Search

Causality is important in many engineering applications for the optimization, robustness, and control of manufacturing processes. Randomized experiments have been the conventional tool for establishing and quantifying causal effects. With great advances in sensorics and information technology, the temptation to use an unprecedented amount of observational data for this purpose has been growing. Most classically trained data scientists are warned against such practice as jumping from correlation to causality is presented as a treacherous leap. Yet causal inference based on observational data has been of great interest in, for example, social and medical studies for which randomized experiments can be infeasible or even unethical. In this Quality Quandaries, we propose a compromise where observational data, with the help of process expertise, is used to establish the set of factors that will be further tested for causality. We demonstrate a practical application of our proposal through a case study in additive manufacturing. [ABSTRACT FROM AUTHOR]

In the era of big data, companies are increasingly driven to amass vast amounts of data, particularly in process industries where advanced sensor technologies are prevalent. However, obtaining accurate labels or product information through quality inspections can be prohibitively expensive. Active learning emerges as a promising approach to optimize data sampling by prioritizing the most informative data points. Nevertheless, active learning strategies heavily rely on predictive models that are iteratively updated. Aligning with the principles of data-centric AI, this study highlights the detrimental effects of passively incorporating all available process variables into a predictive model for guiding data collection. Specifically, in real-time sampling strategies based on online active learning, the inclusion of irrelevant features significantly hampers the efficiency of the learning process. [ABSTRACT FROM AUTHOR]

We present an overview of predictive maintenance (PdM) in industrial operations, highlighting its evolution, benefits, challenges, and potential economic impact. It is assessed that there are considerable benefits associated with PdM but also its practical implementation can prove difficult due to the uncertain benefits of its application, the need for advanced IT infrastructure as well as expert personnel, and, perhaps most importantly, a lack of failure-related data for the PdM model training. We conclude with three PdM case studies to elaborate on some of the issues discussed throughout the paper. [ABSTRACT FROM AUTHOR]

In most data analytic applications in manufacturing, understanding the data‐driven models plays a crucial role in complementing the engineering knowledge about the production process. Identifying relevant input variables, rather than only predicting the response through some "black‐box" model, is of great interest in many applications. There is, therefore, a growing focus on describing the contributions of the input variables to the model in the form of "variable importance", which is readily available in certain machine learning methods such as random forest (RF). Once a ranking based on the importance measure of the variables is established, the question of how many variables are truly relevant in predicting the output variable rises. In this study, we focus on the Boruta algorithm, which is a wrapper around the RF model. It is a variable selection tool that assesses the variable importance measure for the RF model. It has been previously shown in the literature that the correlation among the input variables, which is often a common occurrence in high dimensional data, distorts and overestimates the importance of variables. The Boruta algorithm is also affected by this resulting in a larger set of input variables deemed important. To overcome this issue, in this study, we propose an extension of the Boruta algorithm for the correlated data by exploiting the conditional importance measure. This extension greatly improves the Boruta algorithm in the case of high correlation among variables and provides a more precise ranking of the variables that significantly contribute to the response. We believe this approach can be used in many industrial applications by providing more transparency and understanding of the process. [ABSTRACT FROM AUTHOR]

In many industrial applications, obtaining labeled observations is not straightforward as it often requires the intervention of human experts or the use of expensive testing equipment. In these circumstances, active learning can be highly beneficial in suggesting the most informative data points to be used when fitting a model. Reducing the number of observations needed for model development alleviates both the computational burden required for training and the operational expenses related to labeling. Online active learning, in particular, is useful in high‐volume production processes where the decision about the acquisition of the label for a data point needs to be taken within an extremely short time frame. However, despite the recent efforts to develop online active learning strategies, the behavior of these methods in the presence of outliers has not been thoroughly examined. In this work, we investigate the performance of online active linear regression in contaminated data streams. Our study shows that the currently available query strategies are prone to sample outliers, whose inclusion in the training set eventually degrades the predictive performance of the models. To address this issue, we propose a solution that bounds the search area of a conditional D‐optimal algorithm and uses a robust estimator. Our approach strikes a balance between exploring unseen regions of the input space and protecting against outliers. Through numerical simulations, we show that the proposed method is effective in improving the performance of online active learning in the presence of outliers, thus expanding the potential applications of this powerful tool. [ABSTRACT FROM AUTHOR]

