ASSESSING THE UNRELIABILITY OF THE MEDICAL LITERATURE: A RESPONSE TO "WHY MOST PUBLISHED RESEARCH FINDINGS ARE FALSE"

A recent article in this journal (Ioannidis JP (2005) Why most published research findings are false. PLoS Med 2: e124) argued that more than half of published research findings in the medical literature are false. In this commentary, we examine the structure of that argument, and show that it has three basic components: 1)An assumption that the prior probability of most hypotheses explored in medical research is below 50%. 2)Dichotomization of P-values at the 0.05 level and introduction of a “bias” factor (produced by significance-seeking), the combination of which severely weakens the evidence provided by every design. 3)Use of Bayes theorem to show that, in the face of weak evidence, hypotheses with low prior probabilities cannot have posterior probabilities over 50%. Thus, the claim is based on a priori assumptions that most tested hypotheses are likely to be false, and then the inferential model used makes it impossible for evidence from any study to overcome this handicap. We focus largely on step (2), explaining how the combination of dichotomization and “bias” dilutes experimental evidence, and showing how this dilution leads inevitably to the stated conclusion. We also demonstrate a fallacy in another important component of the argument –that papers in “hot” fields are more likely to produce false findings. We agree with the paper’s conclusions and recommendations that many medical research findings are less definitive than readers suspect, that P-values are widely misinterpreted, that bias of various forms is widespread, that multiple approaches are needed to prevent the literature from being systematically biased and the need for more data on the prevalence of false claims. But calculating the unreliability of the medical research literature, in whole or in part, requires more empirical evidence and different inferential models than were used. The claim that “most research findings are false for most research designs and for most fields” must be considered as yet unproven.

Aggressive vs. conservative phototherapy for infants with extremely low birth weight.

BACKGROUND: It is unclear whether aggressive phototherapy to prevent neurotoxic effects of bilirubin benefits or harms infants with extremely low birth weight (1000 g or less). METHODS: We randomly assigned 1974 infants with extremely low birth weight at 12 to 36 hours of age to undergo either aggressive or conservative phototherapy. The primary outcome was a composite of death or neurodevelopmental impairment determined for 91% of the infants by investigators who were unaware of the treatment assignments. RESULTS: Aggressive phototherapy, as compared with conservative phototherapy, significantly reduced the mean peak serum bilirubin level (7.0 vs. 9.8 mg per deciliter [120 vs. 168 micromol per liter], P<0.01) but not the rate of the primary outcome (52% vs. 55%; relative risk, 0.94; 95% confidence interval [CI], 0.87 to 1.02; P=0.15). Aggressive phototherapy did reduce rates of neurodevelopmental impairment (26%, vs. 30% for conservative phototherapy; relative risk, 0.86; 95% CI, 0.74 to 0.99). Rates of death in the aggressive-phototherapy and conservative-phototherapy groups were 24% and 23%, respectively (relative risk, 1.05; 95% CI, 0.90 to 1.22). In preplanned subgroup analyses, the rates of death were 13% with aggressive phototherapy and 14% with conservative phototherapy for infants with a birth weight of 751 to 1000 g and 39% and 34%, respectively (relative risk, 1.13; 95% CI, 0.96 to 1.34), for infants with a birth weight of 501 to 750 g. CONCLUSIONS: Aggressive phototherapy did not significantly reduce the rate of death or neurodevelopmental impairment. The rate of neurodevelopmental impairment alone was significantly reduced with aggressive phototherapy. This reduction may be offset by an increase in mortality among infants weighing 501 to 750 g at birth. (ClinicalTrials.gov number, NCT00114543.)

Predicting the onset of psychosis in patients at clinical high risk: practical guide to probabilistic prognostic reasoning

Prediction of psychosis in patients at clinical high risk (CHR) has become a mainstream focus of clinical and research interest worldwide. When using CHR instruments for clinical purposes, the predicted outcome is but only a probability; and, consequently, any therapeutic action following the assessment is based on probabilistic prognostic reasoning. Yet, probabilistic reasoning makes considerable demands on the clinicians. We provide here a scholarly practical guide summarising the key concepts to support clinicians with probabilistic prognostic reasoning in the CHR state. We review risk or cumulative incidence of psychosis in, person-time rate of psychosis, Kaplan-Meier estimates of psychosis risk, measures of prognostic accuracy, sensitivity and specificity in receiver operator characteristic curves, positive and negative predictive values, Bayes’ theorem, likelihood ratios, potentials and limits of real-life applications of prognostic probabilistic reasoning in the CHR state. Understanding basic measures used for prognostic probabilistic reasoning is a prerequisite for successfully implementing the early detection and prevention of psychosis in clinical practice. Future refinement of these measures for CHR patients may actually influence risk management, especially as regards initiating or withholding treatment.

