Using heterogeneity in the population structure of U.S. swine farms to compare transmission models for porcine epidemic diarrhoea

Eamon B. O’Dea; Harry Snelson; Shweta Bansal

Most of our predictors were correlated with other predictors as well as with the total positive accessions in each state. For such data, fitting regression models with an elastic net penalty allows groups of correlated variables to be given similar effect sizes whereas other modelling approaches, such as stepwise approaches and the use of a lasso penalty, may lead to one variable in a correlated group being singled out and being given a too-large effect size. 33 In general for elastic net regression, the weight given to the penalty determines whether any variable is selected. Often, the goal of a regression analysis is to obtain a model with good predictive performance and the weight is chosen by cross validation. 33 By contrast, we have no need of a predictive model and are instead more interested in determining what variables are important to include in a model. Stability selection 34 provides a general method of identifying relevant variables. The main idea is to select variables that across many random subsamples of the data are selected with high probability by the elastic net with a given set of weights for the penalty. We use this procedure because it is less likely to select noise variables than is cross validation. 34 Further details are given in the Supplementary Note.