Online active learning is a paradigm in machine learning that aims to select the most informative data points to label from a data stream. The problem of minimizing the cost associated with collecting labeled observations has gained a lot of attention in recent years, particularly in real-world applications where data is only available in an unlabeled form. Annotating each observation can be time-consuming and costly, making it difficult to obtain large amounts of labeled data. To overcome this issue, many active learning strategies have been proposed in the last decades, aiming to select the most informative observations for labeling in order to improve the performance of machine learning models. These approaches can be broadly divided into two categories: static pool-based and stream-based active learning. Pool-based active learning involves selecting a subset of observations from a closed pool of unlabeled data, and it has been the focus of many surveys and literature reviews. However, the growing availability of data streams has led to an increase in the number of approaches that focus on online active learning, which involves continuously selecting and labeling observations as they arrive in a stream. This work aims to provide an overview of the most recently proposed approaches for selecting the most informative observations from data streams in real time. We review the various techniques that have been proposed and discuss their strengths and limitations, as well as the challenges and opportunities that exist in this area of research. [ABSTRACT FROM AUTHOR]

Principal component analysis (PCA) has been a commonly used unsupervised learning method with broad applications in both descriptive and inferential analytics. It is widely used for representation learning to extract key features from a dataset and visualize them in a lower dimensional space. With more applications of neural network-based methods, autoencoders (AEs) have gained popularity for dimensionality reduction tasks. In this paper, we explore the intriguing relationship between PCA and AEs and demonstrate, through some examples, how these two approaches yield similar results in the case of the so-called linear AEs (LAEs). This study provides insights into the evolving landscape of unsupervised learning and highlights the relevance of both PCA and AEs in modern data analysis. [ABSTRACT FROM AUTHOR]

Sensor implementation together with data processing in manufacturing provides continuous measurements enabling process monitoring, optimization, and automation. Additive manufacturing (AM) technologies are shifting from one-off and prototyping production to batch, mass, and continuous production. Current AM technologies are mostly focused on small production devices, without monitoring systems. This study aims to bring AM closer to automation by proposing an online monitor system for bottom-up photopolymerization AM (VPP) systems. The sensor-generated data is used to capture the detachment error of built parts from the build platform that otherwise cannot be observed physically by the machine operator. The detachment itself will not stop the build job, which results in lost material and operating time. The online monitoring procedure consists of two phases, an offline and an online one, respectively. The offline phase is used for training a prediction model to be used in connection with a control chart for online monitoring. The monitoring control chart is advantageous as only the detachment predictions need to be recorded. The research carried out in this article brings novelty in both the vat photopolymerization set-up, as well as in an online monitoring procedure that can be easily extended to similar technologies. [ABSTRACT FROM AUTHOR]

Injection molding is one of the most important processes for the mass production of plastic parts. In recent years, many researchers have focused on predicting the occurrence and intensity of defects in injected molded parts, as well as the optimization of process parameters to avoid such defects. One of the most frequent defects of manufactured parts is blush, which usually occurs around the gate location. In this study, to identify the effective parameters on blush formation, eight design parameters with effect probability on the influence of this defect have been investigated. Using a combination of design of experiments (DOE), finite element analysis (FEA), and ANOVA, the most significant parameters have been identified (runner diameter, holding pressure, flow rate, and melt temperature). Furthermore, to provide an efficient predictive model, machine learning methods such as basic artificial neural networks, their combination with genetic algorithms, and particle swarm optimization have been applied and their performance analyzed. It was found that the basic artificial neural network (ANN), with an average accuracy error of 1.3%, provides the closest predictions to the FEA results. Additionally, the process parameters were optimized using ANOVA and a genetic algorithm, which resulted in a significant reduction in the blush defect area. [ABSTRACT FROM AUTHOR]

