Exact definition of "Specificity of Response" in the multivariate case: Severe consequences for today's practice of PLS/PCR

Ralf Marbach; Pekka Teppola

Exact definition of "Specificity of Response" in the multivariate case: Severe consequences for today's practice of PLS/PCR

Ralf Marbach, Pekka Teppola

BA1404 Optical measurements

Research output: Contribution to conference › Conference article › Scientific

Abstract

A breakthrough in scientific understanding about multivariate calibration & measurement has recently created the method of science-based calibration ("SBC"). Multivariate measurements are not fundamentally different from the univariate case; rather, the two cases correspond one-to-one in every aspect. The significant advantages provided by SBC to the calibration of process analyzers will be shortly reviewed. Following this, the focus will be on the conclusions that can be drawn when applying the scientific understanding generated by SBC to the existing methods of calibration. The results are concerning. A fair estimate is that several ten per cent of PLS and PCR calibrations are not purely measuring what people think they are measuring. (A peer-reviewed publication about these findings is in preparation.) The correct mathematical definitions of sensitivity and specificity are given. Both limits are testable from first principles, i.e., from measurable pieces of spectroscopic data. Applying the limits post factum to today's "statistical" calibration practice reveals why many PLS/PCR results are affected by unspecific correlations (which, unlike spurious correlations, do not disappear when applying PLS or PCR to larger and larger calibration data sets). An overhaul of practices is required. The importance of applying spectroscopic expertise and application knowledge to the calibration is stressed, as is the need to scientifically estimate both the "spectral signal" and the "spectral noise" parts required in calibration. Both estimates are important because proof of specificity is a two-step process. First, the user must prove that the multivariate measurement measures "the right thing;" and then, second, that this "correct" measurement is not affected by any unspecific correlations. Current practices of testing for specificity (ASTM 1655 etc.) are shown to be insufficient and misleading in the case of PLS/PCR

Original language	English
Publication status	Published - 2008
MoE publication type	Not Eligible
Event	1st European Conference on Process Analytics and Control Technology, EuroPACT 2008 - Frankfurt am Main, Germany Duration: 22 Apr 2008 → 25 Apr 2008

Conference

Conference	1st European Conference on Process Analytics and Control Technology, EuroPACT 2008
Country	Germany
City	Frankfurt am Main
Period	22/04/08 → 25/04/08

Fingerprint Dive into the research topics of 'Exact definition of "Specificity of Response" in the multivariate case: Severe consequences for today's practice of PLS/PCR'. Together they form a unique fingerprint.

Cite this

@conference{6c0876e957e141e496ab318134a2d4ae,

title = "Exact definition of {"}Specificity of Response{"} in the multivariate case: Severe consequences for today's practice of PLS/PCR",

abstract = "A breakthrough in scientific understanding about multivariate calibration & measurement has recently created the method of science-based calibration ({"}SBC{"}). Multivariate measurements are not fundamentally different from the univariate case; rather, the two cases correspond one-to-one in every aspect. The significant advantages provided by SBC to the calibration of process analyzers will be shortly reviewed. Following this, the focus will be on the conclusions that can be drawn when applying the scientific understanding generated by SBC to the existing methods of calibration. The results are concerning. A fair estimate is that several ten per cent of PLS and PCR calibrations are not purely measuring what people think they are measuring. (A peer-reviewed publication about these findings is in preparation.) The correct mathematical definitions of sensitivity and specificity are given. Both limits are testable from first principles, i.e., from measurable pieces of spectroscopic data. Applying the limits post factum to today's {"}statistical{"} calibration practice reveals why many PLS/PCR results are affected by unspecific correlations (which, unlike spurious correlations, do not disappear when applying PLS or PCR to larger and larger calibration data sets). An overhaul of practices is required. The importance of applying spectroscopic expertise and application knowledge to the calibration is stressed, as is the need to scientifically estimate both the {"}spectral signal{"} and the {"}spectral noise{"} parts required in calibration. Both estimates are important because proof of specificity is a two-step process. First, the user must prove that the multivariate measurement measures {"}the right thing;{"} and then, second, that this {"}correct{"} measurement is not affected by any unspecific correlations. Current practices of testing for specificity (ASTM 1655 etc.) are shown to be insufficient and misleading in the case of PLS/PCR",

author = "Ralf Marbach and Pekka Teppola",

year = "2008",

language = "English",

note = "1st European Conference on Process Analytics and Control Technology, EuroPACT 2008 ; Conference date: 22-04-2008 Through 25-04-2008",

}

Exact definition of "Specificity of Response" in the multivariate case : Severe consequences for today's practice of PLS/PCR. / Marbach, Ralf; Teppola, Pekka.

