Version log e-book "Modelling survival under chemical stress"
(and list of known errors in the current version, and planned updates for the next version)
Version 1.0 (Date 18 January 2018)
Known errors that will be corrected, and updates that will be included, in the next version:
- On the second page, the correct reference to the book should include the publisher: Jager T and R Ashauer (2018). Modelling survival under chemical stress. A comprehensive guide to the GUTS framework. Version 1.0. Toxicodynamics Ltd., York, UK. Available from Leanpub: https://leanpub.com/guts_book.
- Oops! We forgot to thank Anna-Maija Nyman for pointing out relevant REACH documents for Chapter 6.
- In Table 7.1, k_k must be b_w.
- Expanded the slow and fast kinetics section in the extensions (Section A.7.1). Included slow kinetics for IT and for cases where the exposure concentration is not constant. We will include a new identifiability problem (single-dose runaway) and add examples for these cases.
- The extensions will also be extended with more analytical solutions for damage (e.g., a solution that can be used when exposure changes linearly over time) and for LCx,t and LPx in some cases.
- For the GUTS-RED-IT and GUTS-IT special cases (Eq. 3.20 and 3.22), the background mortality was not included. For the IT models, the background survival probability is calculated separately (Sb = exp(-hb*t)) and multiplied with the chemical-induced survival probability (assuming they are independent causes of death).
- The manner in which the likelihood is defined in various textbooks and papers differs a bit, as does the notation. We think the notation we selected for the book is clear, used consistently, and thus should not lead to problems. However, we will add a few clarifying notes in Section 3.5.1 (and a footnote with link in Section 2.3.1).
- In a frequentist framework, we can estimate the joint-confidence region for the model parameters just as in a Bayesian framework. However, confidence intervals/regions have a different meaning in both frameworks. It is possible to use the frequentist joint-confidence region for the parameters for forward predictions (as is commonly done for Bayes), and thus to make confidence intervals on model predictions. However, when we propagate the 95% joint interval for the parameters (as suggested on Page 34), we end up with a coverage much larger than 95% for a 4-parameter GUTS model (more like 99.5%). In the meantime, we have found out that we need to propagate the parameter sets within the chi2 criterion for df=1 (thanks to a pivotal paper by Kreutz et al, 2012). This will be explained in detail in an update of the book. For the case studies, we only showed the Bayesian perspective, so this finding has no impact. In several places, reference to forward propagation in the frequentist perspective is made, and the text will be revised. This affects for example the text on Page 34 (Section 2.3.4), 77 (Section 4.5), 82 (Section 4.5.2), and some of the ring test results (this method is used by BYOM and Mathematica).
- Connected to the previous points, we have located papers that deal more extensively with identifiability of model parameters and its consequences. These will be added were relevant, and will used to provide some more focus in the book. Identifiability is already touched upon in the book, for example, with the first panel in Fig. 2.10 and Section 2.4.3. We need to stress that identifiability problems, in a Bayesian context, will lead to an improper posterior (that cannot be integrated), unless the posterior is constrained by proper informative priors. This situation occurs in our dieldrin case study for IT (see Fig. 4.11), which implies that all CIs derived from this sample are unreliable. We already point at this issue on Page 75, but this needs to be made stronger. Furthermore, it also occurs for the SD case, as background mortality goes to zero. This is less obvious as we set a minimum for hb of 1e-4 (a uniform prior); without it, the MCMC analysis would not have converged.
- Recently, Andreas Focks and co-workers have published a paper with validation results for GUTS: fitting on data for constant exposure and predicting time-varying exposure (DOI 10.1007/s10646-018-1940-6). This paper will be incorporated in Section 8.6.
- The EFSA opinion on TKTD models has appeared, and we will need to update some parts of the book to reflect the progress made by the EFSA working group.
- ...