During the drying of a surface, the liquid evaporation acts to keep the temperature relatively constant, due to evaporative cooling. As the drying nears completion the liquid film begins to break, exposing areas that are no longer cooled through evaporation, which begin to heat. Although this heating can be measured with an Infrared (IR) camera, the sensitivity is often not sufficient to recognize the point at which the film breaks. Complicating the measurement is the changing emissivity that commonly occurs as objects dry. The sensitivity and emissivity issues can be addressed by analyzing the temperature in the area of interest and computing the coefficient of variance (COV) of the temperature. This technique is compared to temperature and standard deviation measurements made with an IR camera and the COV technique is shown to be superior for determining when the liquid film breaks. The film breakage point is found to vary with temperature and material roughness in two industrially significant applications: the drying of wood flakes and the drying of polymer films. Film breakage in wood flakes is related to detrimental finished quality problems and also to emission problems. The rate at which an adhesive dries affects the roughness of the polymer film and subsequently, the bond strength. The COV technique is used to predict the roughness of the finished polymer film. Use of the COV technique allows the drying of a liquid film to be visualized in a way that has been previously unreported.

ACI Committee 544's repeated drop-weight impact test for concrete is often criticized for large variations within the results. This paper identifies the sources of these large variations and accordingly suggests modifications to the ACI test. The proposed modifications were evaluated and compared to the current ACI test by conducting impact resistance tests on 40 specimens from two batches of polypropylene fiber-reinforced concrete (PPFRC). The results obtained from both methods were statistically analyzed and compared. The variations in the results were investigated within the same batch and between different batches of concrete. The impact resistance of PPFRC specimens tested with the current ACI test exhibited large coefficients of variation (COV) of 58.6% and 50.2% for the first-crack and the ultimate impact resistance, respectively. The corresponding COV for PPFRC specimens tested according to the modified technique were 39.4% and 35.2%, indicating that the reliability of the results was significantly improved. It has been shown that, using the current ACI test, the minimum number of replications needed per each concrete mixture to obtain an error below 10% was 41 compared to 20 specimens for the modified test. Although such a large number of specimens is not good enough for practical and economical reasons, the reduction presents a good step on the development of a standard impact test.

Há pouca literatura acerca dos valores de referência das medidas de dispersão na pesquisa avícola bem como a relação entre tais medidas e o número de aves por repetição e número de repetições e ainda os efeitos de fatores intrínsecos à pesquisa sobre a variabilidade experimental. Com isso, objetiva-se estabelecer faixas de classificação das medidas de dispersão, bem como a relação entre medidas e o número de aves por repetição, número de repetições por tratamento, número de aves abatidas para avaliação de carcaças, número de ovos coletados para análises de qualidade e os efeitos de fatores intrínsecos à pesquisa sobre a variabilidade experimental. Os dados foram obtidos nos trabalhos publicados nos periódicos com classificação Qualis/CAPES nos estratos A1, A2, B1, B2 e B3, na área de zootecnia. Para a determinação das faixas de classificação das medidas de dispersão. Os dados foram correlacionados através da correlação de Spearman, dados qualitativos foram comparados pelo teste de Kruskal-Wallis. As faixas de classificação de coeficientes de variação podem ser utilizadas como balizador da qualidade e confiabilidade dos dados de experimentos com poedeiras comerciais e frangos de corte. A fase de vida ou estágio produtivo de poedeiras e frangos de corte devem ser levados em consideração na comparação de resultados experimentais. O erro padrão da média varia bastante nos parâmetros produtivos de poedeiras e frangos de corte e parece não ser uma medida de dispersão indicada para se comparar a precisão de diferentes experimentos, as suas faixas de classificação servem como referencial da frequência com que se encontra resultados. Coeficientes de variação e erros padrão da média de parâmetros produtivos de frangos de corte e de poedeiras comerciais são menores quanto maior o número de repetições e de aves por repetição. As aves amostradas para o abate nas avaliações de carcaça de frangos de corte devem apresentar o peso médio da parcela. O número de aves amostradas não influencia a variabilidade experimental. Os dias de amostragem para análises de qualidade de ovos não influenciam a variação aleatória. O número ideal de ovos amostrados está entre quatro e cinco ovos por repetição. Em frangos de corte a variabilidade aleatória para variáveis de produção são maiores para aves da linhagem Ross do que para aves da linhagem Cobb. Lotes mistos de frangos de corte apresentam maior variação de ganho de peso do que lotes de machos na fase de criação de um a 21 dias de idade. O rendimento de carcaça e a porcentagem de gordura abdominal apresentam menor variabilidade