- The document discusses nonparametric density estimation for data with non-negative support, such as data from reliability testing.
- It proposes several density estimators based on an approximation lemma, including using asymmetric kernels and distributions placed on lattice points.
- The estimators are motivated by replacing the empirical distribution function in the lemma with a smoothed version, yielding a smoothed density estimator.
- Asymptotic properties of the estimators are established, and simulations compare their performance.
This document summarizes Yogendra Chaubey's upcoming talk on nonparametric density estimation for size biased data. It will highlight recent developments in this area, with an emphasis on density estimation when the data is subject to constraints that traditional estimators may not satisfy. It describes how Hille's approximation lemma can be used to propose alternative smooth density estimators. It will also present the results of a simulation study comparing various nonparametric density estimators and their asymptotic properties.
The document discusses smoothing parameter selection for density estimation from length-biased data using asymmetric kernels. It reviews recent work on applying Bayesian criteria for this purpose. The key points are: 1) Asymmetric kernels are better for density estimation of non-negative data as they avoid positive mass in negative regions. 2) Length-biased data arises in situations where observations are weighted by their values. 3) Estimating the underlying density from length-biased data requires adjusting for the bias. 4) Bayesian methods provide an approach for selecting the smoothing parameter for asymmetric kernel density estimators applied to length-biased data.
A Novel Bayes Factor for Inverse Model Selection Problem based on Inverse Ref...inventionjournals
Statistical model selection problem can be divided into two broad categories based on Forward and Inverse problem. Compared to a wealthy of literature available for Forward model selection, there are very few methods applicable for Inverse model selection context. In this article we propose a novel Bayes factor for model selection in Bayesian Inverse Problem context. The proposed Bayes Factor is specially designed for Inverse problem with the help of Inverse Reference Distribution (IRD). We will discuss our proposal from decision theoretic perspective.
Some sampling techniques for big data analysisJae-kwang Kim
This document describes different sampling techniques for big data analysis, including reservoir sampling and its variants. It provides an example to illustrate simple random sampling and calculates the expected value and variance of sampling errors. It then discusses probability sampling and its advantages over non-probability sampling. The document also introduces survey sampling and challenges in the era of big data, as well as how sampling techniques can still be useful for handling big data. It outlines reservoir sampling and two methods to improve it: balanced reservoir sampling and stratified reservoir sampling. A simulation study is described to compare the performance of these reservoir sampling methods.
The document discusses the benefits of exercise for mental health. Regular physical activity can help reduce anxiety and depression and improve mood and cognitive functioning. Exercise causes chemical changes in the brain that may help protect against mental illness and improve symptoms.
This document summarizes several statistical methods for handling non-ignorable nonresponse in data, including maximum likelihood estimation, partial likelihood approaches, generalized method of moments, and exponential tilting methods. It discusses full likelihood-based maximum likelihood estimation using the observed data likelihood and EM algorithm. Partial likelihood approaches like conditional likelihood and pseudo likelihood are presented as alternatives that use a subset of the observed data.
This document discusses predictive mean matching (PMM) imputation in survey sampling. It begins with an outline and overview of the basic setup, assumptions, and PMM imputation method. It then presents three main theorems: 1) the asymptotic normality of the PMM estimator when the regression parameter β* is known, 2) the asymptotic normality when β* is estimated, and 3) the asymptotic properties of nearest neighbor imputation. The document also discusses variance estimation for the PMM estimator using replication methods like the bootstrap or jackknife. In summary, it provides a theoretical analysis of the asymptotic properties of PMM imputation and approaches for estimating the variance.
This document summarizes Yogendra Chaubey's upcoming talk on nonparametric density estimation for size biased data. It will highlight recent developments in this area, with an emphasis on density estimation when the data is subject to constraints that traditional estimators may not satisfy. It describes how Hille's approximation lemma can be used to propose alternative smooth density estimators. It will also present the results of a simulation study comparing various nonparametric density estimators and their asymptotic properties.
The document discusses smoothing parameter selection for density estimation from length-biased data using asymmetric kernels. It reviews recent work on applying Bayesian criteria for this purpose. The key points are: 1) Asymmetric kernels are better for density estimation of non-negative data as they avoid positive mass in negative regions. 2) Length-biased data arises in situations where observations are weighted by their values. 3) Estimating the underlying density from length-biased data requires adjusting for the bias. 4) Bayesian methods provide an approach for selecting the smoothing parameter for asymmetric kernel density estimators applied to length-biased data.
A Novel Bayes Factor for Inverse Model Selection Problem based on Inverse Ref...inventionjournals
Statistical model selection problem can be divided into two broad categories based on Forward and Inverse problem. Compared to a wealthy of literature available for Forward model selection, there are very few methods applicable for Inverse model selection context. In this article we propose a novel Bayes factor for model selection in Bayesian Inverse Problem context. The proposed Bayes Factor is specially designed for Inverse problem with the help of Inverse Reference Distribution (IRD). We will discuss our proposal from decision theoretic perspective.
Some sampling techniques for big data analysisJae-kwang Kim
This document describes different sampling techniques for big data analysis, including reservoir sampling and its variants. It provides an example to illustrate simple random sampling and calculates the expected value and variance of sampling errors. It then discusses probability sampling and its advantages over non-probability sampling. The document also introduces survey sampling and challenges in the era of big data, as well as how sampling techniques can still be useful for handling big data. It outlines reservoir sampling and two methods to improve it: balanced reservoir sampling and stratified reservoir sampling. A simulation study is described to compare the performance of these reservoir sampling methods.
The document discusses the benefits of exercise for mental health. Regular physical activity can help reduce anxiety and depression and improve mood and cognitive functioning. Exercise causes chemical changes in the brain that may help protect against mental illness and improve symptoms.
This document summarizes several statistical methods for handling non-ignorable nonresponse in data, including maximum likelihood estimation, partial likelihood approaches, generalized method of moments, and exponential tilting methods. It discusses full likelihood-based maximum likelihood estimation using the observed data likelihood and EM algorithm. Partial likelihood approaches like conditional likelihood and pseudo likelihood are presented as alternatives that use a subset of the observed data.
This document discusses predictive mean matching (PMM) imputation in survey sampling. It begins with an outline and overview of the basic setup, assumptions, and PMM imputation method. It then presents three main theorems: 1) the asymptotic normality of the PMM estimator when the regression parameter β* is known, 2) the asymptotic normality when β* is estimated, and 3) the asymptotic properties of nearest neighbor imputation. The document also discusses variance estimation for the PMM estimator using replication methods like the bootstrap or jackknife. In summary, it provides a theoretical analysis of the asymptotic properties of PMM imputation and approaches for estimating the variance.
Bayesian inference for mixed-effects models driven by SDEs and other stochast...Umberto Picchini
An important, and well studied, class of stochastic models is given by stochastic differential equations (SDEs). In this talk, we consider Bayesian inference based on measurements from several individuals, to provide inference at the "population level" using mixed-effects modelling. We consider the case where dynamics are expressed via SDEs or other stochastic (Markovian) models. Stochastic differential equation mixed-effects models (SDEMEMs) are flexible hierarchical models that account for (i) the intrinsic random variability in the latent states dynamics, as well as (ii) the variability between individuals, and also (iii) account for measurement error. This flexibility gives rise to methodological and computational difficulties.
Fully Bayesian inference for nonlinear SDEMEMs is complicated by the typical intractability of the observed data likelihood which motivates the use of sampling-based approaches such as Markov chain Monte Carlo. A Gibbs sampler is proposed to target the marginal posterior of all parameters of interest. The algorithm is made computationally efficient through careful use of blocking strategies, particle filters (sequential Monte Carlo) and correlated pseudo-marginal approaches. The resulting methodology is is flexible, general and is able to deal with a large class of nonlinear SDEMEMs [1]. In a more recent work [2], we also explored ways to make inference even more scalable to an increasing number of individuals, while also dealing with state-space models driven by other stochastic dynamic models than SDEs, eg Markov jump processes and nonlinear solvers typically used in systems biology.
[1] S. Wiqvist, A. Golightly, AT McLean, U. Picchini (2020). Efficient inference for stochastic differential mixed-effects models using correlated particle pseudo-marginal algorithms, CSDA, https://doi.org/10.1016/j.csda.2020.107151
[2] S. Persson, N. Welkenhuysen, S. Shashkova, S. Wiqvist, P. Reith, G. W. Schmidt, U. Picchini, M. Cvijovic (2021). PEPSDI: Scalable and flexible inference framework for stochastic dynamic single-cell models, bioRxiv doi:10.1101/2021.07.01.450748.
On Continuous Approximate Solution of Ordinary Differential EquationsWaqas Tariq
In this work the problem of continuous approximate solution of the ordinary differential equations will be investigated. An approach to construct the continuous approximate solution, which is based on the discrete approximate solution and the spline interpolation, will be provided. The existence and uniqueness of such continuous approximate solution will be pointed out. Its error will be estimated and its convergence will be considered. Finally, with the aid of modern PC and nathematical software three practical computer approaches to perform above construction will be offered.
This document outlines a presentation on linear regression and correlation. It discusses simple linear regression models, how to estimate parameters using least squares, and how to perform hypothesis tests and calculate confidence intervals. It also covers correlation and using transformations to achieve linear regression models. Examples are provided to demonstrate key concepts like fitting a regression line to oxygen purity data and testing hypotheses.
This document provides motivation for using a circular kernel density estimator for nonparametric density estimation of circular data. It describes how a simple approximation theory from linear kernel estimation can be adapted to the circular case by replacing the kernel with a sequence of periodic densities on [-π,π] that converge to a degenerate distribution at θ=0. It shows that the wrapped Cauchy density satisfies the conditions to serve as such a kernel, resulting in the circular kernel density estimator proposed in equation 1.12. This estimator is shown to converge uniformly to the true density f(θ) as the sample size increases, providing theoretical justification for its use in smooth nonparametric density estimation for circular variables.
Big Data analysis involves building predictive models from high-dimensional data using techniques like variable selection, cross-validation, and regularization to avoid overfitting. The document discusses an example analyzing web browsing data to predict online spending, highlighting challenges with large numbers of variables. It also covers summarizing high-dimensional data through dimension reduction and model building for prediction versus causal inference.
We provide an overview of some recent developments in machine learning tools for dynamic treatment regime discovery in precision medicine. The first development is a new off-policy reinforcement learning tool for continual learning in mobile health to enable patients with type 1 diabetes to exercise safely. The second development is a new inverse reinforcement learning tools which enables use of observational data to learn how clinicians balance competing priorities for treating depression and mania in patients with bipolar disorder. Both practical and technical challenges are discussed.
This document proposes an approximate Bayesian inference method for estimating propensity scores under nonresponse. It involves treating the estimating equations as random variables and assigning a prior distribution to the transformed parameters. Samples are drawn from the posterior distribution of the parameters given the observed data to make inferences. The method is shown to be asymptotically consistent and confidence regions can be constructed from the posterior samples. Extensions are discussed to incorporate auxiliary variables and perform Bayesian model selection by assigning a spike-and-slab prior over the model parameters.
