Abstracts

Abstract

Women in the statistics and mathematics fields have to overcome the stigma of being black-box number crunchers who are inept at communicating and applying results from data. Developing the ability to make concise recommendations by leveraging data and statistical models is just as important as developing the skill to build robust, statistically-sound analyses. In order to gain visibility in a data field, it is critical to be able to communicate the business implications from analysis to get the ears of leaders in the business.

Abstract

We focus on developing the theoretical foundations of M-estimation for dependent data. M-estimation is a popular technique to extract a parameter estimate by minimizing a particular loss function. Previous work on the subject focused on specific models of dependence and addressed limited parametric scenario and it lacks any overarching theory to address general problem under dependence. Our goal is to contribute new theory to this important area that will provide unifying tools to address a range of applications. As the first step toward the main result we have proved a general triangular array version of a limit theorem for empirical processes under dependence. This result is then further applied in general empirical process set-up to obtain asymptotics of maximizer of an empirical process. The result is then applied to a number of specific problems including classical cube root asymptotics scenario to obtain new results.

Abstract

RNA-Seq experiments produce millions of discrete sequence reads as a measure of gene expression levels, and enable researchers to investigate complex aspects of the genomic studies. These include but not limited to identification of differentially expressed (DE) genes in two or more treatment conditions and detection of novel transcripts. One of the common assumptions of RNA-Seq data is that, all gene counts follow an overdispersed Poisson or negative binomial (NB) distributions, which may not be appropriate as some genes may have stable transcription levels with no overdispersion. Thus, a more realistic assumption in RNA-Seq data is to consider two sets of genes: overdispersed and non-overdispersed.We propose a two-step integrated approach to detect differentially expressed (DE) genes in RNA-Seq data using standard Poisson model for non-overdispersed genes and NB model for overdispersed genes. We evaluate the proposed approach using two simulated and two real RNA-Seq data sets. We compare the performance of our proposed method with the four popular R-software packages edgeR, DESeq, sSeq, and NBPSeq with their default settings. For both the simulated and real data sets, integrated approaches perform better or at least equally well compared to the regular methods embedded in these R-packages.

Abstract

Modelling of temporal evolution of network data has become a relevant problem for different applications. However, the complexity of the models increase rapidly with the number of nodes making efficient short term prediction of future outcomes of the system a challenge for big network data. Here, we propose an autologistic model for directed binary networks with a fused lasso penalty. This model favors sparse solutions of the coefficients and their differences in consecutive time points, and it is suitable for complex dynamic data where the number of parameters is considerably greater than the number of observations over time. The structure of our model allow us to treat the optimization problem separately for each pair of nodes increasing efficiency of the algorithm through parallel computing. The optimal fused lasso tuning parameters are chosen using BIC. We show the performance of the model on a real trading network from the NYMEX natural gas futures market observed weekly over a period of four years .

Abstract

Abstract

Given genetic data from two meaningful classes of samples, it is often of interest to identify genes (or other predictors) that behave differently in one class than the other. For example, one might consider samples from different tissue types, from a control group versus a treatment group, or from patients of unequal disease severity. A common approach is to study first-order differences; that is, to discover genes whose expression levels, as measured by microarray profiling or RNA sequencing, are higher in samples from the first class than in those from the second. By contrast, we propose a new second-order analysis, Mining of Differential Correlation (MDC).

The MDC procedure seeks to identify a group of genes whose average pairwise correlation amongst samples in the first class is higher than amongst samples in the second class. Our method is an iterative procedure that adaptively updates the size and elements of an initial gene set. These updates are based upon multiple testing of carefully defined p-values, so the final output gene set is statistically meaningful. We investigate the performance of MDC by applying it to simulated data as well as gene expression data taken from The Cancer Genome Atlas and other recent experimental datasets.

Abstract

Nuclear Magnetic Resonance (NMR) spectroscopy is a valuable tool for analyzing the composition of various small non-protein molecules by comparing the experimental spectrum against a library of expected peak locations. Currently, small molecule identification can be time-consuming and labor intensive as spectral results can vary over sample preparation and run conditions, and typically hundreds of molecules are identified simultaneously within a single spectrum with varying overlapping peaks with unknown shifts and concentrations.  We develop a Bayesian statistical learning algorithm to automate the identification of known small molecule entities in a sample.  With rapid real-time data streams available, an algorithm for identifying small molecule entities and handling large-scale data is desirable. We develop a method that is an extension of incremental Bayesian statistical/machine learning methods to identify known and unknown small molecule entities in streaming spectroscopy data, which also has the capability to identify unknown entities, evolve as new entities are added to the library, and detect rare or anomalous events. 

Abstract

For better or for worse, it is not only what you know, but who you know.  A wide and strong network of colleagues is one of the most valuable professional assets regardless of where you choose to practice statistics.  Building a group of to-go-to people may seem daunting if you are just getting started or if you are not the outgoing type, but by taking just a few important steps you can enjoy the many benefits of making professional contacts and even friendships that will stay with you forever.  We tend to think of our network as a means to enhance our own careers, get doors open, receive other special treatment.  But networks are also the means by which you can reach out to others, share opportunities, and mentor younger colleagues.   During the presentation, we discuss  how to build and maintain a solid network, what to expect from your professional contacts and how to contribute to strengthening your network and as a consequence, your own career. 

Abstract

In risk analysis of toxic agents, benchmark dose estimation has proven to be a valuable quantity and is commonly used by federal agencies in health risk assessment. However, one drawback of the benchmark dose estimator is that it is dependent on the assumed underlying risk model. Even when two risk models fit the data equally well, they can produce very different benchmark dose estimators for a given risk level. Herein, we propose a Bayesian model averaging methodology that overcomes this deficiency and does not require any model selection. We illustrate the usefulness of this approach in a simulation study and apply it to a data set from the National Toxicology Program. 

Abstract

In recent years, the availability of infectious disease counts in time and space has increased, and consequently there has been renewed interest in model formulation for such data. In this paper, we describe a model that was motivated by the need to analyze hand, foot and mouth disease (HFMD) collected in China. For these data, the aims of the analysis were to gain insight into the space-time dynamics and to carry out short-term prediction in order to implement public health campaigns in those areas with a large predicted disease burden. The model we develop decomposes disease risk into marginal spatial and temporal components, and a space-time interaction piece. The latter is the crucial element, and we use a tensor product spline model with a Markov random field prior on the coefficients of the basis function. The model is highly parameterized and so computation is carried out using the integrated nested Laplace approximation (INLA) approach, which is fast. A simulation study shows that the model can pick up complex space-time structure and our analysis of HFMD in the central north region of China provides new insights into the dynamics of the disease.

Abstract

National survey data indicated that the number of individuals with disabilities is rising. Those with disabilities experience a large number of barriers to effective patient accessibility/provider contact, patient-provider communication, and provider coordination of care ultimately resulting in patient dissatisfaction.  However, little research has been done on the patient-physician relationship and patient satisfaction issues as perceived by persons with disabilities. Only a limited number of studies have used nationally representative data to examine the health status of individuals with disabilities in comparison to those without disabilities. The Medical Expenditure Panel Survey (MEPS) was used to examine whether disability status influenced patient accessibility/provider contact, patient-provider communication, provider coordination of care and patient satisfaction. The purposes of this study were to determine if there were  relationships between disability status and patient accessibility/provider contact, patient-provider communication, provider coordination of care and patient satisfaction, as well as how disability status affected the likelihood of ineffective patient-provider relationship and patient satisfaction. The MEPS was also used to assess whether disability status was associated with common chronic diseases and lower use of preventive care services. Additional analyses involved correlating disability status with race and ethnicity stratification in terms of patient-physician relationship and patient satisfaction. A retrospective analysis was conducted comparing the health of adults with disabilities to adults with no disabilities using data from the 2005-2009 full year consolidated data files from MEPS. Chi square analyses were performed to determine if there were significant differences in patient-physician relationship and patient satisfaction for persons with disabilities compared with persons without disabilities. A series of logistic regression analyses were conducted examining the likelihood of ineffective patient-physician relationship and patient dissatisfaction with the independent variable of disability status. Those with disabilities were less likely to be able to contact their physician than individuals without disabilities. Adults with disabilities were significantly more likely than persons without disabilities to perceive that their physician did not outline options for treatment or explain treatments in a way they understood, did not involve them in treatment decisions, did not listen to them, did not spend sufficient time with them and did not treat them with respect. Individuals with disabilities were also less likely than individuals without disabilities to receive preventive care services and more likely to be dissatisfied with their health care.  Individuals with disabilities were also more likely to have chronic illnesses than persons without disabilities. This study revealed that adults with disabilities are at an increased risk of experiencing ineffective patient-provider relationship and poor patient satisfaction, thus, compromising their current health status and increasing the possibility of secondary chronic illnesses. This data suggests that adults with disabilities and chronic conditions receive significantly fewer preventive care services and have poorer health status than individuals without disabilities who have the same health conditions. This implies a need for public health interventions and methods for improving the effectiveness of patient-physician relationship and patient satisfaction among the disabled population.