aleatória em lotes mistos do que em lotes de frangos machos. Para frangos de corte não há efeitos aleatórios que possam aumentar a variabilidade experimental em função da área disponível por ave. Para poedeiras quanto maior o período experimental menor tende a ser a variação aleatória. Poedeiras brancas apresentam menor variação aleatória dos parâmetros produtivos do que poedeiras vermelhas. O coeficiente de variação de parâmetros produtivos de poedeiras comerciais aumenta com o aumento da área disponível por ave. / There is little literature on the reference values of spread measures in poultry research and the relationship between such measures and the number of birds per replicate and number of replicates and even the effects of intrinsic factors to research on experimental variability. The aim of this research is to establish ranges for the classification of spread measures, and the relationship between measurements and the number of birds per replicate, number of replicates per treatment, number of birds slaughtered for carcass evaluation, number of eggs collected for quality analysis and impact of research intrinsic factors on experimental variability. The data were obtained in papers published in journals with Qualis/CAPES classification in strata A1, A2, B1, B2 and B3, in the area of animal science. To determine the classification of the spread measures. The data were correlated by Spearman correlation, qualitative data were compared using the Kruskal-Wallis test. The classification ranges of coefficients of variation can be used as an indicator of the quality and reliability of data from experiments with laying hens and broilers. The age or production stage of laying hens and broilers should be considered in the comparison of experimental results. The standard error of means varies greatly in production parameters of laying hens and broilers and seems not to be a measure of dispersion indicated to compare the accuracy of different experiments, their classification ranges serve as a frequency reference which it is results. Coefficients of variation and standard error of means of production parameters of broilers and laying hens are smaller the greater the number of repetitions and birds per replicate. The birds sampled for slaughter in assessing carcass of broilers must present the average weight of the replicate. The number of birds sampled does not influence the experimental variability. The days of sampling for quality analysis of eggs did not influence the random variation. The ideal number of eggs sampled is between four and five eggs per replicate. In broilers random variability for production variables are higher for Ross than for Cobb broilers. Straight run broilers exhibit greater variation in weight gain than lots of males in the creation phase of 1 to 21 days old. The carcass yield and abdominal fat have less random variability in straight run broilers than in lots of male chickens. No random effects that can increase experimental variability depending on the available area per bird on broilers. The higher is experimental period the greater is the random variations. White hens have less random variation in production parameters than brown hens. The coefficient of variation of productive parameters of laying hens increases with increasing the available area per bird.

Energy use index (EUI) is an important measure of energy use which normalizes energy use by dividing by building area. Energy use indices and associated coefficients of variation are computed for major industry categories for electricity and natural gas use in small and medium-sized plants in the U.S. The data is very scattered with the coefficients of variation (CoV) often exceeding the average EUI for an energy type. The combined CoV from all of the industries considered, which accounts for 8,200 plants from all areas of the continental U.S., is 290%. This paper discusses EUIs and their variations based on electricity and natural gas consumption. Data from milder climates appears more scattered than that from colder climates. For example, the ratio of the average of coefficient of variations for all industry types in warm versus cold regions of the U.S. varies from 1.1 to 1.7 depending on the energy sources considered. The large data scatter indicates that predictions of energy use obtained by multiplying standard EUI data by plant area may be inaccurate and are less accurate in warmer than colder climates (warmer and colder are determined by annual average temperature weather data). Data scatter may have several explanations, including climate, plant area accounting, the influence of low cost energy and low cost buildings used in the south of the U.S. This analysis uses electricity and natural gas energy consumption and area data of manufacturing plants available in the U.S. Department of Energy’s national Industrial Assessment Center (IAC) database. The data there come from Industrial Assessment Centers which employ university engineering students, faculty and staff to perform energy assessments for small to medium-sized manufacturing plants. The nation-wide IAC program is sponsored by the U.S. Department of Energy. A collection of six general energy saving recommendations were also written with Texas manufacturing plants in mind. These are meant to provide an easily accessible starting point for facilities that wish to reduce costs and energy consumption, and are based on common recommendations from the Texas A&M University IAC program.

