type="main"> This article evaluates the effect of payment reduction on mortgage default within the context of the Home Affordable Refinance Program. We find that mortgage default is sensitive to payment reduction using univariate, duration and hazard modeling approaches. A relative risk Cox model of default with time-varying covariates estimates that a 10% reduction in mortgage payment is associated with about a 10–11% reduction in monthly default hazard for loans. This finding is robust to the inclusion of empirically important mortgage risk drivers (such as current loan-to-value and FICO score) as well as controlling for selection effects based on observables.

s of Journal Track Articles A Bayesian Approach for Comparing Cross-Validated Algorithms on Multiple Data Sets Giorgio Corani and Alessio Benavoli Machine Learning DOI: 10.1007/s10994-015-5486-z We present a Bayesian approach for making statistical inference about the accuracy (or any other score) of two competing algorithms which have been assessed via cross-validation on multiple data sets. The approach is constituted by two pieces. The first is a novel correlated Bayesian t-test for the analysis of the cross-validation results on a single data set which accounts for the correlation due to the overlapping training sets. The second piece merges the posterior probabilities computed by the Bayesian correlated t-test on the different data sets to make inference on multiple data sets. It does so by adopting a Poisson-binomial model. The inferences on multiple data sets account for the different uncertainty of the cross-validation results on the different data sets. It is the first test able to achieve this goal. It is generally more powerful than the signed-rank test if ten runs of cross-validation are performed, as it is anyway generally recommended. A Decomposition of the Outlier Detection Problem into a Set of Supervised Learning Problems Heiko Paulheim and Robert Meusel Machine Learning DOI: 10.1007/s10994-015-5507-y Outlier detection methods automatically identify instances that deviate from the majority of the data. In this paper, we propose a novel approach for unsupervised outlier detection, which re-formulates the outlier detection problem in numerical data as a set of supervised regression learning problems. For each attribute, we learn a predictive model which predicts the values of that attribute from the values of all other attributes, and compute the deviations between the predictions and the actual values. From those deviations, we derive both a weight for each attribute, and a final outlier score using those weights. The weights help separating the relevant attributes from the irrelevant ones, and thus make the approach well suitable for discovering outliers otherwise masked in high-dimensional data. An empirical evaluation shows that our approach outperforms existing algorithms, and is particularly robust in datasets with many irrelevant attributes. Furthermore, we show that if a symbolic machine learning method is used to solve the individual learning problems, the approach is also capable of generating concise explanations for the detected outliers. Assessing the Impact of a Health Intervention via User-Generated Internet Content Vasileios Lampos, Elad Yom-Tov, Richard Pebody, and Ingemar J. Cox Data Mining and Knowledge Discovery DOI: 10.1007/s10618-015-0427-9 Assessing the effect of a health-oriented intervention by traditional epidemiological methods is commonly based only on population segments that use healthcare services. Here we introduce a complementary framework for evaluating the impact of a targeted intervention, such as a vaccination campaign against an infectious disease, through a statistical analysis of user-generated content submitted on web platforms. Using supervised learning, we derive a nonlinear regression model for estimating the prevalence of a health event in a population from Internet data. This model is applied to identify control location groups that correlate historically with the areas, where a specific intervention campaign has taken place. We then determine the impact of the intervention by inferring a projection of the disease rates that could have emerged in the absence of a campaign. Our case study focuses on the influenza vaccination program that was launched in England during the 2013/14 season, and our observations consist of millions of geo-located search queries to the Bing search engine and posts on Twitter. The impact estimates derived from the application of the proposed statistical framework support conventional assessments of the campaign. Beyond Rankings: Comparing Directed Acyclic Graphs Eric Malmi, Nikolaj Tatti, Aristides Gionis Data Mining and Knowledge Discovery DOI: 10.1007/s10618-015-0406-1 Defining appropriate distance measures among rankings is a classic area of study which has led to many useful applications. In this paper, we propose a more general abstraction of preference data, namely directed acyclic graphs (DAGs), and introduce a measure for comparing DAGs, given that a vertex correspondence between the DAGs is known. We study the properties of this measure and use it to aggregate and cluster a set of DAGs. We show that these problems are NP-hard and present efficient methods to obtain solutions with approximation guarantees. In addition to preference data, these methods turn out to have other interesting applications, such as the analysis of a collection of information cascades in a network. We test the methods on synthetic and real-world datasets, showing that the methods can be used to, e.g., find a set of influential individuals related to a set of topics in a network or to discover meaningful and occasionally surprising clustering structure. XXII Abstracts of Journal Track Articles

