Introduction Bovine TB control programs have substantially reduced or nearly eradicated the disease from farm animals in many industrialized countries. However, bTB is still widespread in Africa, Central and South America, parts of Asia and some Middle East countries. In Spain, herd prevalence has been substantially reduced: from 11.1% in 1986 to 1.3% in 2012. However, in the last years, the decline has only been moderate: from 1.6% in 2007 to 1.3% in 2012. In the South-Central Spain, just recently, conducted risk factor study has identified that previous bTB history, herd size, extensive production system and a high number of fenced big game estates in the neighborhood of the farm was related with bTB persistence. Although most of the published bTB risk factors studies conducted in different countries did not discriminate between transient or persistent infections. The analysis of the causes of persistent infections has received little attention, and differences on the factors that determine both situations might exist. Methods A case-control study on beef farms matched by herd size and geographical location was conducted The model was built following Dohoo et al., 2003 Bivariate analysis between the outcome (i.e., bTB persistent infection vs transient infection) and different predictor variables using a liberal p-value (we used p<0.30). Categorical variables were screened using 2 test, and continuous variables with ANOVA or Kruskal-Wal-lis test. Bartlett's test for inequality of variances was applied to choose between both methods. In the case of non-homogeneity of variances the Kruskal-Wallis test was used. Evaluation of correlations among predictor variables: for those variables associated with the probability of persistence with a p-value lower than 0.30, we calculated the Spearman correlation coefficient, in case of correlation between them (i.e., higher than 0.5), the variable with higher biological signification was retained. Before building the multivariable logistic model and in order to avoid problems derived from non-linear relationships, quantitative variables were reclassified into four categories following their quar-tile distribution. A manual model-building selection was conducted for the development of the multivariable logistic regression model:, as a first step we compared all the possible models with just one variable by the Akaike Information Criteria (AIC) value. To the model with the lowest AIC value and one predictor we included all the remaining covariates and compared them based on the AIC value. This process was repeated until we got the model with lowest AIC. This was considered as the most plausible one, and selected as the final model. Confounding was assessed by monitoring changes on model parameters when adding new variables. If substantial changes where observed on the regression coefficients this was indicative of confusion. Biologically meaningful interactions were included to the final model and retained if the AIC value was reduced. To test the ability of the model to discriminate between cases and controls, we calculated a Receiver Operating Characteristic (ROC) curve, and the area under the curve (AUC). An AUC value greater than 0.8 and between 0.7 and 0.8 were considered as good and moderate discriminate capacities, respectively.