Data‐driven process monitoring and control techniques and their application to industrial chemical processes are gaining popularity due to the current focus on Industry 4.0, digitalization and the Internet of Things. However, for the development of such techniques, there are significant barriers that must be overcome in obtaining sufficiently large and reliable datasets. As a result, the use of real plant and process data in developing and testing data‐driven process monitoring and control tools can be difficult without investing significant efforts in acquiring, treating, and interpreting the data. Therefore, researchers need a tool that effortlessly generates large amounts of realistic and reliable process data without the requirement for additional data treatment or interpretation. In this work, we propose a data generation platform based on the Tennessee Eastman Process simulation benchmark. A graphical user interface (GUI) developed in MATLAB Simulink is presented that enables users to generate massive amounts of data for testing applicability of big data concepts in the realm of process control for continuous time‐dependent processes. An R‐Shiny app that interacts with the data generation tool is also presented for illustration purposes. The app can visualize the results generated by the Tennessee Eastman Process and can carry out a standard fault detection and diagnosis studies based on PCA. The data generator GUI is available free of charge for research purposes at https://github.com/dtuprodana/TEP. [ABSTRACT FROM AUTHOR]

Introduction: The current process for generating evidence in pharmacovigilance has several limitations, which often lead to delays in the evaluation of drug-associated risks. Objectives: In this study, we proposed and tested a near real-time epidemiological surveillance system using sequential, cumulative analyses focusing on the detection and preliminary risk quantification of potential safety signals following initiation of selective serotonin reuptake inhibitors (SSRIs) and serotonin-norepinephrine reuptake inhibitors (SNRIs). Methods: We emulated an active surveillance system in an historical setting by conducting repeated annual cohort studies using nationwide Danish healthcare data (1996–2016). Outcomes were selected from the European Medicines Agency's Designated Medical Event list, summaries of product characteristics, and the literature. We followed patients for a maximum of 6 months from treatment initiation to the event of interest or censoring. We performed Cox regression analyses adjusted for standard sets of covariates. Potential safety signals were visualized using heat maps and cumulative hazard ratio (HR) plots over time. Results: In the total study population, 969,667 new users were included and followed for 461,506 person-years. We detected potential safety signals with incidence rates as low as 0.9 per 10,000 person-years. Having eight different exposure drugs and 51 medical events, we identified 31 unique combinations of potential safety signals with a positive association to the event of interest in the exposed group. We proposed that these signals were designated for further evaluation once they appeared in a prospective setting. In total, 21 (67.7%) of these were not present in the current summaries of product characteristics. Conclusion: The study demonstrated the feasibility of performing epidemiological surveillance using sequential, cumulative analyses. Larger populations are needed to evaluate rare events and infrequently used antidepressants. [ABSTRACT FROM AUTHOR]

The current work addresses an industrial problem related to injection moulding manufacturing with focus on mould wear-out prediction. Real data sets are provided by an industrial partner that uses a multitude of moulds with different shapes and sizes in its production. An analysis of the data is presented and begins with clustering the moulds based on their characteristics and pre-chosen running settings. Using the results of the clustering, the mould wear-out is modelled using Kaplan-Meier survival curves. Furthermore, a random survival forest model is fitted for comparison and model performance is assessed. The main novelty of the case study is the implementation of mould wear-out prediction in real-time with the outcomes presented in terms of conditional survival curves including a proposed early warning system. For visualization and further industrial implementation, an R Shiny dashboard is developed and presented. [ABSTRACT FROM AUTHOR]

The concurrent use of statistical process control and engineering process control involves monitoring manipulated and controlled variables. One multivariate control chart may handle the statistical monitoring of all variables, but observing the manipulated and controlled variables in separate control charts may improve understanding of how disturbances and the controller performance affect the process. In this article, we illustrate how step and ramp disturbances manifest themselves in a single‐input–single‐output system by studying their resulting signatures in the controlled and manipulated variables. The system is controlled by variations of the widely used proportional‐integral‐derivative(PID) control scheme. Implications for applying control charts for these scenarios are discussed. [ABSTRACT FROM AUTHOR]

Manufacturing has been rejuvenated by automation and digitalization. This has brought forth the new industrial era also called Industry 4.0. During the last few years, we have collaborated with companies from various industries that have all been going through this transformation. Through these collaborations, we have collected numerous examples of (sometimes troublesome) experiences with Big Data applications of production analytics. These experiences reflect the current state of production data and the challenges it poses. Our goal in this paper is to share those experiences and lessons learned in dealing with practical issues from data acquisition to data management and finally to data analytics. [ABSTRACT FROM AUTHOR]