Energy use index (EUI) is an important measure of energy use which normalizes energy use by dividing by building area. Energy use indices and associated coefficients of variation are computed for major industry categories for electricity and natural gas use in small and medium-sized plants in the U.S. The data is very scattered with the coefficients of variation (CoV) often exceeding the average EUI for an energy type. The combined CoV from all of the industries considered, which accounts for 8,200 plants from all areas of the continental U.S., is 290%. This paper discusses EUIs and their variations based on electricity and natural gas consumption. Data from milder climates appears more scattered than that from colder climates. For example, the ratio of the average of coefficient of variations for all industry types in warm versus cold regions of the U.S. varies from 1.1 to 1.7 depending on the energy sources considered. The large data scatter indicates that predictions of energy use obtained by multiplying standard EUI data by plant area may be inaccurate and are less accurate in warmer than colder climates (warmer and colder are determined by annual average temperature weather data). Data scatter may have several explanations, including climate, plant area accounting, the influence of low cost energy and low cost buildings used in the south of the U.S. This analysis uses electricity and natural gas energy consumption and area data of manufacturing plants available in the U.S. Department of Energy's national Industrial Assessment Center (IAC) database. The data there come from Industrial Assessment Centers which employ university engineering students, faculty and staff to perform energy assessments for small to medium-sized manufacturing plants. The nation-wide IAC program is sponsored by the U.S. Department of Energy. A collection of six general energy saving recommendations were also written with Texas manufacturing plants in mind. These are meant to provide an easily accessible starting point for facilities that wish to reduce costs and energy consumption, and are based on common recommendations from the Texas A&M University IAC program.

Há pouca literatura acerca dos valores de referência das medidas de dispersão na pesquisa avícola bem como a relação entre tais medidas e o número de aves por repetição e número de repetições e ainda os efeitos de fatores intrínsecos à pesquisa sobre a variabilidade experimental. Com isso, objetiva-se estabelecer faixas de classificação das medidas de dispersão, bem como a relação entre medidas e o número de aves por repetição, número de repetições por tratamento, número de aves abatidas para avaliação de carcaças, número de ovos coletados para análises de qualidade e os efeitos de fatores intrínsecos à pesquisa sobre a variabilidade experimental. Os dados foram obtidos nos trabalhos publicados nos periódicos com classificação Qualis/CAPES nos estratos A1, A2, B1, B2 e B3, na área de zootecnia. Para a determinação das faixas de classificação das medidas de dispersão. Os dados foram correlacionados através da correlação de Spearman, dados qualitativos foram comparados pelo teste de Kruskal-Wallis. As faixas de classificação de coeficientes de variação podem ser utilizadas como balizador da qualidade e confiabilidade dos dados de experimentos com poedeiras comerciais e frangos de corte. A fase de vida ou estágio produtivo de poedeiras e frangos de corte devem ser levados em consideração na comparação de resultados experimentais. O erro padrão da média varia bastante nos parâmetros produtivos de poedeiras e frangos de corte e parece não ser uma medida de dispersão indicada para se comparar a precisão de diferentes experimentos, as suas faixas de classificação servem como referencial da frequência com que se encontra resultados. Coeficientes de variação e erros padrão da média de parâmetros produtivos de frangos de corte e de poedeiras comerciais são menores quanto maior o número de repetições e de aves por repetição. As aves amostradas para o abate nas avaliações de carcaça de frangos de corte devem apresentar o peso médio da parcela. O número de aves