Abstract

The accurate prediction of surgical risk is important to patients and physicians. Logistic regression (LR) models are typically used to estimate these risks. However, in the fields of data mining and machine-learning, alternative classification and prediction algorithms have been developed. This study aimed to compare the performance of LR to machine-learning algorithms in predicting 30-day surgical morbidity in children. All models included procedures and 49 pre-operative patient characteristics and were fit using data from 48,089 cases in the National Surgical Quality Improvement Program-Pediatric. After optimizing model fit using cross-validation in a training dataset, we compared discrimination (c-statistic), calibration intercept and slope, classification accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of the models in the remaining cases. Compared to a LR model with no interactions and linear relationships between predictors and the log-odds of morbidity, ensemble-based methods showed higher accuracy (random forests (RF) 93.6%, boosted classification trees (BCT) 93.6%, LR 93.2%), sensitivity (RF 39.0%, BCT 39.8%, LR 37.5%), specificity (RF 98.7%, BCT 98.5%, LR 98.4%), PPV (RF 73.2%, BCT 71.5%, LR 68.0%), and NPV (RF 94.6%, BCT 94.7%, LR 94.5%). However, only BCT showed superior discrimination (BCT c=0.880, LR c=0.871) (p<.05 for all), and none of the models performed better than a more flexible LR model that incorporated restricted cubic splines and significant interactions (accuracy 93.7%, sensitivity 41.5%, specificity 98.5%, PPV 72.0%, NPV 94.8%, c=0.877). Both LR models showed superior calibration compared to the ensemble-based algorithms. After further validation, the flexible LR model derived in this study could be used to assist with clinical decision-making based on patient-specific surgical risks.

Abstract

In logistic regression models with categorical predictors, if the design matrix used to obtain the linear predictor is constructed so that the model parameters are measuring differences between a reference group and the levels of the factor (e.g. treatment contrasts in R), then the main effects parameters are easily interpreted.  By exponentiating the parameter estimates, we obtain odds ratios, that is the odds of the response variable being one in a level of the predictor divided by the odds of the response variable being one at the reference level.  Under this parameterization, the interaction predictors are easily obtained by multiplication.  Unfortunately, the interaction parameters are not as easily interpreted as the main effect parameters, not to mention inference associated with the interactions.

            Various authors have suggested measures that may be calculated to interpret an interaction in a logistic regression model with categorical predictors, e.g., interaction contrast ratios (ICR) or equivalently relative excess risk due to interaction (RERI), attributable proportion due to interaction (AP) and a synergy index (S) measuring the ratio between combined effect and individual effects.  These measures are said to be additive.  Others are multiplicative.  In as much as most consumers of the information being described by the model are not statisticians, mathematicians nor scientists, the interpretation of interactions using these measures may be difficult. 

            When response variables are quantitative, the predictors are categorical and the link function is the identity link (e.g. ANOVA models), we interpret interactions by forming the difference between differences, construct contrasts and develop confidence intervals for the interaction contrasts.  The method for doing this follows directly from what we know about linear combinations of random variables.  The contrasts are easily interpreted and graphed.  Apparently, in logistic regression, researchers have found it difficult to develop confidence intervals for interaction contrasts, often having to resort to the assumption that odds ratios are approximately risk ratios.  We show that this problem is readily solved in the Bayesian approach using Markov Chain Monte Carlo methods.

Abstract

The health care system has changed dramatically over the years and is more data centric than ever. These days petabytes or Exabyte of data including clinical outcome data, socio-economic data, personal genomic data, and medical imaging data are collected on patients. While organizing, storing and distributing the data in a secured fashion is a daunting task, analyzing these data and integrating the results to assist physicians is even harder. These changes pose an opportunity and a challenge for statisticians. Statisticians are central and essential to uncovering the scientific knowledge hidden in the data tsunami. In this session, we will focus on how statisticians are handling the challenge and transforming patient healthcare.

Abstract

Describing the nature and variability of Indian monsoon rainfall extremes is a topic of much debate in the current literature. We suggest the use of a General Linear Mixed Model (GLMM) specifically, the logit-normal mixed model, to describe the underlying structure of this complex climatic event. Several GLMM algorithms are described and simulations are performed to vet these algorithms before applying them to the Indian rainfall data procured from the National Climatic Data Center. The logit-normal model was applied with fixed covariates of latitude, longitude, elevation, daily minimum and maximum temperatures with a random intercept by weather station. In general, these estimation methods concurred in their suggestion of a strong relationship between the El Nino Southern Oscillation (ENSO) index and extreme rainfall variability estimates. This work provides a valuable starting point for extending GLMM to incorporate the complex dependencies in extreme climate events.

Abstract

Graduate education, especially in Statistics and Biostatistics, in 2014 is much more involved and demanding than it was twenty years ago. While the big data era presents exciting opportunities for statisticians, it also presents significant challenges in graduate training. Specifically, more interdisciplinary training in computer science, machine learning and subject-matter science. Both Masters and Ph.D. students seek out internships, publications, advanced computing skills, teaching, consulting or collaboration with other disciplines, as well as volunteer opportunities in efforts to make themselves more competitive in the job market. As CVs and resumes become packed with experiences and publications, we, as educators are left wondering how to prepare the next global generation of graduate students for successful futures. Each member of this panel will reflect on her own graduate education, discuss and contrast her experiences raising today's generation of graduate students, and discuss emerging challenges and graduate education in Statistics and Biostatistics for the next generation and what it will require. The hope for this session is a frank and open discussion about the past, present and future challenges of graduate education Statistics and Biostatistics.

Abstract

This panel discussion will focus on a series of topical questions addressed by Kiersten and Suzanne, followed by audience Q&A.   The topics will be divided in two broad themes:   1)  A Business Development Dimension (topics to include: How to balance the hard and soft sciences;  How to proactively drive your analytics leadership journey;  The challenges in communicating/messaging analytics to a non-technical audience) and 2) A Personal Development Dimension ( topics to include: How to seek Constructive Feedback; How to identify and find a personal champion;  How to calibrate and balance your high self-expectations across the organization ).  Kiersten and Suzanne will openly discuss their 'moments of truth', hoping that attendees can learn from their experiences and personal insights.

Abstract

Bilaterally Contaminated Normal Model with Nuisance Parameter and Its Applications
Qian Fan*, University of Kentucky
Hongying Dai, Children's Mercy Hospital
Richard Charnigo, University of Kentucky
 
Dai and Charnigo (2008, 2010) have proposed various contaminated density models which can be applied to microarray data and studied the MLRT and D-test statistics for these models under an omnibus null hypothesis of no differential expression. Charnigo et al (2013) subsequently proposed the bilateral contaminated normal model (BCN) without nuisance parameter, which is able to describe both gene under- and over-expression simultaneously. They proved the testing procedure is consistent and the test statistic has a limiting normal distribution under the unilateral null hypothesis of differential expression in one direction only. This presentation extends the previous work and puts forward tests of contamination in the BCN model when the common within-component variance is unknown. This model is more flexible compared to the BCN model without nuisance parameter. Yet the derivation and justification of the testing procedures are challenging. The first step is to do an omnibus test of no contamination, for which we propose a union-intersection test. The next stage is testing unilateral contamination against bilateral contamination. Both test procedures are shown to be asymptotically unbiased and consistent. We investigate their empirical performances via simulations and application to real data.

e-mail: qfa222@uky.edu

Abstract

Biclustering, the process of simultaneously clustering the rows and columns of a data matrix, is a popular and effective tool for finding structure in a high-dimensional dataset.  A variety of biclustering algorithms exist, and they have been applied successfully to data sources ranging from review-website data to gene expression arrays.  Currently, while biclustering appears to work well in practice, there have been no theoretical guarantees about its performance.  We address this shortcoming with a new biclustering algorithm based on profile likelihood and a theorem providing sufficient conditions for consistent biclustering when both dimensions of the data matrix tend to infinity.  This theorem applies to a broad range of data matrices, including binary, count, and continuous data.  We propose a new approximation algorithm to implement profile likelihood biclustering and we show that our algorithm has low computational complexity and performs well in practice.  We demonstrate our results through an empirical study that includes examples from collaborative filtering and microarray analysis.

Abstract

The nonresponse weighting adjustment that consists of multiplying the sampling weight of the respondent by the inverse of the estimated response probability is a common method to reduce the nonresponse bias in sample surveys. However, the response probability estimator from a linear logistic regression using the auxiliary information may not be appropriate for the nonlinear response. We propose the response probability estimator which using the penalized spline regression model and show that it reduces the bias for the point estimator. We discuss the properties of the direct NWA estimator using the penalized spline logistic regression. We describe two methods of the variance estimation. We conduct a simulation study to compare the nonparametric estimator for both linear and nonlinear cases.

Abstract

Grant applications to the National Institutes of Health (NIH) and National Science Foundation (NSF) are submitted as solicited (in response to funding opportunities) or unsolicited (investigator initiated) proposals. This panel will provide suggestions on how to prepare for NIH and NSF grant application submissions; especially for those submitting for the first time. It will cover different stages of grant application process and provide helpful websites and contacts. Information will also be provided for those interested in global health research and collaborations. New policies affecting grant applications are published periodically. As an example, two policies that are published regarding new investigators and multiple investigators will be discussed.

Abstract

In statistical modelling of computer experiments sometimes prior information is available about the underlying function. For example, the physical system simulated by the computer code may be known to be monotone with respect to some or all inputs. We develop a Bayesian approach to Gaussian process modelling capable of incorporating monotonicity information for computer model emulation. A sequential Monte Carlo algorithm is used to sample from the posterior distribution of the process given the simulator output and monotonicity information. The performance of the proposed approach in terms of predictive accuracy and uncertainty quantification is demonstrated in a number of simulated examples as well as a real queueing system.

Abstract

Alignment of functional observations is critical in functional data analysis, and has been extensively studied over the past two decades. An information-geometric framework, referred to as the Extended Fisher-Rao (EFR) method, was recently introduced.  The EFR defines a proper metric on the orbits of functions (quotient space) and all functions are then aligned to a well-defined template.  It was shown that the EFR outperforms previous methods in function registration. However, we note that the EFR has no regularization to control the degree of time warping, which may result in over-alignment in given data.  We propose three forms of regularization terms, referred to as second-order, extrinsic, and intrinsic penalties, in the distance definition.  The new frameworks simplify the EFR by avoiding the notion of quotient space and address issues on over-alignment and outliers, while retaining the efficiency.  Using simulations as well as a real dataset of neural spike trains, we demonstrate that the new methods generate more robust alignment that results in more accurate classification than the EFR method.

Abstract

Thousands of chemicals are utilized in commerce for which adequate toxicity data is lacking.  Unfortunately, standard toxicity assays fail to keep pace with the rate at which new chemicals become commercially available.  Therefore, high throughput in-vitro toxicity screening methods are needed.  In particular, methods for detecting neurotoxicity using multi-well Micro-Electrode Array (MEA) technology are under development.  Electrophysiological data gathered from neuronal cells cultured on MEAs include firing and bursting rates, synchrony and network connectivity.  Harnessing appropriate statistical techniques to quantify chemically induced changes in neuronal cultures is key to the success of in-vitro neurotoxicity screening.  The purpose of this talk is to present the scope and characteristics of MEA data, as well as methods for its visualization and statistical analysis.  

Abstract

Quantile normalization is a powerful method for removing unwanted variability from high-throughput data. It has been widely applied to gene expression microarray data where we assume that observed differences between the distributions of each sample are due to only technical variation unrelated to biological variation. To normalize the samples, the distributions are forced to be the same.  In general, this assumption is justified as only a minority of genes are expected to be differentially expressed between samples, but if the samples are expected to have a high percentage of global differences, it may not be appropriate to use quantile normalization as it may remove interesting global biological variation. We propose a novel method to test for global differences between groups of distributions to guide the choice of a normalization approach. We perform a simulation study to illustrate the bias-variance tradeoff of using normalization methods with and without global adjustments in the context of distributions with and without global differences. In addition, we provide several applications to DNA methylation and gene expression experiments and discuss the importance of batch effects.  An R-package implementing our method is available on Github (https://github.com/stephaniehicks/quantro).

Abstract

Marker selection suffers from the inaccuracy and the lack of reproducibility due to the limitation of sample size. In this study, we conduct integrative analysis and marker selection under the heterogeneity model, which postulates that different datasets have possibly overlapping but not necessarily identical sets of markers. Under certain scenarios, it is reasonable to expect similarity of identified marker sets -- or equivalently, similarity of model sparsity structures -- across multiple datasets. However, the existing methods do not have a mechanism to explicitly promote such similarity. To solve this problem, we develop a novel sparse boosting method. This method uses a BIC/HDBIC criterion to select weak learners and encourage sparsity. A new penalty is introduced to promote similarity of model sparsity structures across datasets. The proposed method has an intuitive formulation and is computationally affordable. Simulation shows that the proposed method outperforms alternatives with more accurate marker identification.

Abstract

Cost efficiency of banks is a key indicator that provides valuable insight to researchers and policymakers about the functioning of the financial intermediation process, and the overall performance of the entire financial system.

This study focuses on the efficiency of the European banking market and uses Bureau van Dijk database of banks' balance sheets and income statements data for 2819 commercial, savings and cooperative banks from 14 European countries for the 2001-2009 period.

Our interest in the subject is twofold. At each nation’s level, cost efficiency influences the relative competitiveness of banks, setting the profile of the national banking industry with direct implications on economic growth.  At the European Union level, the financial, institutional and regulatory integration raise the question about the existence of a common cost frontier.

Among the numerous studies that have explored this subject, two competing approaches that have often produced contradictory results stand out. They differ with respect to the way frontiers are constructed and relative efficiencies determined. These two empirical groups are either developing nation-specific frontiers or assuming a common frontier that may be used to determine and compare efficiencies across countries. There are drawbacks associated with each method. With individual countries frontiers, comparisons across countries are not relevant as the reference system is not the same for everybody. On the other hand, the common frontier setup could be too strong of an assumption as the production technologies might differ substantially across countries.

The objective of this study is to present a Bayesian methodology that nests both of these approaches. A Gibbs sampler is used to estimate a stochastic translog cost frontier with country dummy variables. This is a composite error model that constructs an efficient frontier from which the individual firm deviates due to both inefficiency (incurring higher costs) and measurement error or bad luck (the random aspect).

By varying the degree of precision on the prior distribution of frontier parameters, the model allows for a continuous shift from individual country frontiers to a single, common frontier. This is a flexible way of modeling cost frontiers that nests both approaches (individual multiple frontiers and common single frontier) and it can be applied in numerous other setups.

With an uninformative prior on the variance of the translog parameters, the hybrid model generates the nation-specific frontier results. With a strong prior on the variance of the translog parameters, the hybrid model reproduces the same results as the common frontier approach.

Through the use of an informative prior varied according to our beliefs about the frontiers, the model also allows for different “sub-frontiers” to exist. We find that as the strength of the prior increases and we move from single frontiers for each country towards a common “European frontier", the convergence does not always happen in a direct manner.

The fact that in some cases we need a very strong prior for the parameters' results to start adjusting and move from their single frontier values to their common frontier values might be interpreted as evidence from the data against the idea of a common frontier. This information can be used to separate the countries into subgroups that appear to share a common frontier based on the speed of convergence.

The complexity of the convergence process suggests that our analysis could be improved in future research by running the model for a bigger number of prior values on the variance of the translog parameters.

For a group of selected banks, we compute economies of scale and while the results vary in magnitude depending on the approach, we find more often than not and for all bank sizes that they remain greater than one. Therefore, we can conclude that banks are not operating at their optimum level and could reduce average costs by increasing their output. The subject of economies of scale has been a source of argument for a long time as researchers have found contradictory results over the years. Nevertheless, recent datasets suggest that even large banks have economies of scale greater than one. This study supports these findings.

 Key words:  Stochastic frontier, Bayesian methods, Gibbs sampler, cost efficiency, European banking system

Abstract

Statisticians are extremely respected professionals who play a crucial role in the Department of Health and Human Services and in particular, in the approval process of medical treatments and diagnostics by the Food and Drug Administration. In this presentation the responsibilities of statisticians in the government and the need for communication and collaborative skills in addition to technical statistical skills will be discussed. The statistician in government is a problem solver, who must be interested in science and teaching, and who ideally aspires to leadership positions. The statistician can be a force that spurs innovation, not only in science and medicine, but also in statistical techniques and decision making processes. Potential career ladders of statisticians in the government will be presented. Ample time will be available for questions and comments from the audience.

