Data in Environmental Science and Eco(toxico-)logy

Quantitative Ecotoxicology, Page 159, Example 4.6, Control Mortality

This is example 4.6 on page 159 of Quantitative Ecotoxicology. It is about how to deal with control mortalities.

First we need the data:

The data consists of number of dead animals (DEAD) from all animals (TOTAL) exposed to different concentrations (CONC).
First we create a new column with the proportion of dead animals:

Here is a plot of the data. Note the use of expression() (greek letters in the axis labels).

Control mortality

We can estimate the mean control mortality and the confidence interval for the mean using the t.test function:

These can be also easily extracted from the t.test object:

This gives nearly the same values as in the book. I don’t know what the SAS option OPTC is doing or computing, however it seems like it is the mean +- CI for the control group.

Abbott’s formula

We could adjust for control mortality using Abbott’s formula:

with $p_c$ = the corrected, p = original and $p_0$ = control mortality.

The mean control mortality can be calculated as:

And the corrected mortalities using Abbotts formula as:

Dose-Response-Models

Ignoring control mortality

As in the previous example we can fit a dose-response-model to this data using the drc package:

Using the corrected mortalities

We can also fit a model to the corrected mortalities PROP_c.

Abbotts correction resulted to some negative mortalities, therefore I set the control and all negative mortalities to zero:

Then we fit a dose-response model:

A model without fixed lower limit

The two-parameter log-logistic model from above (mod1) performs quite good. However its lower limit is fixed to 0 and the upper limit to 1.
Since we have a small amount of control mortality we could check if a model with varying lower limit (will be estimated) makes sense.

Let’s fit a three parameter log-logistic function, where the lower limit is an additional parameter:

However looking at the summary we see that the lower limit (c) is basically zero.

Since the lower limit (=control mortality) is so low we could also stick with the simpler model:

suggest that there is no statistically significant better fit of the more complicated model.

All three considered models give nearly the same $LC_{50}$ around 2100: