Professional Documents
Culture Documents
Comprehensive Chemometrics Chemical and Biochemical Data Analysis 2Nd Edition Steven Brown Editor Full Chapter PDF
Comprehensive Chemometrics Chemical and Biochemical Data Analysis 2Nd Edition Steven Brown Editor Full Chapter PDF
https://ebookmass.com/product/statistics-and-data-analysis-for-
nursing-research-2nd-edition-2nd-edition/
https://ebookmass.com/product/advanced-data-analysis-modelling-
in-chemical-engineering-1st-edition-denis-constales/
https://ebookmass.com/product/data-science-with-rust-a-
comprehensive-guide-data-analysis-machine-learning-data-
visualization-more-van-der-post/
https://ebookmass.com/product/introduction-to-chemical-processes-
principles-analysis-synthesis-2nd-edition-regina-murphy/
Chemometrics in Spectroscopy (Second Edition) Howard
Mark
https://ebookmass.com/product/chemometrics-in-spectroscopy-
second-edition-howard-mark/
https://ebookmass.com/product/biochemical-pathways-an-atlas-of-
biochemistry-and-molecular-biology-2nd-edition/
https://ebookmass.com/product/data-assimilation-for-the-
geosciences-from-theory-to-application-2nd-edition-steven-j-
fletcher/
https://ebookmass.com/product/advances-in-oil-water-separation-a-
complete-guide-for-physical-chemical-and-biochemical-processes-
papita-das/
https://ebookmass.com/product/guidelines-for-revalidating-a-
process-hazard-analysis-2nd-edition-ccps-center-for-chemical-
process-safety/
COMPREHENSIVE
CHEMOMETRICS: CHEMICAL
AND BIOCHEMICAL
DATA ANALYSIS
SECOND EDITION
This page intentionally left blank
COMPREHENSIVE
CHEMOMETRICS: CHEMICAL
AND BIOCHEMICAL
DATA ANALYSIS
SECOND EDITION
EDITORS IN CHIEF
Steven Brown
Department of Chemistry and Biochemistry
University of Michigan
USA
Romà Tauler
Department of Environmental Chemistry
Institute of Environmental Assessment and Water Research (IDÆA)
SPANISH COUNCIL OF SCIENTIFIC RESEARCH (CSIC)
Spain
Beata Walczak
Department of Analytical Chemistry
Institute of Chemistry
Silesian University
Poland
VOLUME 1
Elsevier
Radarweg 29, PO Box 211, 1000 AE Amsterdam, Netherlands
The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom
50 Hampshire Street, 5th Floor, Cambridge MA 02139, United States
No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including
photocopying, recording, or any information storage and retrieval system, without permission in writing from the publisher. Details on
how to seek permission, further information about the Publisher’s permissions policies and our arrangements with organizations such as the
Copyright Clearance Center and the Copyright Licensing Agency, can be found at our website: www.elsevier.com/permissions.
This book and the individual contributions contained in it are protected under copyright by the Publisher (other than as may be noted
herein).
Notices
Knowledge and best practice in this field are constantly changing. As new research and experience broaden our understanding, changes in
research methods, professional practices, or medical treatment may become necessary.
Practitioners and researchers may always rely on their own experience and knowledge in evaluating and using any information, methods,
compounds, or experiments described herein. In using such information or methods they should be mindful of their own safety and the
safety of others, including parties for whom they have a professional responsibility.
To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume any liability for any injury and/or
damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods,
products, instructions, or ideas contained in the material herein.
ISBN 978-0-444-64165-6
Steven Brown obtained PhD in analytical chemistry from the University of Washington in 1978,
where he worked with Bruce Kowalski. That year he was appointed Assistant Professor at the University
of California, Berkeley, with a joint appointment at Lawrence Berkeley Laboratory. In 1981, he moved
to Washington State University, and in 1986 to the Department of Chemistry and Biochemistry at the
University of Delaware, where he is presently Willis F. Harrington Professor.
He has served as a Section President of the American Chemical Society from 1982–1984, as Chair
of the Department of Chemistry and Biochemistry at the University of Delaware from 1997–2002, and
as President of the North American Chapter of the International Chemometrics Society from 1986–
1988. He has served on the Editorial Staff of the Journal of Chemometrics since its inception, first as
a founding Editor and then as Editor in Chief before stepping down as Editor in Chief in 2007. He
now serves on the journal’s editorial board.
His research interests concern a wide range of problems in chemometrics and machine learning,
with over 300 publications in the scientific literature. He has edited several books, including the first
edition of the four-volume set Comprehensive Chemometrics, published by Elsevier in 2009. A focus of
his research has been on the development of new instrumental methods through use of multivariate
mathematical methods for multicomponent analysis, including calibration transfer, and the novel
use of data fusion methods. He was winner of the first EAS Award in Chemometrics in 1996 and the Kowalski Award in 2015. His work
has had applications in biomedical analysis, food science, plant science, forensic science, pharmaceutical characterization, and process
chemistry.
Romà Tauler (Barcelona, Spain, 1955) is research professor at IDAEA-CSIC. He obtained his doctorate
in Analytical Chemistry at the University of Barcelona in 1977. He was Assistant and Associate Professor
of the University of Barcelona (1978–2002) and has been CSIC Research Full Professor since 2003. He
is Chief Editor of Journal of Chemometrics and Intelligent Laboratory Systems and of the Comprehensive
Chemometrics major reference work. He has achieved an Award for Achievements in Chemometrics,
Eastern Analytical Symposium, and in 2009 the Kowalski Prize from J. of Chemometrics, Wiley, 2009.
He served as the President of the Catalan Chemistry Society, 2008–2013, and was the recipient of
the Eu-ERC Advanced Grant award Nr. 320337, 2013–2018, CHEMAGEB project. Romà Tauler
has published more than 430 research publications and 60 book chapters, with more than 16,000
citations and h-índex 63. His main research field is Chemometrics and its applications to
Environmental Chemistry, Omic Sciences, and Bioanalytical Chemistry.
v
vi Editors in Chief
Prof. Beata Walczak graduated in chemistry from the Faculty of Mathematics, Physics and Chemistry,
Silesian University, Katowice, Poland, in 1979. Since then she has been working in the Institute of
Chemistry, Silesian University, where now she is the Head of the Department of Analytical Chemistry.
Meanwhile, she stayed as a postdoc at the University of Orleans (France) and at the Graz University of
Technology (Austria). She also held a post of a visiting professor at Vrije Universitiet Brussel
(Belgium), at Rome Univeristy “La Sapienza” (Italy), at AgroParisTech University (France), at the
University of Modena and Reggio Emilia(Italy), and at Radboud University (The Netherlands).
From the early 1990s she has been involved in chemometrics, and her main scientific interests are
in all aspects of data exploration and modeling (dealing with missing and censored data, dealing with
outliers, data representativity, enhancement of instrumental signals, signal warping, data compression,
linear and nonlinear projections, development of modeling approaches, feature selection techniques,
etc.). She authored and coauthored ca. 170 scientific papers, 400 conference papers, and delivered
many invited lectures at the numerous international chemistry meetings. She acted as an editor and
coauthor of the book Wavelets in Chemistry, Vol. 22, in the series “Data Handling in Science and Tech-
nology,” Elsevier, Amsterdam, 2000, and as an coeditor of four-volume Comprehensive Chemometrics,
Elsevier, Amsterdam, 2009. Currently she acts as Editor of the journal Chemometrics and Intelligent
Laboratory Systems and of “Data Handling in Chemistry and Technology” (the Elsevier book series), and also as a member of the editorial
boards of Talanta, Analytical Letters, J. of Chemometrics, and Acta Chromatographica.
SECTION EDITORS
Richard Brereton did his BA, MA, PhD and postdoc in the University of Cambridge, after which time
he moved to the staff of the University of Bristol where he was successively Lecturer, Reader,
Professor and Emeritus. He is Fellow of the Royal Society of Chemistry, Royal Statistical Society
and Royal Society of Medicine and a Chartered Chemist. He is Editor-in-Chief of the journal Heritage
Science, columnist for J Chemometrics, and editorial board member for several journals. He has given
around 200 invited lectures in over 30 countries, been main / sole supervisor of 33 PhD and 6
research masters students, on around 50 conference organising committees, and acted as expert
witness in 13 court cases. He has published over 400 recorded articles, including 8 books, 209 papers
in Web of Knowledge, 20 book chapters, 32 book reviews, 9 conference papers, almost all as main or
sole author. His papers have been cited over 5000 times according to Web of Knowledge, and his
books over 3000 times according to Google Scholar. He has published the most cited paper in J Che-
mometrics over the last decade, and the eighth most cited paper in the Analyst since 2000.
