Academic literature on the topic 'Parametric regression models'

Quantile regression is widely used to estimate conditional quantiles of an outcome variable of interest given covariates. This method can estimate one quantile at a time without imposing any constraints on the quantile process other than the linear combination of covariates and parameters specified by the regression model. While this is a flexible modeling tool, it generally yields erratic estimates of conditional quantiles and regression coefficients. Recently, parametric models for the regression coefficients have been proposed that can help balance bias and sampling variability. So far, however, only models that are linear in the parameters and covariates have been explored. This paper presents the general case of nonlinear parametric quantile models. These can be nonlinear with respect to the parameters, the covariates, or both. Some important features and asymptotic properties of the proposed estimator are described, and its finite-sample behavior is assessed in a simulation study. Nonlinear parametric quantile models are applied to estimate extreme quantiles of longitudinal measures of respiratory mechanics in asthmatic children from an epidemiological study and to evaluate a dose–response relationship in a toxicological laboratory experiment.

There are three common types of regression models: parametric, semiparametric and nonparametric regression. The model should be used to fit the real data depends on how much information is available about the form of the relationship between the response variable and explanatory variables, and the random error distribution that is assumed. Researchers need to be familiar with each modeling approach requirements. In this paper, differences between these models, common estimation methods, robust estimation, and applications are introduced. For parametric models, there are many known methods of estimation, such as least squares and maximum likelihood methods which are extensively studied but they require strong assumptions. On the other hand, nonparametric regression models are free of assumptions regarding the form of the response-explanatory variables relationships but estimation methods, such as kernel and spline smoothing are computationally expensive and smoothing parameters need to be obtained. For kernel smoothing there two common estimators: local constant and local linear smoothing methods. In terms of bias, especially at the boundaries of the data range, local linear is better than local constant estimator.  Robust estimation methods for linear models are well studied, however the robust estimation methods in nonparametric regression methods are limited. A robust estimation method for the semiparametric and nonparametric regression models is introduced.

Proportional hazard regression models are widely used in survival analysis to understand and exploit the relationship between survival time and covariates. For left censored survival times, reversed hazard rate functions are more appropriate. In this paper, we develop a parametric proportional hazard rates model using an inverted Weibull distribution. The estimation and construction of confidence intervals for the parameters are discussed. We assess the performance of the proposed procedure based on a large number of Monte Carlo simulations. We illustrate the proposed method using a real case example.

Model checking for regressions has drawn considerable attention in the last three decades. Compared with global smoothing tests, local smoothing tests, which are more sensitive to high-frequency alternatives, can only detect local alternatives dis- tinct from the null model at a much slower rate when the dimension of predictor is high. When the number of covariates is large, nonparametric estimations used in local smoothing tests lack efficiency. Corresponding tests then have trouble in maintaining the significance level and detecting the alternatives. To tackle the issue, we propose two methods under high but fixed dimension framework. Further, we investigate a model checking test under divergent dimension, where the numbers of covariates and unknown parameters go divergent with the sample size n. The first proposed test is constructed upon a typical kernel-based local smoothing test using projection method. Employed by projection and integral, the resulted test statistic has a closed form that depends only on the residuals and distances of the sample points. A merit of the developed test is that the distance is easy to implement compared with the kernel estimation, especially when the dimension is high. Moreover, the test inherits some feature of local smoothing tests owing to its construction. Although it is eventually similar to an Integrated Conditional Moment test in spirit, it leads to a test with a weight function that helps to collect more information from the samples than Integrated Conditional Moment test. Simulations and real data analysis justify the powerfulness of the test. The second test, which is a synthesis of local and global smoothing tests, aims at solving the slow convergence rate caused by nonparametric estimation in local smoothing tests. A significant feature of this approach is that it allows nonparamet- ric estimation-based tests, under the alternatives, also share the merits of existing empirical process-based tests. The proposed hybrid test can detect local alternatives at the fastest possible rate like the empirical process-based ones, and simultane- ously, retains the sensitivity to high-frequency alternatives from the nonparametric estimation-based ones. This feature is achieved by utilizing an indicative dimension in the field of dimension reduction. As a by-product, we have a systematic study on a residual-related central subspace for model adaptation, showing when alterna- tive models can be indicated and when cannot. Numerical studies are conducted to verify its application. Since the data volume nowadays is increasing, the numbers of predictors and un- known parameters are probably divergent as sample size n goes to infinity. Model checking under divergent dimension, however, is almost uncharted in the literature. In this thesis, an adaptive-to-model test is proposed to handle the divergent dimen- sion based on the two previous introduced tests. Theoretical results tell that, to get the asymptotic normality of the parameter estimator, the number of unknown parameters should be in the order of o(n1/3). Also, as a spinoff, we demonstrate the asymptotic properties of estimations for the residual-related central subspace and central mean subspace under different hypotheses.