markdownabstract__Abstract__ Chapter 1 is the general introduction of the thesis. Chapter 2 presents a proposal on the nomenclature of the aortic root components. Chapters 3 and 4 present the single center results with the subcoronary Ross procedure in the University of Lubeck. Chapter 5 describes the durability of the autograft and homograft as observed in the large population of the German-Dutch Ross registry. Modes of autograft and homograft failure are presented as well as detailed outcomes of the observed reoperations. Chapter 6 challenges a popular belief that the Ross procedure in the presence of a bicuspid aortic valve is associated with reduced durability. Chapter 7 discusses the problem of competing risks when evaluating conduit durability. Chapter 8 presents the results of an effort to improve autograft durability after the Ross procedure. Chapter 9 presents clinical outcomes after the Ross procedure and provides a basis for the further judgment of this procedure and assist physician–patient discussion about the risks, benefits, and expectations after the Ross procedure.Chapter 10 presents a comparison of the survival of Ross patients with the survival of patient with mechanical valves after optimal and intensive anticoagulation monitoring. Chapters 11 and 12 focus on specific aspects of homograft durability. Chapter 13 discusses the results and the knowledge obtained from the above mentioned studies and research questions are summarized and put in perspective. Pitfalls in the methodologies used to evaluate outcomes after the Ross procedure are presented with examples from the European Ross Registry and discussed, which are generalizable in the evaluation of outcomes after complex heart interventions.

including a preoccupation with the frequentist coverage properties of credible intervals (to which I confess with a blush), something that D’Agostini characterizes as “a kind of condensate of frequentistic nonsense” (p. 229). D’Agostini makes no mention of Markov chain Monte Carlo (MCMC) methods, which is rather surprising because the technique was invented by physicists. This omission forces him to stick to analyses with closed-form or easily approximated solutions. For example, Chapter 11, which deals with the combination of experimental results, assumes that independent determinations, say of a physical constant, are normally distributed with the same mean and different nominal precisions, N(μ,pi). To allow for the possibility of large overlooked systematic errors, the true precision is modeled as τ · pi , where τ has a gamma(δ,λ) distribution, with δ and λ given fixed values. However, using MCMC, it is possible to assign fairly diffuse hyperprior distributions to δ and λ and allow the data to identify the distribution of τ , and it is possible to model overlooked systematic error as additive (1/pi +1/τ ), which seems more natural. What is encouraging and instructive for Bayesian statisticians is the seriousness with which D’Agostini takes the task of specifying informative prior distributions. It is tempting to shortchange this issue when teaching entry-level Bayesian statistics, on the assumption that real scientists prefer to shy away from informative priors. As Chapter 13, “Which Priors for Frontier Physics,” suggests, the reverse is the case. It appears that physicists are quite able to specify informative priors, and that not doing so can in many cases produce nonsensical results. According to the author, the audience for this book is practicing physicists and engineers who need to evaluate uncertainty. To that list, I would add upper-level and graduate students in physics and engineering, statisticians and statistics graduate students who collaborate with physicists and engineers, mathematicians who teach statistics courses, and all Bayesian statisticians. The two things I’d want added to the book are at least a heuristic axiomatic derivation of subjective probability and a chapter showing how to use MCMC to permit more flexible modeling.

ii Acknowledgements iii

by Simin Hu An important application of microarray research is to develop cancer diagnostic and prognostic tools based on tumor genetic profiles. For easy interpretation, such studies aim to identify a small fraction of genes to build molecular predictors of clinical outcomes from at least thousands of genes thus require methodologies that can model high dimensional covariates and accomplish variable selection simultaneously. One interesting area is modeling cancer patients’ survival time or time to cancer reoccurrence with gene expression data. In the first part of this dissertation, we propose a new penalized weighted least squares method for model estimation and variable selection in accelerated failure time models. In this method, right censored observations are used as censoring constraints in optimizing the weighted least squares objective function. We also include ridge penalty to deal with singularity caused by collinearity and high dimensionality and use the least absolute shrinkage and selection operator to achieve

[eng] The paper presents an empirical study on the relationship between academic researchers' patenting and publishing activities, based upon a sample of patent applications and scientific papers authored by Italian university professors. By treating patents as discrete events punctuating professors' routine publishing activity over time, we conclude that no major trade-off exists: academic inventors do not publish less than their colleagues with no patents, and do not show any bias towards more applied, less basic science. On the contrary, more productive professors are more likely to end up signing one or more patents; in addition, a temporary increase in a professor's scientific productivity increases the probability of a subsequent patenting event, which suggests that patents are the by-product of fertile research projects. Former scientific collaboration with industry, in the form of co-authored papers, affects positively the probability to patent with Business companies, especially in Biology and Pharmaceutics. This evidence is in line with recent results obtained for the US. [fre] Cet article presente une etude empirique sur la relation entre les activites de depots de brevets et celles de publication des chercheurs academiques. Cette etude se base sur un echantillon de depot de brevets ecrits par des professeurs d'universite italiens. En traitant les brevets comme des evenements discrets caracterisant l'evolution du travail de publication continu des professeurs, nous concluons qu'il n'existe pas d'arbitrage: les inventeurs academiques ne publient pas moins que leurs collegues qui ne brevetent pas. Il n'existe pas, de surcroit, de biais selon la nature plus ou moins appliquee de la recherche. En revanche, les professeurs les plus productifs ont une probabilite plus importante de deposer au moins un brevet. De meme, une augmentation temporaire de la productivite scientifique d'un professeur augmente la probabilite qu'il depose un brevet. Ces resultats laissent supposer que les brevets peuvent etre consideres comme un sous-produit de projets de recherche fertiles. Enfin, les collaborations scientifiques passees avec l'industrie, observees par les publications cosignees, affectent positivement la probabilite de breveter avec une entreprise, surtout dans les biotechnologies et l'industrie pharmaceutique. Ces resultats sont conformes a ceux obtenus recemment aux Etats-Unis.

[1] We examine the potential triggering relationship between large earthquakes and methane mud volcano eruptions. Our data set consists of a 191-year catalog (1810–2001) of eruptions from 77 volcanoes in Azerbaijan, central Asia, supplemented with reports from mud volcano eruptions in Japan, Romania, Pakistan, and the Andaman Islands. We compare the occurrence of historical regional earthquakes (M > 5) with the occurrence of Azerbaijan mud volcano eruptions and find that the number of same-day earthquake/eruption pairs is significantly higher than expected if the eruptions and earthquakes are independent Poisson processes. The temporal correlation between earthquakes and eruptions is most pronounced for nearby earthquakes (within ∼100 km) that produce seismic intensities of Mercalli 6 or greater at the location of the mud volcano. This assumed magnitude/distance relationship for triggering observed in the Azerbaijan data is consistent with documented earthquake-induced mud volcano eruptions elsewhere. We also find a weak correlation that heightened numbers of mud volcano eruptions occur within 1 year after large earthquakes. The distribution of yearly eruptions roughly approximates a Poisson process, although the repose times somewhat favor a nonhomogenous failure rate, which implies that the volcanoes require some time after eruption to recharge. The volcanic triggering likely results from some aspect of the seismic wave's passage, but the precise mechanism remains unclear.