Anna de Juan is an associate professor at the Department of Chemical Engineering and Analytical
Chemistry at the University of Barcelona since 2003, teaching chemometrics at undergraduate and
graduate levels. She holds a degree and PhD in Chemistry from the University of Barcelona (UB)
and her expertise is in Multivariate Curve Resolution (MCR) methods: theoretical development
and application to bioanalytical and analytical problems. Since 2002 she is member of the Editorial
Advisory Board of Chemometrics and Intelligent Laboratory systems and since 2006 of Analytica
Chimica Acta. Recently, it has acted as section editor for the reference work Comprehensive Chemo-
metrics, Elsevier (2009). She has stayed in the framework of research collaborations in Vrije Univer-
siteit Brussel, Brussels (1995), Virginia Commonwealth University, Richmond, US (1998), The
University of Newcastle, Australia (2002), Université des Sciences et Technologies de Lille, France
(2004) and the University of Dalhousie, Canada (2016). In 2004 she received the 4th Chemometrics
Elsevier Award together with Karl Booksh. She has published more than 130 papers in international
journals and books and has given more than 180 presentations in different international confer-
ences, 50 of them plenary or keynote lectures, basically on design of chemometric tools and multi-
variate curve resolution developments and related methods and on applications to process analysis,
hyperspectral image analysis and general analytical applications.
vii
viii Section editors
Rafael Cela Torrijos is a professor of analytical chemistry in the University of Santiago de Compos-
tela, Spain, and the Head of Laboratory for Analytical Chemistry in the Research institute of Chem-
ical and Biological Analyses (IAQBUS), in the same university. Previously, he was at the Universities
of Madrid (Complutense) and Cádiz, belonging to the group of analytical chemists that started the
development of chemometrics by the 1980s in Spain. His research has focused on the analytical
applications of separation science, and particularly, the development and optimization of sample
preparation techniques in chromatographic analysis, including experimental designs and develop-
ment of computer-assisted chromatographic methods, being the author of the Mchrom Scout soft-
ware, distributed by Mestrelab Research S.L. Currently, he is the author or co-author of more than
300 scientific papers and several textbooks.
Riccardo Leardi was born in Novi Ligure (Italy) on October 17, 1959.
In 1983 he graduated cum laude in Pharmaceutical Chemistry and Technology at the Faculty of
Pharmacy of the University of Genova..
His actual position is Associate Professor at the Department of Pharmacy of the School of
Medical and Pharmaceutical Sciences of the University of Genoa. In 2013 he got the qualification
for full professor in Analytical Chemistry.
Since 1985 he has been working in the section of Analytic Chemistry of the Department of Phar-
macy of the University of Genova, and his research field is Chemometrics.
His interests are mainly devoted to problems of classification and regression (applied especially
to food, environmental and clinical data), experimental design, process optimization, multivariate
process monitoring and multivariate quality control.
His original research focused mainly on genetic algorithms, especially in their application to the
problem of variable selection, and three-way methods.
He developed the chemometrical softwares CAT (Chemometric Agile Tool) and BasiCAT, both
freely downloadable from the site http://gruppochemiometria.it/index.php/software.
He is author of almost 150 papers and more than 130 communications in national and interna-
tional meetings, several of which as invited speaker; he was invited to give talks and courses in several
industries and research centers.
He organizes two schools of Chemometrics (Multivariate Analysis; Experimental Design), each
held twice a year at the University of Genoa.
Since November 2002 he started his activity of chemometric consultancy.
Prof. Federico Marini was born in Rome, Italy, in 1977 and he received his MSc (2000) and PhD
(2004) from Sapienza University of Rome. He is currently associate professor of Chemometrics at
Sapienza University of Rome. In 2006, he was awarded the Young Researcher Prize from Italian
Chemical Society and in 2012 he won the Chemometrics and Intelligent Laboratory Systems Award
“for his achievements in chemometrics”. He has been visiting researcher in various Universities
(Copenhagen, Stellenbosch, Silesia, Lille). His research activity is focused on all aspects of chemo-
metrics, ranging from the application of existing methods to real world problems in different fields
to the design and development of novel algorithms. He is author of more than 150 papers in inter-
national journals, and recently he edited and coauthored the book Chemometrics in food chemistry
(Elsevier). He is member of the Editorial boards of Chemolab, Analytica Chimica Acta, J. of Chemo-
metrics, J. of NIR Spectroscopy, J. of Spectral Imaging and he serves as Associate Editor for Chemo-
metrics in Wiley’s Encyclopedia of Analytical Chemistry. He is the past coordinator of the
Chemometric group of the Italian Chemical Society and the coordinator of the Chemometric study
group of EUCheMS.
Section editors ix
Alejandro C. Olivieri was born in Rosario, Argentina (07/28/1958). He obtained his B.Sc. from the
Catholic University (1982), and his Ph.D. from the National University of Rosario (1986). He did
Postdoctoral Research at the University of Illinois, Urbana-Champaign, USA. In 1990, he returned
to the University of Rosario, and joined the National Research Council (CONICET). He founded
a research group in chemometrics in analytical chemistry, publishing about 250 papers in interna-
tional journals, books and book chapters, and supervised nine Ph.D. theses. He received several
national and international awards, including the John Simon Guggenheim Memorial Foundation
fellowship (2001–2002). In 2018 he published “Introduction to multivariate calibration. A practical
approach”, a book which was selected by Choice, from the Association of College and Research
Libraries (ACRL) as one of the Outstanding Academic Titles for 2019.
Dr. William Rayens is Professor and the Dr. Bing Zhang Endowed Department Chair in the Depart-
ment of Statistics at the University of Kentucky. Rayens has an extensive research record focused
primarily on the development of multivariate and multi-way statistical methodologies, mostly
related to problems in chemistry and the neurosciences. He has mentored several Ph.D. students
and has been honored at both the College and the University level as an outstanding teacher. Rayens
also served as Assistant Provost for General Education during which time he was tasked with imple-
menting new general education reforms at the University of Kentucky, the first changes to that
program in almost 30 years. He designed the one-of-a-kind Technologically Enhanced Active
Learning rooms in the University’s multi-million dollar Jacobs Science Building. Dr. Rayens received
his Ph.D. in mathematics from Duke University in 1986.
Luis A. Sarabia received his Ph.D. in Statistics from de University of Valladolid (Spain) in 1979.
Since 1974, he has been teaching Statistics, Mathematics and Design of Experiments mostly to grad-
uate and postgraduate students of Chemistry. In 2000 he reached the position of professor at the
University of Burgos. He is currently director of the Department of Mathematics and Computation
and a member of the research group Chemometrics and Qualimetrics (officially recognized as
consolidated research group, UIC-237). He is author and co-author of around 140 papers, 7 chapter
of book and approximately 160 communications in international meetings. He was co-founder of
Colloquium Chemiometricum Mediterraneum. His research interest are in multivariate statistics,
n-way procedures, evolutionary algorithms for multiresponse optimization, design of experiments,
QbD and PAT with application to regulated chemical analysis, food safety, food characterization and
fraud detection.
This page intentionally left blank
IN MEMORY OF ROGER PHAN TAN LUU
“Science evolves by means of research. Scientific research is based on experiments (Galileo Galilei, Two new
sciences, 1638). With time experiments have becoming more and more complex. So scientists used “experi-
mental design (DOE, design of experiments)” as a useful tool. Only 40 years ago Roger Phan Tan Luu suggested
that DOE is the core of the Methodology of experimental research, much more than a tool, a philosophy.
The research work of Roger Phan Tan Luu includes new tools, the organization of design for complex
experiments, and a lot of applications. To perform research work on a real problem it is necessary to know very
well the problem. Obvious. When the expert of the problem asks for the help of the expert of design, the
problem expert must know just some base elements of the Methodology. The expert of Methodology instead
must study to have a deep knowledge of the problem, in much detail. In this way Roger accumulated knowl-
edge, i.e., experience, in many fields of research. The result of nearly 50 years of activity was a complete scientist,
from theory to new tools, to applications.
A complete scientist is always a teacher. Roger was a great teacher, not only in his Aix-Marseille University, but
throughout the world. He had the rare ability to capture the attention of student, transforming a boring heavy
sequence of theorems into a compelling, instructive show. For the students a pleasure, not a pain.
This description is that of the public man.
The private man was (outside the family) a loner. Only in the last 20 years of the past century did he make
contact in the organization of schools (Eurochemometrics, Erasmus stages) of chemometrics and experimental
design for some people of large affinity. With these people, he opened himself to friendship. As one of his
friends, I received from him very much. It was possible for me to appreciate his generosity, his humor, his speed
to answer to invitation (and of course to work together).
Friendship details are very intimate, not ready to be described. For friends the final separation is a very
hard continuous sentiment. Memories help. As a friend, as a common scientist, always I will remember.