Recently, the machine learning community has expressed strong interest in applying latent variable modeling strategies to causal inference problems with unobserved confounding. Here, I discuss one of the big debates that occurred over the past year, and how we can move forward. I will focus specifically on the failure of point identification in this setting, and discuss how this can be used to design flexible sensitivity analyses that cleanly separate identified and unidentified components of the causal model.
This document summarizes key concepts from Chapter 2 of a book on statistical methods for handling incomplete data. It introduces the likelihood-based approach and defines key terms like the likelihood function, maximum likelihood estimator, Fisher information, and missing at random. The chapter also provides examples of observed likelihood functions for censored regression and survival analysis models with missing data.
Solution of second order nonlinear singular value problemsAlexander Decker
1) The document presents a method for finding solutions to second order nonlinear singular boundary value problems using Taylor series.
2) The method modifies the differential equation to handle singularities, then obtains recurrence relations for the Taylor series coefficients by taking derivatives at the singular point.
3) Examples are provided to demonstrate the method, yielding exact solutions for problems in astronomy modeling gas spheres and other physical applications.
International Journal of Computational Engineering Research(IJCER)ijceronline
The document presents some fixed point theorems for expansion mappings in complete metric spaces. It begins with definitions of terms like metric spaces, complete metric spaces, Cauchy sequences, and expansion mappings. It then summarizes several existing fixed point theorems for expansion mappings established by other mathematicians. The main result proved in this document is Theorem 3.1, which establishes a new fixed point theorem for expansion mappings under certain conditions on the metric space and mapping. It shows that if the mapping satisfies the given inequality, then it has a fixed point. The proof of this theorem constructs a sequence to show that it converges to a fixed point.
This document proposes representing hypothesis testing problems as estimating mixture models. Specifically, two competing models are embedded within an encompassing mixture model with a weight parameter between 0 and 1. Inference is then drawn on the mixture representation, treating each observation as coming from the mixture model. This avoids difficulties with traditional Bayesian testing approaches like computing marginal likelihoods. It also allows for a more intuitive interpretation of the weight parameter compared to posterior model probabilities. The weight parameter can be estimated using standard mixture estimation algorithms like Gibbs sampling or Metropolis-Hastings. Several illustrations of the approach are provided, including comparisons of Poisson and geometric distributions.
We present recent advances and statistical developments for evaluating Dynamic Treatment Regimes (DTR), which allow the treatment to be dynamically tailored according to evolving subject-level data. Identification of an optimal DTR is a key component for precision medicine and personalized health care. Specific topics covered in this talk include several recent projects with robust and flexible methods developed for the above research area. We will first introduce a dynamic statistical learning method, adaptive contrast weighted learning (ACWL), which combines doubly robust semiparametric regression estimators with flexible machine learning methods. We will further develop a tree-based reinforcement learning (T-RL) method, which builds an unsupervised decision tree that maintains the nature of batch-mode reinforcement learning. Unlike ACWL, T-RL handles the optimization problem with multiple treatment comparisons directly through a purity measure constructed with augmented inverse probability weighted estimators. T-RL is robust, efficient and easy to interpret for the identification of optimal DTRs. However, ACWL seems more robust against tree-type misspecification than T-RL when the true optimal DTR is non-tree-type. At the end of this talk, we will also present a new Stochastic-Tree Search method called ST-RL for evaluating optimal DTRs.
Handling missing data with expectation maximization algorithmLoc Nguyen
Expectation maximization (EM) algorithm is a powerful mathematical tool for estimating parameter of statistical models in case of incomplete data or hidden data. EM assumes that there is a relationship between hidden data and observed data, which can be a joint distribution or a mapping function. Therefore, this implies another implicit relationship between parameter estimation and data imputation. If missing data which contains missing values is considered as hidden data, it is very natural to handle missing data by EM algorithm. Handling missing data is not a new research but this report focuses on the theoretical base with detailed mathematical proofs for fulfilling missing values with EM. Besides, multinormal distribution and multinomial distribution are the two sample statistical models which are concerned to hold missing values.
This document discusses Bayesian approaches to combining evidence from multiple data sources or models. It recommends combining data probabilistically using Bayes' theorem rather than averaging. It provides an example of combining three data sources on annual rainfall measurements by treating the data sources as independent measurements and deriving the posterior distribution of rainfall amounts given the data. It also discusses challenges that arise when combining dependent data sources or models, and presents examples of hierarchical modeling approaches.
This document discusses generalized linear mixed models (GLMMs). It begins with examples of GLMM applications and definitions of key terms. The document then covers estimation methods for GLMMs, including maximum likelihood estimation, integrated likelihood, and both deterministic and stochastic approaches. Inference for GLMMs and remaining challenges are also mentioned. The overall document provides an overview of GLMM frameworks, examples, estimation techniques, and open questions.
BINARY TREE SORT IS MORE ROBUST THAN QUICK SORT IN AVERAGE CASEIJCSEA Journal
This document discusses the average case complexity of binary tree sort compared to quicksort. It analyzes the robustness of binary tree sort's O(n log n) average case complexity for non-uniform input distributions like binomial, Poisson, discrete uniform, continuous uniform, and standard normal distributions. Through statistical modeling and simulation experiments on different sample sizes, it is shown that binary tree sort maintains O(n log n) average case complexity even for non-uniform inputs, demonstrating that it is more robust than quicksort in the average case. The document concludes that only algorithms with equal worst case and average case complexities can reliably depend on average complexity measures, otherwise robustness must be verified.
The document presents a dynamic discrete choice model of demand for insecticide treated nets (ITNs) that accounts for time inconsistent preferences and unobserved heterogeneity. The model has three periods where agents make ITN purchase and retreatment decisions. Agents are either time consistent, "naive" time inconsistent, or "sophisticated" time inconsistent. The model is identified in two steps - first when types are directly observed using survey responses, and second when types are unobserved. Identification exploits variation from elicited beliefs about malaria risk. The model can point identify time preference parameters and utility functions up to a normalization.
A empresa CSM fabrica produtos flexíveis industriais e desenvolve soluções para problemas de clientes. Apresenta casos em que criou proteções sanfonadas que podem ser instaladas sem parar a máquina e reduziram em até 90% o tempo de montagem.
O documento discute o uso de recursos computacionais no marketing, mencionando o comércio eletrônico e suas vantagens e desvantagens. Também aborda como as empresas podem se preparar para o sucesso no ambiente digital, como criando banco de dados de clientes e colocando banners em sites relacionados.
Bayesian inference for mixed-effects models driven by SDEs and other stochast...Umberto Picchini
An important, and well studied, class of stochastic models is given by stochastic differential equations (SDEs). In this talk, we consider Bayesian inference based on measurements from several individuals, to provide inference at the "population level" using mixed-effects modelling. We consider the case where dynamics are expressed via SDEs or other stochastic (Markovian) models. Stochastic differential equation mixed-effects models (SDEMEMs) are flexible hierarchical models that account for (i) the intrinsic random variability in the latent states dynamics, as well as (ii) the variability between individuals, and also (iii) account for measurement error. This flexibility gives rise to methodological and computational difficulties.
Fully Bayesian inference for nonlinear SDEMEMs is complicated by the typical intractability of the observed data likelihood which motivates the use of sampling-based approaches such as Markov chain Monte Carlo. A Gibbs sampler is proposed to target the marginal posterior of all parameters of interest. The algorithm is made computationally efficient through careful use of blocking strategies, particle filters (sequential Monte Carlo) and correlated pseudo-marginal approaches. The resulting methodology is is flexible, general and is able to deal with a large class of nonlinear SDEMEMs [1]. In a more recent work [2], we also explored ways to make inference even more scalable to an increasing number of individuals, while also dealing with state-space models driven by other stochastic dynamic models than SDEs, eg Markov jump processes and nonlinear solvers typically used in systems biology.
[1] S. Wiqvist, A. Golightly, AT McLean, U. Picchini (2020). Efficient inference for stochastic differential mixed-effects models using correlated particle pseudo-marginal algorithms, CSDA, https://doi.org/10.1016/j.csda.2020.107151
[2] S. Persson, N. Welkenhuysen, S. Shashkova, S. Wiqvist, P. Reith, G. W. Schmidt, U. Picchini, M. Cvijovic (2021). PEPSDI: Scalable and flexible inference framework for stochastic dynamic single-cell models, bioRxiv doi:10.1101/2021.07.01.450748.
On Continuous Approximate Solution of Ordinary Differential EquationsWaqas Tariq
In this work the problem of continuous approximate solution of the ordinary differential equations will be investigated. An approach to construct the continuous approximate solution, which is based on the discrete approximate solution and the spline interpolation, will be provided. The existence and uniqueness of such continuous approximate solution will be pointed out. Its error will be estimated and its convergence will be considered. Finally, with the aid of modern PC and nathematical software three practical computer approaches to perform above construction will be offered.
This document outlines a presentation on linear regression and correlation. It discusses simple linear regression models, how to estimate parameters using least squares, and how to perform hypothesis tests and calculate confidence intervals. It also covers correlation and using transformations to achieve linear regression models. Examples are provided to demonstrate key concepts like fitting a regression line to oxygen purity data and testing hypotheses.
This document provides motivation for using a circular kernel density estimator for nonparametric density estimation of circular data. It describes how a simple approximation theory from linear kernel estimation can be adapted to the circular case by replacing the kernel with a sequence of periodic densities on [-π,π] that converge to a degenerate distribution at θ=0. It shows that the wrapped Cauchy density satisfies the conditions to serve as such a kernel, resulting in the circular kernel density estimator proposed in equation 1.12. This estimator is shown to converge uniformly to the true density f(θ) as the sample size increases, providing theoretical justification for its use in smooth nonparametric density estimation for circular variables.
Big Data analysis involves building predictive models from high-dimensional data using techniques like variable selection, cross-validation, and regularization to avoid overfitting. The document discusses an example analyzing web browsing data to predict online spending, highlighting challenges with large numbers of variables. It also covers summarizing high-dimensional data through dimension reduction and model building for prediction versus causal inference.
We provide an overview of some recent developments in machine learning tools for dynamic treatment regime discovery in precision medicine. The first development is a new off-policy reinforcement learning tool for continual learning in mobile health to enable patients with type 1 diabetes to exercise safely. The second development is a new inverse reinforcement learning tools which enables use of observational data to learn how clinicians balance competing priorities for treating depression and mania in patients with bipolar disorder. Both practical and technical challenges are discussed.
This document proposes an approximate Bayesian inference method for estimating propensity scores under nonresponse. It involves treating the estimating equations as random variables and assigning a prior distribution to the transformed parameters. Samples are drawn from the posterior distribution of the parameters given the observed data to make inferences. The method is shown to be asymptotically consistent and confidence regions can be constructed from the posterior samples. Extensions are discussed to incorporate auxiliary variables and perform Bayesian model selection by assigning a spike-and-slab prior over the model parameters.