Metodología de análisis estadístico de roturas en redes de distribución de agua

La presente Tesis plantea una metodología de análisis estadístico de roturas de tubería en redes de distribución de agua, que analiza la relación entre las roturas y la presión de agua y que propone la implantación de una gestión de presiones que reduzca el número de roturas que se producen en dichas redes. Las redes de distribución de agua se deterioran y una de sus graves consecuencias es la aparición de roturas frecuentes en sus tuberías. Las roturas llevan asociados elevados costes sociales, económicos y medioambientales y es por ello por lo que las compañías gestoras del agua tratan de reducirlas en la medida de lo posible. Las redes de distribución de agua se pueden dividir en zonas o sectores que facilitan su control y que pueden ser independientes o aislarse mediante válvulas, como ocurre en las redes de países más desarrollados, o pueden estar intercomunicados hidráulicamente. La implantación de una gestión de presiones suele llevarse a cabo a través de las válvulas reductoras de presión (VPR), que se instalan en las cabeceras de estos sectores y que controlan la presión aguas abajo de la misma, aunque varíe su caudal de entrada. Los métodos más conocidos de la gestión de presiones son la reducción de presiones, que es el control más habitual, el mantenimiento de la presión, la prevención y/o alivio de los aumentos repentinos de presión y el establecimiento de un control por alturas. A partir del año 2005 se empezó a reconocer el efecto de la gestión de presiones sobre la disminución de las roturas. En esta Tesis, se sugiere una gestión de presiones que controle los rangos de los indicadores de la presión de cabecera que más influyan en la probabilidad de roturas de tubería. Así, la presión del agua se caracteriza a través de indicadores obtenidos de la presión registrada en la cabecera de los sectores, debido a que se asume que esta presión es representativa de la presión de operación de todas las tuberías porque las pérdidas de carga son relativamente bajas y las diferencias topográficas se tienen en cuenta en el diseño de los sectores. Y los indicadores de presión, que se pueden definir como el estadístico calculado a partir de las series de la presión de cabecera sobre una ventana de tiempo, pueden proveer la información necesaria para ayudar a la toma de decisiones a los gestores del agua con el fin de reducir las roturas de tubería en las redes de distribución de agua. La primera parte de la metodología que se propone en esta Tesis trata de encontrar los indicadores de presión que influyen más en la probabilidad de roturas de tuberías. Para conocer si un indicador es influyente en la probabilidad de las roturas se comparan las estimaciones de las funciones de distribución acumulada (FDAs) de los indicadores de presiones, considerando dos situaciones: cuando se condicionan a la ocurrencia de una rotura (suceso raro) y cuando se calculan en la situación normal de operación (normal operación). Por lo general, las compañías gestoras cuentan con registros de roturas de los años más recientes y al encontrarse las tuberías enterradas se complica el acceso a la información. Por ello, se propone el uso de funciones de probabilidad que permiten reducir la incertidumbre asociada a los datos registrados. De esta forma, se determinan las funciones de distribución acumuladas (FDAs) de los valores del indicador de la serie de presión (situación normal de operación) y las FDAs de los valores del indicador en el momento de ocurrencia de las roturas (condicionado a las roturas). Si las funciones de distribución provienen de la misma población, no se puede deducir que el indicador claramente influya en la probabilidad de roturas. Sin embargo, si se prueba estadísticamente que las funciones proceden de la misma población, se puede concluir que existe una relación entre el indicador analizado y la ocurrencia de las roturas. Debido a que el número de valores del indicador de la FDA condicionada a las roturas es mucho menor que el número de valores del indicador de la FDA incondicional a las roturas, se generan series aleatorias a partir de los valores de los indicadores con el mismo número de valores que roturas registradas hay. De esta forma, se comparan las FDAs de series aleatorias del indicador con la FDA condicionada a las roturas del mismo indicador y se deduce si el indicador es influyente en la probabilidad de las roturas. Los indicadores de presión pueden depender de unos parámetros. A través de un análisis de sensibilidad y aplicando un test estadístico robusto se determina la situación en la que estos parámetros dan lugar a que el indicador sea más influyente en la probabilidad de las roturas. Al mismo tiempo, los indicadores se pueden calcular en función de dos parámetros de cálculo que se denominan el tiempo de anticipación y el ancho de ventana. El tiempo de anticipación es el tiempo (en horas) entre el final del periodo de computación del indicador de presión y la rotura, y el ancho de ventana es el número de valores de presión que se requieren para calcular el indicador de presión y que es múltiplo de 24 horas debido al comportamiento cíclico diario de la presión. Un análisis de sensibilidad de los parámetros de cálculo explica cuándo los indicadores de presión influyen más en la probabilidad de roturas. En la segunda parte de la metodología se presenta un modelo de diagnóstico bayesiano. Este tipo de modelo forma parte de los modelos estadísticos de prevención de roturas, parten de los datos registrados para establecer patrones de fallo y utilizan el teorema de Bayes para determinar la probabilidad de fallo cuando se condiciona la red a unas determinadas características. Así, a través del teorema de Bayes se comparan la FDA genérica del indicador con la FDA condicionada a las roturas y se determina cuándo la probabilidad de roturas aumenta para ciertos rangos del indicador que se ha inferido como influyente en las roturas. Se determina un ratio de probabilidad (RP) que cuando es superior a la unidad permite distinguir cuándo la probabilidad de roturas incrementa para determinados intervalos del indicador. La primera parte de la metodología se aplica a la red de distribución de la Comunidad de Madrid (España) y a la red de distribución de Ciudad de Panamá (Panamá). Tras el filtrado de datos se deduce que se puede aplicar la metodología en 15 sectores en la Comunidad de Madrid y en dos sectores, llamados corregimientos, en Ciudad de Panamá. Los resultados demuestran que en las dos redes los indicadores más influyentes en la probabilidad de las roturas son el rango de la presión, que supone la diferencia entre la presión máxima y la presión mínima, y la variabilidad de la presión, que considera la propiedad estadística de la desviación típica. Se trata, por tanto, de indicadores que hacen referencia a la dispersión de los datos, a la persistencia de la variación de la presión y que se puede asimilar en resistencia de materiales a la fatiga. La segunda parte de la metodología se ha aplicado a los indicadores influyentes en la probabilidad de las roturas de la Comunidad de Madrid y se ha deducido que la probabilidad de roturas aumenta para valores extremos del indicador del rango de la presión y del indicador de la variabilidad de la presión. Finalmente, se recomienda una gestión de presiones que limite los intervalos de los indicadores influyentes en la probabilidad de roturas que incrementen dicha probabilidad. La metodología propuesta puede aplicarse a otras redes de distribución y puede ayudar a las compañías gestoras a reducir el número de fallos en el sistema a través de la gestión de presiones. This Thesis presents a methodology for the statistical analysis of pipe breaks in water distribution networks. The methodology studies the relationship between pipe breaks and water pressure, and proposes a pressure management procedure to reduce the number of breaks that occur in such networks. One of the manifestations of the deterioration of water supply systems is frequent pipe breaks. System failures are one of the major challenges faced by water utilities, due to their associated social, economic and environmental costs. For all these reasons, water utilities aim at reducing the problem of break occurrence to as great an extent as possible. Water distribution networks can be divided into areas or sectors, which facilitates the control of the network. These areas may be independent or isolated by valves, as it usually happens in developing countries. Alternatively, they can be hydraulically interconnected. The implementation of pressure management strategies is usually carried out through pressure-reducing valves (PRV). These valves are installed at the head of the sectors and, although the inflow may vary significantly, they control the downstream pressure. The most popular methods of pressure management consist of pressure reduction, which is the common form of control, pressure sustaining, prevention and/or alleviation of pressure surges or large variations in pressure, and level/altitude control. From 2005 onwards, the effects of pressure management on burst frequencies have become more widely recognized in the technical literature. This thesis suggests a pressure management that controls the pressure indicator ranges most influential on the probability of pipe breaks. Operating pressure in a sector is characterized by means of a pressure indicator at the head of the DMA, as head losses are relatively small and topographical differences were accounted for at the design stage. The pressure indicator, which may be defined as the calculated statistic from the time series of pressure head over a specific time window, may provide necessary information to help water utilities to make decisions to reduce pipe breaks in water distribution networks. The first