124 | Chapter 8 in the mass spectrometer. Peptide peak area ratio analysis was performed using Skyline software. Statistical analysis Categoricalvariablesweredescribedusingpercentages.Thedistributionalcharacteristics of continuous variables were examined using tests for skewness and kurtosis. 21 Many were not normally distributed so all were described using the median and inter-quartile range (IQR). Univariable analysis of the association between the peptides and disease group was based upon the non-parametric Kruskal-Wallis test, and the Mann-Whitney test was used to examine dichotomous covariates. Correlation between continuous variables was assessed using the non-parametric Spearman rank-order correlation coefficient. Inflation of the critical significance level by multiple comparisons was addressed using the sequential rejection modification of the Bonferroni method developed independently by Holm and Simes. 22, 23 Subsequent inferential analysis was based upon the underlying model of the disease process as a sequence of increasingly severe stages, which are irreversible, viz. normal, calcification, CAD. The appropriate statistical model is the continuation ratio regressionmodel 24 (OCR) which is a variation of the Cox proportional hazards model for discrete ordinal outcome data. 25 The results are provided as hazard ratios (HR) with 95% confidence intervals (95% CI). Covariates that had the potential to confound the analysis were also examined using the OCR model. In order to preserve statistical power and in keeping with the development of methods for analysis of observational data, a proximity score was estimated for each peptide following the recommendations of Little andRubin 26 andStuart. 27 Estimationof covariate adjusted HRs used the proximity score as a single covariate. Analysis using the receiver operating characteristic (ROC) curve was used to estimate the predictive accuracy of proteins that showed an association with disease progression. The methods developed by Pepe 28 were used to estimate the area under the ROC curve (AUROC) to compare the peptides with each other, and to examine alternative methods for combining the peptide results. Robust bootstrapped estimation of the standard errors was used to avoid overfitting. The same methods were used to estimate covariate adjusted AUROCs. It became apparent during the analysis that the association between the potential biomarkers and disease progression was negative and so an inverse transformation was used to prepare the ROC curves, resulting in axes which are reversed from the usual form. Because of the semi-continuous nature of the non-negative ‘clumping at zero’ measures of the proteins, analysis of the association between measures of CAD