Aim: To perform a systematic review on the application of artificial intelligence (AI) based knowledge discovery techniques in pharmacoepidemiology. Study Eligibility Criteria: Clinical trials, meta-analyses, narrative/systematic review, and observational studies using (or mentioning articles using) artificial intelligence techniques were eligible. Articles without a full text available in the English language were excluded. Data Sources: Articles recorded from 1950/01/01 to 2019/05/06 in Ovid MEDLINE were screened. Participants: Studies including humans (real or simulated) exposed to a drug. Results: In total, 72 original articles and 5 reviews were identified via Ovid MEDLINE. Twenty different knowledge discovery methods were identified, mainly from the area of machine learning (66/72; 91.7%). Classification/regression (44/72; 61.1%), classification/regression + model optimization (13/72; 18.0%), and classification/regression + features selection (12/72; 16.7%) were the three most frequent tasks in reviewed literature that machine learning methods has been applied to solve. The top three used techniques were artificial neural networks, random forest, and support vector machines models. Conclusions: The use of knowledge discovery techniques of artificial intelligence techniques has increased exponentially over the years covering numerous sub-topics of pharmacoepidemiology. Systematic Review Registration: Systematic review registration number in PROSPERO: CRD42019136552. [ABSTRACT FROM AUTHOR]

As a direct consequence of production systems' digitalization, high‐frequency and high‐dimensional data has become more easily available. In terms of data analysis, latent structures‐based methods are often employed when analyzing multivariate and complex data. However, these methods are designed for supervised learning problems when sufficient labeled data are available. Particularly for fast production rates, quality characteristics data tend to be scarcer than available process data generated through multiple sensors and automated data collection schemes. One way to overcome the problem of scarce outputs is to employ semi‐supervised learning methods, which use both labeled and unlabeled data. It has been shown that it is advantageous to use a semi‐supervised approach in case of labeled data and unlabeled data coming from the same distribution. In real applications, there is a chance that unlabeled data contain outliers or even a drift in the process, which will affect the performance of the semi‐supervised methods. The research question addressed in this work is how to detect outliers in the unlabeled data set using the scarce labeled data set. An iterative strategy is proposed using a combined Hotelling's T2 and Q statistics and applied using a semi‐supervised principal component regression (SS‐PCR) approach on both simulated and real data sets. [ABSTRACT FROM AUTHOR]

Engineering process control and high-dimensional, time-dependent data present great methodological challenges when applying statistical process control (SPC) and design of experiments (DoE) in continuous industrial processes. Process simulators with an ability to mimic these challenges are instrumental in research and education. This article focuses on the revised Tennessee Eastman process simulator providing guidelines for its use as a testbed for SPC and DoE methods. We provide flowcharts that can support new users to get started in the Simulink/Matlab framework, and illustrate how to run stochastic simulations for SPC and DoE applications using the Tennessee Eastman process. [ABSTRACT FROM AUTHOR]

Abstract: The expanding availability of complex data structures requires development of new analysis methods for process understanding and monitoring. In manufacturing, this is primarily due to high‐frequency and high‐dimensional data available through automated data collection schemes and sensors. However, particularly for fast production rate situations, data on the quality characteristics of the process output tend to be scarcer than the available process data. There has been a considerable effort in incorporating latent structure–based methods in the context of complex data. The research question addressed in this paper is to make use of latent structure–based methods in the pursuit of better predictions using all available data including the process data for which there are no corresponding output measurements, ie, unlabeled data. Inspiration for the research question comes from an industrial setting where there is a need for prediction with extremely low tolerances. A semi‐supervised principal component regression method is compared against benchmark latent structure–based methods, principal components regression, and partial least squares, on simulated and experimental data. In the analysis, we show the circumstances in which it becomes more advantageous to use the semi‐supervised principal component regression over these competing methods. [ABSTRACT FROM AUTHOR]

Industrial manufacturing processes often operate under closed-loop control, where automation aims to keep important process variables at their set-points. In process industries such as pulp, paper, chemical and steel plants, it is often hard to find production processes operating in open loop. Instead, closed-loop control systems will actively attempt to minimize the impact of process disturbances. However, we argue that an implicit assumption in most experimental investigations is that the studied system is open loop, allowing the experimental factors to freely affect the important system responses. This scenario is typically not found in process industries. The purpose of this article is therefore to explore issues of experimental design and analysis in processes operating under closed-loop control and to illustrate how Design of Experiments can help in improving and optimizing such processes. The Tennessee Eastman challenge process simulator is used as a test-bed to highlight two experimental scenarios. The first scenario explores the impact of experimental factors that may be considered as disturbances in the closed-loop system. The second scenario exemplifies a screening design using the set-points of controllers as experimental factors. We provide examples of how to analyze the two scenarios. © 2017 The Authors Quality and Reliability Engineering International Published by John Wiley & Sons Ltd [ABSTRACT FROM AUTHOR]

