Question about weighting of data points and model selection

[johannanyf]

Hi!

I have two independent question related to this graph:

1. For me, this fit does not look very good. It seems like the one data point at the second highest concentration is weighed way too much. I would have fit a constant model to that data. This made me wonder how the data points are weighed. It seems to me like for each dose modeling is performed on the average value.

2. I run the same set of data with a different BMR level (10% vs 3 SD) and I got different model selection for that particular "gene"&chemical. Does this mean the model selection depends on my BMR level?

Cheers, Johanna

[auerbachs]

Hi Johanna,

Thanks for the questions . In regard to question 1, BMDExpress 2 does not have a constant model implemented, hence it could not be selected. Implementation of a constant model might be consideration for future releases. The best way to avoid problems like this is to apply a Williams' trend test to the data before performing BMD analysis. Applying the trend test would likely remove this feature as it does not appear to exhibit a dose related trend. In regard to question 2, it is likely that when you change the BMR value (increased to 3SD) the BMD, BMDL or BMDU value was non-convergent (ie -9999) on the original model selected at the 1SD BMR level. Because of non-convergences the software then selected the next best fit model where all values converged.

Hope this helps!

Scott

[johannanyf]

Hi Scott

Thank you for the answers.

To 1: Yes you are right, filtering with Anova or trend test would have get rid of this. I was just shoked by how much one outlier can drive the curve up. But may be it's not the outlier but the fact that no better model was available.

To 2: So the short summary is: Yes, the model selection is influenced by the BMR value. This was just surprising to us, but I understand it now.