Michele Forina
xi
This page intentionally left blank
CONTRIBUTORS TO VOLUME 1
xiii
xiv Contributors to Volume 1
Statistics
Quality of Analytical Measurements: Statistical Methods for Internal Validation
Proficiency Testing in Analytical Chemistry
Quality of Analytical Measurements: Univariate Regression
Robust and Nonparametric Statistical Methods
Bayesian Methodology in Statistics
Robust multivariate statistical methods
An Introduction to the Theory of Sampling: An Essential Part of Total Quality Management
Representative Sampling, Data Quality, Validation: A Necessary Trinity in Chemometrics
Experimental Design
Introduction: Experimental Designs
Screening Strategies
The Study of Experimental Factors
Response Surface Methodology
Experimental Design for Mixture Studies
Nonclassical Experimental Designs
Designing a multicomponent calibration experiment: basic principles and diagonal approach
Analysis of Variance
General Linear Models
Multiset Data Analysis: ANOVA Simultaneous Component Analysis and Related Methods
Regularized Manova
ANOVAdTP
Optimization
Constrained and Unconstrained Optimization
Sequential Optimization Methods
Optimization: Steepest Ascent, Steepest Descent, and Gradient Methods
Multicriteria Decision-Making Methods
Genetic Algorithms in Chemistry
A Guided Tour of Penalties
Particle Swarm Optimization
Linear Soft-Modeling
Linear Soft-Modeling: Introduction
Principal Component Analysis: Concept, Geometrical Interpretation, Mathematical Background, Algorithms, History, Practice
Principal Component Analysis
xv
xvi Subject Classification
Unsupervised Learning
Unsupervised Data Mining: Introduction
Data Mapping: Linear Methods versus Nonlinear Techniques
Tree-Based Clustering and Extensions
Model-Based Clustering
Common Clustering Algorithms
Density-Based Clustering Methods
Data Preprocessing
Preprocessing Methods
Evaluation of Preprocessing Methods
Model-based preprocessing in vibrational spectroscopy
Normalization and Closure
Variable Shift and Alignment
Background Estimation, Denoising, and Preprocessing
Denoising and Signal-to-Noise Ratio Enhancement: Classical Filtering
Denoising and Signal-to-Noise Ratio Enhancement: Derivatives
Denoising and Signal-to-Noise Ratio Enhancement: Splines
Data Quality and Denoising: a Review
Model-Based Preprocessing and Background Elimination: OSC, OPLS, and O2PLS
Applications
Chemometrics in Electrochemistry
Chemometrics in the Pharmaceutical Industry
Environmental Chemometrics
Resampling and Testing in Regression Models with Environmetrical Applications
Application of Chemometrics in the Food Sciences
Chemometrics in Forensics
Chemometric Analysis of Sensory Data
Smart Sensors
Statistical Control of Measures and Processes
Best Practice and Performance of Hardware in Process Analytical Technology (PAT)dA Prerequisite to Avoid
Pitfalls in Data Analytics
Multivariate Statistical Process Control and Process Control, Using Latent Variables
Batch Process Modeling and MSPC
Chemometrics in Raman spectroscopy
Chemometrics in NIR Hyperspectral ImagingdTheory and Applications in the Agricultural Crops and Products Sector
Mass Spectroscopic Imaging: Chemometric Data Analysis
Fast analysis, Processing, and Modeling of Hyperspectral Videos: Challenges and Possible Solutions
Image Processing
Chemometrics Analysis of Big Data
Systems Biology
Analysis of Metabolomics Data
Data Processing for RNA/DNA Sequencing
Analysis of Megavariate Data in Functional Omics
Spectral Map Analysis of Microarray Data
Chemometrics in Flow Cytometry
Chemometrics for QSAR Modeling
Chemoinformatics
High-Performance GRID Computing in Chemoinformatics
This page intentionally left blank
PREFACE
Some 50 years ago, the first publications appeared on the use of computer-aided mathematics to analyze
chemical data. With those publications, the modern field of chemometrics was launched. Both the speed and
power of computers and the sophistication of analytical instrumentation have made great leaps in the inter-
vening time. The ready availability of chemometric software, coupled with the increasing need for rigorous,
systematic examination of ever-larger and more sophisticated sets of measurements from instrumentation has
generated strong interest in reliable methods for converting the mountains of measurements into more
manageable piles of results, and for converting those results into nuggets of useful information. Interest in
application of chemometrics has spread well beyond chemists with a need to understand and interpret their
measurements; now chemometrics is helping to make important contributions in process engineering, in
systems biology, in environmental science, and other disciplines that rely on chemical instrumentation, to name
only a few areas.
In the 12 years since the first edition of this book appeared, there has been considerable change in the fields
of data science and of chemometrics. As applications of chemometrics continue to grow, so too does is the
methodology of chemometrics. After 50 years, chemometrics is a scientific field with mature areas, but it is also
a field where change continues to occur at a rapid pace, driven both by advances in chemical instrumentation
and measurement and by close connection of chemometrics with the data science, machine learning, statistics
and signal processing research communities. The interfacial location of chemometrics, falling between
measurements on the one side and statistical and computational theory and methods on the other, poses
a challenge to the new practitioner: gaining sufficient breadth and depth of understanding in data science and
learning in what ways data science connects with measurement chemistry, in order to use chemometrics
effectively.
The four volumes of Comprehensive Chemometrics, 2nd Ed. are the result of a meeting in Oxford in January,
2017, where the editors planned a revised work that would update most of the work covered in the first edition
and would cover emerging areas of chemometric research, while providing a sampling of current applications.
Our goal was to bring our reference work current with the advances in chemometrics that have occurred since
2006, with a treatment that would serve both the new and the experienced practitioner.
What has resulted from this collaboration is a resource that captures the practice of chemometrics now. The
four volumes in the revised work now include 119 chapters, with 33 new, 35 reprinted and 51 updated chapters,
making this the most wide-reaching and detailed overview of the field of chemometrics ever published.
Comprehensive Chemometrics 2nd Ed. offers depth and rigor to the new practitioner entering the field, and breadth
and varied perspectives on current literature to more experienced practitioners aiming to expand their horizons.
Software and datasets, both of which are especially valuable to those learning the methods, are provided in
some chapters. The coverage is not only comprehensive, it is authoritative as well; authors contributing to
Comprehensive Chemometrics 2nd Ed. are among the most distinguished practitioners of the field.
Comprehensive Chemometrics 2nd Ed. would not have been possible without the considerable help of the
Editorial Board, who assisted in selecting authors and reviewing chapters. For this edition, our Board included
Richard Brereton, Marina Cocchi, Anna De Juan, Riccardo Leardi, Roger Phan Tan Lu, Federico Marini, William
Rayens, Luis Sarabia, Alejandro Olivieri and Rafael Cela.
This new edition would not have been possible without the hard work of the staff at Elsevier. We also owe
thanks to Rachel Conway, Senior Acquisitions Editor at Elsevier, for supporting the project and seeing the
project off, to Sean Simms, who took over the task of Acquisitions Editor in 2019, to Dhivya Karunagaran and
Brinda Subramanian for their help in ensuring that submissions met the requirements for publication, and
xix
xx Preface
especially to Paula Davies, our Content Project Manager, for overseeing the entire project, keeping track of the
due dates and submissions, encouraging authors as needed, and helping us to keep to the production schedule.
Finally, we extend special thanks to all of our authors whose efforts have made the work the valuable reference
that it is.
Steven Brown
Romà Tauler
Beata Walczak
March, 2020
CONTENTS OF ALL VOLUMES
Editors in Chief v
Section Editors vii
In memory of Roger Phan Tan Luu xi
Contributors to Volume 1 xiii
Subject Classification xv
Preface xix
VOLUME 1
xxi
xxii Contents of All Volumes
VOLUME 2
VOLUME 3
VOLUME 4
Index 703
This page intentionally left blank
1.01 Quality of Analytical Measurements: Statistical Methods for
Internal Validationq
M Cruz Ortiz, Departamento de Química, Facultad de Ciencias, Universidad de Burgos, Burgos, Spain
Luis A Sarabia and M Sagrario Sánchez, Departamento de Matemáticas y Computación, Facultad de Ciencias, Universidad de
Burgos, Burgos, Spain
Ana Herrero, Departamento de Química, Facultad de Ciencias, Universidad de Burgos, Burgos, Spain
© 2020 Elsevier B.V. All rights reserved.
This is an update of M.C. Ortiz, L.A. Sarabia, M.S. Sánchez, A. Herrero, 1.02 - Quality of Analytical Measurements: Statistical Methods for Internal
Validation, in Comprehensive Chemometrics, edited by Steven D. Brown, Romá Tauler, Beata Walczak, Elsevier, 2009, https://doi.org/10.1016/B978-
044452701-1.00090-9.
1.01.1 Introduction 3
1.01.2 Confidence and Tolerance Intervals 7
1.01.2.1 Confidence Interval 8
1.01.2.2 Confidence Interval on the Mean of a Normal Distribution 9
1.01.2.2.1 Case 1: Known variance 9
1.01.2.2.2 Case 2: Unknown variance 10
1.01.2.3 Confidence Interval on the Variance of a Normal Distribution 10
1.01.2.4 Confidence Interval on the Difference in Two Means 11
1.01.2.4.1 Case 1: Known variances 11
1.01.2.4.2 Case 2: Unknown variances 11
1.01.2.4.3 Case 3: Confidence interval for paired samples 12
1.01.2.5 Confidence Interval on the Ratio of Variances of Two Normal Distributions 12
1.01.2.6 Confidence Interval on the Median 13
1.01.2.7 Joint Confidence Intervals 13
1.01.2.8 Tolerance Intervals 13
1.01.2.8.1 Case 1: b-content tolerance interval 13
1.01.2.8.2 Case 2: b-expectation tolerance interval 14
1.01.2.8.3 Case 3: Distribution free intervals 14
1.01.3 Hypothesis Tests 15
1.01.3.1 Elements of a Hypothesis Test 15
1.01.3.2 Hypothesis Test on the Mean of a Normal Distribution 19
1.01.3.2.1 Case 1: Known variance 19
1.01.3.2.2 Case 2: Unknown variance 19
1.01.3.2.3 Case 3: The paired t-test 19
1.01.3.3 Hypothesis Test on the Variance of a Normal Distribution 20
1.01.3.4 Hypothesis Test on the Difference in Two Means 20
1.01.3.4.1 Case 1: Known variances 20
1.01.3.4.2 Case 2: Unknown variances 21
1.01.3.5 Test Based on Intervals 22
1.01.3.6 Hypothesis Test on the Variances of Two Normal Distributions 23
1.01.3.7 Hypothesis Test on the Comparison of Several Independent Variances 24
1.01.3.7.1 Case 1: Cochran’s test 24
1.01.3.7.2 Case 2: Bartlett’s test 25
1.01.3.7.3 Case 3: Levene’s test 25
1.01.3.8 Goodness-of-Fit Tests: Normality Tests 26
1.01.3.8.1 Case 1: Chi-square test 26
1.01.3.8.2 Case 2: D’Agostino normality test 27
1.01.4 One-Way Analysis of Variance 28
1.01.4.1 The Fixed Effects Model 28
1.01.4.2 Power of the Fixed Effects ANOVA model 30
1.01.4.3 Uncertainty and Testing of the Estimated Parameters in the Fixed Effects Model 31
1.01.4.3.1 Case 1: Orthogonal contrasts 32
1.01.4.3.2 Case 2: Comparison of several means 32
1.01.4.4 The Random Effects Model 33
q
Change History: October 2019. M. Cruz Ortiz, Luis A. Sarabia, M. Sagrario Sánchez, Ana Herrero added MATLAB live-scripts for the computations; re-written
introduction to tolerance intervals; corrected estimates in Table 13; updated texts; corrected mistakes and updated references.
