Hi Tristan. Thanks for your participation in the webinar and your questions. First, a point of clarification: In response to your question during the webinar about mixtures, I said that SSDs were usually used to derive HC/PC values for single toxicants. While this is the usual application of SSDs, I do acknowledge that they are also used for deriving GVs for mixtures of chemicals (eg. whole effluent toxicity testing) in which case ‘concentration’ is in fact a dilution and thus bounded between 0 and 100%. Although there’s nothing to stop you from fitting an SSD to this proportion data, there is a disconnect between the fact that the SSD models we use are invariably on the range (o,Inf). We (the SSD tools development team) had some discussion around this point following the webinar in which we discussed the use of the beta distribution as an obvious candidate for data measured as %dilution rather than an absolute concentration. Joe Thorley suggested that ssdtools was not currently set up to handle such bounded data. My suggestion, which we’ll pursue further, and requires no modification to exisiting software, is to transform the %diltion data first using: Y = X/(1 – X) where X is the diltion data (on the scale 0 to 1). The resulting Y values are then on the range (0, Inf) and thus conformable with ‘regular’ concnetration data and SSD models. Once the SSD model(s) have been fitted and inference made on Y, we back-transform using X = Y/(1 + Y) to make inference on the original scale.
The second point I wish to make with respect to SSD modelling for mixtures of toxicants is the implicit, but very strong assumption that the chemical composition of the effluent is spatially and temporally invariant. This is a big ask and almost certainly untrue. How do effectively deal with this situation (short of repeating the SSD modelling at regular times and places) is an open question and one we perhaps should also turn our attention to.
Finally, with respect to a ‘turnkey’ software replacement for CETIS/Toxcalc – we’re currently working on new developments in this space – for example, the R package Bayesnec developed by a team including Rebecca and myself, is publicly available on the CRAN website. Furthermore, we recently published a paper in ET&C which introduced the No Significant Effect Concentration (NSEC) and a further publication in IEAM will appear soon. Our challenge now is to integrate these and other aspects of contemporary CR modelling under one unified system/package and, like ssdtools, develop a user-friendly shiny web deployment. This is all food for thought and obviously someone needs to bankroll the task!
Cheers,
David