Zhang, Zhibo. Ph.D., Purdue University, December 2016. Developing Condition-Based Triggers for Bridge Deck Maintenance and Rehabilitation Treatments. Major Professors: Samuel A. Labi and Kumares C. Sinha. The bridges in the U.S. highway system suffer from deficiencies in both their structural condition and functionality. In an effort to improve the condition of bridges, highway agencies continually seek effective and efficient approaches to maintenance and rehabilitation (M&R) treatments for their bridges. However, one drawback to new approaches is that highway agencies have long relied on the subjective judgment of their engineers to determine the time or condition at which to implement the treatments as well as the types of treatments to be applied. The literature shows that previous researchers mainly focused on time-based M&R strategies, but there have been some efforts toward developing condition-based strategies, such as the Indiana Bridge Management System (IBMS). While IBMS and similar systems were laudable efforts, they also were developed on the basis of the judgment and experience of bridge management personnel and were not data-driven. This dissertation proposes condition-based performance thresholds for bridge deck M&R treatments using data-driven analytical methods. The framework was developed for both deterministic and stochastic situations. Under the former, deterministic statistical models for bridge deterioration and costs were developed. The optimization framework was based on life-cycle agency and user costs, and its performance was demonstrated in this dissertation using data from state-owned bridges in the state of Indiana. Separate analyses were conducted with respect to different climate regions and highway functional classes. Sensitivity analysis was performed to investigate

XIAOYAN WANG: Semiparametric Regression Models for Recurrent and Terminal Event Data. (Under the direction of Dr. Jianwen Cai and Dr. Haibo Zhou.) Recurrent events are common in many clinical or observational studies. It is often of interest to evaluate the effects of risk factors on the frequencies of recurrent events. The recurrence of serious events are usually subject to censoring due to the death of a subject which is likely to be informative. In this dissertation, we first consider an accelerated failure time marginal rate model for the cumulative number of the recurrent events over time, while taking the terminal events into account. The marginal approach does not require specifying the dependence structure between the recurrent events and the terminal events, and the mean function incorporates the facts that subjects who die cannot experience any further recurrent events. We develop an estimating procedure for both the regression parameters and the mean function by applying the inverse probability of censoring weighting technique(Robins and Rotnitzky(1992)). The proposed estimators are consistent and asymptotically normal. We investigate the finite-sample properties of the proposed estimators through simulation studies and provide an application to recurrent hospitalization data taken from the Studies of Left Ventricular Dysfunction (SOLVD) Treatment Trial data. Second, we propose a proportional rate model for the recurrent event given the subjects are still alive. Again the dependence between the recurrent event process and the terminal event is unspecified. We consider two estimating procedures for the regression coefficients and the mean function of recurrent events. Asymptotic properties of the proposed estimators are derived. Simulation studies are conducted to assess the finite

Writing a computer program for modeling multi-state disease process for cancer or chronic disease is often an arduous and time-consuming task. We have developed a SAS macro program for estimating the transition parameters in such models using SAS IML. The program is very flexible and enables the user to specify homogeneous and non-homogeneous (i.e. Weibull distribution, log-logistic, etc.) Markov models, incorporate covariates using the proportional hazards form, derive transition probabilities, formulate the likelihood function, and calculate the maximum likelihood estimate (MLE) and 95% confidence interval within a SAS subroutine. The program was successfully applied to an example of a three-state disease model for the progression of colorectal cancer from normal (disease free), to adenoma (pre-invasive disease), and finally to invasive carcinoma, with or without adjusting for covariates. This macro program can be generalized to other k-state models with s covariates.

Word of mouth (WOM) plays an increasingly important role in shaping consumers’ behavior and preferences. In this paper, we examine whether latent personality traits of online users accentuate or attenuate the effectiveness of WOM in social media platforms. To answer this question, we leverage machine-learning methods in combination with econometric techniques utilizing a novel quasi-experiment. Our analysis yields two main results. First, there is a positive and statistically significant effect of the level of personality similarity between two social media users on the likelihood of a subsequent purchase from a recipient of a WOM message after exposure to the WOM message of the sender. In particular, exposure to WOM messages from similar users in terms of personality, rather than dissimilar users, increases the likelihood of a postpurchase by 47.58%. Second, there are statistically significant effects of specific pairwise combinations of personality characteristics of senders and recipients of WOM messag...

