Hypothesis Testing with Left-Censored Data : A Comparison of Naive and Censoring-Aware Methods for Two-Group and Multiple-Group Settings

Sammanfattning

This thesis investigates hypothesis testing for left-censored data, where observations below a limit of detection (LoD) are only partially observed. Through simulation studies for two-group and multiple-group comparisons, conventional substitution methods are compared with censoring-aware methods implemented in R. The results show that the common strategy of replacing censored observations with the LoD or LoD/2 can introduce systematic bias, leading to distorted p-values, inflated Type I error, and unreliable power. In contrast, censoring-aware methods generally provide more stable inference, especially under low to moderate censoring. The findings support the use of statistical methods specifically designed for censored data rather than simple substitution approaches.