In semiconductor industries, highly reliable semiconductor devices are of paramount importance, especially for safety applications. Therefore, a reliable classification of single devices into good ones and bad ones is needed. Filtering bad devices is possible by testing the electrical functionality and imposing specification limits. In contrast to this, good devices might contain a risk, not detectable at that moment but which makes them unreliable in their long-term performance. They are called risk devices or Mavericks. An estimation regarding this risk is mandatory to prevent the delivery of unreliable devices. It is expected that risk devices, compared to good ones, are outliers, visible in at least one test parameter. A screening method applied to this parameter reveals these statistical outliers. Unfortunately, with new technologies and the miniaturization of the devices it is expected that common screening methods on the given test parameters are no longer capable of distinguishing between good and risk devices. Therefore, improved screening is necessary. This can be achieved in two ways. One way is to introduce a data transformation, called Independent Component Analysis (ICA). A second way is to develop a new screening method. The challenge of using ICA, which originally comes from signal processing, is the adaptation of this method to the landscape of semiconductor devices. In this thesis it is shown that increased classification accuracy is achieved when using the ICA transformed data for subsequent screening instead of the original test parameters. Apart from ICA a new screening method is developed. Screening among semiconductor manufacturers is mainly performed using Part Average Testing (PAT). Based on the distribution of the test parameters, which is assumed to be Gaussian, upper and lower PAT limits are calculated. Devices outside these limits are scrapped. Another method to detect outliers, using hypothesis testing, is introduced in this thesis. Therefore, devices are iteratively deleted until the null-hypothesis of a Gaussian distribution is accepted. To decide which devices to delete, the test statistic is optimized, outlier by outlier. In this thesis two novel test statistics are developed and compared to commonly known ones. To summarize, this thesis provides novel techniques which help semiconductor manufacturers to sell reliable devices.