Button to scroll to the top of the page.

Updates

Campus health and safety are our top priorities. Get the latest from UT on COVID-19.

Get help with Zoom and more.

Summer 2020 Colloquium: Graduate Portfolio in Applied Statistical Modeling

Speaker

Date

Time

Title and Zoom link

 Kirsten Tuggle

Aug 13

3:30 - 4:00pm

"A Multidimensional Splitting Technique for Gaussian Mixture Model Navigation Filters"

Zoom link: https://utexas.zoom.us/j/93292685387

Meeting ID: 932 9268 5387

 Bethany Wood Aug 13 4:00 - 4:30pm

"Beyond the Federal Poverty Level: Measuring the Effect of Family Wealth on Psychological Distress over Time"

Zoom link: https://utexas.zoom.us/j/93292685387

Meeting ID: 932 9268 5387

 Minghao Gu Aug 13 4:30 - 5:00 pm

"Unsupervised learning applications in e-commerce fraud detection "

Zoom link: https://utexas.zoom.us/j/93292685387

Meeting ID: 932 9268 5387

 Weiwen Zeng Aug 14 11:00 - 11:30am

"Reducing Disparities for Latinx Families of Children with ASD: Evidence from A Two-site RCT Study at 4 Months Post Intervention"

Zoom link: https://utexas.zoom.us/j/93097287039

Meeting ID: 930 9728 7039

 Ibrahim Bicak Aug 14 11:30am - Noon

"Course Repetition in College-level Mathematics Courses Among Community College Transfer Students"

Zoom link: https://utexas.zoom.us/j/93097287039

Meeting ID: 930 9728 7039

       

 

Kirsten Tuggle

Orbit determination, ground tracking, as well many other aerospace and non-aerospace applications require very accurate estimates of system states paired with accurate characterizations of the uncertainties associated with those estimates. Furthermore, this is often in the presence of nonlinear dynamics and measurement models. Nonlinear minimum mean square error estimation (MMSE) techniques are often applied, and due to constrained computation, approximations to the true Bayesian solution must be used. In doing so, a choice is made among the competing notions of computational efficiency and estimation performance. For example, the Extended Kalman Filter is computationally simple but with the possibility of degraded performance in comparison to the more computationally taxing Particle Filter (a sequential Monte Carlo method). Gaussian mixture models (GMMs) offer a promising balance between these notions. Effectiveness of GMM estimation filters relies on efficient management of mixture components over nonlinear dynamics and observations. Traditionally, this is done by re-approximating mixture components as their own mixtures (splitting), thereby increasing the total number of components. Nearly all existing methods perform one-dimensional splitting, meaning components are iteratively added along single dimensions. Doing so can be computationally inefficient, and in effect it gives greater mixture weights components is directions split earlier. Here, one of the first multidimensional splitting techniques is developed for a novel GMM filter. We rely on linear-Gaussian MMSE principles to produce a static splitting library that can applied to not only our version but a wide class of GMM filters.

 

 Bethany Wood

Poverty has been linked to several adverse mental health outcomes across the lifespan, including depression, psychological distress, and anxiety. Most studies that examine poverty look at income and the Federal Poverty Level, however, these studies fail to account for debt, assets, home equity, and other important factors of a person’s financial status. Objectives:The purpose of this study is to observe how a holistic measure of financial status (including debt, home equity, and assets) affects psychological distress across the lifespan. Method: Data was from the publicly available Panel Study of Income Dynamics (PSID). The dependent variable, psychological distress, was measured by the continuous Kessler Psychological Distress Scale (K6) across eight waves from 2001-2017. The predictor was a measure of family wealth, including home equity, family’s income, debt, savings, and reported assets. Data cleaning was conducted in Stata version 14, and longitudinal growth modeling was conducted in SAS version 9.4. Results: The final sample included 35,771 adults across the eight waves of data. Family wealth significantly predicted psychological distress across time (p<0.001). This was true even when accounting for race, sex, education, and other control variables. Implications: The results of this study provide a deeper understanding of wealth and the effects of poverty on psychological distress across the lifespan. Future research should continue to examine mental health outcomes with wealth, rather than simply income, as a social determinant to promote policy changes that mitigate the psychological distress associated with poverty, debt, and financial status.