Women with ductal carcinoma in situ (DCIS) are at substantially increased risk for a second breast cancer, but few strong predictors for these subsequent tumors have been identified. We used Cox regression modeling to examine the association between mammographic density at diagnosis of DCIS of 504 women from the National Surgical Adjuvant Breast and Bowel Project B-17 trial and risk of subsequent breast cancer events. In this group of patients, mostly 50 years old or older, approximately 6.6% had breasts categorized as highly dense (i.e., > or =75% of the breast occupied by dense tissue). After adjusting for treatment with radiotherapy, age, and body mass index, women with highly dense breasts had 2.8 (95% confidence interval [CI] = 1.3 to 6.1) times the risk of subsequent breast cancer (DCIS or invasive), 3.2 (95% CI = 1.2 to 8.5) times the risk of invasive breast cancer, and 3.0 (95% CI = 1.2 to 7.5) times the risk of any ipsilateral breast cancer, compared with women with less than 25% of the breast occupied by dense tissue. Our results provide initial evidence that the risk of second breast cancers may be increased among DCIS patients with highly dense breasts.

With the rapid development of data collection and aggregation technologies in many scientific disciplines, it is becoming increasingly ubiquitous to conduct large-scale or online regression to analyze real-world data and unveil real-world evidence. In such applications, it is often numerically challenging or sometimes infeasible to store the entire dataset in memory. Consequently, classical batch-based estimation methods that involve the entire dataset are less attractive or no longer applicable. Instead, recursive estimation methods such as stochastic gradient descent that process data points sequentially are more appealing, exhibiting both numerical convenience and memory efficiency. In this paper, for scalable estimation of large or online survival data, we propose a stochastic gradient descent method which recursively updates the estimates in an online manner as data points arrive sequentially in streams. Theoretical results such as asymptotic normality and estimation efficiency are established to justify its validity. Furthermore, to quantify the uncertainty associated with the proposed stochastic gradient descent estimator and facilitate statistical inference, we develop a scalable resampling strategy that specifically caters to the large-scale or online setting. Simulation studies and a real data application are also provided to assess its performance and illustrate its practical utility.

With the rapid developing of the Internet, more and more business websites are appearing all around the world. As the proportion of C2C business mode increasing, the problem about personal credit evaluation in the business websites also becomes more critical. The deal methods are very different between the C2C mode and the traditional enterprises in evaluating the credit rank due to the feature of customers in the business websites. In this paper, we construct a support vector machines (SVMs) ensemble method based on fuzzy integral to evaluate personal credit under the environment of electronic commerce. This method aggregates the outputs of separate component SVMs with importance of each component SVM. By comparing the experimental result SVMs ensemble with the single SVM, the neural network ensemble, the proposed method outperforms the single SVM, and neural network ensemble in terms of classification accuracy. It proves that the proposed method is stable, highly accurate, strong robust and feasible. It is useful for providing a sound credit assessment system.

With the increased availability of survival datasets, that comprise both molecular information (e.g., gene expression), and clinical information (e.g., patient survival), numerous genes are proposed as prognostic biomarkers. Despite efforts and money invested, very few of these biomarkers have been clinically validated and are used routinely. A high false discovery rate is assumed to be largely responsible for this, in particular as the number of tested genes is extremely high relative to the number of patients followed. Here, after describing the historical methodologies on which recent developments have often been based, this review describes studies that have been performed in the last few years. The concepts will be illustrated for a renal cancer dataset, and the corresponding scripts are provided (Supporting Information). These new developments belong to three main fields of applications. First, variable selection concerns various improvements to lasso penalization. Second, accurate definition of p‐values and control of the false discovery rate have also been the subject of many studies. Third, the incorporation of biological knowledge, often through the form of networks or pathways, can be used as an a priori and/or to reduce dimensionality. These new and promising developments deserve benchmarking by independent groups not involved in their development, with various independent datasets. Further work on the methodologies is also still required.