Recently, the machine learning community has expressed strong interest in applying latent variable modeling strategies to causal inference problems with unobserved confounding. Here, I discuss one of the big debates that occurred over the past year, and how we can move forward. I will focus specifically on the failure of point identification in this setting, and discuss how this can be used to design flexible sensitivity analyses that cleanly separate identified and unidentified components of the causal model.
This document summarizes key concepts from Chapter 2 of a book on statistical methods for handling incomplete data. It introduces the likelihood-based approach and defines key terms like the likelihood function, maximum likelihood estimator, Fisher information, and missing at random. The chapter also provides examples of observed likelihood functions for censored regression and survival analysis models with missing data.
Solution of second order nonlinear singular value problemsAlexander Decker
1) The document presents a method for finding solutions to second order nonlinear singular boundary value problems using Taylor series.
2) The method modifies the differential equation to handle singularities, then obtains recurrence relations for the Taylor series coefficients by taking derivatives at the singular point.
3) Examples are provided to demonstrate the method, yielding exact solutions for problems in astronomy modeling gas spheres and other physical applications.
International Journal of Computational Engineering Research(IJCER)ijceronline
The document presents some fixed point theorems for expansion mappings in complete metric spaces. It begins with definitions of terms like metric spaces, complete metric spaces, Cauchy sequences, and expansion mappings. It then summarizes several existing fixed point theorems for expansion mappings established by other mathematicians. The main result proved in this document is Theorem 3.1, which establishes a new fixed point theorem for expansion mappings under certain conditions on the metric space and mapping. It shows that if the mapping satisfies the given inequality, then it has a fixed point. The proof of this theorem constructs a sequence to show that it converges to a fixed point.
This document proposes representing hypothesis testing problems as estimating mixture models. Specifically, two competing models are embedded within an encompassing mixture model with a weight parameter between 0 and 1. Inference is then drawn on the mixture representation, treating each observation as coming from the mixture model. This avoids difficulties with traditional Bayesian testing approaches like computing marginal likelihoods. It also allows for a more intuitive interpretation of the weight parameter compared to posterior model probabilities. The weight parameter can be estimated using standard mixture estimation algorithms like Gibbs sampling or Metropolis-Hastings. Several illustrations of the approach are provided, including comparisons of Poisson and geometric distributions.
We present recent advances and statistical developments for evaluating Dynamic Treatment Regimes (DTR), which allow the treatment to be dynamically tailored according to evolving subject-level data. Identification of an optimal DTR is a key component for precision medicine and personalized health care. Specific topics covered in this talk include several recent projects with robust and flexible methods developed for the above research area. We will first introduce a dynamic statistical learning method, adaptive contrast weighted learning (ACWL), which combines doubly robust semiparametric regression estimators with flexible machine learning methods. We will further develop a tree-based reinforcement learning (T-RL) method, which builds an unsupervised decision tree that maintains the nature of batch-mode reinforcement learning. Unlike ACWL, T-RL handles the optimization problem with multiple treatment comparisons directly through a purity measure constructed with augmented inverse probability weighted estimators. T-RL is robust, efficient and easy to interpret for the identification of optimal DTRs. However, ACWL seems more robust against tree-type misspecification than T-RL when the true optimal DTR is non-tree-type. At the end of this talk, we will also present a new Stochastic-Tree Search method called ST-RL for evaluating optimal DTRs.
Handling missing data with expectation maximization algorithmLoc Nguyen
Expectation maximization (EM) algorithm is a powerful mathematical tool for estimating parameter of statistical models in case of incomplete data or hidden data. EM assumes that there is a relationship between hidden data and observed data, which can be a joint distribution or a mapping function. Therefore, this implies another implicit relationship between parameter estimation and data imputation. If missing data which contains missing values is considered as hidden data, it is very natural to handle missing data by EM algorithm. Handling missing data is not a new research but this report focuses on the theoretical base with detailed mathematical proofs for fulfilling missing values with EM. Besides, multinormal distribution and multinomial distribution are the two sample statistical models which are concerned to hold missing values.
This document discusses Bayesian approaches to combining evidence from multiple data sources or models. It recommends combining data probabilistically using Bayes' theorem rather than averaging. It provides an example of combining three data sources on annual rainfall measurements by treating the data sources as independent measurements and deriving the posterior distribution of rainfall amounts given the data. It also discusses challenges that arise when combining dependent data sources or models, and presents examples of hierarchical modeling approaches.
This document discusses generalized linear mixed models (GLMMs). It begins with examples of GLMM applications and definitions of key terms. The document then covers estimation methods for GLMMs, including maximum likelihood estimation, integrated likelihood, and both deterministic and stochastic approaches. Inference for GLMMs and remaining challenges are also mentioned. The overall document provides an overview of GLMM frameworks, examples, estimation techniques, and open questions.
BINARY TREE SORT IS MORE ROBUST THAN QUICK SORT IN AVERAGE CASEIJCSEA Journal
This document discusses the average case complexity of binary tree sort compared to quicksort. It analyzes the robustness of binary tree sort's O(n log n) average case complexity for non-uniform input distributions like binomial, Poisson, discrete uniform, continuous uniform, and standard normal distributions. Through statistical modeling and simulation experiments on different sample sizes, it is shown that binary tree sort maintains O(n log n) average case complexity even for non-uniform inputs, demonstrating that it is more robust than quicksort in the average case. The document concludes that only algorithms with equal worst case and average case complexities can reliably depend on average complexity measures, otherwise robustness must be verified.
The document presents a dynamic discrete choice model of demand for insecticide treated nets (ITNs) that accounts for time inconsistent preferences and unobserved heterogeneity. The model has three periods where agents make ITN purchase and retreatment decisions. Agents are either time consistent, "naive" time inconsistent, or "sophisticated" time inconsistent. The model is identified in two steps - first when types are directly observed using survey responses, and second when types are unobserved. Identification exploits variation from elicited beliefs about malaria risk. The model can point identify time preference parameters and utility functions up to a normalization.
A empresa CSM fabrica produtos flexíveis industriais e desenvolve soluções para problemas de clientes. Apresenta casos em que criou proteções sanfonadas que podem ser instaladas sem parar a máquina e reduziram em até 90% o tempo de montagem.
O documento discute o uso de recursos computacionais no marketing, mencionando o comércio eletrônico e suas vantagens e desvantagens. Também aborda como as empresas podem se preparar para o sucesso no ambiente digital, como criando banco de dados de clientes e colocando banners em sites relacionados.
O documento discute o comércio eletrônico, definindo-o como uma nova forma de conduzir negócios através da Internet e analisando suas oportunidades e desafios em comparação ao comércio tradicional, como a segurança das transações online e a necessidade de oferecer informações detalhadas sobre os produtos.
O documento discute os conceitos e aplicações de e-commerce, m-commerce e b2b, b2c. Aborda como o e-commerce envolve vendas, marketing e transações online entre empresas e consumidores. Também explica como o m-commerce permite compras via celular e apresenta casos de uso como pagamentos e informações de produtos. Por fim, detalha os modelos b2b para troca de informações entre empresas e b2c para venda direta a consumidores.
Gerenciamento de Projetos para engenharias e arquiteturasFladhimyr Castello
Apresentação sobre gerenciamento de projetos com o objetivo de apresentar a que situações o gerenciamento de projetos se propõe a resolver.
São apresentados os problemas comuns do gerenciamento de projetos bem como seus benefícios.
É apresentado o relatório do caos e o estudo benchmarking do PMI Brasil.
Também é abordado um pouco a respeito das certificações e o processo da certificação PMP.
Foco em projetos de construção civil.
A palestra "Faça Apresentações! Não Faça Slides!" é resultado de um trabalho de modelagem dos padrões linguísticos (PNL) de um dos maiores oradores mais admirados do mundo, Steve Jobs.
Na palestra "Faça Apresentações! Não Faça Slides!", Victor Gonçalves apresenta dicas para realizar apresentações inspiradoras e que proporcionem experiências memoráveis a seu público.
This document discusses statistical inference for multiple regression analysis. It begins by recapping the assumptions of the classical linear model (CLM), which are used to derive the sampling distributions of the OLS estimators. It then explains that the sampling distributions are needed to test hypotheses about the population parameters. Under the CLM assumptions, specifically the additional normality assumption, the sampling distributions are normal. This allows hypotheses about a single parameter to be tested using the t-distribution. The document provides an example of testing the null hypothesis that a parameter equals zero against a one-sided alternative hypothesis.
This document presents a talk on kernel density estimation using Bernstein polynomials and for circular data. It begins with an introduction to kernel density estimation and the empirical distribution function. It then discusses an approximation lemma due to Feller that motivates the Bernstein polynomial density estimator. The talk outlines how the kernel density estimation can be extended to multivariate density estimation and adapted for circular density estimation using Bernstein polynomials. It concludes with connecting circular kernel density estimation to orthogonal polynomials on a unit circle and provides examples of applying it to turtle direction and ant movement data.
Min-based qualitative possibilistic networks are one of the effective tools for a compact representation of
decision problems under uncertainty. The exact approaches for computing decision based on possibilistic
networks are limited by the size of the possibility distributions. Generally, these approaches are based on
possibilistic propagation algorithms. An important step in the computation of the decision is the
transformation of the DAG (Direct Acyclic Graph) into a secondary structure, known as the junction trees
(JT). This transformation is known to be costly and represents a difficult problem. We propose in this paper
a new approximate approach for the computation of decision under uncertainty within possibilistic
networks. The computing of the optimal optimistic decision no longer goes through the junction tree
construction step. Instead, it is performed by calculating the degree of normalization in the moral graph
resulting from the merging of the possibilistic network codifying knowledge of the agent and that codifying
its preferences.
This document discusses unifying Bayesian, frequentist, and fiducial statistical inferences through the concept of confidence distributions (CDs). CDs provide a sample-dependent distribution function for estimating parameters of interest that can be used to construct confidence intervals. Several examples are given that show how normalized likelihood functions, p-value functions, and other approaches can all be considered CDs. The document argues this provides a first-level conceptual union of the different statistical paradigms. A second-level union is proposed through viewing CDs, bootstrap distributions, fiducial distributions, and Bayesian posteriors as estimating parameters through artificial or "fake" data samples, representing uncertainty in a similar way across paradigms.
The document discusses various methods for constructing confidence intervals for estimating multinomial proportions. It aims to analyze the propensity for aberrations (i.e. unrealistic bounds like negative values) in the interval estimates across different classical and Bayesian methods. Specifically, it provides the mathematical conditions under which each method may produce aberrant interval limits, such as zero-width intervals or bounds exceeding 0 and 1, especially for small sample counts. The document also develops an R program to facilitate computational implementation of the various methods for applied analysis of multinomial data.
Two parameter entropy of uncertain variableSurender Singh
This document introduces a two parameter entropy measure for uncertain variables based on two parameter probabilistic entropy. It begins by reviewing concepts of uncertainty theory such as uncertainty space, uncertain variables, and independence of uncertain variables. It then defines entropy of an uncertain variable using Shannon entropy. Previous work on one parameter entropy of uncertain variables is summarized. The document proposes a new two parameter entropy for uncertain variables, providing its definition and examples of calculating it for specific uncertainty distributions. Properties of the two parameter entropy are discussed.
This document outlines 5 problems related to Bayesian statistics and hypothesis testing. Problem 1 involves using a Bayesian t-test to analyze orbital data from Mercury that did not match Newtonian predictions. Problem 2 asks which of several hypotheses would be appropriate to test precisely versus not. Problem 3 provides two scale distribution models and asks about the prior and Bayes factor for testing between them. Problem 4 gives a scenario about machine part production and asks for the Bayes factor for whether the machine is operating properly or failed. Problem 5 asks which predictive density model would be best for predicting a new data point from among several location-scale models.
A Study of Some Tests of Uniformity and Their PerformancesIOSRJM
This document discusses a study that evaluates the performance of eleven tests for uniformity. The tests include Kolmogorov-Smirnov, Anderson-Darling, Cramer-von Mises, Watson, Sukhatme, probability product, Kuiper, Gini, ZhangA, ZhangC tests. Through simulation, the power of these tests is assessed under various sample sizes and five alternative distributions. The results are displayed in tables and graphs. The document concludes that the performance of the tests depends on the sample size and nature of the alternative distribution.
A derivative free high ordered hybrid equation solverZac Darcy
Generally a range of equation solvers for estimating the solution of an equation contain the derivative of
first or higher order. Such solvers are difficult to apply in the instances of complicated functional
relationship. The equation solver proposed in this paper meant to solve many of the involved complicated
problems and establishing a process tending towards a higher ordered by alloying the already proved
conventional methods like Newton-Raphson method (N-R), Regula Falsi method (R-F) & Bisection method
(BIS). The present method is good to solve those nonlinear and transcendental equations that cannot be
solved by the basic algebra. Comparative analysis are also made with the other racing formulas of this
group and the result shows that present method is faster than all such methods of the class.
Similarity Measure Using Interval Valued Vague Sets in Multiple Criteria Deci...iosrjce
IOSR Journal of Mathematics(IOSR-JM) is a double blind peer reviewed International Journal that provides rapid publication (within a month) of articles in all areas of mathemetics and its applications. The journal welcomes publications of high quality papers on theoretical developments and practical applications in mathematics. Original research papers, state-of-the-art reviews, and high quality technical notes are invited for publications.
The document introduces factor of safety and probability of failure in engineering design. It discusses using sensitivity studies to systematically vary parameters over their credible ranges to determine the influence on factor of safety. This allows a more rational assessment of design risks than relying on a single calculated factor of safety. The document then provides an introduction to probability theory and statistical concepts used in probabilistic analyses, including random variables, probability distributions, sampling techniques, and calculating the probability of failure for a slope design example.
1) The author proposes using stochastic dominance to estimate equivalence scales, where the scaling factor is chosen such that the income distribution of household type i is equal to the distribution of the reference household scaled by the factor.
2) Estimating these stochastic equivalence scales using Polish household expenditure data from 2005-2010, the author finds that for most household types, the estimated scaling factors satisfy first-order stochastic dominance, implying households are indifferent between distributions.
3) However, the author notes higher-order stochastic dominance could incorporate different social welfare perspectives by placing more weight on lower incomes.
Talk at 2013 WSC, ISI Conference in Hong Kong, August 26, 2013Christian Robert
Those are the slides for my conference talk at 2013 WSC, in the "Jacob Bernoulli's "Ars Conjectandi" and the emergence of probability" session organised by Adam Jakubowski
This document presents an exponential-Lindley additive failure rate model (ELAFRM) by combining the hazard functions of an exponential distribution and a Lindley distribution. The key properties of the ELAFRM are derived, including the probability density function, cumulative distribution function, hazard function, moments, and graphical representations. Estimation of the model parameters is also discussed. The document proposes this new ELAFRM distribution and analyzes its mathematical properties.
A DERIVATIVE FREE HIGH ORDERED HYBRID EQUATION SOLVERZac Darcy
Generally a range of equation solvers for estimating the solution of an equation contain the derivative of
first or higher order. Such solvers are difficult to apply in the instances of complicated functional
relationship. The equation solver proposed in this paper meant to solve many of the involved complicated
problems and establishing a process tending towards a higher ordered by alloying the already proved
conventional methods like Newton-Raphson method (N-R), Regula Falsi method (R-F) & Bisection method
(BIS). The present method is good to solve those nonlinear and transcendental equations that cannot be
solved by the basic algebra. Comparative analysis are also made with the other racing formulas of this
group and the result shows that present method is faster than all such methods of the class.
On Extension of Weibull Distribution with Bayesian Analysis using S-Plus Soft...Dr. Amarjeet Singh
This document discusses Bayesian analysis of the Weibull distribution using different priors in S-Plus software. It proposes a new lifetime distribution that is an extension of the Weibull distribution and derives the Weibull distribution as a special case of the proposed model. It describes the likelihood and posterior distributions for the Weibull parameters under different prior distributions and discusses Bayesian estimation methods and a simulation study to compare the mean square error of the estimators.
This document presents an asymptotic expansion of the posterior density in high dimensional generalized linear models. The main results are:
1) The authors prove a third order correct asymptotic expansion of the posterior density for generalized linear models with canonical link functions when the number of regressors grows with sample size.
2) This asymptotic expansion is then used to derive moment matching priors in the generalized linear model setting.
3) The expansion assumes the number of regressors grows such that p6+εn /n → 0 as n → ∞ for some small ε > 0, which is stronger than prior work requiring only p4n log(pn)/n → 0.
ISM_Session_5 _ 23rd and 24th December.pptxssuser1eba67
The document discusses random variables and their probability distributions. It defines discrete and continuous random variables and their key characteristics. Discrete random variables can take on countable values while continuous can take any value in an interval. Probability distributions describe the probabilities of a random variable taking on different values. The mean and variance are discussed as measures of central tendency and variability. Joint probability distributions are introduced for two random variables. Examples and homework problems are also provided.
Remote-sensing data offer unprecedented opportunities to address Earth-system-science challenges, such as understanding the relationship between the atmosphere and Earth's surface using physics, chemistry, biology, mathematics, and computing. Statistical methods have often been seen as a hybrid of the latter two, so that a lot of attention has been given to computing estimates but far less to quantifying the uncertainty of the estimates. In my "bird's-eye view," I shall give a way to look at the problem using conditional probability models and three states of knowledge. Examples will be given of analyzing remotely sensed data of a leading greenhouse gas, carbon dioxide.
it describes the bony anatomy including the femoral head , acetabulum, labrum . also discusses the capsule , ligaments . muscle that act on the hip joint and the range of motion are outlined. factors affecting hip joint stability and weight transmission through the joint are summarized.
हिंदी वर्णमाला पीपीटी, hindi alphabet PPT presentation, hindi varnamala PPT, Hindi Varnamala pdf, हिंदी स्वर, हिंदी व्यंजन, sikhiye hindi varnmala, dr. mulla adam ali, hindi language and literature, hindi alphabet with drawing, hindi alphabet pdf, hindi varnamala for childrens, hindi language, hindi varnamala practice for kids, https://www.drmullaadamali.com
How to Build a Module in Odoo 17 Using the Scaffold MethodCeline George
Odoo provides an option for creating a module by using a single line command. By using this command the user can make a whole structure of a module. It is very easy for a beginner to make a module. There is no need to make each file manually. This slide will show how to create a module using the scaffold method.
How to Add Chatter in the odoo 17 ERP ModuleCeline George
In Odoo, the chatter is like a chat tool that helps you work together on records. You can leave notes and track things, making it easier to talk with your team and partners. Inside chatter, all communication history, activity, and changes will be displayed.
How to Fix the Import Error in the Odoo 17Celine George
An import error occurs when a program fails to import a module or library, disrupting its execution. In languages like Python, this issue arises when the specified module cannot be found or accessed, hindering the program's functionality. Resolving import errors is crucial for maintaining smooth software operation and uninterrupted development processes.
Executive Directors Chat Leveraging AI for Diversity, Equity, and InclusionTechSoup
Let’s explore the intersection of technology and equity in the final session of our DEI series. Discover how AI tools, like ChatGPT, can be used to support and enhance your nonprofit's DEI initiatives. Participants will gain insights into practical AI applications and get tips for leveraging technology to advance their DEI goals.
Walmart Business+ and Spark Good for Nonprofits.pdfTechSoup
"Learn about all the ways Walmart supports nonprofit organizations.
You will hear from Liz Willett, the Head of Nonprofits, and hear about what Walmart is doing to help nonprofits, including Walmart Business and Spark Good. Walmart Business+ is a new offer for nonprofits that offers discounts and also streamlines nonprofits order and expense tracking, saving time and money.
The webinar may also give some examples on how nonprofits can best leverage Walmart Business+.
The event will cover the following::
Walmart Business + (https://business.walmart.com/plus) is a new shopping experience for nonprofits, schools, and local business customers that connects an exclusive online shopping experience to stores. Benefits include free delivery and shipping, a 'Spend Analytics” feature, special discounts, deals and tax-exempt shopping.
Special TechSoup offer for a free 180 days membership, and up to $150 in discounts on eligible orders.
Spark Good (walmart.com/sparkgood) is a charitable platform that enables nonprofits to receive donations directly from customers and associates.
Answers about how you can do more with Walmart!"
1. On Nonparametric Density Estimation for Size Biased
Data
Yogendra P. Chaubey
Department of Mathematics and Statistics
Concordia University, Montreal, Canada H3G 1M8
E-mail: yogen.chaubey@concordia.ca
Talk to be presented at
5e Recontre scientifiques Sherbrooke-Montpelier:
Colloque de statistique et de biostatistique
June 10-11, 2015
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 1 / 75
2. Abstract
This talk will highlight some recent developments in the area of
nonparametric functional estimation with emphasis on nonparametric
density estimation for size biased data. Such data entail constraints that
many traditional nonparametric density estimators may not satisfy. A
lemma attributed to Hille, and its generalization [see Lemma 1, Feller
(1965) An Introduction to Probability Theory and Applications, §VII.1)] is
used to propose estimators in this context from two different perspectives.
After describing the asymptotic properties of the estimators, we present
the results of a simulation study to compare various nonparametric density
estimators. The optimal data driven approach of selecting the smoothing
parameter is also outlined.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 2 / 75
3. Outline
1 1. Introduction/Motivation
1.1 Kernel Density Estimator
1.2. Smooth Estimation of Densities on R+
2 2. An Approximation Lemma and Some Alternative Smooth Density
Estimators
2.1 Some Alternative Smooth Density Estimators on R+
2.2 Asymptotic Properties of the New Estimator
2.3 Extensions Non-iid cases
3 3. Estimation of Density in Length-biased Data
3.1 Smooth Estimators Based on the Estimators of G
3.2 Smooth Estimators Based on the Estimators of F
4 4. Choice of Smoothing Parameters
4.1 Unbiased Cross Validation
4.2 Biased Cross Validation
5 5. A Comparison Between Different Estimators: Simulation Studies
5.1 Simulation for χ2
2
5.2 Simulation for χ2
6
5.3 Simulation for Some Other Standard Distributions
6 6. References
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 3 / 75
4. 1. Introduction/Motivation
1.1 Kernel Density Estimator
Consider X as a non-negative random variable with density f(x) and
distribution function
F(x) =
x
0
f(t)dt for x > 0. (1.1)
Such random variables are more frequent in practice in life testing and
reliability.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 4 / 75
5. Based on a random sample (X1, X2, ..., Xn), from a univariate
density f(.), the empirical distribution function (edf) is defined as
Fn(x) =
1
n
n
i=1
I(Xi ≤ x). (1.2)
edf is not smooth enough to provide an estimator of f(x).
Various methods (viz., kernel smoothing, histogram methods, spline,
orthogonal functionals)
The most popular is the Kernel method (Rosenblatt, 1956).
[See the text Nonparametric Functional Estimation by Prakasa Rao
(1983) for a theoretical treatment of the subject or Silverman (1986)].
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 5 / 75
6. ˆfn(x) =
1
n
n
i=1
kh(x − Xi) =
1
nh
n
i=1
k
x − Xi
h
) (1.3)
where the function k(.) called the Kernel function has the following
properties;
(i)k(−x) = k(x)
(ii)
∞
−∞
k(x)dx = 1
and
kh(x) =
1
h
k
x
h
h is known as bandwidth and is made to depend on n, i.e. h ≡ hn,
such that hn → 0 and nhn → ∞ as n → ∞.
Basically k is a symmetric probability density function on the entire
real line. This may present problems in estimating the densities of
non-negative random variables.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 6 / 75
7. Kernel Density Estimators for Suicide Data
0 200 400 600
0.0000.0020.0040.006
x
Default
SJ
UCV
BCV
Figure 1. Kernel Density Estimators for Suicide Study Data
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 7 / 75
8. 1.2. Smooth Estimation of densities on R+
ˆfn(x) might take positive values even for x ∈ (−∞, 0], which is not
desirable if the random variable X is positive. Silverman (1986)
mentions some adaptations of the existing methods when the support
of the density to be estimated is not the whole real line, through
transformation and other methods.
1.2.1 Bagai-Prakasa Rao Estimator
Bagai and Prakasa Rao (1996) proposed the following adaptation of
the Kernel Density estimator for non-negative support [which does
not require any transformation or corrective strategy].
fn(x) =
1
nhn
n
i=1
k
x − Xi
hn
, x ≥ 0. (1.4)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 8 / 75
9. Here k(.) is a bounded density function with support (0, ∞), satisfying
∞
0
x2
k(x)dx < ∞
and
hn is a sequence such that hn → 0 and nhn → ∞ as n → ∞.
The only difference between ˆfn(x) and fn(x) is that the former is
based on a kernel possibly with support extending beyond (0, ∞).
One undesirable property of this estimator is that that for x such that
for X(r) < x ≤ X(r+1), only the first r order statistics contribute
towards the estimator fn(x).
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 9 / 75
10. Bagai-Prakasa Rao Density Estimators for Suicide Data
0 200 400 600
0.0000.0020.0040.006
x
Default
SJ
UCV
BCV
Figure 2. Bagai-Prakasa Rao Density Estimators for Suicide Study Data
Silverman (1986)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 10 / 75
11. 2.1 An approximation Lemma
The following discussion gives a general approach to density
estimation which may be specialized to the case of non-negative data.
The key result for the proposal is the following Lemma given in Feller
(1965, §VII.1).
Lemma 1: Let u be any bounded and continuous function and
Gx,n, n = 1, 2, ... be a family of distributions with mean µn(x) and
variance h2
n(x) such that µn(x) → x and hn(x) → 0. Then
˜u(x) =
∞
−∞
u(t)dGx,n(t) → u(x). (2.1)
The convergence is uniform in every subinterval in which hn(x) → 0
uniformly and u is uniformly continuous.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 11 / 75
12. This generalization may be adapted for smooth estimation of the
distribution function by replacing u(x) by the empirical distribution
function Fn(x) as given below ;
˜Fn(x) =
∞
−∞
Fn(t)dGx,n(t). (2.2)
Note that Fn(x) is not a continuous function as desired by the above
lemma, hence the above lemma is not directly used in proposing the
estimator but it works as a motivation for the proposal. It can be
considered as the stochastic adaptation in light of the fact that the
mathematical convergence is transformed into stochastic convergence
that parallels to that of the strong convergence of the empirical
distribution function as stated in the following theorem.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 12 / 75
13. Theorem 1: Let hn(x) be the variance of Gx,n as in Lemma 1 such
that hn(x) → 0 as n → ∞ for every fixed x as n → ∞, then we have
sup
x
| ˜Fn(x) − F(x)|
a.s.
→ 0 (2.3)
as n → ∞.
Technically, Gx,n can have any support but it may be prudent to
choose it so that it has the same support as the random variable
under consideration; because this will get rid of the problem of the
estimator assigning positive mass to undesired region.
For ˜Fn(x) to be a proper distribution function, Gx,n(t) must be a
decreasing function of x, which can be shown using an alternative
form of ˜Fn(x) :
˜Fn(x) = 1 −
1
n
n
i=1
Gx,n(Xi). (2.4)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 13 / 75
14. Equation (2.4) suggests a smooth density estimator given by
˜fn(x) =
d ˜Fn(x)
dx
= −
1
n
n
i=1
d
dx
Gx,n(Xi). (2.5)
The potential of this lemma for smooth density estimation was
recognized by Gawronski (1980) in his doctoral thesis written at Ulm.
Gawronski and Stadmuller (1980, Skand. J. Stat.) investigated mean
square error properties of the density estimator when Gx,n is obtained
by putting Poisson weight
pk(xλn) = e−λnx (λnx)k
k!
(2.6)
to the lattice points k/λn, k = 0, 1, 2, ...
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 14 / 75
15. Other developments:
This lemma has been further used to motivate the Bernstein
Polynomial estimator (Vitale, 1975) for densities on [0, 1] by Babu,
Canty and Chaubey (1999). Gawronski (1985, Period. Hung.)
investigates other lattice distributions such as negative binomial
distribution.
Some other developments:
Chaubey and Sen (1996, Statist. Dec.): survival functions, though in a
truncated form.
Chaubey and Sen (1999, JSPI): Mean Residual Life; Chaubey and Sen
(1998a, Persp. Stat., Narosa Pub.): Hazard and Cumulative Hazard
Functions; Chaubey and Sen (1998b): Censored Data;
(Chaubey and Sen, 2002a, 2002b): Multivariate density estimation
Smooth density estimation under some constraints: Chaubey and
Kochar (2000, 2006); Chaubey and Xu (2007, JSPI).
Babu and Chaubey (2006): Density estimation on hypercubes. [see
also Prakasa Rao(2005)and Kakizawa (2011), Bouezmarni et al. (2010,
JMVA) for Generalised Bernstein Polynomials and Bernstein copulas]
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 15 / 75
16. A Generalised Kernel Estimator for Densities with
Non-Negative Support
Lemma 1 motivates the generalised kernel estimator of F¨oldes and
R´ev´esz (1974):
fnGK(x) =
1
n
n
i=1
hn(x, Xi)
Chaubey et al. (2012, J. Ind. Stat. Assoc.) show the following
adaptation using asymmetric kernels for estimation of densities with
non-negative support.
Let Qv(x) represent a distribution on [0, ∞) with mean 1 and
variance v2, then an estimator of F(x) is given by
F+
n (x) = 1 −
1
n
n
i=1
Qvn
Xi
x
, (2.7)
where vn → 0 as n → ∞.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 16 / 75
17. Obviously, this choice uses Gx,n(t) = Qvn (t/x) which is a decreasing
function of x.
This leads to the following density estimator
d
dx
(F+
n (x)) =
1
nx2
n
i=1
Xi qvn
Xi
x
, (2.8)
where qv(.) denotes the density corresponding to the distribution
function Qv(.).
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 17 / 75
18. However, the above estimator may not be defined at x = 0, except in
cases where limx→0
d
dx (F+
n (x)) exists. Moreover, this limit is
typically zero, which is acceptable only when we are estimating a
density f with f(0) = 0.
Thus with a view of the more general case where 0 ≤ f(0) < ∞, we
considered the following perturbed version of the above density
estimator:
f+
n (x) =
1
n(x + n)2
n
i=1
Xi qvn
Xi
x + n
, x ≥ 0 (2.9)
where n ↓ 0 at an appropriate (sufficiently slow) rate as n → ∞. In
the sequel, we illustrate our method by taking Qv(.) to be the
Gamma (α = 1/v2, β = v2) distribution function.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 18 / 75
19. Remark:
Note that if we believe that the density is zero at zero, we set n ≡ 0,
however in general, it may be determined using the cross-validation
methods. For n > 0, this modification results in a defective distribution
F+
n (x + n). A corrected density estimator f∗
n(x) is therefore proposed:
f∗
n(x) =
f+
n (x)
cn
, (2.10)
where cn is a constant given by
cn =
1
n
n
i=1
Qvn
Xi
n
.
Note that, since for large n, n → 0, f∗
n(x) and f+
n (x) are asymptotically
equivalent, we study the asymptotic properties of f+
n (x) only.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 19 / 75
20. Next we present a comparison of our approach with some existing
estimators.
Kernel Estimator.
The usual kernel estimator is a special case of the representation
given by Eq. (2.5), by taking Gx,n(.) as
Gx,n(t) = K
t − x
h
, (2.11)
where K(.) is a distribution function with mean zero and variance 1.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 20 / 75
21. Transformation Estimator of Wand et al.
The well known logarithmic transformation approach of Wand,
Marron and Ruppert (1991) leads to the following density estimator:
˜f(L)
n (x) =
1
nhnx
n
i=1
k(
1
hn
log(Xi/x)), (2.12)
where k(.) is a density function (kernel) with mean zero and variance
1.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 21 / 75
22. This is easily seen to be a special case of Eq. (2.5), taking Gx,n again
as in Eq. (2.11) but applied to log x. This approach, however, creates
problem at the boundary which led Marron and Ruppert (1994) to
propose modifications that are computationally intensive.
Estimators of Chen and Scaillet.
