Comparing Posteriors: Estimating Practical Differences Between Models

Is not the first time Max Kuhn appears in this blog and this time with a lecture (in the last New York R Conference) about advices to estimate what is the best model based on R statistics. Sure we can get good advices  to find the best model possible for our data sets.

2018 New York R Conference Highlights

On April 20 this year the New York R Conference has been celebrated with a great  success.
Just look to the great atmosphere in the video of the conference.

Checking the slopes in Validation Groups

This is an study to develop calibrations for meat in a reflectance instrument from 1100 to 1650 nm. Normally meat are measured in transmitance but this is an approach to do it in reflectance.

 

I have just 64 samples with fat laboratory data. I split the spectra into 4 sets of 16 samples and merge 3 of leaving the other three for external validation. So I have 48 samples for training and 16 for validation and I can develop four calibrations and validate with 4 external sets.

 

Considering that we have few samples are in the training set, I have to use few terms. The SEPs for external or Cross Validation are quite high , but the idea here is to see the changes in the slope for the four validation sets.

 

The reason is that we have few samples and the slope value will stabilize as soon as more samples are included into the calibration and validation sets.

 

 

 

To improve the SEP we have to check the sample presentation method for this product and the procedure to obtain the laboratory reference method.

Monitoring the performance with the histogram

NIR can be used to detect levels food additives and check if they are in the right limits.

In this cases there are several types of doughs, and they use two levels of additive concentration depending on the type. So we have always the same reference data.

A calibration is developed and we have new data to validate. NIR will give  results which I expect to be covering the reference value with a Gauss distribution.

Using the Monitor function I can see the prediction distribution vs. the reference distribution and check if the expectations are fine.

 

In the case of the higher concentration is fine, and in the lower concentration is skewed (that is why the S/I adjust is suggested).This can be a first approach to continue with this application with mor accurate reference values.