Abstract

Understanding the effects of phenological events, due to both natural and man-made causes, is critical for research in global climate modeling and agriculture, among many others. During the last two decades, remote sensing data on satellite derived biophysical variables (such as chlorophyll content) have become widely available through the launch of satellites such as MERIS (MEdium Resolution Imaging Spectrometer). Weekly MTCI aggregates from 2003 to 2007 were used to model phenological changes in southern India. Three modeling techniques -- a classical time series model with seasonality represented by Fourier terms, a method integrating the decomposition of time series into season, trend, and white noise components with methods for detecting significant changes (BFAST), and a hierarchical model incorporating spatially distributed covariates -- were used to extract the phenological variables of onset of greenness, peak of greenness, and end of senescence using an iterative search. The advantages and shortcomings of the three methods are compared and discussed in terms of phenological variable extraction and efficacy across spatial locations.

Abstract

All participants in the ongoing STEM education discussion agree that, in addressing national priorities, a key concern is the critical transition of students from high school (or community college) to a four-year college program in the mathematical sciences in particular. Failure in college-level mathematics and statistics courses may discourage students from pursuing STEM majors or perhaps lead to complete college drop-out. In fact, even a mediocre performance in these courses often restricts student career choices to fields outside of STEM disciplines. This presentation is aimed at summarizing a statistical investigation of student self-efficacy and self-confidence in mathematics and statistics, particularly with regard to gender differences and the impact of mathematical preparation at high school or community college levels. Interestingly, preliminary results indicate much higher levels of self-efficacy/-confidence in statistics compared to mathematics. Additionally, for statistics these results are consistent across gender, which is seemingly not the case for mathematics.

Abstract

Receiver operating characteristic (ROC) analysis is widely used to evaluate the performance of diagnostic tests with continuous or ordinal responses. A popular study design for assessing the accuracy of diagnostic tests involves multiple readers interpreting multiple diagnostic test results, called the multi-reader, multi-test design. Although several different approaches to analyzing data from this design exist, few methods have discussed the sample size and power issues. In this article, we develop a power formula to compare the correlated areas under the ROC curves (AUC) in a multi-reader, multi-test design. We present a nonparametric approach to estimate and compare the correlated AUCs by extending DeLong et al.’s (1988) approach. A power formula is derived based on the asymptotic distribution of the nonparametric AUCs. Simulation studies are conducted to demonstrate the performance of the proposed power formula and an example is provided to illustrate the proposed procedure.

Abstract

This poster provides information about the Caucus of Women in Statistics.  I will present our objectives, as well as ways to get involved.  There will be some information about the leaders of this organization.  We would like to recruite you, reach out to women in statistics, remind everyone that there are still gender issues that need to be addressed, and find ways to address them.  We  value the many great accomplishments of women in statistics and celebrate them.  We have come a long way, but we still have a long way to go.  A recurring theme is that we need more women in leadership and other visible positions - we hope that the CWIS can help change the tide by being more visible and outspoken, reminding people of the gender differences that persist.  Our hope is to make a difference for women, ourselves and those who come after us, in their professional life. 

Abstract

Interest in the ocean has grown over time, and in the last three centuries ocean temperature data was gathered at particular locations and depths around the globe, and is currently available via the World Ocean Database (WOD). We develop a hierarchical Bayesian model to describe the relationship between ocean temperature and its depth. The proposed model is useful for interpolating an entire profile of temperature at a new location and identifying the number of mixed layers in the part of the ocean. Spatial random effects are modeled through a generalization of "kernel convolution". The model is fitted to the data gathered at 31 sites in the North Atlantic Ocean region with latitude between 41.5 and 47 and longitude between -27.5 and -22 in July 2012.

Abstract

          We analyze the annual audit fees for 1947 companies from 2002 through 2009 as functions of company size and degree of industry regulation.  Bayesian models with different variance-covariance structures across the 8 years of repeated measures are used and model fits are compared and contrasted with traditional parametric mixed models fitted using likelihood methods.  Markov Chain Monte Carlo (MCMC) methods are used for the simulations as they provide estimates of the posterior distributions for model parameters which are not dependent on parametric assumptions such as normality or constant variance and are generally robust to choices of priors, most especially non-informative priors.  In cases where an analytic representation of the posterior distribution is not tractable, an additional Metropolis-Hastings (MH) step is incorporated at the expense of greater computational complexity.  Within the fitted model structures, we compare mean responses between large companies and small companies and regulated and non-regulated industries and analyze the trend in mean audit fee ratios across the eight years of observed data.  Of greater importance in these analyses is any trend(s), before and after regulatory requirements changed at year 2004, in variability in audit-fee ratios and the degree of dependence between adjacent years, as these are measures of changing economic volatility in company audit expenses.

Abstract

Large scale surveys in mammalian tissue culture cells suggest that the protein expressed at the median abundance is present at 8,000–16,000 molecules per cell and that differences in mRNA expression between genes explain only 10–40% of the differences in protein levels. We find, however, that these surveys have significantly underestimated protein abundances and the relative importance of transcription. Using individual measurements for 61 housekeeping proteins to rescale whole proteome data from Schwanhausser et al. (2011), we find that the median protein detected is expressed at 170,000 molecules per cell and that our corrected protein abundance estimates show a higher correlation with mRNA abundances than do the uncorrected protein data. In addition, we estimated the impact of further errors in mRNA and protein abundances using direct experimental measurements of these errors. The resulting analysis suggests that mRNA levels explain at least 56% of the differences in protein abundance for the 4,212 genes detected by Schwanhausseret al. (2011), though because one major source of error could not be estimated the true percent contribution should be higher. We also employed a second, independent strategy to determine the contribution of mRNA levels to protein expression. We show that the variance in translation rates directly measured by ribosome profiling is only 12% of that inferred by Schwanhausser et al. (2011), and that the measured and inferred translation rates correlate poorly (R2 = 0.13). Based on this, our second strategy suggests that mRNA levels explain ∼81% of the variance in protein levels. We also determined the percent contributions of transcription, RNA degradation, translation and protein degradation to the variance in protein abundances using both of our strategies. While the magnitudes of the two estimates vary, they both suggest that transcription plays a more important role than the earlier studies implied and translation a much smaller role. Finally, the above estimates only apply to those genes whose mRNA and protein expression was detected. Based on a detailed analysis by Hebenstreit et al. (2012), we estimate that approximately 40% of genes in a given cell within a population express no mRNA. Since there can be no translation in the absence of mRNA, we argue that differences in translation rates can play no role in determining the expression levels for the ∼40% of genes that are non-expressed.

Abstract

Identifying differential expressed (DE) genes across various conditions or genotypes is the most typical approach to studying the regulation of gene expression. An estimate of gene specific variance is needed in many methods for DE detection, including linear models (eg. for transformed and normalized microarray data) or generalized linear models (eg. for count data in RNAseq). Due to a common limit in sample size, the variance estimate can be unstable. Shrinkage estimates using empirical Bayes methods have proven useful in improving the detection of DE. The most widely used empirical Bayes methods shrink gene specific variance estimates by borrowing information across genes within the same experiments. In these methods, genes are considered exchangeable or exchangeable conditioning on expression level. We propose, with the increasing accumulation of expression data, borrowing information on historical data on the same gene can provide better estimate of gene specific variance, thus further improve DE detection.  Specifically, we show that the variation of gene expression is truly gene specific and reproducible between different experiments. We present a new method to establish gene specific prior on the variance of expression using existing public data, and illustrate how to shrink the variance estimate and detect differential expression. We demonstrate improvement in DE detection under our strategy compared to leading DE detection methods in resampled real expressed data.

Abstract

It is often important to study the association between two continuous variables. We propose a novel regression framework for assessing conditional associations on quantiles. General methodology is developed which permits covariate effects on both the marginal quantile models for the two variables and their quantile associations. The proposed quantile copula models have straightforward interpretation, facilitating a comprehensive view of association structure which is much richer than that based on standard product moment and rank correlations. The resulting estimators are shown to be uniformly consistent and weakly convergent as a process of the quantile index. Simple variance estimators are presented which perform well in numerical studies. Extensive simulations and a real data example demonstrate the practical utility of the proposed methodology.