Chen’s (2000) estimator is of the form
ˆfC(x) =
1
n
n
i=1
gx,n(Xi), (2.13)
where gx,n(.) is the Gamma(α = a(x, b), β = b) density with b → 0
and ba(x, b) → x.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 22 / 75
23. This also can be motivated from Eq. (2.1) as follows: take
u(t) = f(t) and note that the integral f(t)gx,n(t)dt can be
estimated by n−1 n
i=1 gx,n(Xi). This approach controls the
boundary bias at x = 0; however, the variance blows up at x = 0, and
computation of mean integrated squared error (MISE) is not
tractable. Moreover, estimators of derivatives of the density are not
easily obtainable because of the appearance of x as argument of the
Gamma function.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 23 / 75
24. Scaillet’s (2004) estimators replace the Gamma kernel by inverse
Gaussian (IG) and reciprocal inverse Gaussian (RIG) kernels. These
estimators are more tractable than Chen’s; however, the IG-kernel
estimator assumes value zero at x = 0, which is not desirable when
f(0) > 0, and the variances of the IG as well as the RIG estimators
blow up at x = 0.
Bouezmarni and Scaillet (2005), however, demonstrate good
finite-sample performance of these estimators.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 24 / 75
25. It is interesting to note that one can immediately define a
Chen-Scaillet version of our estimator, namely,
f+
n,C(x) =
1
n
n
i=1
1
x
qvn
Xi
x
.
On the other hand, our version (i.e., perturbed version) this estimator
would be
ˆf+
C (x) =
1
n
n
i=1
gx+ n,n(Xi);
that should not have the problem of variance blowing up at x = 0.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 25 / 75
26. It may also be remarked that the idea used here may be extended to
the case of densities supported on an arbitrary interval
[a, b], − ∞ < a < b < ∞, by choosing for instance a Beta kernel
(extended to the interval [a, b]) as in Chen (1999). Without loss of
generality, suppose a = 0 and b = 1. Then we can choose, for
instance, qv(.) as the density of Y/µ, where
Y ∼ Beta(α, β), µ = α/(α + β), such that α → ∞ and β/α → 0, so
that Var(Y/µ) → 0.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 26 / 75
27. 2.2 Asymptotic Properties of the New Estimator
2.2.1 Asymptotic Properties of ˜F+
n (x)
The strong consistency holds in general for the estimator ˜F+
n (x). We can
easily prove the following theorem parallel to the strong convergence of the
empirical distribution function.
Theorem:
If λn → ∞ as n → ∞ we have
sup
x
| ˜F+
n (x) − F(x)|
a.s.
→ 0.
as n → ∞.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 27 / 75
28. We can also show that for large n, the smooth estimator can be arbitrarily
close to the edf by proper choice of λn, as given in the following theorem.
Theorem:Assuming that f has a bounded derivative, and λn = o(n),
then for some δ > 0, we have, with probability one,
sup
x≥0
| ˜F+
n (x) − Fn(x)| = O n−3/4
(log n)1+δ
.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 28 / 75
29. 2.2.2 Asymptotic Properties of ˜f+
n (x)
Under some regularity conditions, they obtained
Theorem:
sup
x≥0
| ˜f+
n (x) − f(x)|
a.s.
−→ 0
as n → ∞.
Theorem:
(a) If nvn → ∞, nv3
n → 0, nvnε2
n → 0 as n → ∞, we have
√
nvn(f+
n (x) − f(x)) → N 0, I2(q)
µf(x)
x2
, for x > 0.
(b) If nvn
2
n → ∞ and nvnε4
n → 0 as n → ∞, we have
nvn
2
n(f+
n (0) − f(0)) → N 0, I2(q)f(0) .
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 29 / 75
30. 2.3. Extensions to Non-iid cases
We can extend the technique to non-iid cases where a version of Fn(x) is
available.
Chaubey, Y.P., Dewan, I. and Li, J. (2012) – Density estimation for
stationary associated sequences. Comm. Stat.- Simula. Computa.
41(4), 554- 572 –Using generalised kernel approach Chaubey
Chaubey, Yogendra P., Dewan, Isha and Li, Jun (2011) – Density
estimation for stationary associated sequences using Poisson weights.
Statist. Probab. Lett. 81, 267-276.
Chaubey, Y.P. and Dewan, I. (2010). A review for smooth estimation
of survival and density functions for stationary associated sequences:
Some recent developments – J. Ind. Soc. Agr. Stat. 64(2), 261-272.
Chaubey, Y.P., La¨ıb, N. and Sen, A. (2010). Generalised kernel
smoothing for non-negative stationary ergodic processes – Journal of
Nonparametric Statistics, 22, 973-997
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 30 / 75
31. 3. Estimation of Density in Length-biased Data
In general, when the probability that an item is sampled is
proportional to its size, size biased data emerges.
The density g of the size biased observation for the underlying density
f, is given by
g(x) =
w(x)f(x)
µw
, x > 0, (3.1)
where w(x) denotes the size measure and µw = w(x)f(x).
In the area of forestry, the size measure is usually proportional to
either length or area (see Muttlak and McDonald, 1990).
Another important application occurs in renewal theory where
inter-event times data are of this type if they are obtained by
sampling lifetimes in progress at a randomly chosen point in time (see
Cox, 1969).
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 31 / 75
32. Here we will talk about the length biased case where we can write
f(x) =
1
x
g(x)/µ. (3.2)
In principle any smooth estimator of the density function g may be
transformed into that of the density function f as follows:
ˆf(x) =
1
x
ˆg(x)/ˆµ, (3.3)
where ˆµ is an estimator of µ.
Note that 1/µ = Eg(1/X), hence a strongly consistent estimator of µ
is given by
ˆµ = n{
n
i=1
X−1
i }−1
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 32 / 75
33. Bhattacharyya et al. (1988) used this strategy in proposing the
following smooth estimator of f,
ˆfB(x) = ˆµ(nx)−1
n
i=1
kh(x − Xi). (3.4)
Also since, F(x) = µ Eg(X−11(X≤x)), Cox (1969) proposed the
following as an estimator of the distribution function F(x) :
ˆFn(x) = ˆµ
1
n
n
i=1
X−1
i 1(Xi≤x). (3.5)
So there are two competing strategies for density estimation for LB
data. One is to estimate g(x) and then use the relation (3.3) (i.e.
smooth Gn as in Bhattacharyya et al. (1988)). The other is to
smooth the Cox estimator ˆFn(x) directly and use the derivative as the
smooth estimator of f(x).
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 33 / 75
34. Jones (1991) studied the behaviour of the estimator fB(x) in contrast
to smooth estimator obtained directly by smoothing the estimator
Fn(x), by Kernel method:
ˆfJ (x) = n−1
ˆµ
n
i=1
X−1
i kh(x − Xi). (3.6)
He noted that this estimator is a proper density function when
considered with the support on the whole real line, where as ˆfB(y)
may be not. He compared the two estimators based on simulations,
and using the asymptotic arguments, concluded that the latter
estimator may be preferable in practical applications.
Also using Jensen’s inequality we find that
Eg(ˆµ) ≥ 1/Eg{
1
n
n
i=1
X−1
i } = µ,
hence the estimator ˆµ may be positively biased which would transfer
into increased bias in the above density estimators.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 34 / 75
35. If g(x)/x is integrable, the deficiency of fB(x) of not being a proper
density may be corrected by considering the alternative estimator
ˆfa(x) =
ˆg(x)/x
(ˆg(x)/x)dx
, (3.7)
and this may also eliminate the increase in bias to some extent.
However, in these situations, since X is typically a non-negative
random variable, the estimator must satisfy the following two
conditions:
(i) ˆg(x) = 0 for x ≤ 0,
(ii) ˆg(x)/x is integrable.
Here both of the estimators ˆfB(x) and ˆfJ (x) do not satisfy these
properties.
We have a host of alternatives, those based on smoothing Gn and
those based on smoothing Fn, that we are going to talk about next.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 35 / 75
36. 3.1.1 Poisson Smoothing of Gn
Here, we would like to see the application of the weights generated by
the Poisson probability mass function as motivated in Chaubey and
Sen (1996, 2000). However, a modification is necessary in the present
situation which is also outlined here.
Using Poisson smoothing, an estimator of g(x) may be given by,
˜gnP (x) = λn
∞
k=0
pk(λnx) Gn
k + 1
λn
− Gn
k
λn
, (3.8)
however, note that limx→0˜gnP (x) = λnGn(1/λn) which may
converge to 0 as λn → ∞, however for finite samples it may not be
zero, hence the density f at x = 0 may not be defined. Furthermore,
˜gnP (x)/x is not integrable.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 36 / 75
37. A simple modification by attaching the weight pk(λnx) to
Gn((k − 1)/λn), rather than to Gn(k/λn), the above problem is
avoided. This results in the following smooth estimator of G(x) :
˜Gn(x) =
k≥0
pk(xλn)Gn
k − 1
λn
, (3.9)
The basic nature of the smoothing estimator is not changed, however
this provides an alternative estimator of the density function as its
derivative is given by
˜gn(x) = λn
k≥1
pk(xλn) Gn
k
λn
− Gn
k − 1
λn
, (3.10)
such that ˜gn(0) = 0 and that ˜gn(x)/x is integrable.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 37 / 75
38. Since,
∞
0
˜gn(x)
x
dx = λn
k≥1
[Gn
k
λn
− Gn
k − 1
λn
]
∞
0
pk(xλn)
x
dx
= λn
k≥1
[Gn
k
λn
− Gn
k − 1
λn
]
1
k
= λn
k≥1
1
k(k + 1)
Gn
k
λn
,
The new smooth estimator of the length biased density f(x) is given
by
˜fn(x) =
k≥1
pk−1(xλn)
k Gn
k
λn
− Gn
k−1
λn
k≥1
1
k(k+1) Gn
k
λn
. (3.11)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 38 / 75
39. The corresponding smooth estimator of the distribution function
F(x) is given by
˜Fn(x) =
k≥1(1/k)Wk(xλn)[Gn
k
λn
− Gn
k−1
λn
]
k≥1
1
k(k+1)Gn
k
λn
(3.12)
where
Wk(λnx) =
1
Γ(k)
λnx
0
e−y
yk−1
dy =
j≥k
pj(λnx).
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 39 / 75
40. An equivalent expression for the above estimator is given by
˜Fn(x) =
k≥1 Gn
k
λn
Wk(λnx)
k −
Wk+1(λnx)
k+1
k≥1
1
k(k+1) Gn
k
λn
= 1 +
k≥1 Gn
k
λn
Pk(λnx)
k+1 −
Pk−1(λnx)
k
k≥1
1
k(k+1) Gn
k
λn
,
where Pk(µ) =
k
j=0
pj(µ)
denotes the cumulative probability corresponding to the Poisson(µ)
distribution.