2008. Paper presented at 1st European Conference on Process Analytics and Control Technology, EuroPACT 2008, Frankfurt am Main, Germany.

Research output: Contribution to conference › Conference article › Scientific

TY - CONF

T1 - Exact definition of "Specificity of Response" in the multivariate case

T2 - 1st European Conference on Process Analytics and Control Technology, EuroPACT 2008

AU - Marbach, Ralf

AU - Teppola, Pekka

PY - 2008

Y1 - 2008

N2 - A breakthrough in scientific understanding about multivariate calibration & measurement has recently created the method of science-based calibration ("SBC"). Multivariate measurements are not fundamentally different from the univariate case; rather, the two cases correspond one-to-one in every aspect. The significant advantages provided by SBC to the calibration of process analyzers will be shortly reviewed. Following this, the focus will be on the conclusions that can be drawn when applying the scientific understanding generated by SBC to the existing methods of calibration. The results are concerning. A fair estimate is that several ten per cent of PLS and PCR calibrations are not purely measuring what people think they are measuring. (A peer-reviewed publication about these findings is in preparation.) The correct mathematical definitions of sensitivity and specificity are given. Both limits are testable from first principles, i.e., from measurable pieces of spectroscopic data. Applying the limits post factum to today's "statistical" calibration practice reveals why many PLS/PCR results are affected by unspecific correlations (which, unlike spurious correlations, do not disappear when applying PLS or PCR to larger and larger calibration data sets). An overhaul of practices is required. The importance of applying spectroscopic expertise and application knowledge to the calibration is stressed, as is the need to scientifically estimate both the "spectral signal" and the "spectral noise" parts required in calibration. Both estimates are important because proof of specificity is a two-step process. First, the user must prove that the multivariate measurement measures "the right thing;" and then, second, that this "correct" measurement is not affected by any unspecific correlations. Current practices of testing for specificity (ASTM 1655 etc.) are shown to be insufficient and misleading in the case of PLS/PCR

AB - A breakthrough in scientific understanding about multivariate calibration & measurement has recently created the method of science-based calibration ("SBC"). Multivariate measurements are not fundamentally different from the univariate case; rather, the two cases correspond one-to-one in every aspect. The significant advantages provided by SBC to the calibration of process analyzers will be shortly reviewed. Following this, the focus will be on the conclusions that can be drawn when applying the scientific understanding generated by SBC to the existing methods of calibration. The results are concerning. A fair estimate is that several ten per cent of PLS and PCR calibrations are not purely measuring what people think they are measuring. (A peer-reviewed publication about these findings is in preparation.) The correct mathematical definitions of sensitivity and specificity are given. Both limits are testable from first principles, i.e., from measurable pieces of spectroscopic data. Applying the limits post factum to today's "statistical" calibration practice reveals why many PLS/PCR results are affected by unspecific correlations (which, unlike spurious correlations, do not disappear when applying PLS or PCR to larger and larger calibration data sets). An overhaul of practices is required. The importance of applying spectroscopic expertise and application knowledge to the calibration is stressed, as is the need to scientifically estimate both the "spectral signal" and the "spectral noise" parts required in calibration. Both estimates are important because proof of specificity is a two-step process. First, the user must prove that the multivariate measurement measures "the right thing;" and then, second, that this "correct" measurement is not affected by any unspecific correlations. Current practices of testing for specificity (ASTM 1655 etc.) are shown to be insufficient and misleading in the case of PLS/PCR

M3 - Conference article

Y2 - 22 April 2008 through 25 April 2008

ER -