Covariates for adjustment were selected a priori either because they represented important demographic variables (age, gender, and race) or because prior reports3,10 indicated an association with mortality in IPF (age, smoking status, baseline FVC, Dlco, BAL fluid neutrophil percentage, and alveolar-arterial oxygen gradient). Before study inclusion, Pearson product moment correlations between continuous covariates were evaluated to avoid colinearity. None had correlation coefficients > 0.60. We evaluated for multiplicative interactions on the basis of smoking status, which was selected as a candidate effect modifier a priori on the basis of prior research3,8,19 and treatment status. Standard methods do not exist for deriving receiver operating characteristic (ROC) curves for time-to-event data24; thus, we used logistic regression for this part of the analyses. Complementary analyses included a series of logistic regression models to predict 1-year mortality, using odds ratios and 95% confidence intervals (CIs) to estimate relative risk. Our regression models sequentially included the simple clinical, clinical and SP-A and SP-D, and clinical and biomarker (including SP-A, SP-D and BAL fluid neutrophil percentage) models, with evaluation of the discriminatory capability of the models using the C statistic, or the area under the ROC curve (AROC). Two-tailed p values < 0.05 were considered statistically significant. Analyses were performed with a statistical software package (Stata, version 9; StataCorp; College Station, TX).