amostradas não influencia a variabilidade experimental. Os dias de amostragem para análises de qualidade de ovos não influenciam a variação aleatória. O número ideal de ovos amostrados está entre quatro e cinco ovos por repetição. Em frangos de corte a variabilidade aleatória para variáveis de produção são maiores para aves da linhagem Ross do que para aves da linhagem Cobb. Lotes mistos de frangos de corte apresentam maior variação de ganho de peso do que lotes de machos na fase de criação de um a 21 dias de idade. O rendimento de carcaça e a porcentagem de gordura abdominal apresentam menor variabilidade aleatória em lotes mistos do que em lotes de frangos machos. Para frangos de corte não há efeitos aleatórios que possam aumentar a variabilidade experimental em função da área disponível por ave. Para poedeiras quanto maior o período experimental menor tende a ser a variação aleatória. Poedeiras brancas apresentam menor variação aleatória dos parâmetros produtivos do que poedeiras vermelhas. O coeficiente de variação de parâmetros produtivos de poedeiras comerciais aumenta com o aumento da área disponível por ave. / There is little literature on the reference values of spread measures in poultry research and the relationship between such measures and the number of birds per replicate and number of replicates and even the effects of intrinsic factors to research on experimental variability. The aim of this research is to establish ranges for the classification of spread measures, and the relationship between measurements and the number of birds per replicate, number of replicates per treatment, number of birds slaughtered for carcass evaluation, number of eggs collected for quality analysis and impact of research intrinsic factors on experimental variability. The data were obtained in papers published in journals with Qualis/CAPES classification in strata A1, A2, B1, B2 and B3, in the area of animal science. To determine the classification of the spread measures. The data were correlated by Spearman correlation, qualitative data were compared using the Kruskal-Wallis test. The classification ranges of coefficients of variation can be used as an indicator of the quality and reliability of data from experiments with laying hens and broilers. The age or production stage of laying hens and broilers should be considered in the comparison of experimental results. The standard error of means varies greatly in production parameters of laying hens and broilers and seems not to be a measure of dispersion indicated to compare the accuracy of different experiments, their classification ranges serve as a frequency reference which it is results. Coefficients of variation and standard error of means of production parameters of broilers and laying hens are smaller the greater the number of repetitions and birds per replicate. The birds sampled for slaughter in assessing carcass of broilers must present the average weight of the replicate. The number of birds sampled does not influence the experimental variability. The days of sampling for quality analysis of eggs did not influence the random variation. The ideal number of eggs sampled is between four and five eggs per replicate. In broilers random variability for production variables are higher for Ross than for Cobb broilers. Straight run broilers exhibit greater variation in weight gain than lots of males in the creation phase of 1 to 21 days old. The carcass yield and abdominal fat have less random variability in straight run broilers than in lots of male chickens. No random effects that can increase experimental variability depending on the available area per bird on broilers. The higher is experimental period the greater is the random variations. White hens have less random variation in production parameters than brown hens. The coefficient of variation of productive parameters of laying hens increases with increasing the available area per bird.

A new confidence interval for the coefficient of dispersion (mean absolute deviation from the median divided by median) is proposed and is shown to perform better than the BCa bootstrap confidence interval.

