There is hardly a mention of censoring. This is unfortunate, as censored data, especially rightcensored observations, are common in regulatory risk assessment and there is not yet a general agreement on how, or even whether, such data should be analyzed for SSD purposes. This issue is mentioned here for only one software package included in the list of software, namely Mosaic_SSD. Among the references that discuss this issue are Brock et al (2011), EFSA (2013), Aldenberg (2015) and Green et al (2018).
- Aldenberg T 2015. Some Species Sensitivity Distribution Statistics Revisited – A Governmental Perspective. Environmental Toxicology and Chemistry 34: 2442-2444.
- Brock TCM, Arts GHP, Ten Hulscher TEM, De Jong FMW, Luttik R, Roex EWM, Smit CE and Van Vliet PJM 2011. Aquatic effect assessment for plant protection products: a Dutch proposal that addresses the requirements of the Plant Protection Product Regulation and
- Cade BS 2015. Model averaging and muddled multimodel inferences. Ecology 96: 2370-2382. https://doi.org/10.1890/14-1639.1
- Green JW, Springer TA, Holbech H 2018. Statistical Analysis of Ecotoxicity Studies. Wiley, Chapter 12
RESOLUTION
See companion Issue#38.
Also, note Aldenberg (2015) does not discuss censored data.
RESOLUTION
· Response to reviewer:
We are a little perplexed by the statement “There is hardly a mention of censoring”. The section “Left-tail weighting of the SSD” is devoted to the use of censoring as a means of increasing the weight in the left tail of the SSD. However, we appreciate that this is perhaps not the type of censoring the reviewer had in mind. The other kind of censoring is the <DL type that is a consequence of the measurement process. This type of data is frequently encountered in ecotox. and it is well accommodated for in existing software packages including MOSAIC, ssdtools, and SSD Toolbox. Consistent with the stated aims and scope of the MS, we do not feel any additional discussion is warranted.
I would just like to point out that ssdtools handles censored data correctly (if the distributions have the same number of parameters because the sample size is irrelevant for comparing model weights) – it’s on the to do list to enforce this condition in ssdtools.
Also see Wheeler and Bailer paper that I also put in another comment.
Wheeler and Bailer (2009) don’t mention censoring.
See the paper by Nysen et al (2016) that I refer to in a separate comment who show how this can be done and is used.
If we want to include censored endpoints, we could do it via maximum likelihood but it might be easier to use Bayesian methods. See Angelina and my manscript on using Bayesian methods.
Aldenberg (2015) does not discuss censored data.
We agree that censoring is an important topic, however it is not a focal point of the MS. Our mention and use of censoring has been in the context of the stated and discussed problem of left-tail weighting not censoring per se. One of the strategies for increasing the weight of observations in the left-tail of the SSD is to censor some fraction of observations in the right tail of the SSD. It was never our intent to have a fuller discourse on the treatment of censored data. While the concepts are related, there’s a subtle distinction between how to incorporate and use censored toxicity data in SSD fitting and the process of deliberately censoring some data for left-tail weighting.
My view is we maybe add some text in the MS to this effect.
Finally – I disagree with the assertion “there is not yet a general agreement on how, or even whether, such data should be analyzed for SSD purposes”. As we point out, MOSAIC and SSD Tollbox do handle censoreed data and they do that in a statistically legitimate way.