Abstract

The likelihood ratio test and the signed likelihood root (r) are two of the most commonly used statistics for inference in parametric models, with the r often viewed as the most reliable. A third-order adjusted signed likelihood root is available from likelihood theory, but the formulas and development methods are not always easily implemented and understood. Using a log-model Taylor expansion, we develop simple second order additive adjustments to the r. For a scalar linear interest parameter, the adjustment to r are easy to calculate and easy to explain. In the curve interest parameter case, adjustment is available but need more calculations. The theory is developed and simulations are conducted to indicate the repetition accuracy.

Abstract

A risk predictive model that can dynamically predict an individual’'s cause-specific cumulative incidence probabilities is crucial to risk stratification, drug effect evaluation and treatment assignment. For data containing no competing risks, landmark Cox models have served for this purpose. In this study, we extended the landmark method to the Fine-Gray proportional subdistribution hazards (PSH) model for data with competing risks. The proposed landmark PSH model is robust against the violation of the PSH assumption and can directly estimate the conditional cumulative incidence probabilities at a fixed landmark point. We further developed a more comprehensive landmark PSH supermodel which enables the user to complete a dynamic prediction involving a number of landmark points in one step. Through simulations we evaluated the prediction performance of the proposed landmark PSH models by estimating the time-dependent Brier scores. The proposed models were applied to a breast cancer trial to predict the dynamic cumulative incidence probabilities of developing locoregional recurrence. 

Abstract

Roll call data are used to make inference about a legislator’s most preferred policy, referred to as his or her ideal point. We develop a Bayesian hierarchical spatial model to account for censoring bias from the Hastert rule in the United States Congress, which restricts voting to bills supported by a majority of the majority party in an attempt to limit the power of the minority. Our model can be summarized as a latent factor model that allows for the position of each legislator in policy space to vary according with whether voting on a specific bill has been censored through the Hastert rule. Zero-inflated prior allows us to assess the effect of the Hastert rule of the voting patterns of individual legislators. The model is illustrated using data from the 110th House of Representatives.

Abstract

Statisticians working in behavioral research have long been interested in estimating the occurences of sensitive behaviors in a population. In this poster session, we consider a questionnaire that asks three binary sensitive questions of each individual participating in a survey. We regard the three responses as dependent Bernoulli random variables governed by a Markov dependency model. We employ a Markov Chain Monte Carlo approach to obtain posterior probabilities for the sensitive questions and consider problems like induced priors and Bayesian sample size determination. The construction of a Bayesian logistic regression framework is outlined.

Abstract

There is a need for new computational tools to address modern data problems, such as teaching students true data analysis, making transparent data journalism, and providing user rewards to participants in citizen science projects. Novices (people with little to no knowledge of statistics or computer science) should be able to start using a simple platform that will help scaffold them into more sophisticated analysis and computation. Such a system should encourage computational thinking, exploratory data analysis, flexible data visualization, and reproducible and publishable research. This framework was inspired by the struggle to find an appropriate computational tool for high school students to use for data analysis that we have experienced in the NSF-funded project, Mobilize. Over the years of the grant, we have iterated through a series of tools (standard R, Deducer, R within RStudio, and several artisanal dashboards) and our experience has given us some insight into what makes a tool succeed or flop in this sort of situation. Inspired by the need for a new tool, we have begun creating a new platform. Instead of asking students to learn on a toy system and make the big leap to a ``real'' programming language, we are attempting to scaffold the transition visually; from pieces that illustrate the processes that are taking place to the underlying code that makes it happen.

Abstract

The use of hyperspectral sensors in remote sensing of gas plumes has proven to be important for a wide variety of military and environmental applications.  For example, hyperspectral imagery can be used to detect chemical warfare agents, and to monitor gas plumes in the atmosphere. A hyperspectral sensor captures a hyperspectral image over the electromagnetic spectrum. Hence, a hyperspectral image is a cube of data. These images typically consist of thousands of pixels. Currently, algorithms for detecting, quantifying, and identifying the constituents of chemical plumes in a hyperspectral cube involve analyzing such images one pixel at a time. However, analyzing pixels individually is inefficient because it ignores the spatial relationships among neighboring pixels. As a novel approach, we consider a variant of the least absolute shrinkage and selection operator (LASSO), called the Fused LASSO, which analyses pixels simultaneously, and allows us to borrow strength or information from nearby pixels. As an illustration, we apply the Fused Lasso to a hyperspectral image which contains plume-present pixels, as well as pixels with no plume, and by using confusion matrices and the Frobenius norm, we show that borrowing information across nearby pixels does substantially better than methods that analyze pixels individually, such as the BIC, and the LASSO.

Abstract

This panel will explore different styles of successful leadership. Women inherently develop as people leaders or managers. They traditionally have primary responsibility for managing the activities and growth of the family. They are also often expected to be the caretaker of their parents as they age. Women or girls are often raised to have empathetic skills so that they can be good mothers and/or caretakers. The panel will discuss the ways the traditional female role helps make women successful leaders. Other leadership styles contrary to the traditional female role will also be discussed.

Abstract

The relatively new upsurge of online education promises to reshape how educators deliver knowledge. As promising as the new model is, it is still in the budding stages. Online education poses challenges that were not present in face to face delivery of education. This paper compares learning outcomes of undergraduate students at the University of New England, a top twenty medical school in the nation. The paper explores strengths and weaknesses of the two modes of delivery, differences of learning outcomes by gender, final grades, midterm and final exam grades, year in college, major and measures the transition over time. This study compares the gender differences on learning outcomes, with a special focus on which model works best for females.

Abstract

We focus on improving the current methodology for estimating transmission parameters by applying a Bayesian statistical framework with a probabilistic model of disease transmission and generalizing this formulation to any disease. This method takes into account the intrinsic stochasticity of disease transmission and provides more robust parameter estimates.We then use adaptive control techniques with the updated parameters in order to obtain the best policies to minimize the cost infection. Increasing estimation accuracy through the adoption of the Bayesian updating framework will equip policymakers with better tools for mitigating the effects of an epidemic

Abstract

Gene-by-environment interactions are a hypothesized contributing factor to the problem of "missing heritability", and so are the focus of major research efforts. However, there is no clear bridge between the biological and the statistical formulations of gene-by-environment interaction, and due to this deficiency, no statistical standard exists to detect such interactions in a robust, scale-invariant manner. We propose a stricter, more precise definition of statistical interaction and develop a rigorous, generalized method that detects linkage between genotype and gene expression in various environments using this definition. Our method calculates each gene's posterior probability of self-linkage in each of an arbitrary number of conditions through an application of the local false discovery rate, while correcting for confounding non-local linkages via surrogate variable analysis. This method is more robust to monotonic transformation of the data due to a reformulation of the statistical definition of interaction. We validate the results obtained from this method using a specially designed diploid dataset, and extend the method to more complex problems and data.

Abstract

Consider a queueing system with two different types of customers. Each customer incurs a cost depending on its type and waiting time in the queue. It has been studied in many researches when the cost functions are linear, now we are considering under nonlinear cost functions what the best static policy should be, and how the best static policy performs comparing to the generalized $c\mu$ rule which is proved to be asymptotic optimal policy under heavy traffic with convex cost functions. Then we give some sufficient conditions when priority policy performs better than non-priority policy, and when non-priority policy performs better. At last, we compare costs of using the static policies and using the generalized $c\mu$ rule by simulation and conclude that the cost of best static policy does not have significant difference compare to the cost of using the generalized $c\mu$ rule.

Abstract
Regular imputation methods have been used to deal with non-response in several types of survey data. However, in some of these studies, the assumption of missing at random is not valid since the probability of missing depends on the response variable. We propose an imputation method for multivariate data sets when there is non-ignorable missingness. A Dirichlet process mixture of multivariate normals is fit to the observed data under a Bayesian framework to provide flexibility. We provide some guidelines on how to alter the estimated distribution using the posterior samples of the mixture model and obtain imputed data under different scenarios. Lastly, we apply the method to a real data set. 

Abstract

As a student or recent graduate, knowing what types of jobs to apply for and what to expect in your first year is difficult and for some, full of confusion. This session will describe some of the many varied career options available to recent graduates in statistics and biostatistics. The speakers will discuss their transition from student life into a work environment, aspects about their position they found difficult or surprising, aspects about their position they enjoy, how they feel their PhD experience did or did not prepare them for the work expected of them, what they see for themselves in the the future and advice for current students and/or jobseekers.

