Journal of Data Science

This paper aims to determine the effects of socioeconomic and healthcare factors on the performance of controlling COVID-19 in both the Southern and Southeastern United States. This analysis will provide government agencies with information to determine what communities need additional COVID-19 assistance, to identify counties that effectively control COVID-19, and to apply effective strategies on a broader scale. The statistical analysis uses data from 328 counties with a population of more than 65,000 from 13 states. We define a new response variable by considering infection and mortality rates to capture how well each county controls COVID-19. We collect 14 factors from the 2019 American Community Survey Single-Year Estimates and obtain county-level infection and mortality rates from USAfacts.org. We use the least absolute shrinkage and selection operator (LASSO) regression to fit a multiple linear regression model and develop an interactive system programmed in R shiny to deliver all results. The interactive system at https://asa-competition-smu.shinyapps.io/COVID19/ provides many options for users to explore our data, models, and results.

Linear regression models are widely used in empirical studies. When serial correlation is present in the residuals, generalized least squares (GLS) estimation is commonly used to improve estimation efficiency. This paper proposes the use of an alternative estimator, the approximate generalized least squares estimators based on high-order AR(p) processes (GLS-AR). We show that GLS-AR estimators are asymptotically efficient as GLS estimators, as both the number of AR lag, p, and the number of observations, n, increase together so that $p=o({n^{1/4}})$ in the limit. The proposed GLS-AR estimators do not require the identification of the residual serial autocorrelation structure and perform more robust in finite samples than the conventional FGLS-based tests. Finally, we illustrate the usefulness of GLS-AR method by applying it to the global warming data from 1850–2012.

Statistical learning methods have been growing in popularity in recent years. Many of these procedures have parameters that must be tuned for models to perform well. Research has been extensive in neural networks, but not for many other learning methods. We looked at the behavior of tuning parameters for support vector machines, gradient boosting machines, and adaboost in both a classification and regression setting. We used grid search to identify ranges of tuning parameters where good models can be found across many different datasets. We then explored different optimization algorithms to select a model across the tuning parameter space. Models selected by the optimization algorithm were compared to the best models obtained through grid search to select well performing algorithms. This information was used to create an R package, EZtune, that automatically tunes support vector machines and boosted trees.

Identifying treatment effect modifiers (i.e., moderators) plays an essential role in improving treatment efficacy when substantial treatment heterogeneity exists. However, studies are often underpowered for detecting treatment effect modifiers, and exploratory analyses that examine one moderator per statistical model often yield spurious interactions. Therefore, in this work, we focus on creating an intuitive and readily implementable framework to facilitate the discovery of treatment effect modifiers and to make treatment recommendations for time-to-event outcomes. To minimize the impact of a misspecified main effect and avoid complex modeling, we construct the framework by matching the treated with the controls and modeling the conditional average treatment effect via regressing the difference in the observed outcomes of a matched pair on the averaged moderators. Inverse-probability-of-censoring weighting is used to handle censored observations. As matching is the foundation of the proposed methods, we explore different matching metrics and recommend the use of Mahalanobis distance when both continuous and categorical moderators are present. After matching, the proposed framework can be flexibly combined with popular variable selection and prediction methods such as linear regression, least absolute shrinkage and selection operator (Lasso), and random forest to create different combinations of potential moderators. The optimal combination is determined by the out-of-bag prediction error and the area under the receiver operating characteristic curve in making correct treatment recommendations. We compare the performance of various combined moderators through extensive simulations and the analysis of real trial data. Our approach can be easily implemented using existing R packages, resulting in a straightforward optimal combined moderator to make treatment recommendations.

Due to long-standing federal restrictions on cannabis-related research, the implications of cannabis legalization on traffic and occupational safety are understudied. Accordingly, there is a need for objective and validated measures of acute cannabis impairment that may be applied in public safety and occupational settings. Pupillary response to light may offer an avenue for detection that outperforms typical sobriety tests and tetrahydrocannabinol concentrations. We developed a video processing and analysis pipeline that extracts pupil sizes during a light stimulus test administered with goggles utilizing infrared videography. The analysis compared pupil size trajectories in response to a light for those with occasional, daily, and no cannabis use before and after smoking. Pupils were segmented using a combination of image pre-processing techniques and segmentation algorithms which were validated using manually segmented data and found to achieve 99% precision and 94% F-score. Features extracted from the pupil size trajectories captured pupil constriction and rebound dilation and were analyzed using generalized estimating equations. We find that acute cannabis use results in less pupil constriction and slower pupil rebound dilation in the light stimulus test.