 

 Minghao Gu

Fraud detection is widely required in government sectors, financial institutions, and e-commerce businesses to protect the security of customers and company and avoid financial losses. Its importance can never be overemphasized. While the majority of the fraud detections are done by the supervised learning algorithms like logistic regression, random forest, and boosting trees, unsupervised learning algorithms have caught much attention in recent studies, because they do not require ground truth to be provided. The underlying hypothesis is that transactions with fraudulent behavior are rare and unique, and often considered as outliers in the dataset. Therefore by applying certain transformations and boundaries, we can measure the “outlierness” of each data point without knowing the label. This is the premise of the research. The dataset used in the research is five months of transactions from an e-commerce website with 190 original features and 600,000 transactions, with a fraud rate of 0.78%. The goal of this study is to implement and determine the effectiveness of various unsupervised learning methods in this dataset and compare the results between each method. For evaluation metrics, the area under precision-recall curve (AUC-PR) and confusion matrix are used. It is found out that with the 10% contamination, DBSCAN(Density-Based Spatial Clustering) and HBOS(Histogram-Based Outlier Score) has the highest recall rate close to 40%. However, the major drawback for all unsupervised learning algorithms is a relatively low precision score among 2-3%. 

 

 Weiwen Zeng

Latinx children with Autism Spectrum Disorder (ASD) experience persistent disparities in diagnostic and healthcare services (Liptak et al, 2008). Parents Taking Action (PTA), a culturally tailored parent education intervention program, was developed to address such disparities. A two-site randomized waitlist-control study was implemented between 2014-2017 to examine the efficacy of the intervention in empowering Latinx mothers of children with ASD. The purpose of this study is to understand whether the intervention effects sustained by examining the differences from baseline to follow-ups. Methods:93 (intervention: N=41, control: N=52) Latinx mothers-child dyads were randomized. We conducted repeated measures ANCOVAs for mother and child outcomes with time, treatment, and an interaction term time*treatment in all models. Maternal education was adjusted for since it was the only significant between-group demographic variable. Parent outcomes examined include: the mother’s confidence in and frequency of using evidence-based treatment strategies for their child, and the family outcomes scale (FOS). Child outcomes included the number of total services received, the child’s social communication questionnaire (SCQ) scores, and challenging behaviors. Findings:Overall, we found significant intervention effects in mothers’ confidence in and frequency of using EB strategies, and in total number of services received by their children with ASD. We did not observe significant treatment differences in the FOS and SCQ scores, neither did we find significant differences in the child’s challenging behaviors (SIB-R). We also observed significant treatment differences across the two study sites.  Conclusion and implications:The results of this study demonstrated the effects overtime of PTA, a culturally tailored intervention for Latinx mothers of children with ASD. This study demonstrates that outcomes from a culturally tailored approach to autism intervention can hold up overtime and can be utilized to address disparities in advocacy and services for racial/ethnic minority children with ASD and their families.

 

 Ibrahim Bicak

I examine math course redundancy experienced by community college transfer students, describing their course repetition patterns in college-level mathematics sequences across community colleges and public universities. This study will analyze how course repetition patterns among community college transfer students in Texas predict student college outcomes (cumulative GPA, bachelor's degree attainment, time to degree, and excess credits). Using data from the Texas Educational Research Center, I leverage longitudinal statewide administrative records for the 2011-2012 and 2012-2013 community college entrants (n=36,079) who transferred to a university within six years of college entry and track students' academic progress over six years. Math course repetitions appear to have some consequences for community college transfer students. Regression results indicate that course repetitions lower the probability of earning a bachelor’s degree.  Moreover, course repetitions increase time to degree and excess credits when students earn a bachelor’s degree. I discuss implications of the study for statewide policy initiatives and institutional practices.