Master of Science
Department of Statistics
Weixing Song
In the classical Tobit regression model, the regression error term is often assumed to have a zero mean normal distribution with unknown variance, and the regression function is assumed to be linear. If the normality assumption is violated, then the commonly used maximum likelihood estimate becomes inconsistent. Moreover, the likelihood function will be very complicated if the regression function is nonlinear even the error density is normal, which makes the maximum likelihood estimation procedure hard to implement. In the full nonparametric setup when both the regression function and the distribution of the error term [epsilon] are unknown, some nonparametric estimators for the regression function has been proposed. Although the assumption of knowing the distribution is strict, it is a widely adopted assumption in Tobit regression literature, and is also confirmed by many empirical studies conducted in the econometric research. In fact, a majority of the relevant research assumes that [epsilon] possesses a normal distribution with mean 0 and unknown standard deviation. In this report, we will try to develop a semi-parametric estimation procedure for the regression function by assuming that the error term follows a distribution from a class of 0-mean symmetric location and scale family. A minimum distance estimation procedure for estimating the parameters in the regression function when it has a specified parametric form is also constructed. Compare with the existing semiparametric and nonparametric methods in the literature, our method would be more efficient in that more information, in particular the knowledge of the distribution of [epsilon], is used. Moreover, the computation is relative inexpensive. Given lots of application does assume that [epsilon] has normal or other known distribution, the current work no doubt provides some more practical tools for statistical inference in Tobit regression model.

The central objective of this work is to develop statistical tools for semi-parametric regression models with generalized log-gamma errors under the presence of censored and uncensored observations. The estimates of the parameters are obtained through the multivariate version of Newton-Raphson algorithm and an adequate combination of Fisher Scoring and Backffitting algorithms. Through analytical tools and using simulations the properties of the penalized maximum likelihood estimators are studied. Some diagnostic techniques such as quantile and deviance-type residuals as well as local influence measures are derived. The methodologies are implemented in the statistical computational environment R. The package sglg is developed. Finally, we give some applications of the models to real data.
O objetivo central do trabalho é proporcionar ferramentas estatísticas para modelos de regressão semiparamétricos quando os erros seguem distribução log-gamma generalizada na presença de observações censuradas ou não censuradas. A estimação paramétrica e não paramétrica são realizadas através dos procedimentos Newton - Raphson, escore de Fisher e Backfitting (Gauss - Seidel). As propriedades assintóticas dos estimadores de máxima verossimilhança penalizada são estudadas em forma analítica, bem como através de simulações. Alguns procedimentos de diagnóstico são desenvolvidos, tais como resíduos tipo componente do desvio e resíduo quantílico, bem como medidas de influ\\^encia local sob alguns esquemas usuais de perturbação. Todos procedimentos do presente trabalho são implementados no ambiente computacional R, o pacote sglg é desenvolvido, assim como algumas aplicações a dados reais são apresentadas.

In a regression context, the aim is to analyse a response variable of interest conditional to a set of covariates. In many applications the response variable is discrete. Examples include the event of surviving a heart attack, the number of hospitalisation days, the number of times that individuals benefit of a health service, and so on. This thesis advances the methodology and the application of regression models with discrete response. First, we present a difference-in-differences approach to model a binary response in a health policy evaluation framework. In particular, generalized linear mixed methods are employed to model multiple dependent outcomes in order to quantify the effect of an adopted pay-for-performance program while accounting for the heterogeneity of the data at the multiple nested levels. The results show how the policy had a positive effect on the hospitals' quality in terms of those outcomes that can be more influenced by a managerial activity. Next, we focus on regression models for count response variables. In a parametric framework, Poisson regression is the simplest model for count data though it is often found not adequate in real applications, particularly in the presence of excessive zeros and in the case of dispersion, i.e. when the conditional mean is different to the conditional variance. Negative Binomial regression is the standard model for over-dispersed data, but it fails in the presence of under-dispersion. Poisson-Inverse Gaussian regression can be used in the case of over-dispersed data, Generalised-Poisson regression can be employed in the case of under-dispersed data, and Conway-Maxwell Poisson regression can be employed in both cases of over- or under-dispersed data, though the interpretability of these models is ot straightforward and they are often found computationally demanding. While Jittering is the default non-parametric approach for count data, inference has to be made for each individual quantile, separate quantiles may cross and the underlying uniform random sampling can generate instability in the estimation. These features motivate the development of a novel parametric regression model for counts via a Discrete Weibull distribution. This distribution is able to adapt to different types of dispersion relative to Poisson, and it also has the advantage of having a closed form expression for the quantiles. As well as the standard regression model, generalized linear mixed models and generalized additive models are presented via this distribution. Simulated and real data applications with different type of dispersion show a good performance of Discrete Weibull-based regression models compared with existing regression approaches for count data.

This thesis analyses, derives and evaluates specification tests of Generalized Auto-Regressive Conditional Heteroskedasticity (GARCH) regression models, both univariate and multivariate. Of particular interest, in the first half of the thesis, is the derivation of robust test procedures designed to assess the Constant Conditional Correlation (CCC) assumption often employed in multivariate GARCH (MGARCH) models. New asymptotically valid conditional moment tests are proposed which are simple to construct, easily implementable following the full or partial Quasi Maximum Likelihood (QML) estimation and which are robust to non-normality. In doing so, a non-normality robust version of the Tse's (2000) LM test is provided. In addition, a new and easily programmable expressions of the expected Hessian matrix associated with the QMLE is obtained. The finite sample performances of these tests are investigated in an extensive Monte Carlo study, programmed in GAUSS.In the second half of the thesis, attention is devoted to nonparametric testing of GARCH regression models. First simultaneous consistent nonparametric tests of the conditional mean and conditional variance structure of univariate GARCH models are considered. The approach is developed from the Integrated Generalized Spectral (IGS) and Projected Integrated Conditional Moment (PICM) procedures proposed recently by Escanciano (2008 and 2009, respectively) for time series models. Extending Escanciano (2008), a new and simple wild bootstrap procedure is proposed to implement these tests. A Monte Carlo study compares the performance of these nonparametric tests and four parametric tests of nonlinearity and/or asymmetry under a wide range of alternatives. Although the proposed bootstrap scheme does not strictly satisfy the asymptotic requirements, the simulation results demonstrate its ability to control the size extremely well and therefore the power comparison seems justified. Furthermore, this suggests there may exist weaker conditions under which the tests are implementable. The simulation exercise also presents the new evidence of the effect of conditional mean misspecification on various parametric tests of conditional variance. The testing procedures are also illustrated with the help of the S&P 500 data. Finally the PICM and IGS approaches are extended to the MGARCH case. The procedure is illustrated with the help of a bivariate CCC-GARCH model, but can be generalized to other MGARCH specifications. Simulation exercise shows that these tests have satisfactory size and are robust to non-normality. The marginal mean and variance tests have excellent power; however the covariance marginal tests lack power for some alternatives.

Submitted by RAFAEL ESPIGOLAN (espigolan@yahoo.com.br) on 2017-03-17T22:04:14Z No. of bitstreams: 1 Tese_Rafael_Espigolan.pdf: 1532864 bytes, checksum: c79ad7471b25137c47529f25762a83a2 (MD5)
Approved for entry into archive by Juliano Benedito Ferreira (julianoferreira@reitoria.unesp.br) on 2017-03-22T12:50:50Z (GMT) No. of bitstreams: 1 espigolan_r_dr_jabo.pdf: 1532864 bytes, checksum: c79ad7471b25137c47529f25762a83a2 (MD5)
Made available in DSpace on 2017-03-22T12:50:50Z (GMT). No. of bitstreams: 1 espigolan_r_dr_jabo.pdf: 1532864 bytes, checksum: c79ad7471b25137c47529f25762a83a2 (MD5) Previous issue date: 2017-02-23
Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP)
O melhoramento genético animal visa melhorar a produtividade econômica das futuras gerações de espécies domésticas por meio da seleção. A maioria das características de interesse econômico na pecuária é de expressão quantitativa e complexa, isto é, são influenciadas por vários genes e afetadas por fatores ambientais. As análises estatísticas de informações de fenótipo e pedigree permite estimar os valores genéticos dos candidatos à seleção com base no modelo infinitesimal. Uma grande quantidade de dados genômicos está atualmente disponível para a identificação e seleção de indivíduos geneticamente superiores com o potencial de aumentar a acurácia de predição dos valores genéticos e, portanto, a eficiência dos programas de melhoramento genético animal. Vários estudos têm sido conduzidos com o objetivo de identificar metodologias apropriadas para raças e características específicas, o que resultará em estimativas de valores genéticos genômicos (GEBVs) mais acurados. Portanto, o objetivo deste estudo foi verificar a possibilidade de aplicação de modelos semiparamétricos para a seleção genômica e comparar a habilidade de predição com os modelos paramétricos para dados reais (características de carcaça, qualidade da carne, crescimento e reprodutiva) e simulados. As informações fenotípicas e de pedigree utilizadas foram fornecidas por onze fazendas pertencentes a quatro programas de melhoramento genético animal. Para as características de carcaça e qualidade da carne, o banco de dados continha 3.643 registros para área de olho de lombo (REA), 3.619 registros para espessura de gordura (BFT), 3.670 registros para maciez da carne (TEN) e 3.378 observações para peso de carcaça quente (HCW). Um total de 825.364 registros para peso ao sobreano (YW) e 166.398 para idade ao primeiro parto (AFC) foi utilizado para as características de crescimento e reprodutiva. Genótipos de 2.710, 2.656, 2.749, 2.495, 4.455 e 1.760 animais para REA, BFT, TEN, HCW, YW e AFC foram disponibilizados, respectivamente. Após o controle de qualidade, restaram dados de, aproximadamente, 450.000 polimorfismos de base única (SNP). Os modelos de análise utilizados foram BLUP genômico (GBLUP), single-step GBLUP (ssGBLUP), Bayesian LASSO (BL) e as abordagens semiparamétricas Reproducing Kernel Hilbert Spaces (RKHS) e Kernel Averaging (KA). Para cada característica foi realizada uma validação cruzada composta por cinco “folds” e replicada aleatoriamente trinta vezes. Os modelos estatísticos foram comparados em termos do erro do quadrado médio (MSE) e acurácia de predição (ACC). Os valores de ACC variaram de 0,39 a 0,40 (REA), 0,38 a 0,41 (BFT), 0,23 a 0,28 (TEN), 0,33 a 0,35 (HCW), 0,36 a 0,51 (YW) e 0,49 a 0,56 (AFC). Para todas as características, os modelos GBLUP e BL apresentaram acurácias de predição similares. Para REA, BFT e HCW, todos os modelos apresentaram ACC similares, entretanto a regressão RKHS obteve o melhor ajuste comparado ao KA. Para características com maior quantidade de registros fenotípicos comparada ao número de animais genotipados (YW e AFC) o modelo ssGBLUP é indicado. Considerando o desempenho geral, para todas as características estudadas, a regressão RKHS é, particularmente, uma alternativa interessante para a aplicação na seleção genômica, especialmente para características de baixa herdabilidade. No estudo de simulação, genótipos, pedigree e fenótipos para quatro características (A, B, C e D) foram simulados utilizando valores de herdabilidade baseados nos obtidos com os dados reais (0,09, 0,12, 0,36 e 0,39 para cada característica, respectivamente). O genoma simulado consistiu de 735.293 marcadores e 1.000 QTLs distribuídos aleatoriamente por 29 pares de autossomos, com comprimento variando de 40 a 146 centimorgans (cM), totalizando 2.333 cM. Assumiu-se que os QTLs explicavam 100% da variação genética. Considerando as frequências do alelo menor maiores ou iguais a 0,01, um total de 430.000 marcadores foram selecionados aleatoriamente. Os fenótipos foram obtidos pela soma dos resíduos (aleatoriamente amostrados de uma distribuição normal com média igual a zero) aos valores genéticos verdadeiros, e todo o processo de simulação foi replicado 10 vezes. A ACC foi calculada por meio da correlação entre o valor genético genômico estimado e o valor genético verdadeiro, simulados da 12a a 15a geração. A média do desequilíbrio de ligação, medido entre os pares de marcadores adjacentes para todas as características simuladas foi de 0,21 para as gerações recentes (12a, 13a e 14a), e 0,22 para a 15a geração. A ACC para as características simuladas A, B, C e D variou de 0,43 a 0,44, 0,47 a 0,48, 0,80 a 0,82 e 0,72 a 0,73, respectivamente. Diferentes metodologias de seleção genômica implementadas neste estudo mostraram valores similares de acurácia de predição, e o método mais adequado é dependente da característica explorada. Em geral, as regressões RKHS obtiveram melhor desempenho em termos de ACC com menor valor de MSE em comparação com os outros modelos.
Animal breeding aims to improve economic productivity of future generations of domestic species through selection. Most of the traits of economic interest in livestock have a complex and quantitative expression i.e. are influenced by a large number of genes and affected by environmental factors. Statistical analysis of phenotypes and pedigree information allows estimating the breeding values of the selection candidates based on infinitesimal model. A large amount of genomic data is now available for the identification and selection of genetically superior individuals with the potential to increase the accuracy of prediction of genetic values and thus, the efficiency of animal breeding programs. Numerous studies have been conducted in order to identify appropriate methodologies to specific breeds and traits, which will result in more accurate genomic estimated breeding values (GEBVs). Therefore, the objective of this study was to verify the possibility of applying semi-parametric models for genomic selection and to compare their ability of prediction with those of parametric models for real (carcass, meat quality, growth and reproductive traits) and simulated data. The phenotypic and pedigree information used were provided by farms belonging to four animal breeding programs which represent eleven farms. For carcass and meat quality traits, the data set contained 3,643 records for rib eye area (REA), 3,619 records for backfat thickness (BFT), 3,670 records for meat tenderness (TEN) and 3,378 observations for hot carcass weight (HCW). A total of 825,364 records for yearling weight (YW) and 166,398 for age at first calving (AFC) were used as growth and reproductive traits of Nelore cattle. Genotypes of 2,710, 2,656, 2,749, 2,495, 4,455 and 1,760 animals were available for REA, BFT, TEN, HCW, YW and AFC, respectively. After quality control, approximately 450,000 single nucleotide polymorphisms (SNP) remained. Methods of analysis were genomic BLUP (GBLUP), single-step GBLUP (ssGBLUP), Bayesian LASSO (BL) and the semi-parametric approaches Reproducing Kernel Hilbert Spaces (RKHS) regression and Kernel Averaging (KA). A five-fold cross-validation with thirty random replicates was carried out and models were compared in terms of their prediction mean squared error (MSE) and accuracy of prediction (ACC). The ACC ranged from 0.39 to 0.40 (REA), 0.38 to 0.41 (BFT), 0.23 to 0.28 (TEN), 0.33 to 0.35 (HCW), 0.36 to 0.51 (YW) and 0.49 to 0.56 (AFC). For all traits, the GBLUP and BL models showed very similar prediction accuracies. For REA, BFT and HCW, models provided similar prediction accuracies, however RKHS regression had the best fit across traits considering multiple-step models and compared to KA. For traits which have a higher number of animals with phenotypes compared to the number of those with genotypes (YW and AFC), the ssGBLUP is indicated. Judged by overall performance, across all traits, the RKHS regression is particularly appealing for application in genomic selection, especially for low heritability traits. Simulated genotypes, pedigree, and phenotypes for four traits A, B, C and D were obtained using heritabilities based on real data (0.09, 0.12, 0.36 and 0.39 for each trait, respectively). The simulated genome consisted of 735,293 markers and 1,000 QTLs randomly distributed over 29 pairs of autosomes, with length varying from 40 to 146 centimorgans (cM), totaling 2,333 cM. It was assumed that QTLs explained 100% of genetic variance. Considering Minor Allele Frequencies greater or equal to 0.01, a total of 430,000 markers were randomly selected. The phenotypes were generated by adding residuals, randomly drawn from a normal distribution with mean equal to zero, to the true breeding values and all simulation process was replicated 10 times. ACC was quantified using correlations between the predicted genomic breeding value and true breeding values simulated for the generations of 12 to 15. The average linkage disequilibrium, measured between pairs of adjacent markers for all simulated traits was 0.21 for recent generations (12, 13 and 14), and 0.22 for generation 15. The ACC for simulated traits A, B, C and D ranged from 0.43 to 0.44, 0.47 to 0.48, 0.80 to 0.82 and 0.72 to 0.73, respectively. Different genomic selection methodologies implemented in this study showed similar accuracies of prediction, and the optimal method was sometimes trait dependent. In general, RKHS regressions were preferable in terms of ACC and provided smallest MSE estimates compared to other models.
FAPESP: 2014/00779-0
FAPESP: 2015/13084-3

The dependence between survival times and covariates is described e.g. by proportional hazard models. We consider partly parametric Cox models and discuss here the estimation of interesting parameters. We represent the ma- ximum likelihood approach and extend the results of Huang (1999) from linear to nonlinear parameters. Then we investigate the least squares esti- mation and formulate conditions for the a.s. boundedness and consistency of these estimators.

In doctoral thesis redundant systems and degradation models are considered. To ensure high reliability of important elements of the system, the stand-by units can be used. These units are commuted and operate instead of the main failed unit. The stand-by units can function in the different conditions: “hot”, “cold” or “warm” reserving. In the thesis systems with “warm” stand-by units are analyzed. Hypotheses of smooth commuting are formulated and goodness-of-fit tests for these hypotheses are constructed. Nonparametric and parametric point and interval estimation procedures are given. Modeling and statistical estimation of reliability of systems from failure time and degradation data are considered.
Daktaro disertacijos tyrimo objektai yra rezervuotos sistemos ir degradaciniai modeliai. Norint užtikrinti svarbių sistemos elementų aukštą patikimumą, naudojami jų rezerviniai elementai, kurie gali būti įjungiami sugedus šiems pagrindiniams elementams. Rezerviniai elementai gali funkcionuoti skirtinguose režimuose: „karštame“, „šaltame“ arba „šiltame“. Disertacijoje yra nagrinėjamos sistemos su „šiltai“ rezervuotais elementais. Darbe suformuluojama rezervinio elemento „sklandaus įjungimo“ hipotezė ir konstruojami statistiniai kriterijai šiai hipotezei tikrinti. Nagrinėjami neparametrinio ir parametrinio taškinio bei intervalinio vertinimo uždaviniai. Disertacijoje nagrinėjami pakankamai bendri degradacijos modeliai, kurie aprašo elementų gedimų intensyvumą kaip funkciją kiek naudojamų apkrovų, tiek ir degradacijos lygio, kuri savo ruožtu modeliuojama naudojant stochastinius procesus.