Abstract

Professional physician networks can potentially influence clinical practices and quality of care.. Currently used models for modeling physician networks assume independence of relationships between different pairs of actors (dyads) conditional on actor-specific effects. However, biased dyadic-level (e.g., reciprocity) and individual-level effects (e.g., same-gender, same residency location) may be obtained if effects involving 3 or more actors are not accounted for in the analysis of the network. To address this deficiency, we develop a new model that accounts for inter-dyad dependence involving multiple (≥ 3) actors thereby allowing for effects such as “transitivity” and feasible implementation on both directed and undirected networks. The new methodology is motivated by two real life physician networks: a directed physician influential conversation network (N=33) and an undirected physician network obtained from patient visit data (N=135). In both cases we find extensive evidence of triadic dependence that if not accounted for confounds the effect of reciprocity and nodal attributes (spatial proximity, complementary expertise). The results of our analysis suggest alternative conclusions to those from incumbent models.

Abstract

Studies which investigate the association between human health and long-term exposure to air pollution typically model the spatial structure in the health outcome using standard spatial regression methods such as kriging, the inclusion of spatial random effects or penalized splines. The goal of these methods is to reduce bias from spatial confounding, which occurs when we cannot distinguish the effect of air pollution from residual spatial variation in the health outcome due to unmeasured spatially varying confounders. These methods, however, will only reduce this bias if the spatial variability in the unconfounded part of the exposure is at a smaller scale than that of the unmeasured confounders. Thus, it is important to consider the spatial scale of both the exposure and the confounding variables. 

Using the health data from the Medicare billings claims and pollution data from the EPA monitoring network, we propose to estimate the health effects associated with long-term exposure to coarse thoracic particulate matter using a two stage approach which considers the spatial scales involved. First, we model the spatial trend in the pollution data using nonparametric methods and specifically a `bias corrected and estimated' generalized cross-validation criterion for selecting the bandwidth. Second, we fit a health model to determine the effect of pollution, which includes a kernel density to account for the spatial structure in the health data. The degrees of freedom for this smooth function are provided by the bandwidth matrix from the first stage, thus allowing the spatial variability of the air pollution data to capture the spatial variability of the unmeasured confounding variables which would effect the health model. 

Abstract

Nowadays, generalized linear mixed models are widely used for data analysis in various sciences. In such models, it is often adopted a parametric approach and accordingly, a normal distribution is used for random effects. Although, this assumption leads us to simple calculations but it can be inappropriate. To alleviate this obstacle, this paper adopts a nonparametric Bayesian framework, which relies on the Dirichlet process. Hence we  consider a flexible class of distribution for random effects. Then, the methodology is extended for longitudinal and spatial data. It should be noted that inference is implemented via the Bayesian approach in which for model parameters, the posterior distribution is sampled based on Markov chain Monte Carlo methods such as the blocked Gibbs sampling algorithm. at the end, the methodology is illustrated  assessed by a simulation example and two applied examples  related standard penetration resistance of soil in Chabahar and the number of economic activists province during the years 1384 to 1389. 

Abstract

Mentoring is important at every stage of life, both in our personal and our professional lives. This panel will focus on how our academic and professional journeys are shaped by good mentoring, whether formal or informal.  

With collective experience in government, industry and academia, the panelists will describe their key mentoring experiences and how those experiences have helped to advance their careers or the careers of others.

The panelists will discuss different types of mentoring relationships, including that of an “executive sponsor” (a personal advocate).  They will also address finding mentors, expectations of a mentor-protege relationship, and mentoring opportunities outside of the workplace, e.g. through professional societies.

Abstract

In survival analysis, the failure time of an event is interval-censored when it is only known to occur between two consecutive observation times. Most existing methods for interval-censored data only account for a single cause of failure. However, in some situations a subject may fail due to more than one causes. We propose estimation procedures that account for both interval censoring and competing risks, adopting the modeling framework of the proportional subdistribution hazards model. The proposed estimating equations effectively utilize the ordering of the event time pairs, and the technique of inverse probability weighting is used to account for the missing mechanism. Simulation studies show that the proposed methods perform well under realistic scenarios.

Abstract

Meta-analysis combines evidence from multiple studies with a common research question to derive a stronger conclusion. It is particularly useful in the health sciences, where it can provide better understanding of the efficacy of a treatment. While an ideal meta-analysis would combine the datasets of each study, a more widely used method performs an analysis on aggregate data extracted from published studies. A common problem is that the summary statistics necessary for the meta-analysis are often unavailable from publications, and common practice is to compute best-guesses using other information in the paper, such as Kaplan Meier plots. However, treating these best-guesses as observed summary statistics leads to unjustified certainty, and possibly to inaccurate conclusions and unfounded policy recommendations. 

We introduce the Uncertain Reading Estimated Events (UR-EE) Bayesian random effects model to incorporate the uncertainty that arises during data extraction in the meta-analysis of odds ratios. We re-evaluate published meta-analyses that compare two treatments (percutaneous coronary intervention and coronary artery bypass grafting) for unprotected left main coronary artery stenosis in three outcomes: (1) mortality, (2) major adverse cardiovascular or cerebrovascular events and (3) target vessel revascularization after one year. The added uncertainty results in an increase in the standard deviation of the log-odds ratio by up to 28%. Alternatively, uncertainty from data extraction is equivalent to a reduction in the meta-analysis effective sample size by up to 38%.

Abstract

The application of statistical analyses has drastically changed over the last years due to continued developments in computer science and information technology. The availability of powerful computers and powerful statistical software in academic research now allow the possibility of using complex models and techniques for the estimation of sample parameters. Applied researchers rely on theoretical and empirical bases to select the required sample size to achieve a desired statistical power; however, within Item Response Theory (IRT), the selection of sample size for accurate person and item parameters estimates is still not well established. Using a simulation approach, this study expands De Ayala and Sava-Bolesta’s (1999) study to the evaluation of the accuracy and precision of item parameter estimates for graded responses using Likert-type data. Manipulated factors include sample size ratio (SSR) to total number of parameters (2.5:1, 5:1, 10:1, and 20:1) and population distribution shape (normal, uniform, and skewed). A total of 5000 samples were generated for each condition in the simulation, which provides a maximum SE of .007 and a 95% CI width ± .013 The simulation study was conducted using SAS 9.3; the IRTGEN SAS macro (Whittaker, Fitzpatrick, Williams, & Dodd, nd) was used to generate data for the graded response model and MULTILOG 7.03 (Thissen, 2003) was used to calibrate the item parameters.

Abstract

Biostatistics concepts are important topics for future health professionals and researchers that are often lacking in health professions curricula, biomedical graduate programs, and post-graduate training programs (i.e. residencies and fellowship programs).  Curricula for these programs are already full without room to add additional courses, although the importance of understanding biostatistics concepts has been recognized.

This presentation describes the development of five e-modules introducing common concepts in biostatistics. (1) Summarizing Data (2) Study Design (3) Variability (4) Risk (5) Statistical Inference Basics.  These topics were selected based on their wide applicability to programs and are designed to be used in any combination ranging from one module to all five.   Each e-module is able to be used with a flipped classroom, where the specific classroom activity or discussion topic is determined by the subject matter faculty member.

This series of modules can be used in whole or in part during a semester course or as additional training not included as part of a formal course. The modules themselves contain biostatistical concepts and examples, along with links for interactive tools or games and practice questions to reinforce concepts. The flipped classroom portion is left to the subject-matter expert so it can be tailored to the discipline specific content, but examples are provided.

Abstract

Graduate school have you stressed? This panel will provide personal stories and advice for not only surviving but thriving in graduate school. Tips and resources for stress management, writing a dissertation, working out advisor conflict, time management, and social networking will be discussed.

Abstract

RNA viruses are challenging for protein and nucleotide sequence based methods of analysis because of their high mutation rates and complex secondary structures. With new DNA and RNA sequencing technologies, viral sequence data from both individuals and populations are becoming easier and cheaper to obtain. Thus, there is a critical need for methods that can identify alleles whose frequencies change over time or due to a treatment. We have developed a novel statistical approach for identifying evolved nucleotides in a viral genome without relying on sequence annotation or the nature of the change. Its findings reveal nucleotides that have similar patterns of change. Our approach models allelic variances under a Bayesian Dirichlet mixture distribution. With a multi-stage clustering procedure we have developed an efficient clustering scheme that distinguishes treatment causal changes from variation within viral populations. Our method has been applied to an established HIV data set where we are able to replicate and improve previous findings. We believe our approach can be broadly applied and is particularly useful for the cases that are recalcitrant to traditional sequence analysis.