The properties of above estimators can be established in an analogous
way to those in the regular case.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 40 / 75
41. 3.1.2 Gamma Smoothing of Gn
The smooth estimator using the log-normal density may typically have
a spike at zero, however the gamma density may be appropriate, since
it typically has the density estimator ˆg(x) such that ˆg(0) = 0, so that
no perturbation is required. The smooth density estimator in this case
is simply given by
g+
n (x) =
1
nx2
n
i=1
Xi qvn
Xi
x
, (3.13)
where qv(.) denotes the density corresponding to a
Gamma(α = 1/vn, β = vn). Since
∞
0
gn(x)
x dx = 1
n
n
i=1
1
Xi
, this
gives the density estimator
fn(x) =
1
x3
n
i=1 Xiqvn
Xi
x
n
i=1
1
Xi
. (3.14)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 41 / 75
42. Note that the above estimator may be reasonable for a density f(x) with
f(0) = 0, but may need a modification for general density functions. We
use the idea of perturbation as before to arrive at an acceptable estimator
through this approach as given by
ˆf+
n (x) =
1
(x+εn)3
n
i=1 Xiqvn
Xi
x+εn
n
i=1
1
Xi
(3.15)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 42 / 75
43. 3.2.1 Poisson Smoothing of Fn
smoothing Fn directly using Poisson weights, an estimator of f(x)
may be given by,
˜fnP (x) = λn
∞
k=0
pk(λnx) Fn
k + 1
λn
− Fn
k
λn
. (3.16)
No modifications are necessary.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 43 / 75
44. 3.2.2 Gamma Smoothing of Fn
The gamma based smooth estimate of F(x) is given by
F+
n (x) = 1 −
n
i=1
1
Xi
Qvn (Xi
x )
n
i=1
1
Xi
, (3.17)
and that for the density f in this case is simply given by
˜f+
n (x) =
1
(x+εn)2
n
i=1 qvn ( Xi
x+εn
)
n
i=1
1
Xi
. (3.18)
where qv(.) denotes the density corresponding to a
Gamma(α = 1/vn, β = vn).
Note that the above estimator is computationally intensive as two
smoothing parameters have to be computed using bivariate cross
validation.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 44 / 75
45. 4. Choice of Smoothing Parameters
4.1 Unbiased Cross Validation
This criterion evolves from minimizing an unbiased estimate of the
integrated squared error (ISE) of fn with respect to f given by
ISE(fn; f) =
∞
0
(fn(x) − f(x))2
dx
=
∞
0
f2
n(x) − 2
∞
0
fn(x)f(x)dx +
∞
0
f2
(x)dx
The last term being free of the smoothing parameter (b), we minimize
UCV (b) =
∞
0
f2
n(x; b, D)dx − 2
n
i=1
fn−1(Xi; b, Di)/Zi,
that estimates the first two terms of ISE(fn; f), where Zi =
j=i
Xi
Xj
and fn(x; b, D) denotes the estimator of f(x) that involves data D.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 45 / 75
46. 4. Choice of Smoothing Parameters
4.2 Biased Cross Validation
This criterion involves minimization of an estimate of the asymptotic
mean integrated squared error (AMISE).
For AMISE( ˆfn), we use the formula as obtained in Chaubey et al.
(2012), that is given by
AMISE[ ˆf+
n (x)] =
I2(q)µ
nvn
∞
0
f(x)
(x + εn)2
dx
+
∞
0
[v2
nf(x) + (2v2
nx + εn)f (x) + v2
n
x2
2
f (x)]2
dx
(4.1)
where I2(q) lim
v→0
v
∞
0 (qvn (t))2dt.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 46 / 75
47. 4. Choice of Smoothing Parameters
4.2 Biased Cross Validation
The AMISE( ˜fn), is given by in Chaubey et al. (2012), that is given
by
AMISE[ ˜f+
n ] =
∞
0
[(xv2
n + εn)f (x) +
x2
2
f (x)v2
n]2
dx
+
I2(q)µ
nvn
∞
0
f(x)
(x + εn)2
dx.
(4.2)
The Biased Cross Validation (BCV) criterion is defined as
BCV (vn, εn) = AMISE(fn). (4.3)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 47 / 75
48. 5. A Simulation Study
Here we consider parent distributions to estimate as exponential (χ2
2),
χ2
6, lognormal, Weibull and mixture of exponential densities.
Since the computation is very extensive for obtaining the smoothing
parameters, we compute approximations to MISE and MSE by
computing
ISE(fn, f) =
∞
0
[fn(x) − f(x)]2
dx
and
SE (fn(x), f(x)) = [fn(x) − f(x)]2
for 1000 samples.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 48 / 75
49. Here, MISE give us the global performance of density estimator.
MSE let us to see how the density estimator performs locally at the
points in which we might be interested. It is no doubt that we
particularly want to know the behavior of density estimators near the
lower boundary. We illustrate only MISE values.
Optimal values of smoothing parameters are obtained using either
BCV or UCV criterion, that roughly approximates Mean Integrated
Squared Error.
For Poisson smoothing as well as for Gamma smoothing BCV
criterion is found to be better, where as for Chen and Scaillet
method, use of BCV method is not tractable as the derivative of their
density is not explicitly available..
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 49 / 75
50. Next table gives the values of MISE for exponential density using new
estimators as compared with Chen’s and Scaillet estimators. Note
that we include the simulation results for Scaillet’s estimator using
RIG kernel only.
Inverse Gaussian kernel is known not to perform well for direct data
[see Kulasekera and Padgett (2006)]. Similar observations were noted
for LB data.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 50 / 75
55. Table : Simulated MSE for χ2
6
Sample Size Estimator
x
0 0.1 1 4 10
n=30
I 0.0017 0.0018 0.0018 0.0019 0.0001
II 0.0018 0.0017 0.0011 0.0017 0.0002
III 5.6 × 10−5
6.7 × 10−5
0.0006 0.0017 0.0002
IV 0.0016 0.0016 0.0012 0.0012 0.0001
V 0.0000 2.6 × 10−5
0.0017 0.0008 0.0001
VI 0.0011 0.0010 0.0019 0.0012 8.5 × 10−5
VI* 0.0015 0.0021 0.0020 0.0012 7.9 × 10−5
VII 0.0000 3.6 × 10−7
0.0058 0.0008 0.0001
VII* 0.0000 3.6 × 10−7
0.0058 0.0008 0.0001
n=50
I 0.0012 0.0013 0.0015 0.0012 0.0001
II 0.0013 0.0012 0.0008 0.0011 0.0001
III 4.6 × 10−5
5.7 × 10−5
0.0006 0.0005 .0001
IV 0.0011 0.0011 0.0010 0.0008 8.4 × 10−5
V 0.0000 6.7 × 10−5
0.0012 0.0005 0.0001
VI 0.0005 0.0005 0.0016 0.0008 5.5 × 10−5
VI* 0.0006 0.0015 0.0016 0.0008 5.3 × 10−5
VII 0.0000 4.3 × 10−6
0.0037 0.0004 7.9 × 10−5
VII* 0.0000 4.3 × 10−6
0.0037 0.0004 7.9 × 10−5
I-Chen-1, II-Chen-2, III-RIG, IV-Poisson(F), V-Poisson(G), VI-Gamma(F), VI*-Corrected Gamma(F), VII-Gamma(G),
VII*-Corrected Gamma(G)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 55 / 75
56. For the exponential density, ˆfC2 has smaller MSEs at the boundary
and MISEs than ˆfC1. This means ˆfC2 performs better locally and
globally than ˆfC1. Similar result holds in direct data.
Poisson weight estimator based on Fn is found to be better than that
based on Gn.
Although Poisson weight estimator based on Gn has relatively smaller
MISEs, it has large MSEs at the boundary as well, just like Scaillet
estimator.
Scaillet estimator has huge MSEs at the boundary and the largest
MISEs.
Corrected Gamma estimators have similar and smaller MISE values as
compared to the corresponding Poisson weight estimators.
For χ2
6, all estimators have comparable global results. Poison weight
estimators based on Fn or Gn have similar performances and may be
slightly better than the others.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 56 / 75
57. We have considered following additional distributions for simulation as well:
(i). Lognormal Distribution
f(x) =
1
√
2πx
exp{−(log x − µ)2
/2}I{x > 0};
(ii). Weibull Distribution
f(x) = αxα−1
exp(−xα
)I{x > 0};
(iii). Mixtures of Two Exponential Distribution
f(x) = [π
1
θ1
exp(−x/θ1) + (1 − π)
1
θ2
exp(−x/θ2]I{x > 0}.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 57 / 75
62. Table : Simulated MSE for Weibull with α = 2
Sample Size Estimator
x
0 0.1 1 2 3
n=30
I 0.0856 0.1343 0.0588 .0030 1.9 × 10−4
II 0.0949 0.0555 0.0398 .0116 1.4 × 10−4
III 0.0025 0.0802 0.0394 .0095 4.6 × 10−4
IV 0.0844 0.0548 0.0280 0.0086 6.9 × 10−4
V 0.0068 0.0636 0.0186 0.0031 3.1 × 10−5
VI* 0.0019 0.1049 0.0682 0.0053 0.0022
VII* 0.0000 0.2852 0.0336 0.0011 1.8 × 10−4
n=50
I 0.0644 0.0576 0.0349 0.0020 1.0 × 10−4
II 0.0679 0.0431 0.0223 0.0077 7.1 × 10−4
III 0.0021 0.0208 0.0218 0.0063 2.2 × 10−4
IV 0.0682 0.0427 0.0217 0.0059 3.7 × 10−4
V 0.0025 0.0453 0.0138 0.0018 1.6 × 10−5
VI* 1.1 × 10−6
0.0763 0.0560 0.0048 0.0018
VII* 0.0000 0.1865 0.0251 0.0008 1.4 × 10−4
I-Chen-1, II-Chen-2, III-RIG, IV-Poisson(F), V-Poisson(G), VI*-Corrected Gamma(F), VII*-Corrected Gamma(G)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 62 / 75
63. Table : Simulated MISE for Mixture of Two Exponential Distributions with
π = 0.4, θ1 = 2 and θ2 = 1
Distribution Estimator
Sample Size
30 50 100 200 300 500
Mixture
Chen-1 0.22876 0.17045 0.08578 0.06718 0.05523 0.03811
Chen-2 0.17564 0.15083 0.07331 0.08029 0.04931 0.03808
RIG 0.25284 0.20900 0.13843 0.10879 0.09344 0.07776
Poisson(F) 0.06838 0.05746 0.04116 0.02612 0.01896 0.01179
Poisson(G) 0.11831 0.09274 0.06863 0.05019 0.03881 0.03044
Gamma*(F) 0.04147 0.02645 0.01375 0.00758 0.00532 0.00361
Gamma*(G) 0.02534 0.01437 0.01091 0.01223 0.01132 0.00994
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 63 / 75
64. Table : Simulated MSE for Mixtures of Two Exponential Distributions with
π = 0.4, θ1 = 2 and θ2 = 1
Sample Size Estimator
x
0 0.1 1 2 10
n=30
I 0.3499 0.3075 0.0249 0.0037 2.6 × 10−6
II 0.3190 0.3181 0.0245 0.0071 1.3 × 10−5
III 0.5610 0.4423 0.0564 0.0056 2.9 × 10−6
IV 0.3778 0.1907 0.0057 0.0027 1.7 × 10−6
V 0.6409 0.3237 0.0156 0.0043 2.1 × 10−6
VI* 0.0652 0.0549 0.0098 0.0006 1.1 × 10−4
VII* 0.0696 0.0539 0.0065 0.0009 1.4 × 10−5
n=50
I 0.3158 0.7921 0.0128 0.0023 1.1 × 10−6
II 0.2848 0.7600 0.0143 0.0051 2.3 × 10−6
III 0.5582 0.8473 0.0364 0.0041 1.3 × 10−6
IV 0.3840 0.1633 0.0051 0.0020 1.0 × 10−6
V 0.6228 0.2673 0.0121 0.0028 1.3 × 10−6
VI* 0.0489 0.0414 0.0066 0.0004 7.7 × 10−5
VII* 0.0500 0.0336 0.0030 0.0007 9.2 × 10−6
I-Chen-1, II-Chen-2, III-RIG, IV-Poisson(F), V-Poisson(G), VI*-Corrected Gamma(F), VII*-Corrected Gamma(G)
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 64 / 75
65. The basic conclusion is that the smoothing based on Fn using
Poisson weights or corrected Gamma perform in a similar way and
produce better boundary correction as compared to Chen or Scaillet
asymmetric kernel estimators.
