0

Are binomial GLMs better than nonlinear regression for model-fitting, and predicting ED50s and other effective dose point intercepts?

In toxicology it is typical to run an experiment with a concentration-response design. For example, various concentrations of a toxicant as the independent variable and the response (perhaps survivorship) as the dependent variable. To model the response, often nonlinear regression (such as a four-parameter logistic curve), or a binomial GLM is used. EC values (also called LD, IC, ED) can easily be extracted from any of these models, but my question is if one model produces a more reliable fit for predicting EC values. My experience with GLMs is that they sometimes don't uphold the S-shape well.

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
Tom
  • 216
  • 1
  • 9
  • 2
    I think this really needs more information e.g. about the nature of the data gathered. Are you talking about how the [effective dose](https://en.wikipedia.org/wiki/Effective_dose_(pharmacology)) might be measured, or rather about how it might be predicted on the basis of other information? – Silverfish Apr 28 '16 at 11:39
  • Sure. In toxicology it is typical to run an experiment with a concentration-response design. For example, various concentrations of a toxicant as the independent variable and the response (perhaps survivorship) as the dependant variable. To model the response, often nonlinear regression (such as a four-parameter logistics curve), or binomial GLMs are used. EC values (also called LD, IC, ED) can easily by extracted from the models, but my questions is if one model produces a more reliable fit for predicting EC values. My experience with GLMs is that they sometimes don't uphold the S-shape well. – Tom Apr 30 '16 at 00:45
  • Thanks - I suggest you edit this context in to your question body. There's an "edit" button at the bottom of your post. – Silverfish Apr 30 '16 at 01:19
  • Thanks. I didn't see that. The post is updated. – Tom May 01 '16 at 02:50
  • 2
    It's not an either-or situation: GLMs can handle nonlinear responses directly (as in the example at http://stats.stackexchange.com/a/64039) or broadly and flexibly, such as by incorporating splines within the regressors. The question of "better" will surely depend on what kind of "nonlinear regression" you have in mind and how well it models the dose-response *variability* as well as the dose-response curve. – whuber May 01 '16 at 17:35
  • Thanks whuber. The answer you gave in your example is great, and I am getting much better fits, both in the curve and the variability, by adding nonlinear components into the model. Now I just need to find a reference to justify it. – Tom Feb 05 '17 at 22:59

0 Answers0