yes / ACI Committee 544’s repeated drop-weight impact test for concrete is often criticized for large variations within the results. This paper identifies the sources of these large variations and accordingly suggests modifications to the ACI test. The proposed modifications were evaluated and compared to the current ACI test by conducting impact resistance tests on 40 specimens from two batches of polypropylene fiber-reinforced concrete (PPFRC). The results obtained from both methods were statistically analyzed and compared. The variations in the results were investigated within the same batch and between different batches of concrete. The impact resistance of PPFRC specimens tested with the current ACI test exhibited large coefficients of variation (COV) of 58.6% and 50.2% for the first-crack and the ultimate impact resistance, respectively. The corresponding COV for PPFRC specimens tested according to the modified technique were 39.4% and 35.2%, indicating that the reliability of the results was significantly improved. It has been shown that, using the current ACI test, the minimum number of replications needed per each concrete mixture to obtain an error below 10% was 41 compared to 20 specimens for the modified test. Although such a large number of specimens is not good enough for practical and economical reasons, the reduction presents a good step on the development of a standard impact test.

Street trees are a subpopulation of the urban forest resource and exist in the rights-of-way adjacent to public roads in a municipality. Benefit-cost analyses have shown that the annual benefits provided by the average street tree far outweigh the costs of planting and maintenance. City and municipal foresters spend a majority of their time and resources managing street tree populations. Sample street tree inventories are a common method of estimating municipal street tree populations for the purposes of making urban forest policy, planning, and management decisions. i-Tree Streets is a suite of software tools capable of producing estimates of street tree abundance and value from a sample of street trees taken along randomly selected sections (segments) of public streets. During sample street tree inventories conducted by Virginia Tech Urban Forestry, it was observed that the lengths of the sample streets recommended by i-Tree varied greatly within most municipalities leading to concern about the impact of street length variation on sampling precision. This project was conducted to improve i-Tree Streets by changing the recommended sampling protocol without altering the software. Complete street tree censuses were obtained from 7 localities and standardized using GIS. The effects of standardizing street segments to 3 different lengths prior to sampling on the accuracy and precision of i-Tree Streets estimates were investigated though computer simulations and analysis of changes in variation in number of trees per street segment as a basis for recommending procedural changes. It was found that standardizing street segments significantly improved the precision of i-Tree Streets estimates. Based on the results of this investigation, it is generally recommended that street segments be standardized to 91m (300 ft) prior to conducting a sample inventory. Standardizing to 91m will significantly reduce the number of trees, the number of street segments, and the percentage of total street segments that must be sampled to achieve an estimate with a 10% relative standard error. The effectiveness of standardization and the associated processing time can be computed from municipal attributes before standardization so practitioners can gauge the marginal gains in field time versus costs in processing time. Automating standardization procedures or conducting an optimization study of segment length would continue to increase the efficiency and marginal gains associated with street segment standardization. / Master of Science