part of the methodology presented in this Thesis provides the pressure indicators which have the greatest impact on the probability of pipe breaks to be determined. In order to know whether a pressure indicator influences the probability of pipe breaks, the proposed methodology compares estimates of cumulative distribution functions (CDFs) of a pressure indicator through consideration of two situations: when they are conditioned to the occurrence of a pipe break (a rare event), and when they are not (a normal operation). Water utilities usually have a history of failures limited to recent periods of time, and it is difficult to have access to precise information in an underground network. Therefore, the use of distribution functions to address such imprecision of recorded data is proposed. Cumulative distribution functions (CDFs) derived from the time series of pressure indicators (normal operation) and CDFs of indicator values at times coincident with a reported pipe break (conditioned to breaks) are compared. If all estimated CDFs are drawn from the same population, there is no reason to infer that the studied indicator clearly influences the probability of the rare event. However, when it is statistically proven that the estimated CDFs do not come from the same population, the analysed indicator may have an influence on the occurrence of pipe breaks. Due to the fact that the number of indicator values used to estimate the CDF conditioned to breaks is much lower in comparison with the number of indicator values to estimate the CDF of the unconditional pressure series, and that the obtained results depend on the size of the compared samples, CDFs from random sets of the same size sampled from the unconditional indicator values are estimated. Therefore, the comparison between the estimated CDFs of random sets of the indicator and the estimated CDF conditioned to breaks allows knowledge of if the indicator is influential on the probability of pipe breaks. Pressure indicators depend on various parameters. Sensitivity analysis and a robust statistical test allow determining the indicator for which these parameters result most influential on the probability of pipe breaks. At the same time, indicators can be calculated according to two model parameters, named as the anticipation time and the window width. The anticipation time refers to the time (hours) between the end of the period for the computation of the pressure indicator and the break. The window width is the number of instantaneous pressure values required to calculate the pressure indicator and is multiple of 24 hours, as water pressure has a cyclical behaviour which lasts one day. A sensitivity analysis of the model parameters explains when the pressure indicator is more influential on the probability of pipe breaks. The second part of the methodology presents a Bayesian diagnostic model. This kind of model belongs to the class of statistical predictive models, which are based on historical data, represent break behavior and patterns in water mains, and use the Bayes’ theorem to condition the probability of failure to specific system characteristics. The Bayes’ theorem allows comparing the break-conditioned FDA and the unconditional FDA of the indicators and determining when the probability of pipe breaks increases for certain pressure indicator ranges. A defined probability ratio provides a measure to establish whether the probability of breaks increases for certain ranges of the pressure indicator. The first part of the methodology is applied to the water distribution network of Madrid (Spain) and to the water distribution network of Panama City (Panama). The data filtering method suggests that the methodology can be applied to 15 sectors in Madrid and to two areas in Panama City. The results show that, in both systems, the most influential indicators on the probability of pipe breaks are the pressure range, which is the difference between the maximum pressure and the minimum pressure, and pressure variability, referred to the statistical property of the standard deviation. Therefore, they represent the dispersion of the data, the persistence of the variation in pressure and may be related to the fatigue in material resistance. The second part of the methodology has been applied to the influential indicators on the probability of pipe breaks in the water distribution network of Madrid. The main conclusion is that the probability of pipe breaks increases for the extreme values of the pressure range indicator and of the pressure variability indicator. Finally, a pressure management which limits the ranges of the pressure indicators influential on the probability of pipe breaks that increase such probability is recommended. The methodology presented here is general, may be applied to other water distribution networks, and could help water utilities reduce the number of system failures through pressure management.