This article presents a unifying classification for functional regression modeling, and more specifically for modeling the link between two variables X and Y, when the explanatory variable (X) is of a functional nature. It first provides a background on the proposed classification of regression models, focusing on the regression problem and defining parametric, semiparametric, and nonparametric models, and explains how semiparametric modeling can be interpreted in terms of dimension reduction. It then gives four examples of functional regression models, namely: functional linear regression model, additive functional regression model, smooth nonparametric functional model, and single functional index model. It also considers a number of new models, directly adapted to functional variables from the existing standard multivariate literature.

Parametric bootstrapping (BS) provides an attractive alternative, both theoretically and numerically, to asymptotic theory for estimating sampling distributions. This chapter summarizes its use not only for calculating confidence intervals for estimated parameters and functions of parameters, but also to obtain log-likelihood-based confidence regions from which confidence bands for cumulative distribution and regression functions can be obtained. All such BS calculations are very easy to implement. Details are also given for calculating critical values of EDF statistics used in goodness-of-fit (GoF) tests, such as the Anderson-Darling A2 statistic whose null distribution is otherwise difficult to obtain, as it varies with different null hypotheses. A simple proof is given showing that the parametric BS is probabilistically exact for location-scale models. A formal regression lack-of-fit test employing parametric BS is given that can be used even when the regression data has no replications. Two real data examples are given.

This chapter introduces embedded models. This is a special case of a parametric model which cannot be obtained simply by setting the parameters to particular values in a simple way. An example is the regression function y = b[1−exp(−ax)], which is always curved when a and b have fixed values. But letting a tend to zero and b tend to infinity simultaneously, whilst keeping ab = c fixed, yields y = cx, a straight-line special case. When this is the true model, fitting the original two-parameter model leads to very unstable and individually meaningless estimates of a and b. Such embedded models are actually very common in the literature, leading to confusion in interpretation of results when undetected. In this chapter, embeddedness is defined and a large number of regression embedded model examples given. Detection and removal of embeddedness by reparametrization is discussed. Two real data numerical examples are given.

Combustion of a single fuel droplet has long been studied because it establishes the background of understanding the behavior of spray combustion. Ignition is the most critical process during its life time. According to a practical ignition criterion, many parametric studies are performed to examine the effects on ignition under different droplet conditions and different ambient conditions. In addition, two liquid-phase models, infinite conductivity model and conduction limit model, are discussed to demonstrate the heating effect of droplet itself. Dual-period concept is introduced to clarify the dominant factor that governs the ignition process. A special hybrid numerical scheme, ICED-ALE, is used to resolve the difficulties arising from the rapid transition of ignition and regression behavior of a fuel droplet. The numerical predictions show good agreement with the experimental data.

Optimal control algorithms such as distributed model predictive control (DMPC) offer tremendous potential in reducing energy consumption of building operations. Heating, ventilation and air-conditioning (HVAC) systems which form a major part of the building operations contain a large number of interconnected subsystems. One of the challenges associated with implementing DMPC is the development of reliable models of individual subsystems for prediction, especially for large scale systems. In this paper an automated method is proposed to develop linear parametric black box models for individual building HVAC subsystems. The modeling method proposed identifies the significant inputs, and the upstream and downstream neighbors of each subsystem before performing regression analysis to determine the model parameters. Automation of the model development makes the implementation of the model-based control algorithms much more feasible. The modeling method is then verified through an EnergyPLus model, and using data of a real office building.

