Tools ModelRobustness

From Eigenvector Research Documentation Wiki
Revision as of 08:43, 29 July 2010 by imported>Jeremy
Jump to navigation Jump to search

Table of Contents | Previous | Next | Index

Model Robustness Tool

You use the Model Robustness tool to measure the sensitivity of a regression model to artifacts in new spectroscopic measurements. To open the Model Robustness tool, on the Analysis window, click Tools > Model Robustness, and then click Shifts or Interferences.

Shifts

The Shifts option measures the sensitivity of a regression model to shifts in x-axis data that are caused by instrument instability-that is, if you have an instrument that is not particularly stable or reproducible over time, what is the impact on predictions using the given model? The Shifts plot is a three-dimensional plot that details the RMSEP (Root Mean Squared Error of Prediction) for the model as a function of shift, where shift is described in terms of the number of variables and the Smoothing window.

Example of a Shifts plot

Tools ModelRobustness.29.1.1.jpg


Consider the figure above, which shows the model robustness for a regression model with an RMSEC (Root Mean Squared Error of Calibration) of approximately 0.5. As shown in this figure:

  • 1 - Without shift and without smoothing of the variables, the RMSEP indicates that you have test data that is identical to your calibration data and you have performance that is on par for the RMSEC of the model
  • 2 - Shifting a spectrum over by simply one variable increases the RMSEP for the model by almost twelve orders of magnitude, from 0.5 to almost 60.
  • 3 - With a combination of shifting and smoothing, the impact on the model is lessened somewhat.

Interferences

The Interferences option measures the sensitivity of a regression model to the location and width of a new peak in test data-that is, if you have a chemical entity that is present in the test data but that was not reflected in the calibration data, what is the impact on predictions using the given model? The Interferences plot is a three-dimensional plot that details the RMSEP (Root Mean Squared Error of Prediction) for the model as a function of a new peak, where the peak is described in terms of its width and location.

Example of an Interferences plot

Tools ModelRobustness.29.1.2.jpg


Consider the figure above, which shows the model robustness for a regression model with an RMSEC (Root Mean Squared Error of Calibration) of approximately 0.5. As shown in this figure, the RMSEP for the model can be impacted in one of three ways:

  • 1 - An interferant area where there is virtually no impact on the RMSEP for the model, no matter how wide the interfering peak is.
  • 2 - An interferant area where there is a slight impact on the RMSEP for the model, but the impact is lessened as the width of the peak increases.
  • 3 - An interference area where there is a significant impact on the RMSEP for the model, but the impact is lessened as the width of the peak increases.