Análise de portfólio: uma perspectiva bayesiana

Este trabalho tem com objetivo abordar o problema de alocação de ativos (análise de portfólio) sob uma ótica Bayesiana. Para isto foi necessário revisar toda a análise teórica do modelo clássico de média-variância e na sequencia identificar suas deficiências que comprometem sua eficácia em casos reais. Curiosamente, sua maior deficiência não esta relacionado com o próprio modelo e sim pelos seus dados de entrada em especial ao retorno esperado calculado com dados históricos. Para superar esta deficiência a abordagem Bayesiana (modelo de Black-Litterman) trata o retorno esperado como uma variável aleatória e na sequência constrói uma distribuição a priori (baseado no modelo de CAPM) e uma distribuição de verossimilhança (baseado na visão de mercado sob a ótica do investidor) para finalmente aplicar o teorema de Bayes tendo como resultado a distribuição a posteriori. O novo valor esperado do retorno, que emerge da distribuição a posteriori, é que substituirá a estimativa anterior do retorno esperado calculado com dados históricos. Os resultados obtidos mostraram que o modelo Bayesiano apresenta resultados conservadores e intuitivos em relação ao modelo clássico de média-variância.

The STARD statement for reporting studies of diagnostic accuracy: Explanation and elaboration

The quality of reporting of studies of diagnostic accuracy is less than optimal. Complete and accurate reporting is necessary to enable readers to assess the potential for bias in the study and to evaluate the generalisability of the results. A group of scientists and editors has developed the STARD (Standards for Reporting of Diagnostic Accuracy) statement to improve the reporting the quality of reporting of studies of diagnostic accuracy. The statement consists of a checklist of 25 items and flow diagram that authors can use to ensure that all relevant information is present. This explanatory document aims to facilitate the use, understanding and dissemination of the checklist. The document contains a clarification of the meaning, rationale and optimal use of each item on the checklist, as well as a short summary of the available evidence on bias and applicability. The STARD statement, checklist, flowchart and this explanation and elaboration document should be useful resources to improve reporting of diagnostic accuracy studies. Complete and informative reporting can only lead to better decisions in healthcare.

Distribution modeling of nonlinear inverse controllers under a Bayesian framework

The inverse controller is traditionally assumed to be a deterministic function. This paper presents a pedagogical methodology for estimating the stochastic model of the inverse controller. The proposed method is based on Bayes' theorem. Using Bayes' rule to obtain the stochastic model of the inverse controller allows the use of knowledge of uncertainty from both the inverse and the forward model in estimating the optimal control signal. The paper presents the methodology for general nonlinear systems and is demonstrated on nonlinear single-input-single-output (SISO) and multiple-input-multiple-output (MIMO) examples. © 2006 IEEE.

A Bayesian approach to modeling the conditional density of the inverse controller

The inverse controller is traditionally assumed to be a deterministic function. This paper presents a pedagogical methodology for estimating the stochastic model of the inverse controller. The proposed method is based on Bayes' theorem. Using Bayes' rule to obtain the stochastic model of the inverse controller allows the use of knowledge of uncertainty from both the inverse and the forward model in estimating the optimal control signal. The paper presents the methodology for general nonlinear systems. For illustration purposes, the proposed methodology is applied to linear Gaussian systems. © 2004 IEEE.

Reverse Engineering of Modified Genes by Bayesian Network Analysis Defines Molecular Determinants Critical to the Development of Glioblastoma