The analysis of split plot experiments can be challenging due to a complicated error structure resulting from restrictions on complete randomization. Similarly, standard visualization methods do not provide the insight practitioners desire to understand the data, think of explanations, generate hypotheses, build models, or decide on next steps. This article demonstrates the effective use of trellis plots in the preliminary data analysis for split plot experiments to address this problem. Trellis displays help to visualize multivariate data by allowing for conditioning in a general way. They can also be used after the statistical analysis for verification, clarification, and communication. [ABSTRACT FROM PUBLISHER]

Most of today's complex systems and processes involve several stages through which input or the raw material has to go before the final product is obtained. Also in many cases factors at different stages interact. Therefore, a holistic approach for experimentation that considers all stages at the same time will be more efficient. However, there have been only a few attempts in the literature to provide an adequate and easy-to-use approach for this problem. In this paper, we present a novel methodology for constructing two-level split-plot and multistage experiments. The methodology is based on the Kronecker product representation of orthogonal designs and can be used for any number of stages, for various numbers of subplots and for different number of subplots for each stage. The procedure is demonstrated on both regular and nonregular designs and provides the maximum number of factors that can be accommodated in each stage. Furthermore, split-plot designs for multistage experiments with good projective properties are also provided. [ABSTRACT FROM AUTHOR]

A basic assumption when using principal component analysis (PCA) for inferential purposes, such as in statistical process control (SPC), is that the data are independent in time. In many industrial processes, frequent sampling and process dynamics make this assumption unrealistic rendering sampled data autocorrelated (serially dependent). PCA can be used to reduce data dimensionality and to simplify multivariate SPC. Although there have been some attempts in the literature to deal with autocorrelated data in PCA, we argue that the impact of autocorrelation on PCA and PCA-based SPC is neither well understood nor properly documented. This article illustrates through simulations the impact of autocorrelation on the descriptive ability of PCA and on the monitoring performance using PCA-based SPC when autocorrelation is ignored. In the simulations, cross-correlated and autocorrelated data are generated using a stationary first-order vector autoregressive model. The results show that the descriptive ability of PCA may be seriously affected by autocorrelation causing a need to incorporate additional principal components to maintain the model's explanatory ability. When all variables have equal coefficients in a diagonal autoregressive coefficient matrix, the descriptive ability is intact, while a significant impact occurs when the variables have different degrees of autocorrelation. We also illustrate that autocorrelation may impact PCA-based SPC and cause lower false alarm rates and delayed shift detection, especially for negative autocorrelation. However, for larger shifts, the impact of autocorrelation seems rather small. © 2015 The Authors. Quality and Reliability Engineering International published by John Wiley & Sons Ltd. [ABSTRACT FROM AUTHOR]

The origins of quality engineering are in manufacturing, where quality engineers apply basic statistical methodologies to improve the quality and productivity of products and processes. In the past decade, people have discovered that these methodologies are effective for improving almost any type of system or process, such as financial, health care, and supply chains. This paper begins with a review of key advances and trends within quality engineering over the past decade. The second part uses the first part as a foundation to outline new application areas for the field. It also discusses how quality engineering needs to evolve in order to make significant contributions to these new areas. © 2015 The Authors Quality and Reliability Engineering International Published by John Wiley & Sons Ltd. [ABSTRACT FROM AUTHOR]

One of the basic assumptions for traditional univariate and multivariate control charts is that the data are independent in time. For the latter, in many cases, the data are serially dependent (autocorrelated) and cross-correlated because of, for example, frequent sampling and process dynamics. It is well known that the autocorrelation affects the false alarm rate and the shift-detection ability of the traditional univariate control charts. However, how the false alarm rate and the shift-detection ability of the Hotelling T² control chart are affected by various autocorrelation and cross-correlation structures for different magnitudes of shifts in the process mean is not fully explored in the literature. In this article, the performance of the Hotelling T² control chart for different shift sizes and various autocorrelation and cross-correlation structures are compared based on the average run length using simulated data. Three different approaches in constructing the Hotelling T² chart are studied for two different estimates of the covariance matrix: (i) ignoring the autocorrelation and using the raw data with theoretical upper control limits; (ii) ignoring the autocorrelation and using the raw data with adjusted control limits calculated through Monte Carlo simulations; and (iii) constructing the control chart for the residuals from a multivariate time series model fitted to the raw data. To limit the complexity, we use a first-order vector autoregressive process and focus mainly on bivariate data. [ABSTRACT FROM AUTHOR]