Les examens de laboratoire jouent un rôle essentiel dans la pratique médicale. Ils sont utilisés à des fins diagnostique, pronostique, thérapeutique ou encore pour dépister des maladies spécifiques dans des populations présumées en bonne santé (Tietz, 1986). Quotidiennement, les laboratoires de biologie clinique réalisent des millions de tests fournissant autant de données à interpréter. Les responsables des soins de santé ont toujours été préoccupés par la qualité globale du travail réalisé dans les laboratoires. En dépit de procédures de contrôle de qualité interne rigoureuses, les résultats obtenus pour un même échantillon dans deux laboratoires peuvent occasionnellement différer de façon notoire. Il serait inacceptable cependant quun résultat dun test biologique soit considéré comme « normal » dans un laboratoire et « anormal » dans lautre. Les programmes dEvaluation Externe de la Qualité (EEQ) ont précisément comme objectif de contrôler la performance analytique des laboratoires de biologie clinique à une grande échelle et généralement par le biais dun organisme externe. Le but de lEEQ est de sassurer que les résultats des tests soient compatibles quel que soit le laboratoire qui réalise les analyses, en dautres termes de garantir la comparabilité des laboratoires et par là-même dadministrer les meilleurs soins aux patients (Libeer, 1993). Les protocoles EEQ consistent à organiser des enquêtes dans lesquelles les laboratoires participants doivent réaliser des analyses sur un même échantillon contrôle comme sil sagissait danalyses de routine. Il faut ensuite renvoyer les résultats de ces analyses au centre EEQ avec des informations détaillées sur les techniques de dosage utilisées. Les résultats sont alors soumis à une analyse statistique. En réalité, lanalyse statistique poursuit trois objectifs : (1) obtenir une estimation robuste de la concentration du constituant (moyenne) dans léchantillon contrôle et de la variabilité entre les laboratoires (écart-type), une estimation robuste étant nécessaire car les résultats EEQ contiennent souvent des valeurs aberrantes qui peuvent sérieusement affecter la moyenne et de lécart-type ; (2) évaluer la qualité des laboratoires en mettant en évidence les résultats « hors-limites » et les laboratoires « peu performants » et ce, à court et à long termes ; (3) évaluer et comparer la précision analytique des techniques/équipements utilisés par les participants (Albert, 1997). Depuis plusieurs dizaines dannées, des programmes EEQ ont été mis en place dans de nombreux pays et ils contrôlent la plupart des analyses de biologie clinique classiques comme le glucose, le cholestérol ou le calcium, ou encore les enzymes et les hormones. Ce travail a débuté lorsque lélectrophorèse de protéines fut introduite dans le panel des analyses de laboratoire soumises à lEEQ. Contrairement aux autres tests contrôlés jusqualors, lélectrophorèse de protéines fournit non pas une valeur mais cinq fractions, respectivement, lalbumine, les globulines α1, α2, β et , dont la somme fait 100% et dont linterprétation doit se faire globalement. En dautres termes, les données électrophorétiques obtenues dans lEEQ doivent être analysées par des méthodes de la statistique multivariée (Zhang et al, 2008). Ceci nécessite ladaptation à lenvironnement EEQ de méthodes multivariées existantes (telles quon les trouve dans les ouvrages de référence) ou le développement dapproches nouvelles. Dans cette thèse, nous nous sommes fixé comme but dapporter une solution théorique et pratique au problème de lanalyse et de linterprétation des résultats délectrophorèses dans le domaine de lEvaluation Externe de la Qualité. Dans lintroduction de ce travail, un bref rappel est fait des protocoles EEQ et de leur mise en pratique notamment en ayant recours à lInternet. Les méthodes univariées de statistique robuste permettant destimer la moyenne et lécart-type des résultats EEQ sont décrites et la notion de valeur « hors-limites » est définie. Ceci permet dapprécier la qualité dun laboratoire par rapport à ses pairs sur base du résultat quil a fourni. Le concept de coefficient de variation (CV) est aussi présenté comme un moyen de mesurer et de comparer la reproductibilité analytique des techniques de laboratoire. Finalement, on rappelle quelques notions relatives aux électrophorèses de protéines, leur utilité clinique et les méthodes de dosage utilisées. Les bases de données EEQ qui servent à illustrer la méthodologie statistique développée dans ce travail sont décrites en détail. Elles proviennent principalement des enquêtes de lEvaluation Externe de la Qualité réalisées entre 2004 et 2008 en France et en Belgique. La première partie de cette thèse concerne le problème de lévaluation de la performance des laboratoires pour le dosage de lélectrophorèse de protéines. La façon la plus simple consiste à appliquer les méthodes univariées classiques à chaque fraction de lélectrophorèse. Cette façon de procéder cependant ignore que les fractions doivent être interprétées globalement car elles sont corrélées. De plus, elles sont linéairement dépendantes (car leur somme est égale à 100%), ce qui conduit à une matrice de variances-covariances singulière et donc non inversible. La première approche multivariée que nous proposons vise à retirer une des cinq fractions et ainsi éviter le problème de