Nomenclature
1 L a Confidence level
1 L b Power
CCa Limit of decision
CCb Capability of detection
Fn1, n2 F distribution with n1 and n2 degrees of freedom (d.f.)
H0 Null hypothesis
H1 Alternative hypothesis
N(m,s) Normal distribution with mean m and standard deviation s
NID(m,s) (Normally and Independently Distributed) independent random variables equally distributed as normal with mean
m and standard deviation s
s Sample standard deviation
s2 Sample variance
tn Student’s t distribution with n degrees of freedom (d.f.)
x Sample mean
V(X) Variance of the random variable X
a Significance level, probability of type I error
b Probability of type II error
D Bias (systematic error)
3 Random error
m Mean
n Degree(s) of freedom, d.f.
s Standard deviation
s2 Variance
sR Reproducibility (as standard deviation)
sr Repeatability (as standard deviation)
cn2 c2 (chi-square) distribution with n degrees of freedom
Quality of Analytical Measurements: Statistical Methods for Internal Validation 3
1.01.1 Introduction
Every day millions of analytical determinations are made in thousands of laboratories all around the world. These measurements
are necessary for assessment of merchandise in the commercial interchanges, supporting health care, maintaining security, for
quality control of water and environment, for characterization of raw materials and manufactured products, and for forensic anal-
yses. Practically, any aspect of the contemporary social activity is somehow supported in the analytical measurements. The cost of
these measurements is high, but the cost of the decisions made based on incorrect results is much greater. For example, a test that
wrongly shows the presence of a forbidden substance in a food destined for human consumption can result in an expensive claim,
the confirmation of the presence of an abuse drug can lead to serious judicial sentences or doping in the sport practice may result in
severe sanctions. The importance of providing a correct result is evident but it is equally important to be able to prove that the result
is correct.
Once an analytical problem is posed to a laboratory and the analytical method is selected, the next step is the in-house validation
of the method. This is the process of defining the analytical requirements to respond to the problem and to confirm that the consid-
ered method has performance characteristics consistent with those required. The results of the validation experiments must be eval-
uated in order to ensure that the method meets the measurement required specification.
The set of operations to determine the value of an amount (measurand) suitably defined is called the measurement. The method
of measurement is the sequence of operations that is used when conducting the measurements. It is documented with enough
details so that the measurement may be done without additional information.
Once a method is designed or selected, it is necessary to evaluate its performance characteristics and to identify the factors that
can change these characteristics and to what extent they can change. If, in addition, the method is developed to solve a particular
analytical problem, it is necessary to verify that the method is fit for purpose.1 This process of evaluation is called validation of the
method. It implies the determination of several parameters that characterize the method performance: decision limit, capability of
detection, selectivity, specificity, ruggedness, and accuracy (trueness and precision). In any case, they are the measurements them-
selves which allow evaluation of the performance characteristics of the method and its fit for purpose. In addition, when using the
method, the obtained measurements are also the ones that will be used to make decisions on the analyzed sample, for example,
whether the amount of an analyte fulfills a legal specification. Therefore, it is necessary to suitably model the data that a method
provides. In what follows we will consider that the data provided by the analytical method are real numbers; other possibilities exist,
for example, the count of bacteria or impacts in a detector take only (discrete) natural values, or also, sometimes, the data resulting
from an analysis are qualitative, for example, the identification of an analyte through its m/z ratios in a mass spectrometry-
chromatography analysis.
With regard to the analytical measurement, it is admitted that the value, x, provided by the method of analysis consists of three
terms, the true value of the parameter m, a systematic error (bias) D, and a random error 3 with zero mean, in an additive way as
expressed in Eq. (1):
x¼mþDþ3 (1)
All the possible measurements that a method can provide when analyzing a sample constitute the population of the measure-
ments. This is indeed a theoretical situation because it is being assumed that there are infinitely many samples and that the method
of analysis remains unalterable. In these conditions, the model of the analytical method, Eq. (1), is mathematically a random vari-
able, X, with mathematical expectation m þ D and variance equal to the variance of 3; in statistical notation, E(X) ¼ m þ D and
V(X) ¼ V(3), respectively.
A random variable, and thus the analytical method, is described by its cumulative distribution function FX(x), that is, the prob-
ability that the method provides measurements less than or equal to x for any value x. Symbolically, this is written as FX(x) ¼ pr
{X x} for any real value x. In most of the applications, it is assumed that FX(x) is differentiable, which implies, among other
things, that the probability of obtaining exactly a specific value is zero. In the case of a differentiable cumulative distribution func-
tion, the derivative of FX(x) is the probability density function (pdf) fX(x). Any function f(x) such that it is positive, f(x) 0, and the
area under the function is 1, !Rf(x)dx ¼ 1, is the pdf of a random variable. The probability that the random variable X takes values in
the interval [a, b] is the area under the pdf over the interval [a, b], that is,
Z b
prfX ˛ ½a; bg ¼ f ðxÞ dx (2)
a
and the mean and variance of X are written as in Eqs. (3), (4), respectively
Z
Eð X Þ ¼ x f ðxÞdx (3)
R
Z
V ðXÞ ¼ ðx Eð X ÞÞ2 f ðxÞdx (4)
R
In general, mean and variance do not characterize in a unique way a random variable and therefore neither the method of anal-
ysis. Fig. 1 shows the pdf of four random variables with the same mean 6.00 and standard deviation 0.61.
4 Quality of Analytical Measurements: Statistical Methods for Internal Validation
1.2 1.2
(A) (B)
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
4 5 6 7 8 4 5 6 7 8
1.2 1.2
(C) (D)
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
4 5 6 7 8 4 5 6 7 8
Fig. 1 Probability density functions of four random variables with mean 6 and variance 0.375. (A) Uniform in [4.94, 7.06]; (B) Symmetric triangular
in [4.5, 7.5]; (C) Normal N(6, 0.61); (D) Weibull with shape 1.103 and scale 0.7 shifted to give a mean of 6. Dotted vertical lines mark the interval
[5.0, 7.0].
These four distributions, uniform or rectangular (Fig. 1A), triangular (Fig. 1B), normal (Fig. 1C), and Weibull (Fig. 1D), are
frequent in the scope of analytical determinations, and they appear in Appendix E of the EURACHEM/CITAC Guide1 and also
they are used in metrology.2
If the only available information regarding a quantity X is the lower limit, l, and the upper limit, u, but the quantity could be
anywhere in between, with no idea of whether any part of the range is more likely, then a rectangular distribution in the interval [l, u]
would be assigned to X. This is so because it is the pdf that maximizes the “information entropy” of Shannon, in other words the pdf
that adequately characterizes the incomplete knowledge about X. Frequently, in reference material, the certified concentration is
expressed in terms of a number and unqualified limits (e.g., 1000 2 mg L 1). In this case, a rectangular distribution should be
used (Fig. 1A).
When the available information concerning X includes the knowledge that values close to c (between l and u) are more likely
than those near the bounds, the adequate distribution is a triangular one (Fig. 1B), with the maximum of its pdf in c.
If a good location estimate, m, and a scale estimate, s, are the only information available regarding X, then, according to the prin-
ciple of maximum entropy, a normal probability distribution N(m,s) (Fig. 1C) would be assigned to X (remember that m and s may
have been obtained from repeated applications of a measurement method).
Finally, the Weibull distribution (Fig. 1D) is very versatile; it can mimic the behavior of other distributions such as the normal or
exponential. It is adequate for the analysis of reliability of processes, and in chemical analysis it is useful in describing the behavior
of the figures of merit of a long-term procedure. For example, the distribution of the capability of detection CCb3 is a Weibull one or
the distribution of the determinations of ammonia in water by UV-vis spectroscopy during 350 different days in Aldama.4
In the four cases given in Fig. 1, the probability of obtaining values between 5 and 7 has been computed with Eq. (2). For the
uniform distribution (Fig. 1A) this probability is 0.94, whereas for the triangular distribution (Fig. 1B) is 0.88, for the normal distri-
bution (Fig. 1C) is 0.90, and for the Weibull distribution (Fig. 1D), 0.93. Sorting in decreasing order of the proportion of values that
each distribution accumulates in the interval [5.0, 7.0] we have uniform, Weibull, normal, and triangular although the triangular
and normal distributions tend to give values symmetrically around the mean and the Weibull distribution does not. If another
interval is considered, say [5.4, 6.6], the distributions accumulate probabilities of 0.57, 0.64, 0.67, and 0.54, respectively, in which
Quality of Analytical Measurements: Statistical Methods for Internal Validation 5
Table 1 Values of b such that p ¼ pr{X < b} where X is each one of the random variables defined in the caption of Fig. 1.