With increasing use of fiber reinforced polymer (FRP) composites in several industries such as aviation, automotive and construction, effective reliability analysis of composites has become imperative in recent years. Periodic inspection by robust non-destructive evaluation (NDE) techniques and accurate health prognosis is essential for condition-based maintenance (CBM) of the safety-critical components and structures. Prediction of future damage level in composites often becomes challenging due to lack of physics-based damage growth models for unknown materials which leaves us to rely solely on the NDE data for prognosis. In this study, delamination growth in glass fiber reinforced polymer (GFRP) joints, under Mode I cyclic loading, was monitored by guided waves(GW) using miniature surface-mounted piezoelectric wafers(PZT). Data-driven prognosis techniques such as Kalman filter and particle filter were implemented on the indirect CBM data obtained from GW signals to predict future delamination area and validated against optical transmission scans (OTS) of damaged samples. A comparison of data-driven prognosis methods with static regression versus dynamic update of estimated parameters is presented in this paper. Results show that even when a simple logarithmic fit is assumed, use of NDE data to estimate function parameters in a stochastic framework outperforms the static regression approach leading to a robust sensor-aided reliability analysis.

While there are many validated prognostic classifiers used in practice, often their accuracy is modest and heterogeneity in clinical outcomes exists in one or more risk subgroups. Newly available markers, such as genomic mutations, may be used to improve the accuracy of an existing classifier by reclassifying patients from a heterogenous group into a higher or lower risk category. The statistical tools typically applied to develop the initial classifiers are not easily adapted toward this reclassification goal. In this article, we develop a new method designed to refine an existing prognostic classifier by incorporating new markers. The two-stage algorithm called Boomerang first searches for modifications of the existing classifier that increase the overall predictive accuracy and then merges to a prespecified number of risk groups. Resampling techniques are proposed to assess the improvement in predictive accuracy when an independent validation data set is not available. The performance of the algorithm is assessed under various simulation scenarios where the marker frequency, degree of censoring, and total sample size are varied. The results suggest that the method selects few false positive markers and is able to improve the predictive accuracy of the classifier in many settings. Lastly, the method is illustrated on an acute myeloid leukemia data set where a new refined classifier incorporates four new mutations into the existing three category classifier and is validated on an independent data set.

When fitting a Cox proportional hazards model with missing covariates, it is inefficient to exclude observations with missing values in the analysis. Furthermore, if the missing-data mechanism is not Missing Completely At Random(MCAR), it may lead to biased parameter estimation. Many approaches have been suggested to handle the Cox proportional hazards model when covariates are sometimes missing, but they are based on the selection model. This paper suggest an approach to handle Cox proportional hazards model with missing covariates by using the pattern-mixture model (Little, 1993). The pattern-mixture model is expressed by the joint distribution of survival time and the missing-data mechanism. In the pattern-mixture model, many models can be considered by setting up various restrictions, and different results under various restrictions indicate the sensitivity of the model due to missing covariates. A simulation study was conducted to show the sensitivity of parameter estimation under different restrictions in a pattern-mixture model. The proposed approach was also applied to mouse leukemia data.

When data are collected longitudinally, measurement times often vary among patients. This is of particular concern in clinic-based studies, for example retrospective chart reviews. Here, typically no two patients will share the same set of measurement times and moreover, it is likely that the timing of the measurements is associated with disease course; for example, patients may visit more often when unwell. While there are statistical methods that can help overcome the resulting bias, these make assumptions about the nature of the dependence between visit times and outcome processes, and the assumptions differ across methods. The purpose of this paper is to review the methods available with a particular focus on how the assumptions made line up with visit processes encountered in practice. Through this we show that no one method can handle all plausible visit scenarios and suggest that careful analysis of the visit process should inform the choice of analytic method for the outcomes. Moreover, there are some commonly encountered visit scenarios that are not handled well by any method, and we make recommendations with regard to study design that would minimize the chances of these problematic visit scenarios arising.