Skip to main content
Fig. 5 | BMC Medical Research Methodology

Fig. 5

From: To tune or not to tune, a case study of ridge logistic regression in small or sparse datasets

Fig. 5

Boxplots showing distribution of calibration slopes over 1000 generated datasets in scenarios with the expected value of Y, E(Y) = 0.1, the number of predictors K = 5, noise absent or present, the sample size of N {100, 250, 500, 1000} considering A) moderate (a = 0.5) and B) strong (a = 1) predictors. Datasets where calibration slopes were larger than 5 are not shown. The whiskers extend no more than 1.5-times the interquartile range from the box. EPV, indicated in the top right, denotes the events per variable ratio. Further results from other scenarios are contained in Table S8 and S9. OP, prediction oracle; D, deviance; GCV, generalized cross-validation; CE, classification error; RCV50, repeated 10-fold cross-validated deviance with θ = 0.5; RCV95, repeated 10-fold cross-validated deviance with θ = 0.95; AIC, Akaike’s information criterion; IP, shrinkage based on informative priors; WP, shrinkage based on weakly informative priors; FLIC, Firth’s logistic regression with intercept-correction

Back to article page