The article offers the author's insights on the statistical analysis of categorical data. The author discusses the usage of kappa statistic in agreement studies based on the comprehensive review of Professor Jeoren de Mast and colleagues. The author mentions that de Mast also presents the confusing behavior of using kappa statistic in fictitious datasets.

In many industrial applications, autocorrelated data are becoming increasingly common due to, for example, on-line data collection systems with high-frequency sampling. Therefore, the basic assumption of independent observations for process capability analysis is not valid. The purpose of this article is to compare decision methods using the process capability index Cpk when data are autocorrelated. This is done through a case study followed by a simulation study. In the simulation study the actual significance level and power of the decision methods are investigated. The outcome of the article is that two methods appeared to be better than the others. [ABSTRACT FROM AUTHOR]

We propose a new methodology for designing computer experiments that was inspired by the split-plot designs that are often used in physical experimentation. The methodology has been developed for a simulation model of a surgical unit in a Danish hospital. We classify the factors as controllable and uncontrollable on the basis of their characteristics in the physical system. The experiments are designed so that, for a given setting of the controllable factors, the various settings of the uncontrollable factors cover the design space uniformly. Moreover the methodology allows for overall uniform coverage in the combined design when all settings of the uncontrollable factors are considered at once. [ABSTRACT FROM AUTHOR]

The simultaneous optimization of multiple responses in a dynamic system is challenging. When a response has a known gradient, it is often easily improved along the path of steepest ascent. On the contrary a stochastic approximation technique may be used when the gradient is unknown or costly to obtain. We consider the problem of optimizing multiple responses in which the gradient is known for only one response. We propose a hybrid approach for this problem, called simultaneous perturbation stochastic approximation steepest ascent, SPSA-SA or SP(SA)2 for short. SP(SA)2 is an SPSA technique that leverages information about the known gradient to constrain the perturbations used to approximate the others. We apply SP(SA)2 to the cross-layer optimization of throughput, packet loss, and end-to-end delay in a mobile ad hoc network (MANET), a self- organizing wireless network. The results show that SP(SA)2 achieves higher throughput and lower packet loss and end-to-end delay than the steepest ascent, SPSA, and the Nelder-Mead stochastic approximation approaches. It also reduces the cost in the number of iterations to perform the optimization. [ABSTRACT FROM AUTHOR]

Computer networks often display nonlinear behavior when examined over a wide range of operating conditions. There are few strategies available for modeling such behavior and optimizing such systems as they run. Profile-driven regression is developed and applied to modeling and runtime optimization of throughput in a mobile ad hoc network, a self-organizing collection of mobile wireless nodes without any fixed infrastructure. The intermediate models generated in profile-driven regression are used to fit an overall model of throughput, and are also used to optimize controllable factors at runtime. Unlike others, the throughput model accounts for node speed. The resulting optimization is very effective; locally optimizing the network factors at runtime results in throughput as much as six times higher than that achieved with the factors at their default levels. [ABSTRACT FROM AUTHOR]

Robust parameter design is one of the important tools used in Design for Six Sigma. In this article, we present an application of the generalized linear mixed model (GLMM) approach to robust design and analysis of signal–response systems. We propose a split-plot approach to the signal–response system characterized by two variance components—within-profile variance and between-profile variance. We demonstrate that explicit modeling of variance components using GLMMs leads to more precise point estimates of important model coefficients with shorter confidence intervals. Copyright © 2009 John Wiley & Sons, Ltd. [ABSTRACT FROM AUTHOR]

In this paper we present a modeling framework for analyzing computer models with two types of variations. The paper is based on a case study of an orthopedic surgical unit, which has both controllable and uncontrollable factors. Our results show that this structure of variation can be modeled effectively with linear mixed effects models and generalized additive models. Copyright © 2009 John Wiley & Sons, Ltd. [ABSTRACT FROM AUTHOR]