Abstract

Data exploration includes generating an overview of variables contained in a dataset and an understanding of the most important relationships among the existing variables. Data needs to be prepared for modeling while the models themselves are used for gaining insights into the main aspects of the discovered relationships. It involves use of many statistical techniques that require the judicious application of a proficient analyst. The task becomes much more difficult for large datasets with many variables. I present a robust automation framework that runs a set of exploratory statistical analyses on a given dataset. The statistics range from univariate analyses establishing the necessary metadata information to multivariate analyses discovering the relationships between the target variables and potential predictors. All the results are sorted by a new "interestingness" index comprised of suitable statistics. This allows the most relevant relationships to be selected. The most important aspects of the discovered relationships are further analyzed and presented as insights and expressed using a non-technical language allowing non-experts to gain a better understanding of their data.

Abstract

 In health services research, it is common to encounter semicontinuous data characterized by a point mass at zero followed by a right-skewed continuous distribution with positive support. Examples include health expenditures, in which the zeros represent a subpopulation of patients who do not use health services, while the continuous distribution describes the level of expenditures among health services users. Semicontinuous data are typically analyzed using two-part mixtures that separately model the probability of health services use and the distribution of positive expenditures among users. However, because the second part conditions on a nonzero response, conventional two-part models do not provide a marginal interpretation of covariate effects on the overall population of health service users and non-users, even though this is often of greatest interest to investigators. We propose a marginalized two-part model that yields more interpretable effect estimates in two-part models by parameterizing the model in terms of the marginal mean. This model maintains many of the important features of conventional two-part models, such as capturing zero-inflation and skewness, but allows investigators to examine covariate effects on the overall marginal mean, a target of primary interest in many applications. Using a simulation study, we examine properties of the maximum likelihood estimates from this model.We illustrate the approach by evaluating the effect of a behavioral weight loss intervention on health care expenditures in the Veterans Affairs health care system.

Abstract

Case-control studies are designed towards studying associations between risk factors and a single, primary outcome. Information about additional, secondary outcomes is also collected,  but association studies targeting such secondary outcomes should account for the case-control sampling scheme, or otherwise results may be biased. Often, one uses inverse probability weighted (IPW) estimators to estimate population effects in such studies. However, these estimators are  inefficient relative to estimators that make additional assumptions about the data generating mechanism. 

We propose a class of estimators for the effect of risk factors on a secondary outcome in case-control studies,  when the mean is modeled using either the identity or the log link.  The proposed estimator combines IPW with a mean zero control function that depends explicitly on a model for the primary disease outcome. The efficient estimator in our class of estimators reduces to standard IPW when the model for the primary disease outcome is unrestricted, and is more efficient than standard IPW when the model is either parametric or semiparametric. 

Abstract

The American Statistical Association presidency has been held by 12 women:  Helen Walker, Gertrude Cox, Margaret Martin, Barbara Bailer, Janet Norwood, Katherine Wallman, Lynne Billard, Sallie Keller, Mary Ellen Bock, Sally Morton, Nancy Geller, and Marie Davidian.  This panel will introduce you to women who have taken on leadership of the American Statistical Association.  Come learn about how they entered the field of statistics/biostatistics, what launched and propelled their careers, what hurdles they encountered, what tools they used to stay inspired, and what advice they have for women in all stages of their careers.

Abstract
Internships can be a valuable mechanism for gaining exposure to the field of statistics and obtaining practical experience,
particularly early in a career. In addition to honing technical skills, interns also have the ability to develop their soft skills
in areas such as communication, teamwork, leadership, etc. Many companies, government agencies, and other
organizations offer internships to identify promising employees, to increase the available talent pool in a given field,
and to promote educational opportunities for students and early career participants. This session will discuss how
internships are structured, how to identify internship opportunities, and how to make the most out of an internship
experience from the perspective of three different presenters. Susan Paddock, Senior Statistician at the RAND Corporation and Head of the Statistics Group, will describe the
opportunities available through RAND's Summer Associate Program and discuss examples of summer projects.
Jennifer Van Mullekom, Senior Consulting Statistician in DuPont's Applied Statistics Group, will discuss DuPont's
extended internship program as well as projects from past interns.
Joanne Wendelberger, Group Leader of the Statistical Sciences Group, will describe internship opportunities at
Los Alamos National Laboratory.

Abstract

Researchers often analyze data assuming models with order-restricted parameters. In Bayesian hypothesis testing, incorporating inequality constraints on model parameters has lead to "informative hypotheses'' and associated priors. It is well known in the frequentist context that specifying inequality constraints in the alternative hypothesis increases power by eliminating ad hoc multiple tests and the associated inflated family-wise error. In Bayesian hypothesis testing, similar improvements are seen in operating characteristics. Additionally, due to model parameter estimation, model dimensionality is a key component to the sample size requirement. The effect of sample size in these problems has received little attention. In this poster session, we investigate informative hypotheses using Bayesian methods. We explore Bayesian sample size determination techniques for a variety of settings in the informative hypothesis context, including model dimension and order violations.

Abstract

A central goal in the biological and biomedical sciences is to identify locations along the genome that can explain variation in quantitative traits. Over the last decade, improvements in sequencing technologies coupled with the active development of association mapping methods have made it possible to detect locations (single nucleotide polymorphisms, or SNPs) linked to quantitative traits using data from randomly-sampled individuals.  However, existing methods are severely limited by either their inability to consider complex, but biologically-realistic, scenarios or by their disregard of relationships imposed by the evolutionary history among SNPs during analysis.  By using the evolutionary history among SNPs to incorporate a variance-covariance structure into analysis, the proposed association mapping method relaxes the assumption of independence among observations assumed by most classical statistical methods.  Other previous methods have shown increased performance by accounting for the evolutionary history among SNPs during analysis, but are not flexible enough for application to complex data types.  By using a broad-scale estimate of the evolutionary history among SNPs in a flexible modeling framework, the proposed method is computationally feasible and can be used to detect SNPs associated with quantitative trait variation in a variety of complex scenarios.

Abstract

The second Strategic Highway Research Program (SHRP2) is a partnership of federal, state, and private transportation organizations which aims to increase the safety, efficiency, productivity, and reliability of highway systems in the United States.  The Institute for Transportation at Iowa State University plays an important role in this program by investigating driver behavior when navigating rural curves, as curves have been shown to have a crash rate of nearly three times that of tangent sections.  A key element in this research is determining how drivers change their behavior when navigating a curve.  In this poster, I focus on driver behavior when approaching a curve and specifically on pinpointing at what point before a curve drivers react to the curve.  I use frequentist and Bayesian methods to fit change point models to over 100 driver traces in order to determine how far upstream from a curve drivers should change their behavior in order to navigate the curve safely.  This research has important implications for placement of chevrons and other signage on rural roads that could help reduce rate of accidents on rural highways.   

Abstract

Escalation in frequency of  drug self-administration is a hallmark of developing drug dependence. In NIH pre-clinical studies with species other than humans, the body of evidence implicates several processes that might promote this escalation, including development of tolerance to drug effects, sensitization, and possibly an allostatic reduction in the reward function subserved by each successive drug self-administration occasion.  We posit that parametric mathematical models should be useful in the prediction of drug dose-effect curves (DECs) in response to these processes. To begin, extending this line of research to human clinical research, the focal point of our research starts with estimates of the probability of being a case of drug dependence, expressed as a function of the count of days or occasions of drug use. In these first steps, the DECs are plotted with the estimated probability of being a DD case on the y-axis and with the most recent count of days or occasions of use of the drug on the x-axis, based on an assumption of no feedback loops. That is, we begin with a simple model that ignores a likely feedback loop we mentioned in prior work -- namely, the likely influence of the drug dependence state on what happens next, including the possibility that becoming drug dependence starts to drive up the count of occasions of drug use. To illustrate our approach, we turn to recent empirical data from the  National Surveys on Drug Use and Health (NSDUH) data and we generate DECs for different psychoactive drug subtypes, aiming to shed light on drug-specific variations in the DEC curves well as potential male-female variations.

