Meja zaznavnosti v bioloških poskusih: primerjava statističnih pristopov

The focus of this thesis is statistical treatment of datasets containing missing data due to limit of detection. Our motivation comes from secondary data analysis of the KARISMA trial, which investigates the effect of different tamoxifen doses on blood hormone concentrations. Commonly used approaches for handling values below the limit of detection in linear regression are described and compared, including substitution, complete case analysis, missingness indicators, Tobit regression, accelerated failure time model, and single and multiple imputation. Methods are compared on simulated data under log-normal distribution, with approximately half of the values fa ling below the limit of detection in independent and/or dependent variables. Methods suited for the scenario where the independent and the dependent variable have values fa ling below the limit of detection are then applied to estradiol blood hormone concentration data from the KARISMA trial to assess how the choice of method influences estimated effects in a real-world se ting. The results show that accelerated failure time and Tobit models outperform substitution and complete case analysis when the dependent variable has values below limit of detection, whereas for the independent variable with values below limit of detection, the missingness indicator approach offers a compromise between bias and efficiency, even though the complete case analysis is less biased. When data is missing both in independent and dependent variables, a combined approach gives best results.