Random variable
P Uniform Triangular Normal Weibull
n, Minimum b among the four distributions; m, Maximum b among the four distributions.
the difference among values is larger than before and, in addition, sorted the distributions as normal, triangular, uniform, and
Weibull.
If for each of those variables, value b is determined so that there is a fixed probability, p, of obtaining values below b (i.e., the
value b such that p ¼ pr{X < b} for each distribution X), the results of Table 1 are obtained. For example, second row, 5% of the
times the uniform distribution at hand gives values less than b ¼ 5.05, less than 4.97 if it is the triangular distribution, and so on. In
the table, the extreme values among the four distributions for each probability p have been identified, and large differences are
observed caused by the form in which the values far from 6 are distributed (notice the differences in Fig. 1 for the normal, the trian-
gular, or the uniform distribution) and also due to the asymmetry of the Weibull distribution.
Therefore, the mean and variance of a random variable give very limited information on the values provided by the random
variable, unless additional information is at hand about the form of its density (pdf). For example, if one knows that the distribu-
tion is uniform or symmetrical triangular or normal, the random variable is completely characterized by its mean and variance.
In practice, the pdf of a method of analysis is unknown. We only have a finite number, n, of measurements, which are the
outcomes obtained when applying repeatedly (n times) the same method to the same sample. These n measurements constitute
a statistical sample of the random variable X defined by the method of analysis.
Fig. 2 shows histograms of 100 results obtained when applying four methods of analysis, named A, B, C, and D, to aliquot parts
of a sample to determine an analyte. Clearly, the four methods behave differently.
From the experimental data, the (sample) mean and variance are computed as
Pn
x
x ¼ i¼1 i (5)
n
Pn
xÞ2
i¼1 ðxi
s2 ¼ (6)
n1
x and s2 are estimates of the mean and variance of the distribution of X. These estimates with the data in Fig. 2 are shown in Table 2.
According to the model of Eq. (1), Eð X Þ ¼ m þ Dxx, that is, the sample mean estimates the true value m plus the bias D.
Assuming that the true value is m ¼ 6 and subtracting it from the sample means in the first row of Table 2, the bias estimated
for methods A and B would be 0.66 and 0.16 for methods C and D. The bias of a method is one of its performance characteristics
and must be evaluated during the validation of the method. In fact, technical guides, for example, the one by the International Orga-
nization for Standardization (ISO), state that, for a method, better trueness means less bias. To estimate the bias, it is necessary to
have samples with known concentration m (e.g., certified material, spiked samples).
The value of the variance is independent of the true content, m, of the sample. For this reason, to estimate the variance, it is only
necessary to have replicated measurements on aliquot parts of the same sample. The second row of Table 2 shows that methods B
and C have the same variance, 1.26, which is 5 times greater than the one of methods A and D, 0.25. The dispersion of the data
obtained with a method is the precision of the method and constitutes another performance characteristic to be determined in
the validation of the method. In agreement with model in Eq. (1), a measure of the dispersion is the variance V(X), which is esti-
mated by means of s2.
In some occasions, for evaluating trueness and precision, it is more descriptive to use statistics other than mean and variance. For
example, when the distribution is rather asymmetric, as in Fig. 1D, it is more reasonable to use the median than the mean. The
median is the value in which the distribution accumulates 50% of the probability, 5.83 for the pdf in Fig. 1D and 6.00 for the other
three distributions, which are symmetric around their mean. In practice, it is frequent to see the presence of anomalous data
(outliers) that influence the mean and above all the variance, which is improperly increased; in these cases, it is advisable to use
robust estimates of central tendency and spread (dispersion).5–7 Details can be found in the chapter of the present book devoted
to robust procedures.
Fig. 2 and Table 2 show that the two characteristics of a measurement method, trueness and precision, are independent to one
another, in the sense that a method with better trueness (less bias), methods C and D, can be more, case D, or less, case C, precise.
Analogously, methods A and B have an appreciable bias but A is more precise than B. A method is said to be accurate when it is
precise and fulfills trueness.
6 Quality of Analytical Measurements: Statistical Methods for Internal Validation
40 40
(A) (B)
30 30
20 20
10 10
0 0
3 4 5 6 7 8 9 10 3 4 5 6 7 8 9 10
40 40
(C) (D)
30 30
20 20
10 10
0 0
3 4 5 6 7 8 9 10 3 4 5 6 7 8 9 10
Fig. 2 Frequency histograms of 100 measures obtained with four different analytical methods, named (A), (B), (C), and (D), on aliquot parts of
a sample. Dotted vertical lines mark the interval [5.0, 7.0].
Method
A B C D
Histograms are estimates of the pdf and allow evaluation of the performance of each method in a more detailed way than when
only considering trueness and precision. For example, the probability of obtaining values in any interval can be estimated with the
histogram. The third row in Table 2 shows the frequencies for the interval [5.0, 7.0]. Method D (best trueness and precision among
the four) provides 98% of the values in the interval, whereas method B (worst trueness and precision) provides only 56% of the
values in the interval. Nonetheless, trueness and precision should be jointly considered. See how according to data in Table 2
the effect of increasing the precision, using method A instead of B when the bias is “high” is an increase of 14% of results of the
measurement method to be in the interval [5.0, 7.0], whereas when the bias is small (C and D), there is an increase of 40%.
This behavior should be taken into account when optimizing a method and also in the ruggedness analysis, which is another perfor-
mance characteristic to be validated according to most of the guides. As can be seen in the fourth row of Table 2, if the method that
provides more results below 6 is needed, C would be the method selected.
Quality of Analytical Measurements: Statistical Methods for Internal Validation 7
The previous explanations show the usefulness of knowing the pdf of the results of a method of analysis. As in practice we have
only a limited number of results, two basic strategies are possible to estimate the pdf: (1) to assess that the experimental data are
compatible with a known distribution (e.g., normal) and then use the corresponding pdf; (2) to estimate the pdf by a data-driven
technique based on a computer-intensive method such as the kernel method8 or by using other methods such as adaptive or penal-
ized likelihood.9,10 The data of Fig. 2 can be adequately modeled by a normal distribution, according to normality hypothesis tests
whose details are explained later in Section “Goodness-of-Fit Tests: Normality Tests”. The fitted normal distributions are used to
compute the probabilities of obtaining values in the interval [5.0, 7.0] or less than 6, last two rows in Table 2. When comparing
these values with those computed with the empirical histograms (compare rows 3 and 5, and rows 4 and 6), there are no appreciable
differences and the normal pdf can be used instead.
In the validation of an analytical method and during its later use, statistical methodological strategies are needed to make deci-
sions from the available experimental data. The knowledge of these strategies supposes a way of thinking and acting that, subor-
dinated to the chemical knowledge, makes it objective both the analytical results and their comparison with those of other
researchers and/or other analytical methods.
Ultimately, a good method of analysis is a serious attempt to come close to the true value of the measurement, always unknown.
For this reason, the result of a measurement has to be accompanied by an evaluation of uncertainty or its degree of reliability. This is
done by means of a confidence interval. When the requirement is to establish the quality of an analytical method, then its capability
of detection, precision, etc. must be compared with those corresponding to other methods. This is formalized with a hypothesis test.
Confidence intervals and test of hypothesis are the basic tools in the validation of analytical methods.
In this introduction, the word sample has been used with two different meanings. Usually, there is no confusion because the
context allows one to distinguish whether it is a sample in the statistical or chemical sense.
In Chemistry, according to the International Union for Pure and Applied Chemistry (IUPAC) (Page 50 in Section 18.3.2 of
Inczédy et al.11), “sample” should be used only when it is a part of a selected material of a great amount of material. This meaning
coincides with that of a statistical sample and implies the existence of sampling error, that is, error caused by the fact that the sample
can be more or less representative of the amount in the material. For example, suppose that we want to measure the amount of
pesticide that remains in the ground of an arable land after a certain time. We take several samples “representative” of the ground
of the parcel (statistical sampling) and this introduces an uncertainty in the results characterized by a variance (theoretical) ss 2.
Afterward, the quantity of pesticide in each chemical sample is determined by an analytical method, which has its own uncertainty,
characterized by sm2, in such a way that the uncertainty in the quantity of pesticide in the parcel is ss 2 þ sm2 provided that the
method gives results independent of the location of the sample. Sometimes, when evaluating whether a method is adequate for
a task, the sampling error can be an important part of the uncertainty in the result and, of course, should be taken into account
to plan the experimentation.
When the sampling error is negligible, for example, when a portion is taken from a homogeneous solution, the IUPAC recom-
mends using words such as test portion, aliquot, or specimen.
In summary, there is a clear link between measurement method and a random variable which is why the probability is the
natural form of expressing experimental uncertainty. This is thus the focus of the present article that is organized as follows:
Section “Confidence and Tolerance Intervals” describes confidence intervals to measure bias and precision under the normality
hypothesis and tolerance intervals, useful in evaluating the fit for purpose of a method. Also, a nonparametric interval on the
median is described.
