In cases where the two reviewers disagreed in the scoring of any ADR, they met to reach an agreement, or the case was referred to a senior researcher for review. To meet the objective of the study,GW-572016 a sequential series of logistic regression models was developed using ADR as the dependent variable. We used a combination of medical literature guidance and forward selection method to determine variable selection. At the univariate analysis, variables with a probability value of P,0.05 were entered in the multivariate logistic regression analysis. To prevent model over fitting, the maximum number of variables entered in the multivariate regression models was one variable for every eight ADR events. All selected variables were tested for multicollinearity to avoid any strong correlation between the variables. The presence of collinearity was examined by the evaluation of variance inflation factors and magnitude of standard errors. Variables with more than AB1010 missing values were not included in the analysis. All other missing data were imputed using the multiple imputation technique with 5 imputations. Models were developed in accordance with the chronology in which patient data are available in clinical practice. Models were consecutively extended with data from patient demographic, physical examination, comorbid conditions, laboratory test and medications used during hospital stay. In each model, variables with probability value of P.0.10 were excluded from further analysis. All models were adjusted for age, sex and estimated GFR. Results from the multivariate logistic regression were expressed in terms of the odd ratio for a particular variable with accompanying 95% confidence interval. Improvement in model performance was tested using measures for calibration and discrimination. For each model, the calibration was measured using the Hosmer and Lemeshow goodness of fit test. Calibration determines the differences between observed and predicted outcomes for groups of patients, with better model having smaller differences between predicted and observed outcomes. The discriminatory power of each model was assessed using the concordance statistics.