Location
Logan, UT
Start Date
5-17-2022 1:00 PM
Description
In this paper, a simulation study was conducted to assess whether it is ideal to address the issue of non-detects in data using a traditional substitution approach for non-detects, imputation, or a non-imputation based approach. Simulated data used were simple nested designs motivated by a real-life data in a study of bumble bee activity in a commercial cherry orchard by Kuivila et al. (2021). The simulated data were generated at different thresholds or censoring levels and at different effect sizes. For each simulated data, seven popular existing techniques to handle non-detects were applied: (i) Zero substitution, (ii) Substitution with half Limit of Detection (LOD/2), (iii) Substitution with LOD/√ 2, (iv) Multiple Imputation (MI), (v) Regression on Order Statistics (ROS) (Imputation approach), and (vi) Maximum Likelihood Estimation (MLE) (likelihood estimation approach) and (vii) Kaplan-Meier (KM). Multiple Imputation (MI) was not applicable as the design of the simulated data violated the assumption of having a multivariate distribution. By comparative analysis of the simulated data, substituting with LOD/2 seemed appropriate for the design simulated, as it outperformed the other techniques (i.e ROS, MLE, KM, LOD/√ 2, and zero substitution) by yielding a lower Type I error, lower bias, and a better power across increasing effect sizes.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Included in
Handling Non-detects with Imputation in a Nested Design: A Simulation Study
Logan, UT
In this paper, a simulation study was conducted to assess whether it is ideal to address the issue of non-detects in data using a traditional substitution approach for non-detects, imputation, or a non-imputation based approach. Simulated data used were simple nested designs motivated by a real-life data in a study of bumble bee activity in a commercial cherry orchard by Kuivila et al. (2021). The simulated data were generated at different thresholds or censoring levels and at different effect sizes. For each simulated data, seven popular existing techniques to handle non-detects were applied: (i) Zero substitution, (ii) Substitution with half Limit of Detection (LOD/2), (iii) Substitution with LOD/√ 2, (iv) Multiple Imputation (MI), (v) Regression on Order Statistics (ROS) (Imputation approach), and (vi) Maximum Likelihood Estimation (MLE) (likelihood estimation approach) and (vii) Kaplan-Meier (KM). Multiple Imputation (MI) was not applicable as the design of the simulated data violated the assumption of having a multivariate distribution. By comparative analysis of the simulated data, substituting with LOD/2 seemed appropriate for the design simulated, as it outperformed the other techniques (i.e ROS, MLE, KM, LOD/√ 2, and zero substitution) by yielding a lower Type I error, lower bias, and a better power across increasing effect sizes.