Section “Hypothesis Test” is devoted to making decisions based on experimental data that, as such, are affected by uncertainty. In
this section, the computation of the power of a test is systematically proposed as a key element to evaluate the quality of the decision
at the desired significance level. A brief incursion into tests based on intervals is also made as they solve the problem of deciding
whether an interval of values is acceptable, for example, a relative error less than 10% in absolute value. The section ends with some
goodness-of-fit tests to evaluate the compatibility of a theoretical probability distribution with some experimental data.
Section “One-Way Analysis of Variance” is dedicated to the analysis of variance (ANOVA) for both fixed and random effects, and
in Section “Statistical Inference and Validation” some more specific questions related to the usual parameters of the analytical
method validation and their relation with the developed statistical methodologies are analyzed.
Mathematical proofs are not covered in this article and, to be operative from a practical point of view, several examples have
been included so that the reader can verify the understanding of the formulas and the argumentation for their thoughtful use.
This aspect is completed with the inclusion of an Appendix where some essential aspects related to the effectiveness of the statistical
models and the limits laws are described. The Appendix also contains the necessary sentences, in MATLAB code, to repeat all the
calculations proposed along the article. The same sentences are also available as supplementary material in the form of MATLAB
.mlx live scripts (at least release R2016a is needed to read and execute them).
There are some important questions when evaluating a method, for example, “in a given sample, what is the maximum value that it
provides?” that, due to the random character of the results, cannot be answered with just a number.
In order to include the degree of certainty in the answer, the question should be reformulated as: What is the maximum value, U,
that will be obtained 95% of the times that the method is used in the sample? The answer to the question thus posed would be
8 Quality of Analytical Measurements: Statistical Methods for Internal Validation
Known Known
variance variances
Fig. 3 Diagram summarizing the different cases for computing confidence intervals.
a tolerance interval, and to build it the probability distribution must be known. For instance, let us suppose that it is a N(m,s) and
we denote by z0.05 the critical value of a N(0,1) ¼ Z distribution, the one that accumulates probability 0.95. Then, a possible answer
is U ¼ m þ z0.05s because then the probability that the analytical method gives values greater than U is pr{method > U} ¼ pr
{N(m,s) > m þ z0.05s}, which, according to the result in Appendix, is equal to pr{Z > z0.05} ¼ 0.05. In general, for any percentage
of results 100(1 a)%, the maximum value provided by the method would be
U ¼ m þ za s (7)
with a probability a that the aforementioned assertion is false.
If, instead, the interest was in the value L so that the 100(1 a)% of the results are greater than L, then the answer would be
L ¼ m za s (8)
Finally, the interval [L, U] that contains 100(1 a)% of the values obtained with the method would be
h i
½L; U ¼ m za=2 s; m þ za=2 s (9)
An analytical example where one of these tolerance intervals with a normal distribution N(m,s) needs to be computed would be:
An analytical method gives values (mg L 1) that follow a N(9, 0.5) distribution when measuring a standard with 9 mg L 1. To assess
whether the method is still properly working, tenstandards are
ffi included in the daily sequence of determinations. The probability
pffiffiffiffiffi
distribution
pffiffiffiffiffiffi of the mean of these ten values is a N 9; 0:5= 10 . Following Eq. (9), the tolerance interval at 95% level is 9 1:96
0:5= 10 ¼ 9 0.31 mg L 1. Consequently, if one day a mean of, say, 9.5 mg L 1 is obtained, the method does not work properly
because 9.5 does not belong to the tolerance interval and the method should be revised, at the risk of doing this revision uselessly
5% of the times. Notice that
ffi the tolerance interval is always the same, built at the desired confidence level 100(1 a)% with the
pffiffiffiffiffi
distribution N 9; 0:5= 10 and it is not updated daily with the new samples.
Different to Eq. (9), two variants of tolerance intervals, namely the b-content and the b-expectation tolerance intervals, are
explained in Section “Tolerance Intervals” due to their relevance in the context of validation of analytical methods. In any case,
any of them is completely different from the confidence intervals introduced and developed in the following sections (from Section
“Confidence Interval” to Section “Joint Confidence Intervals”).
After explaining all the studied cases, the section finishes with a comparative analysis of both concepts (tolerance and confidence
intervals).
In the general case, with a random variable X, obtaining a confidence interval for X from a sample x1, x2,., xn consists of obtain-
ing two functions l(x1, x2,., xn) and u(x1, x2,., xn) such that
prfX ˛ ½lug ¼ prfl X ug ¼ 1 a (10)
1 a is the confidence level and a is the significance level, meaning that the statement that the value of X is between l and u will be
false 100a% of the times.
In the next sections this idea will be particularized for some different cases, according to the random variable X of interest. Fig. 3
is a diagram that summarizes the cases studied in the following sections. All the examples are written in MATLAB live-script file
Intervals_section1022_live.mlx, in the supplementary material, so that they can be easily repeated or adapted for the reader’s
own data.
Therefore, according to Eq. (10), the confidence interval on the mean that is obtained from Eq. (12) is
z psffiffiffi ; X
X = þ z psffiffiffi = (13)
a2 n a2 n
Analogously, the confidence intervals at confidence level 100(1 a)% for the maximum and minimum values of the mean are
computed from Eqs. (14), (15), respectively
pr m X þ za psffiffiffi ¼ 1 a (14)
n
pr X za psffiffiffi m ¼ 1 a (15)
n
and, thus, the corresponding one-sided intervals would be N; X za psffiffi ; N .
þ za psffiffi and X
n n
In an experimental context, when measuring n aliquot parts of a test sample, we obtain n values x1, x2,., xn. Their sample mean x
and is also an estimate of the true value m.
is the particular value taken by the random variable X
Example 1: Suppose that an analytical method follows a N(m,4) and we have a sample of size 10 with values 98.87, 92.54, 99.42,
105.66, 98.70, 97.23, 98.44, 103.73, 94.45 and 101.08. With this sample, the mean is 99.01 and using Eq. (13), the interval at
pffiffiffiffiffiffi pffiffiffiffiffiffi
95% confidence level is 99:01 1:96 4 =
10 ; 99:01 þ 1:96 4=
10 ¼ [96.53, 101.49].
For the interpretation of this interval, notice that with different samples of size 10 (same analytical method), different intervals
will be obtained at the same 95% confidence level. The endpoints of these intervals are nonrandom values, and the unknown mean
value, which is also a specific value, will or will not belong to the interval. Therefore, the affirmation “the interval contains the
mean” is a deterministic assertion that is true or false for each of the intervals. What one knows is that it is true for
100(1 a)% of those intervals. In our case, as 95% of the constructed intervals will contain the true value, we say, at 95% confi-
dence level, that the interval [96.53, 101.49] contains m.
This is the interpretation with the frequentist approach adopted in this article, that is to say that the information on random
variables is obtained by means of samples of them and that the parameters to be estimated are not known but are fixed amounts
(e.g., the amount of analyte in a sample, m, is estimated by the measurement results obtained by analyzing it n times). With
a Bayesian approach to the problem, a probability distribution is attributed to the amount of analyte m and once fixed an interval
of interest [a,b], the “a priori” distribution of m, the experimental results, and the Bayes’ theorem are used to calculate the probability
10 Quality of Analytical Measurements: Statistical Methods for Internal Validation
a posteriori that m belongs to the interval [a,b]. It is shown that, although in most practical cases the uncertainty intervals obtained
from repeated measurements using either theory may be similar, their interpretation is completely different. The works by Lira and
Wöger12 and Zech13 are devoted to compare both approaches from the point of view of the experimental data and their uncertainty.
Also, an introduction to Bayesian methods for analyzing chemical data can be seen in Armstrong and Hibbert.14,15
deviation. Analogously, the one-sided intervals at the 100(1 a)% confidence level come from
pr m X þ ta;n psffiffiffi ¼ 1 a (17)
n
pr X ta;n psffiffiffi m ¼ 1 a (18)
n
Example 2: Suppose that the probability distribution of an analytical method is a normal, but its standard deviation is unknown.
With the data of Example 1, the sample standard deviation, s, is computed as 3.90. As t0.025,9 ¼ 2.262 (see Appendix), the confi-
dence interval at 95% level is [99.01 2.26 1.24, 99.01 þ 2.26 1.24] ¼ [96.21, 101.81]. The 95% confidence interval on the
minimum of the mean (i.e., the 95.0% lower confidence bound) is made up, according to Eq. (18), by all the values greater than
96.74 ¼ 99.01 1.83 1.24. The corresponding interval on the maximum (upper confidence bound for mean), Eq. (17), will be
made up by the values less than 101.28 ¼ 99.01 þ 1.83 1.24.
The length of the confidence intervals from Eqs. (12)–(15) is a function of the sample size and tends towards zero when the
sample size tends to infinity. This functional relation permits the computation of the sample size needed to obtain an interval
2
2 za=2 s
of given length, d. It will suffice to consider 2d ¼ za=2 psffiffin and take as n the nearest integer greater than d . For example, if
we want a 95% confidence interval with length d less than 2, in the hypothesis of Example 1, we will need a sample size greater
than or equal to 62. 2
2 ta=2;n s
The same argument can be applied when the standard deviation is unknown. However, in this case, to compute n by d
it is necessary to have an initial estimation of s, which, in general, is obtained in a pilot study with size n0 , in such a way that in the
previous expression the d.f., n, are n0 1. An alternative is to define the desired length of the interval in standard deviation units
(remember that the standard deviation is unknown). For instance, in Example 2, if we want d ¼ 0.5s, we will need a sample size
greater than (4za/2)2 ¼ 61.5; note the substitution of ta/2,n by za/2, which is mandatory because we do not have the sample size
needed to compute ta/2,n, which is precisely what we want to estimate.
where c2a/2, n is the critical value of a c2 distribution with n ¼ n 1 d.f. at significance level a/2. As in the previous case for the
sample mean, we should distinguish between the random variable sample variance S2 and one of its values, s2, computed with Eq.