The objectives of this project were to develop procedures and models, based on neural networks, for quality sorting of agricultural produce. Two research teams, one in Purdue University and the other in Israel, coordinated their research efforts on different aspects of each objective utilizing both melons and tomatoes as case studies. At Purdue: An expert system was developed to measure variances in human grading. Data were acquired from eight sensors: vision, two firmness sensors (destructive and nondestructive), chlorophyll from fluorescence, color sensor, electronic sniffer for odor detection, refractometer and a scale (mass). Data were analyzed and provided input for five classification models. Chlorophyll from fluorescence was found to give the best estimation for ripeness stage while the combination of machine vision and firmness from impact performed best for quality sorting. A new algorithm was developed to estimate and minimize training size for supervised classification. A new criteria was established to choose a training set such that a recurrent auto-associative memory neural network is stabilized. Moreover, this method provides for rapid and accurate updating of the classifier over growing seasons, production environments and cultivars. Different classification approaches (parametric and non-parametric) for grading were examined. Statistical methods were found to be as accurate as neural networks in grading. Classification models by voting did not enhance the classification significantly. A hybrid model that incorporated heuristic rules and either a numerical classifier or neural network was found to be superior in classification accuracy with half the required processing of solely the numerical classifier or neural network. In Israel: A multi-sensing approach utilizing non-destructive sensors was developed. Shape, color, stem identification, surface defects and bruises were measured using a color image processing system. Flavor parameters (sugar, acidity, volatiles) and ripeness were measured using a near-infrared system and an electronic sniffer. Mechanical properties were measured using three sensors: drop impact, resonance frequency and cyclic deformation. Classification algorithms for quality sorting of fruit based on multi-sensory data were developed and implemented. The algorithms included a dynamic artificial neural network, a back propagation neural network and multiple linear regression. Results indicated that classification based on multiple sensors may be applied in real-time sorting and can improve overall classification. Advanced image processing algorithms were developed for shape determination, bruise and stem identification and general color and color homogeneity. An unsupervised method was developed to extract necessary vision features. The primary advantage of the algorithms developed is their ability to learn to determine the visual quality of almost any fruit or vegetable with no need for specific modification and no a-priori knowledge. Moreover, since there is no assumption as to the type of blemish to be characterized, the algorithm is capable of distinguishing between stems and bruises. This enables sorting of fruit without knowing the fruits' orientation. A new algorithm for on-line clustering of data was developed. The algorithm's adaptability is designed to overcome some of the difficulties encountered when incrementally clustering sparse data and preserves information even with memory constraints. Large quantities of data (many images) of high dimensionality (due to multiple sensors) and new information arriving incrementally (a function of the temporal dynamics of any natural process) can now be processed. Furhermore, since the learning is done on-line, it can be implemented in real-time. The methodology developed was tested to determine external quality of tomatoes based on visual information. An improved model for color sorting which is stable and does not require recalibration for each season was developed for color determination. Excellent classification results were obtained for both color and firmness classification. Results indicted that maturity classification can be obtained using a drop-impact and a vision sensor in order to predict the storability and marketing of harvested fruits. In conclusion: We have been able to define quantitatively the critical parameters in the quality sorting and grading of both fresh market cantaloupes and tomatoes. We have been able to accomplish this using nondestructive measurements and in a manner consistent with expert human grading and in accordance with market acceptance. This research constructed and used large databases of both commodities, for comparative evaluation and optimization of expert system, statistical and/or neural network models. The models developed in this research were successfully tested, and should be applicable to a wide range of other fruits and vegetables. These findings are valuable for the development of on-line grading and sorting of agricultural produce through the incorporation of multiple measurement inputs that rapidly define quality in an automated manner, and in a manner consistent with the human graders and inspectors.