Abstract

Keynote

Abstract

Cross-validation methodology has long been a popular method for selecting tuning parameters in non and semiparametric models. However, it suffers from high variability and tends to overfit the data. We consider possible improvements of leave-one-out cross validation in kernel density estimation, and propose a subsampling-extrapolation procedure that reduces the variability of the conventional bandwidth selector dramatically. It is based on first evaluating the risk at a subsample size m (m<sample size n), and then extrapolating the optimal bandwidth from m to n. We have noticed that the proposed first-order extrapolated bandwidth selector is equivalent to the rescaled bagging CV method in Hall and Robinson (2009) if one sets the bootstrap size equal to the subsample size. However, our simple expression of the U-statistic form risk estimate enables us to compute the aggregated risk much more efficiently than bootstrapping. In addition, we also consider a second-order extrapolation as an alternative to improve the performance of approximating the true optimal bandwidth. Extensive simulation study has shown that the subsampling-extrapolation bandwidth selectors outperform the conventional bandwidth selector across a wide selection of distributions. To select the optimal subsample size in the subsampling stage, we propose a Nested Cross Validation method which is a data-driven, automatic selection criterion.

Abstract

Many of the time series data exhibit nonlinearity. Nonlinear autoregressive models are very popular to use in time series analysis because of its great exibility. In such modeling process, very often one needs to include many lagged variables in the model to capture the persistence of a time series. Sometimes, the lag length can be very long, or even close to the length of time series. Such "curse of dimensionality" problem is challenging for nonparametric modelling and there is a need to select signicant explanatory lagged variables. Single-index model is an appealing and fundamental tool for handling "curse of dimensionality". In this paper we consider nonlinear single-index time series models. In addition we propose a method to select signicant explanatory lagged variables. We apply polynomial spline basis function expansion and smoothly clipped absolute deviation penalty to perform estimation and lag selection in the framework of high-dimensional time series. Under stationary and strong mixing conditions, the resulting estimators enjoy the "oracle" property even when the number of index parameters tends to innity as the sample size increases. An ecient iterative algorithm has been developed to identify the lags and estimate the coecients simultaneously. Both numerical studies and real data application conrm a good performance of the proposed method.

Abstract

Large entry-level courses are commonplace at public 2- and 4-year institutions of higher education (IHEs) across the United States. Low pass rates in these entry-level courses, coupled with tight budgets, have put pressure on IHEs to look for ways to teach more students more effectively at a lower cost. Efforts to improve student outcomes in such courses are often called ``course redesigns.'' The difficulty arises in trying to determine the impact of a particular course redesign; true random-controlled trials are expensive and time-consuming, and few IHEs have the resources or patience to implement them. As a result, almost all evaluations of efforts to improve student success at scale rely on observational studies. At the same time, standard multilevel models may be inadequate to extract meaningful information from the complex and messy sets of student data available to evaluators because they throw away information by treating all passing grades equally. We propose a new Bayesian approach that keeps all grading information: a partially ordered multinomialprobit model with random effects fit using a Markov Chain Monte Carlo algorithm, and a logit model that can be fit with importance sampling. Simulation studies show that the Bayesian Partially Ordered Probit/Logit Models work well, and the parameter estimation is precise in large samples. We also compared this model with standard models considering Mean Squared Error and the area under the Receiver Operating Characteristic (ROC) curve. We applied these new models to evaluate the impact of a course redesign at a large public university using the students' grade data from the Fall semester of 2012 and the Spring semester of 2013.

Abstract

The area under the curve (AUC) summary measure has had extensive use in the pharmaceutical sciences and is gaining popularity in animal sciences research as well. Summary measures are used to summarize multiple measurements on a subject into a single measurement, then a typical linear model-based analysis follows using the summary measure as either an independent or dependent variable. Analyses that use summary measures induce measurement error into the model, which should be accounted for. However, these methods require an estimate of the measurement error in the AUC. A simulation study is used to explore two different measurement error estimates, based on leave-one-out cross validation and bootstrapping. An application to an equine data set that uses the AUC summary measures demonstrates the potential advantages of accounting for measurement error in an AUC-based analysis. 

Abstract

The field of Statistics is growing; however, there remains a shortage of women and minorities who earn advanced degrees and attain senior positions.  This panel will examine best practices for promoting inclusiveness in Statistics.  Graduation rates of underrepresented groups in doctoral Statistics programs will be presented.  The panel will also discuss factors contributing to the success of women and minorities in the profession, specifically in academia. Initiatives to increase doctoral recipients and senior faculty from underrepresented groups will be highlighted.

Abstract

We study Bayesian hierarchical models in environmental applications where likelihood calculations are infeasible for use in inference procedures such as Markov chain Monte Carlo. In this case, an approximation to the likelihood may be used. One such instance is when the likelihood depends on output from a computationally demanding computer model. A statistical emulator may be used to approximate the computer model output. In this work, we study the discrepancy between posterior distributions obtained by using the ‘true’ model versus the approximating model. We quantify the model error in the resulting posterior distributions using an estimate of the Kullback-Leibler divergence. Our estimator is general enough to be used in situations where the true model likelihood can only be evaluated a moderate number of times. We illustrate our methodology with ecological applications. 

Abstract
Many scientists are drawn to statistics and biostatistics by their passion to make a difference by using their skills in mathematics to tackle applied problems. We discuss a variety of aspects of working in interdisciplinary teams, including expectations and roles of statisticians and collaborators, responsibility for data management, authorship and publication, funding expectations, grant and proposal writing, and combining methodology with collaboration. The panel has participated in a broad range of interdisciplinary work. Dr. Wendelberger has experience with interdisciplinary teaming in a government laboratory setting at Los Alamos, in an industrial research setting at General Motors, and in an academic consulting laboratory at the University of Wisconsin. Dr. Wilson has worked collaboratively at the National Institutes of Health with clinical scientists, at a small defense contractor with military personnel, at a national laboratory with engineers, and in Washington with policy analysts. Dr. Stinnett has worked extensively with physician researchers in several departments at Duke University School of Medicine. Dr. Gaydos leads a cross-disciplinary team at Eli Lilly fosued on clinical development scenario planning and decision making through the use of modeling and simulation, novel trial designs and analyses to increase clinical development efficiencies.

Abstract

In isotonic regression, the mean function is assumed to be monotone in- creasing (or decreasing) but otherwise unspecified. The classical isotonic estimator is known to be inconsistent at boundaries; this is called the “spiking” problem. A penalty on the range of the regression function is proposed to correct the spiking problem for univariate and multivariate isotonic models. The penalized estimator is shown to be consistent everywhere for a wide range of sizes of the penalty parameter. For the univariate case, the optimal penalty is shown to depend on the derivatives of the true regression function at the boundaries. Pointwise confidence intervals are constructed using the penalized estimator and bootstrapping ideas; these are shown through simulations to behave well in moderate sized samples. Simulation studies also show that the power of the hypothesis test of constant ver- sus increasing regression function improves substantially compared to the power of the test with unpenalized alternative, and also compares favorably to tests using parametric alternatives.

Abstract

This paper considers testing serial correlation in partially linear additive errors-in-variables model. Based on the empirical likelihood based approach, a test statistic was proposed, and it was shown to follow asymptotically a chi-square distribution under the null hypothesis of no serial correlation. Finally, some simulation studies  are conducted to illustrate the performance of the proposed method.

Abstract

This panel aims to provide useful resources and discussions on how to write a successful proposal to apply for NSF Early Faculty Development (CAREER) award, the most prestigious awards in support of junior faculty who exemplify the role of teacher-scholars in research and education. The panelists will share their views and experiences, from the perspectives of awardees and reviewers, on various aspects of the award application. Specifically, the panel discussion will cover the proposal preparation, writing tips, budget, review criteria and selection process.  

Abstract

Biomedical research is increasingly data-driven and has become a fertile ground for innovation in statistical methodologies. Here we show an example where an interest of identifying genes associated with phenotype interactions using microarray data motivated the development of a novel method for identifying differentially altered (DA) biomolecules using observational high-throughput data. Most methods for DA bio-molecule detection can be considered as single model approaches since they rely on the ranking of the bio-molecules based on statistics derived from a single model for two or more group comparisons, with or without adjustment for other covariates. Such single model approaches are conceptually flawed since they unavoidably result in model misspecification. We show the evidence that DA bio-molecule detection based on high throughput data intrinsically requires a multi-model handling. To properly control for sample heterogeneity and to provide a flexible and coherent framework for identifying simultaneously DA bio-molecules associated with a single or multiple sample characteristics and/or their interactions, we developed a Bayesian model averaging approach with an empirical prior model probability specification. We demonstrated through simulated microarray data that this approach improves the performance of DE gene detection comparing to the single model approaches. Flexibility of this approach is illustrated through analysis of gene expression microarray data and metabolomics data.