The smoothing based on Gn may have large local MSE near the
boundary and hence is not preferable over smoothing of Fn. A similar
message is given in Jones and Karunamuni (1997, Austr. J. Stat.).
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 65 / 75
66. References
Babu, G.J., Canty, A.J. and Chaubey, Y.P.(2002). Application of
Bernstein polynomials for smooth estimation of a distribution and
density function. J. Statist. Plann. Inference 105, no. 2, 377-392.
Babu, G.J. and Chaubey, Yogendra P. (2006). Smooth estimation of
a distribution and density function on a hypercube using Bernstein
polynomials for dependent random vectors. Statistics & Probability
Letters 76 959-969.
Bagai, I. and Prakasa Rao, B.L.S. (1996). Kernel Type Density
Estimates for Positive Valued Random Variables. Sankhya A57
56–67.
Bhattacharyya, B.B., L.A. Franklin, G.D. Richardson (1988). A
comparison of nonparametric unweighted and length biased density
estimation of fibres.Communications in Statistics- Theory Method
17(11) 3629–3644.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 66 / 75
67. Bouezmarni, T., Rombouts, J.V.K. and Taamouti, A. (2010).
Asymptotic Properties of the Bernstein Density Copula for Dependent
Data. Journal of Multivariate Analysis, 101, 1-10.
Bouezmarni, T. and Scaillet, O. (2005). Consistency of Asymmetric
Kernel Density Estimators and Smoothed Histograms with
Application to Income Data. Econometric Theory, 21, 390-412.
Chaubey, Y.P. and Dewan, I. and Li, J. (2010b). Smooth estimation
of survival and density functions for a stationary associated process
using Poisson weights Statistics and Probability Letters, 81, 267-276.
Chaubey, Y.P. and Kochar, S. (2000). Smooth estimation of
stochastically ordered Survival functions. J. Indian Statistical
Association, 38, 209-225.
Chaubey, Y. P. and Kochar, S. C. (2006). Smooth estimation of
uniformly stochastically ordered survival functions. Journal of
Combinatorics, Information and System Sciences, 31 1-13.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 67 / 75
68. Chaubey, Y.P. and Li, J. (2013). Asymmetric kernel density estimator
for length biased data. In Contemporary Topics in Mathematics and
Statistics with Applications, Vol. 1, (Ed.: A. Adhikari, M.R. Adhikari
and Y.P. Chaubey), Pub: Asian Books Pvt. Ltd., New Delhi, India.
Chaubey, Y. P. , Li, J., Sen, A. and Sen, P. K. (2012). A New
Smooth Density Estimator for Non-Negative Random Variables.
Journal of the Indian Statistical Association 50, 83-104.
Chaubey, Y. P. and Sen, P. K. (1996). On Smooth Estimation of
Survival and Density Function. Statistics and Decision, 14, 1-22.
Chaubey, Y.P., Sen, P.K. (1998a). On smooth estimation of hazard
and cumulative hazard functions. In Frontiers of Probability and
Statistics, S.P. Mukherjee et al. (eds.) Narosa: New Delhi; 91–99.
Chaubey, Y. P. and Sen, P. K. (1998b). On Smooth Functional
Estimation under Random Censorship. In Frontiers in Reliability 4.
Series on Quality, Reliability and Engineering Statistics (A. P. Basu et
al., eds.), 83-97. World Scientific, Singapore.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 68 / 75
69. Chaubey Y.P. and Sen P. K. (1999). On smooth estimation of mean
residual life. Journal of Statistical Planning and Inference 75 223–236.
Chaubey Y.P. and Sen, P.K. (2002a). Smooth isotonic estimation of
density, hazard and MRL functions. Calcutta Statistical Association
Bulletin, 52, 99-116.
Chaubey, Y. P. and Sen, P. K. (2002b). Smooth estimation of
Multivariate Survival and Density Functions. Jour. Statist. Plann.
and Inf., 103, 361-376.
Chaubey, Y. P. , Sen, A. and Sen, P. K. (2007a). A New Smooth
Density Estimator for Non-Negative Random Variables. Technical
Report No. 01/07, Department of Mathematics and Statistics,
Concordia University, Montreal.
Chaubey, Y. P. and Xu, H. (2007b) Smooth estimation of survival
functions under mean residual life ordering. Jour. Statist. Plann. Inf.
137, 3303-3316.
Chaubey, Y. P. and Sen, P. K. (2009). On the Selection of
Smoothing Parameter in Poisson Smoothing of Histogram Estimator:
Computational Aspects. Pak. J. Statist., 25(4), 385-401.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 69 / 75
70. Chaubey, Y. P. , Sen, P. K. and Li, J. (2010a). Smooth Density
Estimation for Length Biased Data. Journal of the Indian Society of
Agricultural Statistics, 64(2), 145-155.
Chen, S. X. (1999). Beta kernel estimators for density functions.
Computational Statistics and Data Analysis, 31, 131–145.
Chen, S. X. (2000). Probability Density Function Estimation Using
Gamma Kernels. Annals of the Institute of Statistical Mathematics,
52, 471-480.
Cox, D.R. (1969). Some Sampling Problems in Technology. In New
Developments in Survey Sampling, U. L. Johnson and H. Smith
(eds.), New York: Wiley Interscience.
Feller, W. (1965). An Introduction to Probability Theory and Its
Applications, Vol. II. New York: Wiley.
F¨oldes, A. and R´ev´esz, P. (1974). A general method for density
estimation. Studia Sci. Math. Hungar. (1974), 8192.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 70 / 75
71. Gawronski, W. (1980). Verallgemeinerte Bernsteinfunktionen und
Schtzung einer Wahrscheinlichkeitsdichte, Universit¨at Ulm,
(Habilitationsschrift).
Gawronski, W. (1985). Strong laws for density estimators of
Bernstein type. Periodica Mathematica Hungarica 16, 23-43
Gawronski, W. and Stadm¨uler, U. (1980). On Density Estimation by
Means of Poisson’s Distribution. Scandinavian Journal of Statistics,
7, 90-94.
Gawronski, W. and Stadm¨uler, U. (1981). Smoothing of Histograms
by Means of Lattice and Continuous Distributions. Metrika, 28,
155-164.
Jones, M.C. (1991). Kernel density estimation for length biased data.
Biometrika 78 511–519.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 71 / 75
72. Jones, M.C. and Karunamuni, R.J. (1997). Fourier series estimation
for length biased data. Australian Journal of Statistics, 39, 5768.
Kakizawa, Y. (2011). A note on generalized Bernstein polynomial
density estimators. Statistical Methodology, 8, 136-153.
Kulasekera, K. B. and Padgett, W. J. (2006). Bayes Bandwidth
Selection in Kernel Density Estimation with Censored Data. Journal
of Nonparametric Statistics, 18, 129-143.
Marron, J. S., Ruppert, D. (1994). Transformations to reduce
boundary bias in kernel density estimation. J. Roy. Statist. Soc. Ser.
B 56, 653-671.
Muttlak, H.A. and McDonald, L.L. (1990). Ranked set sampling with
size-biased sampling with applications to wildlife populations and
human families. Biometrics 46, 435-445.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 72 / 75
73. Prakasa Rao, B. L. S. (1983). Nonparametric Functional Estimation.
Academic Press:New York.
Prakasa Rao, B. L. S. (2005). Estimation of distributions and density
functions by generalized Bernstein polynomials. Indian J. Pure and Applied
Math. 36, 63-88.
Rosenblatt, M. (1956). Remarks on some nonparametric estimates of
density functions. Ann.Math. Statist. 27 832–837.
Scaillet, O. (2004). Density Estimation Using Inverse and Reciprocal Inverse
Gaussian Kernels. Journal of Nonparametric Statistics, 16, 217-226.
Silverman, B. W. (1986). Density Estimation for Statistics and Data
Analysis. Chapman and Hall: London.
Vitale, R. A. (1975). A Bernstein polynomial approach to density
estimation. In Statistical Inference and Related Topics (ed. M. L. Puri), 2
87–100. New York: Academic Press.
Wand, M.P., Marron, J.S. and Ruppert, D. (1991). Transformations in
density estimation. Journal of the American Statistical Association, 86,
343-361.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 73 / 75
74. Acknowledgement of Collaborators
Most of the results in this talk are taken from the following papers
Chaubey, Y.P. and Li, J. (2013). Asymmetric kernel density estimator
for length biased data. In Contemporary Topics in Mathematics and
Statistics with Applications, Vol. 1, (Ed.: A. Adhikari, M.R. Adhikari
and Y.P. Chaubey), Pub: Asian Books Pvt. Ltd., New Delhi, India.
Chaubey, Y. P. , Li, J., Sen, A. and Sen, P. K. (2012). A New
Smooth Density Estimator for Non-Negative Random Variables.
Journal of the Indian Statistical Association 50, 83-104.
Chaubey, Y. P. and Sen, P. K. (1996). On Smooth Estimation of
Survival and Density Function. Statistics and Decision, 14, 1-22.
Chaubey, Y. P. , Sen, P. K. and Li, J. (2010a). Smooth Density
Estimation for Length Biased Data. Journal of the Indian Society of
Agricultural Statistics, 64(2), 145-155.
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 74 / 75
75. Talk slides are available on SlideShare:
http://www.slideshare.net/YogendraChaubey/Slides-CSM
THANKS!!
Yogendra Chaubey (Concordia University) Department of Mathematics & Statistics June 11, 2015 75 / 75