(6) from sample x1, x2, ., xn.
The intervals for the maximum and minimum of the variance at 100(1 a)% confidence level are obtained from Eqs. (20),
(21), respectively.
( )
2 ðn 1ÞS2
pr s ¼1a (20)
c21a;n
Quality of Analytical Measurements: Statistical Methods for Internal Validation 11
( )
ðn 1ÞS2
pr s2 ¼1a (21)
c2a;n
Example 3: Knowing that the n ¼ 10 data of Example 2 come from a normal distribution with both mean and variance unknown, the
95% confidence interval on s2 is found from Eq. (19) as [7.21, 50.81] because s2 ¼ 15.25, c20.025, 9 ¼ 19.02, and c20.975, 9 ¼ 2.70. If
the analyst is interested in obtaining a confidence interval for the maximum variance, the 95% upper confidence interval is found
from Eq. (20) as [0, 41.27] because c20.95, 9 ¼ 3.33, that is, the upper bound for the variance is 41.27 with 95% confidence. Notice
the lower bound in 0. To obtain confidence intervals on the standard deviation, it suffices to take the square root of the aforemen-
tioned intervals because this operation is a monotonically increasing transformation; therefore, the intervals at 95% confidence
level on the standard deviation are [2.69, 7.13] and [0, 6.42], respectively.
The sample size, n, needed so that s2/s2 is between 1 k and 1 þ k is given by the nearest integer greater than 1 þ
h pffiffiffiffiffiffiffiffiffiffiffi . i2
12 z
=
a=2 1 þ k þ 1 k . For example, for k ¼ 0.5, such that the length of the confidence interval verifies 0.5 < s2/s2 < 1.5,
we would need n ¼ 40 data (at least). Just for comparative purposes, we will admit in the example that with the sample of size
40 we obtain the same variance s2 ¼ 15.25. As c20.025, 39 ¼ 58.12, and c20.975, 39 ¼ 23.65, the two-sided interval at 95% confidence
level is now [10.23, 25.15], which verifies the required specifications.
where X 1 and X
2 are the random variables of the sample mean, which take the values x1 and x2 . The reader can easily write the
expressions analogous to Eqs. (14), (15) for the one-sided intervals.
When the assumption s21 ¼ s22 is not reasonable, we can still obtain an interval on the difference m1 m2 by using the fact that
1 X 2 ðm1 m2 Þ
the statistic Xp ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2
is distributed approximately as a t with d.f. given by,
s1 =n1 þs2 =n2
2
s21 =n1 þ s22 =n2
n¼ 2 2 (25)
ðs21 =n1 Þ ðs22 =n2 Þ
n1 1 þ n2 1
The 100(1 a)% confidence interval is obtained from the following equation:
8 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 9
< s 2 s 2 2 2 =
pr X1 X
2 ta=2;n 1
þ 2
m1 m2 X 2 þ ta=2;n s1 þ s2
1 X ¼1a (26)
: n1 n2 n1 n2 ;
Example 5: We want to compute a confidence interval on the difference of two means with unknown and not equal variances, with
the results that come from an experiment carried out with four aliquot samples by two different analysts. The first analyst obtains
x1 ¼ 3:285, and the second x2 ¼ 3:257. The variances were s12 ¼ 3.33 10 5 and s22 ¼ 9.17 10 5, respectively. Assuming that
s12 s s22, Eq. (25) gives n ¼ 4.9, so the d.f. to apply Eq. (26) are 5 and t0.025,5 ¼ 2.571. Thus, the 95% confidence interval is
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð3:285 3:257Þ 2:571 3:3310 þ 9:1710
5 5
4 4 , that is, [0.014, 0.042]. So, at 95% of confidence, the two analysts provide unequal
measurements because zero is not in the interval.
The confidence intervals for the maximum and the minimum are obtained by considering the last or the first term respectively in
Eq. (26) and replacing ta/2,n by ta,n.
where d and sd are the mean and standard deviation of the differences di and n ¼ n 1 are the d.f. of the t distribution.
where F1 a/2, n1, n2 and Fa/2, n1, n2 are the critical values (upper tail) of an F distribution with n1 ¼ n2 1 d.f. in the numerator and
n2 ¼ n1 1 d.f. in the denominator. The Appendix contains a description of some relevant properties of the F distribution.
We can also compute one-sided confidence intervals. The 100(1 a)% upper or lower confidence bound on s12/s22 is obtained
from Eqs. (29), (30), respectively. Remember that, when computing the intervals by using Eq. (29), the lower bound is always 0.
2
s S2
pr 12 Fa;n1 ;n2 12 ¼ 1 a (29)
s2 S2
S2 s2
pr F1a;n1 ;n2 12 12 ¼ 1 a (30)
S2 s2
Example 6: In this example, we compute a two-sided 95% confidence interval for the ratio of the variances in Example 4 (n1 ¼ n2 ¼ 8,
s12 ¼ 3.89, s22 ¼ 4.02). The resulting interval is [0.20 (3.89/4.02), 4.99 (3.89/4.02)] ¼ [0.19, 4.83]. As 1 belongs to this
interval, we can admit that both variances are equal.
Quality of Analytical Measurements: Statistical Methods for Internal Validation 13
Hence, the 95% confidence interval on the median is made by the values that are between the values in position 2 and 8, that is,
[94.45, 101.08].
1 1
pr Nð0; 1Þ pffiffiffi þ k1 pr Nð0; 1Þ pffiffiffi k1 ¼ b (33)
n n
Therefore
sffiffiffiffiffiffiffiffiffiffiffiffiffi
n1
k ¼ k1 (34)
c2g;n1
cg,2 n 1 is the point exceeded with probability g when using the c2 distribution with n 1 d.f.
2
Example 7: With the data in Example 1, and b ¼ g ¼ 0.95, we have
x ¼ 99:01, s ¼ 3.91, k1 ¼ 2.054, and c0.95, 9¼ 3.33; thus, accord-
ing to Eq. (34), k ¼ 3.379 and, as a consequence, the interval [99.01 3.38 3.91, 99.01 þ 3.38 3.91] ¼ [85.79, 112.23]
contains 95% of the results of the method 95% of the times that the procedure is repeated with a sample of size 10.
methods, Rebafka et al.33 proposed the use of a bootstrap technique to calculate b-expectation tolerance intervals, whereas Fernholz
and Gillespie34 studied the estimation of the b-content tolerance intervals by using bootstrap.
To summarize this whole section about tolerance and confidence intervals, it is worth pointing out some comparative aspects
because there is a tendency to confuse both concepts that have nothing in common but the word interval. The difference between
them is clear: the confidence interval is the set that is supposed to contain (with a 100(1 a)% confidence) the true value of the
unknown parameter; the tolerance interval is the set that contains a value which is taken by the random variable in a percentage of b,
with a given confidence g.
In particular, confidence intervals must be used in the process of evaluating trueness and precision of a method when there is no
need to fulfill external requirements but just to compare with other methods or to quantify uncertainty and bias of the results ob-
tained with it.
A usual error is to mistakenly consider a confidence interval as a tolerance interval when the difference between them is impor-
tant. For instance, with the data of Example 7, notice that to compute the confidence interval, the standard deviation of the mean is
pffiffiffi
estimated as s= n ¼ 1:24, whereas the standard deviation of the individual results of the method is estimated as s ¼ 3.91, very
different.
Also, it is important to remember that when the sample size n tends to infinity, the length of a confidence interval tends toward
zero, independently of the chosen confidence level. For example, with the confidence intervals for the mean, in the limit we will
have x ¼ m thus the estimator and the true parameter will be equal for sure (1 a ¼ 1). On the contrary, the length of a b-content
tolerance interval does not tend towards zero when increasing the sample size but to the interval that contains for sure (1 g ¼ 1)
the 100b % of the values.
There are other aspects of the determination of the uncertainty that are of practical interest, for example, the problem that arises
by the fact that any uncertainty interval, particularly an expanded uncertainty interval, should be restricted to the range of feasible
values of the measurand. Cowen and Ellison35 analyzed how to modify the interval when the data are close to a natural limit in
a feasible range such as 0 or 100% mass or mole fraction.
This section is devoted to the introduction of a statistical methodology to decide whether an affirmation is false, for example, the
affirmation “this method of analysis applied to this sample of reference provides the certified value”. If, on the basis of the exper-
imental results, it is decided that it is false, we will conclude that the method has bias. The affirmation is customarily called hypoth-
esis and the procedure of decision making is called hypothesis testing. A statistical hypothesis is an asseveration on the probability
distribution that follows a random variable. Sometimes one has to decide on a parameter, for example, whether the mean of
a normal distribution is a specific value. In other occasions it may be required to decide on other characteristics of the distribution,
for example, whether the experimental data are compatible with the hypothesis that they come from a normal or uniform
distribution.