In this study we have identified key genes that are critical in development of astrocytic tumors. Meta-analysis of microarray studies which compared normal tissue to astrocytoma revealed a set of 646 differentially expressed genes in the majority of astrocytoma. Reverse engineering of these 646 genes using Bayesian network analysis produced a gene network for each grade of astrocytoma (Grade I–IV), and ‘key genes’ within each grade were identified. Genes found to be most influential to development of the highest grade of astrocytoma, Glioblastoma multiforme were: COL4A1, EGFR, BTF3, MPP2, RAB31, CDK4, CD99, ANXA2, TOP2A, and SERBP1. All of these genes were up-regulated, except MPP2 (down regulated). These 10 genes were able to predict tumor status with 96–100% confidence when using logistic regression, cross validation, and the support vector machine analysis. Markov genes interact with NFkβ, ERK, MAPK, VEGF, growth hormone and collagen to produce a network whose top biological functions are cancer, neurological disease, and cellular movement. Three of the 10 genes - EGFR, COL4A1, and CDK4, in particular, seemed to be potential ‘hubs of activity’. Modified expression of these 10 Markov Blanket genes increases lifetime risk of developing glioblastoma compared to the normal population. The glioblastoma risk estimates were dramatically increased with joint effects of 4 or more than 4 Markov Blanket genes. Joint interaction effects of 4, 5, 6, 7, 8, 9 or 10 Markov Blanket genes produced 9, 13, 20.9, 26.7, 52.8, 53.2, 78.1 or 85.9%, respectively, increase in lifetime risk of developing glioblastoma compared to normal population. In summary, it appears that modified expression of several ‘key genes’ may be required for the development of glioblastoma. Further studies are needed to validate these ‘key genes’ as useful tools for early detection and novel therapeutic options for these tumors.

Global, regional, and national incidence, prevalence, and years lived with disability for 310 diseases and injuries, 1990–2015 : a systematic analysis for the Global Burden of Disease Study 2015

Índice acumulativo de riesgo (nice) como prueba diagnóstica con base en tomografía corneal por elevación en la prevención de ectasia corneal post-lasik

Antecedentes La ectasia corneal post-lasik (ECPL) es una complicación infrecuente, pero devastadora en la cirugía lasik (queratomileusis asistida con éxcimer láser) para el tratamiento de la miopía con o sin astigmatismo. Con base en la tomografía corneal por elevación por imágenes de Scheimpflug (Sistema Pentacam HR, Oculus Wetzlar, Alemania), se propone un novedoso índice acumulativo de riesgo para ser utilizado como prueba diagnóstica de tamizaje y así prevenir esta complicación. Metodología Se realizó un estudio observacional analítico, de corte transversal tipo pruebas diagnósticas, con el fin de evaluar las características operativas del índice NICE teniendo como estándar de referencia el módulo de Belin-Ambrosio (Pentacam HR) utilizando un modelo de regresión logística binaria, tablas de contingencia y estimando el área bajo la curva ROC. Resultados Se evaluaron 361 ojos de los cuales el 59,3% provenían de pacientes de sexo femenino, la edad media global fue de 30 años (RIC 11,0). El modelo logístico binario aplicado se construyó con base en cuatro variables independientes cuantitativas (K2, PAQUI, EP e I-S) y una cualitativa (SEXO), y se determinó su relación con la variable dependiente, NICE (puntaje final). Las variables predictoras fueron estadísticamente significativas clasificando adecuadamente el 92,9% de los ojos evaluados según presencia o ausencia de riesgo. El coeficiente de Nagelkerke fue de 74,4%. Conclusiones El índice acumulativo de riesgo NICE es una herramienta diagnóstica novedosa en la evaluación de candidatos a cirugía refractiva lasik para prevenir la ectasia secundaria.

Construction of nonparametric Bayesian models from parametric Bayes equations

We consider the general problem of constructing nonparametric Bayesian models on infinite-dimensional random objects, such as functions, infinite graphs or infinite permutations. The problem has generated much interest in machine learning, where it is treated heuristically, but has not been studied in full generality in non-parametric Bayesian statistics, which tends to focus on models over probability distributions. Our approach applies a standard tool of stochastic process theory, the construction of stochastic processes from their finite-dimensional marginal distributions. The main contribution of the paper is a generalization of the classic Kolmogorov extension theorem to conditional probabilities. This extension allows a rigorous construction of nonparametric Bayesian models from systems of finite-dimensional, parametric Bayes equations. Using this approach, we show (i) how existence of a conjugate posterior for the nonparametric model can be guaranteed by choosing conjugate finite-dimensional models in the construction, (ii) how the mapping to the posterior parameters of the nonparametric model can be explicitly determined, and (iii) that the construction of conjugate models in essence requires the finite-dimensional models to be in the exponential family. As an application of our constructive framework, we derive a model on infinite permutations, the nonparametric Bayesian analogue of a model recently proposed for the analysis of rank data.

