192559_en-US

liliconcom

Effect in a DSD having issues with LOD runs

Hi,

 

I'll have a discussion with a customer soon on the resulting effect in a DSD experiment, when some runs will have an lod issue (limit of detection). I know GenReg can handle this censored situation, however how does it affect the power and predictive performance of the DSD havin 1 or even more runs with lod?

 

Any experience or rule of thumb would be highly appreciated. I thought about checking the Design power through evaluate design after the runs have been executed. However, evaluate design does not take into account the DSD structure, so I'm not sure if this is misleading anyhow, and even more with LOD in place. 

 

Thanks,

Martin

Re: Effect in a DSD having issues with LOD runs

Hi Martin,

It will negatively affect power. Fundamentally, there is less information in a

But I don't think there are any rules of thumb about this. And determining the power will be very difficult. I don't think the power simulation capabilties in JMP Pro could be used in this situation.

Having said all that, analysing in Gen Reg with the

Cheers,

Phil

Re: Effect in a DSD having issues with LOD runs

Just to add to what Phil said. Take a pragmatic approach: learn what you can from the data collected and if needed augment your data to learn more (which might be adding replicates to increase confidence or adding runs to explore a wider space or higher order model or something else). Treat learning from data as a continual (iterative) process not a one-shot thing.

Re: Effect in a DSD having issues with LOD runs

Hey Martin,

 

Here is a discovery talk on the subject

 

https://community.jmp.com/t5/Discovery-Summit-2014/Using-JMP-to-Analyze-Data-From-a-Designed-Experiment-With-Some/ta-p/21864

 

This came from my team at US Synthetic. It has video and .pdf report. No data were made available as it was fairly sensitive at the time even anonymized. I can give some additional background if you find it helpful. 

 

J

Re: Effect in a DSD having issues with LOD runs

A quick add on to Jason's link. Normal Distribution was added to Parametric Survival (with some other distributions, even All Distributions) in a version later than what was used for the Discovery Talk. Also, one can now use censored data for Generalize Regression, so that may provide an additional modeling option..

 

I like the solution in the paper. Treating as censored data makes sense to for below level of detection  (now that I think of it and after reading the discover paper to remind me).

 

hopefully that will help if the "Sweet Spot" is near level of detection.