The hypotheses are not affirmations about the sample but about the distribution from which those values come, that is to say, m
is the value, unknown, of the pH of the solution that will be the same as the value provided by the procedure if the bias is zero (see
the model of Eq. (1)). In general, to test a hypothesis, the analyst must consider the experimental goal and define, accordingly, the
null hypothesis for the test, as in Eq. (37). Hypothesis-testing procedures rely on using the information in a random sample; if this
information is inconsistent with the null hypothesis, we would conclude that the hypothesis is false. If there is not enough evidence
to prove falseness, the test defaults to the decision of not rejecting the null hypothesis though this does not actually prove that it is
correct. It is therefore critical to choose carefully the null hypothesis in each problem.
In practice, to test a hypothesis, we must take a random sample, compute an appropriate test statistic from the sample data, and
then use the information contained in this statistic to make a decision. However, as the decision is based on a random sample, it is
subject to error. Two kinds of potential errors may be made when testing hypothesis. If the null hypothesis is rejected when it is true,
then a type I error has been made. A type II error occurs when the researcher accepts the null hypothesis when it is false. The situation
is described in Table 3.
In Example 8, if the experimental data lead to rejection of the null hypothesis H0 being true, our (wrong) conclusion is that the
pH of the solution is less than 2. A type I error has been made and the analyst will use the solution in the procedure when in fact it is
not chemically valid. If, on the contrary, the experimental data lead to acceptance of the null hypothesis when it is false, the analyst
will not use the solution when in fact the pH is less than 2 and a type II error has been made. Note that both types of error have to be
considered because their consequences are very different. In the case of type I error, an unsuitable solution is accepted, the procedure
will be inadequate, and the analytical result will be wrong with the subsequent damages that it may cause (e.g., the loss of a client, or
a mistaken environmental diagnosis). On the contrary, the type II error implies that a valid solution is not used with the correspond-
ing extra cost of the analysis. It is clear that the analyst has to specify the assumable risk of making these errors, and this is done in
terms of the probability that they will occur.
The probabilities of occurrence of type I and II errors are denoted by specific symbols, defined in Eq. (39). The probability a of
the test is called the significance level, and the power of the test is 1 b, which measures the probability of correctly rejecting the
null hypothesis.
a ¼ prftype I errorg ¼ prfreject H0 j H0 is trueg
(39)
b ¼ prftype II errorg ¼ prfaccept H0 j H0 is falseg
In Eq. (39), the symbol “|” indicates that the probability is calculated under that condition. In the example we are following,
a will be calculated with the normal distribution of mean 2 and standard deviation 0.55.
Statistically expressed, with the n ¼ 10 results in Example 8 (sample mean x ¼ 1:708), one wants to decide about the value of
the mean of a normal distribution with known variance and one-sided alternative hypothesis (a one-tail test).
pffiffiffi
With ffi premises, the related statistic is written in Table 4 (second row) and gives Zcalc ¼ ðx m0 Þ=ðs= n Þ ¼ ð1:708 2:0Þ=
these
pffiffiffiffiffi
0:55= 10 ¼ 1:679.
In addition, the analyst must assume the risk a, say 0.05. This means that the decision rule that is going to apply to the exper-
imental results will accept an inadequate (chemical) solution 5% of the times. Therefore, the critical or rejection region is written in
Table 4, second row, as CR ¼ {Zcalc < 1.645}, meaning that the null hypothesis will be rejected for the samples of size 10 that
provide values of the statistic less than 1.645. In the example, the actual value Zcalc ¼ 1.679 belongs to the critical region; thus,
the decision is to reject the null hypothesis at 5% significance level.
Given the present facilities of computation, instead of the CR, the available statistical software calculates the so-called P-value,
which is the probability of obtaining the current value of the statistic under the null hypothesis H0. In our case, P-value ¼ pr
{Z 1.679} ¼ 0.0466. When the P-value is less than the significance level a, the null hypothesis is rejected because this is the
same as saying that the value of the statistic belongs to the critical region.
The next question that immediately arises is about the power of the applied decision rule (statistic and critical region). To calcu-
late b, defined in Eq. (39), it is necessary to exactly specify the meaning of the alternative hypothesis. In our case, what is meant by
pH smaller than 2. From a mathematical point of view, the answer is clear: any number less than 2, for example, 1.9999 which
clearly does not make sense from the point of view of the analyst. In this context, sometimes due to previous knowledge, in other
cases because of the regulatory stipulations or simply by the detail of the standardized work procedure, the analyst can decide the
value of pH that is considered to be less than 2.00, for example, a pH less than 1.60. This is the same as assuming that “pH equal to
2” is any smaller value whose distance to 2 is less than 0.40. In these conditions,
jdj pffiffiffi
b ¼ pr Nð0; 1Þ < za n (40)
s
“‘Sir: You will cause the following persons to leave the State of
Missouri, within a reasonable time after the receipt of this order, and
reside, during the war or until permitted to return, at some place
north of Indianapolis, Indiana, and east of Illinois. They will be
required to report to you, by letter, once a month, and are not
permitted to leave the State by way of St. Louis, but directed to go by
Macon City and Hannibal, Missouri. Rev. B. H. Spencer, * * * * * * *.
“‘By command of Brigadier-General Merrill.
“‘Rev. B. H. Spencer:
“‘Sir: The above is a true copy of Gen. Merrill’s order to me. You
will obey said order within six days from this date. You will report to
these headquarters on the day of departure.
“‘By order of J. G. Kettle, Col. Commanding.
“‘Rev. B. H. Spencer:
“On the 25th of January, 1863, I preached the sermon alluded to;
and then, in company with four others, made my report to military
headquarters at Danville, Mo. But, in consequence of an accident on
the railroad, I was permitted to remain with my family until the 28th
of that month, when, with a sad heart, I was compelled to leave my
distressed wife and six little children and go into a land of strangers,
and remain in exile for ten long months.
“Dr. H. W. Pitman, Rev. D. W. Nowlin, Rev. J. D. Gregory and Rev.
Wm. A. Taylor were banished in company with me. We had no trial,
either civil or military, nor would they condescend to tell us what
were the charges against us, or whether, indeed, there were any. Nor
to this day—September 7th, 1869—have we found out why it was
done, except through private and unofficial sources. The information
thus received as to the cause of my banishment was as I expected—I
was banished because I was a Southern Methodist preacher! One of
the officers was asked by one of my friends: ‘What are the charges
against Spencer?’ He answered, ‘I never heard that there are any;
but he is a man of influence, and, if disposed, can do a great deal of
harm!’ Another officer was asked by another friend, and he replied,
‘The fact that he is a Southern Methodist preacher is all I want to
know!’ There never was a more clear case of ecclesiastical
persecution than was my banishment. Certain men sought to
produce secession, treason and rebellion in the M. E. Church, South,
by way of showing how they professed to hate these things in the
nation; I opposed them, and they became my enemies and had me
banished. If any one doubts this let him attend to the following
documents:
“‘Dear Sir—I have just received yours of 22d inst., and must
acknowledge I am utterly at a loss to comprehend it.
“‘I want to say, once for all, to yourself, as also to Doct. Pitman and
Judge Nowlin, that I had no hand in your banishment whatever,
either as a private citizen or as an officer; that I never had, either
directly or indirectly, an intimation that such a thing was
contemplated. An order was issued by General McKean, who is
Commanding General of this district, headquarters at Palmyra, to J.
G. Lane, Provost-Marshal of Wellsville district, to take the testimony
of the loyal men of Montgomery county in relation to the propriety
of your return home. Lane was removed from office and his district
thrown into mine, and the order was sent to me by General McKean,
which I executed by taking the evidence of loyal men, both at High
Hill and Montgomery City, as well as Danville. The evidence was
sworn to and sent by order of the commanding General to his
headquarters.
“‘Now, sir, I have given you the facts in regard to everything I have
had to do with this case. And, although you protest against any
intention to insult or offend in your communication, I must frankly
admit that the whole tenor of your letter seems to savor of both.
‘How can you consent, without just cause, to keep one in exile who is
in feeble health,’ &c., is one extract from your letter. ‘Will you not
then allow me to come home at once?’ is another. Now, sir, you must
know that I have no direct control of this matter! Why ask me such
questions? Why not ask me, as a private citizen, to use my influence
to obtain a revocation of the order? The authorities that issued the
order of your banishment have never asked, neither have I given, my
opinion as to the propriety of the order. Notwithstanding I consider
your letter as invidious, and, as I understand it, full of insinuations
against me, yet, under the circumstances, I will allow humanity to
step in, discard all feeling that your letter may have excited, and give
you the best advice I am capable of.
“‘Judge Nowlin, Doct. Pitman and yourself get up a letter, directed
to Brig.-Gen. McKean, Palmyra, Mo., through me as Provost-Marshal
of Montgomery county. Take humanity for your text; appeal to him
through the tears of your wife and helpless children; let Government
officers alone; agree to report to me once a week in person, if it
should be considered necessary; give every assurance that your lips
will be sealed in future as to the utterance of treason, directly or
indirectly; send the letter to me and I will forward it, with such
recommendation as I may deem proper and right, and, if that fails, I
am at the end of my row. The success of this thing will very much
depend on keeping my advice to yourselves. I may be mistaken, but I
believe your liberation may be effected in that way. Give my respects
to Judge Nowlin and Doctor Pitman.
“‘Yours, &c.,
“‘A. C. Stewart, Prov.-Marshal.’
“‘Most respectfully,
“‘B. H. Spencer.’
“‘H. S. Lane.’