The aim of the present study was to evaluate empirically confusion matrices in device validation. We compared the confusion matrix method to linear regression and error indices in the validation of a device measuring feeding behaviour of dairy cattle. In addition, we studied how to extract additional information on classification errors with confusion probabilities. The data consisted of 12 h behaviour measurements from five dairy cows; feeding and other behaviour were detected simultaneously with a device and from video recordings. The resulting 216 000 pairs of classifications were used to construct confusion matrices and calculate performance measures. In addition, hourly durations of each behaviour were calculated and the accuracy of measurements was evaluated with linear regression and error indices. All three validation methods agreed when the behaviour was detected very accurately or inaccurately. Otherwise, in the intermediate cases, the confusion matrix method and error indices produced relatively concordant results, but the linear regression method often disagreed with them. Our study supports the use of confusion matrix analysis in validation since it is robust to any data distribution and type of relationship, it makes a stringent evaluation of validity, and it offers extra information on the type and sources of errors.

The possibility of an association of early pregnancy loss (EPL) with residential exposure to ELF magnetic fields was investigated in a case-control study. Eighty-nine cases and 102 controls were obtained from the data of an earlier study aimed at investigating the occurrence of EPL in a group of women attempting to get pregnant. Magnetic-field exposure was characterized by measurements in residences. Strong magnetic fields were measured more often in case than in control residences. In an analysis based on fields measured at the front door, a cutoff score of 0.5 A/m (0.63 microT) resulted in an odds ratio of 5.1 (95% confidence interval 1.0-25). The results should be interpreted cautiously due to the small number of highly exposed subjects and other limitations of the data.