Dose response and lethal dose 50 analysis I am going to perform an experiment to test for the pathogenicity of several bacterial strains. For this I will infect several animals (e.g. 5 per dose per bacterial strain) with increasing doses of bacteria (e.g. 10^5, 10^6, etc) and then obtain a binary response variable, which basically is whether the animal survived or died after exposure to a given bacterial strain/dose combination. I will then determine the dose response curves and their respective lethal dose 50 (LD50). 
Before I start doing this experiment I have been trying to understand what is the best method to analyse this data. I have been thinking that the best approach for me would be to use a binomial GLM, however I have started to read a bit more of the ecotoxicology literature and found the R package drc (by Christian Ritz & Jens C. Streibig) that implements several models for non-linear regression. Given that my knowledge of nonlinear regression is low, I am now wondering which of the two methods would be better to analyse the data (drc or glm), so any help on making a decision would be greatly appreciated.
 A: Both analyses (non-linear regression and binomial GLM) have their advantages and disadvantages. They will produce quite similar LD50s if the data aren't variable but can produce wildly different results if the data are quite variable or if the bottom of the curve is not found (e.g. 100% lethality in the highest treatment). 
I'd recommend trying both to see what makes the most sense with your data. 
Personally I find nonlinear regression gives a more realistic fit of the data given the experiment is a controlled lab-based assay, and therefore can better meet some of the requirements such as normality of residuals, homogeneity of variances, minimal deviation from the model (i.e run/replicate tests), reasonably balanced sample units. Usually at least 5 treatments levels are needed for model convergence. Appropriate constraints are also needed. In addition to the drc package, a very easy to use program is Graphpad Prism which requires no coding. I'd recommend using it's free online Help-section to better understand non-linear regression.
On the other hand, binomial GLMs (logistic regressions) need to meet fewer criteria and may be stronger for an unbalanced designs and smaller sample sizes, utilising the underlying probability distribution (the binomial) to inform estimates. Random variation, which is common in biological data, can also be incorporated using a binomial GLMM. Also they also don't need artificial constraints like non-linear regression often do. 
Hope this helps. 
A: Let's assume non-linear regression means non-linear least squares which is a common way of measuring pharmacokinetic curves. That is the intent of the package "drc" which has been renamed "drm".
However this is not the right methodology to measure fatality.
Pharmacokinetic curves predicts the concentration of drug, or it's conjugates or metabolites, in the right medium (blood, urine) as a function of time and/or subject characteristics. They typically require multiple measurements over time, multiple administrations of drug (possibly at several levels), and the resulting curves may be "fed forward" in a heirarchy or network to predict, e.g. symptom resolution, tumor response, or death.
If you are administering one-and-done dosing with no in-vivo markers of drug activity like repeated biospecimen sampling, a typical failure analysis will be the right way to go. You should plot the survival curves of the species using a stratified Kaplan-Meier curve and test either with a parametric survival or Cox-model / log-rank test. These confer more power than a simple logistic regression because we additionally consider the time until death.