Bayes and empirical-Bayes multiplicity adjustment in the variable-selection problem

This paper studies the multiplicity-correction effect of standard Bayesian variable-selection priors in linear regression. Our first goal is to clarify when, and how, multiplicity correction happens automatically in Bayesian analysis, and to distinguish this correction from the Bayesian Ockham's-razor effect. Our second goal is to contrast empirical-Bayes and fully Bayesian approaches to variable selection through examples, theoretical results and simulations. Considerable differences between the two approaches are found. In particular, we prove a theorem that characterizes a surprising aymptotic discrepancy between fully Bayes and empirical Bayes. This discrepancy arises from a different source than the failure to account for hyperparameter uncertainty in the empirical-Bayes estimate. Indeed, even at the extreme, when the empirical-Bayes estimate converges asymptotically to the true variable-inclusion probability, the potential for a serious difference remains. © Institute of Mathematical Statistics, 2010.

Bayes quadratic unbiased estimator of spatial covariance parameters

This work presents Bayes invariant quadratic unbiased estimator, for short BAIQUE. Bayesian approach is used here to estimate the covariance functions of the regionalized variables which appear in the spatial covariance structure in mixed linear model. Firstly a brief review of spatial process, variance covariance components structure and Bayesian inference is given, since this project deals with these concepts. Then the linear equations model corresponding to BAIQUE in the general case is formulated. That Bayes estimator of variance components with too many unknown parameters is complicated to be solved analytically. Hence, in order to facilitate the handling with this system, BAIQUE of spatial covariance model with two parameters is considered. Bayesian estimation arises as a solution of a linear equations system which requires the linearity of the covariance functions in the parameters. Here the availability of prior information on the parameters is assumed. This information includes apriori distribution functions which enable to find the first and the second moments matrix. The Bayesian estimation suggested here depends only on the second moment of the prior distribution. The estimation appears as a quadratic form y'Ay , where y is the vector of filtered data observations. This quadratic estimator is used to estimate the linear function of unknown variance components. The matrix A of BAIQUE plays an important role. If such a symmetrical matrix exists, then Bayes risk becomes minimal and the unbiasedness conditions are fulfilled. Therefore, the symmetry of this matrix is elaborated in this work. Through dealing with the infinite series of matrices, a representation of the matrix A is obtained which shows the symmetry of A. In this context, the largest singular value of the decomposed matrix of the infinite series is considered to deal with the convergence condition and also it is connected with Gerschgorin Discs and Poincare theorem. Then the BAIQUE model for some experimental designs is computed and compared. The comparison deals with different aspects, such as the influence of the position of the design points in a fixed interval. The designs that are considered are those with their points distributed in the interval [0, 1]. These experimental structures are compared with respect to the Bayes risk and norms of the matrices corresponding to distances, covariance structures and matrices which have to satisfy the convergence condition. Also different types of the regression functions and distance measurements are handled. The influence of scaling on the design points is studied, moreover, the influence of the covariance structure on the best design is investigated and different covariance structures are considered. Finally, BAIQUE is applied for real data. The corresponding outcomes are compared with the results of other methods for the same data. Thereby, the special BAIQUE, which estimates the general variance of the data, achieves a very close result to the classical empirical variance.

Improving Multi-class Text Classification with Naive Bayes

There are numerous text documents available in electronic form. More and more are becoming available every day. Such documents represent a massive amount of information that is easily accessible. Seeking value in this huge collection requires organization; much of the work of organizing documents can be automated through text classification. The accuracy and our understanding of such systems greatly influences their usefulness. In this paper, we seek 1) to advance the understanding of commonly used text classification techniques, and 2) through that understanding, improve the tools that are available for text classification. We begin by clarifying the assumptions made in the derivation of Naive Bayes, noting basic properties and proposing ways for its extension and improvement. Next, we investigate the quality of Naive Bayes parameter estimates and their impact on classification. Our analysis leads to a theorem which gives an explanation for the improvements that can be found in multiclass classification with Naive Bayes using Error-Correcting Output Codes. We use experimental evidence on two commonly-used data sets to exhibit an application of the theorem. Finally, we show fundamental flaws in a commonly-used feature selection algorithm and develop a statistics-based framework for text feature selection. Greater understanding of Naive Bayes and the properties of text allows us to make better use of it in text classification.