singularité, ensuite à obtenir une estimation robuste du vecteur moyen et de la matrice de variances-covariances par la technique du déterminant de covariance minimum (MCD) publiée par Rousseuw et Van Driessen (1999). On utilise alors la distance de Mahalanobis pour identifier les profils électrophorétiques « hors-limites » et détecter les laboratoires dont la qualité est insatisfaisante (Zhang et al. 2008). Appliquée aux bases de données EEQ, cette méthode simple savère ne pas être optimale car elle donne des corrélations peu fiables et décèle trop de laboratoires « hors-limites ». Cest la raison pour laquelle, une approche nouvelle est proposée dans laquelle une transformation log-ratio (Egozcue et al. 2003) est appliquée aux profils électrophorétiques avant de les analyser statistiquement. Cette méthode transforme les cinq fractions électrophorétiques en quatre variables indépendantes et sans dimension. La technique MCD est alors appliquée pour obtenir des estimations robustes du vecteur moyen et de la matrice de dispersion. Les estimations sont utilisées pour calculer la distance de Mahalanobis et mettre en lumière les laboratoires « hors-limites ». Appliquée aux bases de données EEQ, cette seconde approche est meilleure que la première, non seulement dun point de vue théorique mais aussi pratique, en détectant un nombre plus raisonnable de laboratoires peu performants. Des méthodes de représentations graphiques des profils électrophorétiques sont aussi proposées au moyen du "MCD z-score plot" ou du "star plot" obtenu à partir de lanalyse en composantes principales (ACP) robuste. La seconde partie de la thèse sattache à lévaluation et à la comparaison de la précision analytique des techniques de dosage délectrophorèses utilisées par les laboratoires participants à lEEQ. Ceci nous a conduit à rechercher des méthodes dextension du coefficient de variation (CV) classique au cas multivariée. Les coefficients de variation multivariés publiés dans la littérature sont passés en revue, en ce compris ceux de Reyment (1960) et de Van Valen (1974, 2005). Reyment fut le premier à proposer une définition du CV multivarié et à donner une formule permettant de calculer l'erreur type de l'estimation. Van Valen suggéra une définition plus générale car applicable en toute circonstances mais qui malheureusement ne tient pas compte explicitement des corrélations entre les variables. Par ailleurs, nous avons exploité une idée de Voinov et Nikulin (1996) pour développer un CV multivarié basée sur la distance de Mahalanobis et qui est invariant par rapport à léchelle utilisée. Nous lavons appliqué aux données de lEEQ franco-belge de 2004 et avons ainsi pu classer les techniques électrophorétiques par ordre décroissant de précision analytique (Zhang et al. 2010). Malheureusement, cette approche comme celle du CV de Reyment requiert une matrice de covariance non singulière et nest donc pas applicable pour des groupes deffectif faible (n < 5) de laboratoires utilisant la même technique. Nous avons dès lors proposé une définition originale et tout à fait générale du CV multivarié, notée CVm (Albert et Zhang, 2010). Cette nouvelle formulation jouit de propriétés intéressantes; elle est simple et facile à calculer, ne requiert aucune inversion de matrice (contrairement aux autres techniques) mais uniquement le calcul de formes quadratiques. On nimpose aucune restriction sur le nombre dobservations ni sur le nombre de variables du problème. Nous avons appliqué cette méthode CVm aux bases de données EEQ délectrophorèses et nous avons pu démontrer en pratique sa grande flexibilité. Elle nous a permis de classer toutes les techniques électrophorétiques même pour des groupes de participants de petite taille (n = 2) et de mettre en évidence les plus précises, comme celle de lélectrophorèse capillaire de zone (CZE) complètement automatisée. Nous avons aussi appliqué la nouvelle méthode à des données de cytométrie de flux récoltées dans le cadre dune enquête EEQ réalisée en Belgique en 2010 ainsi quà un échantillon de données de « microarray » publié dans la littérature (Golub et al. 1999), confirmant ainsi son applicabilité à des domaines variés. Enfin, nous présentons quelques développements théoriques personnels sur le CV de Reyment afin de corriger la définition initiale, entachée à notre sens dune erreur de dimension, son estimation ainsi que la formule de lerreur type. En résumé, la méthodologie statistique développée dans ce travail propose une solution complète à lanalyse des données électrophorétiques ou de tout autre profil de tests de laboratoire récoltés dans le cadre denquêtes de lEvaluation Externe de la Qualité.