Obesity rates continue to exhibit an upward trajectory, particularly in the US, and is the underlying cause of several comorbidities, including but not limited to high blood pressure, high cholesterol, diabetes, heart disease, stroke, and cancers. To monitor obesity, body mass index (BMI) and proportion body fat (PBF) are two commonly used measurements. Although BMI and PBF changes over time in an individual’s lifespan and their relationship may also change dynamically, existing work has mostly remained cross-sectional, or separately modeling BMI and PBF. A combined longitudinal assessment is expected to be more effective in unravelling their complex interplay. To mitigate this, we consider Bayesian cross-domain latent growth curve models within a structural equation modeling framework, which simultaneously handles issues such as individually varying time metrics, proportion data, and potential missing not at random data for joint assessment of the longitudinal changes of BMI and PBF. Through simulation studies, we observe that our proposed models and estimation method yielded parameter estimates with small bias and mean squared error in general, however, a mis-specified missing data mechanism may cause inaccurate and inefficient parameter estimates. Furthermore, we demonstrate application of our method to a motivating longitudinal obesity study, controlling for both time-invariant (such as, sex), and time-varying (such as diastolic and systolic blood pressure, biceps skinfold, bioelectrical impedance, and waist circumference) covariates in separate models. Under time-invariance, we observe that the initial BMI level and the rate of change in BMI influenced PBF. However, in presence of time-varying covariates, only the initial BMI level influenced the initial PBF. The added-on selection model estimation indicated that observations with higher PBF values were less likely to be missing.

The COVID-19 outbreak of 2020 has required many governments to develop and adopt mathematical-statistical models of the pandemic for policy and planning purposes. To this end, this work provides a tutorial on building a compartmental model using Susceptible, Exposed, Infected, Recovered, Deaths and Vaccinated (SEIRDV) status through time. The proposed model uses interventions to quantify the impact of various government attempts made to slow the spread of the virus. Furthermore, a vaccination parameter is also incorporated in the model, which is inactive until the time the vaccine is deployed. A Bayesian framework is utilized to perform both parameter estimation and prediction. Predictions are made to determine when the peak Active Infections occur. We provide inferential frameworks for assessing the effects of government interventions on the dynamic progression of the pandemic, including the impact of vaccination. The proposed model also allows for quantification of number of excess deaths averted over the study period due to vaccination.

Inspired by the impressive successes of compress sensing-based machine learning algorithms, data augmentation-based efficient Gibbs samplers for Bayesian high-dimensional classification models are developed by compressing the design matrix to a much lower dimension. Ardent care is exercised in the choice of the projection mechanism, and an adaptive voting rule is employed to reduce sensitivity to the random projection matrix. Focusing on the high-dimensional Probit regression model, we note that the naive implementation of the data augmentation-based Gibbs sampler is not robust to the presence of co-linearity in the design matrix – a setup ubiquitous in $n<p$ problems. We demonstrate that a simple fix based on joint updates of parameters in the latent space circumnavigates this issue. With a computationally efficient MCMC scheme in place, we introduce an ensemble classifier by creating R (∼25–50) projected copies of the design matrix, and subsequently running R classification models with the R projected design matrix in parallel. We combine the output from the R replications via an adaptive voting scheme. Our scheme is inherently parallelizable and capable of taking advantage of modern computing environments often equipped with multiple cores. The empirical success of our methodology is illustrated in elaborate simulations and gene expression data applications. We also extend our methodology to a high-dimensional logistic regression model and carry out numerical studies to showcase its efficacy.

Multi-touch attribution (MTA) estimates the relative contributions of the multiple ads a user may see prior to any observed conversions. Increasingly, advertisers also want to base budget and bidding decisions on these attributions, spending more on ads that drive more conversions. We describe two requirements for an MTA system to be suitable for this application: First, it must be able to handle continuously updated and incomplete data. Second, it must be sufficiently flexible to capture that an ad’s effect will change over time. We describe an MTA system, consisting of a model for user conversion behavior and a credit assignment algorithm, that satisfies these requirements. Our model for user conversion behavior treats conversions as occurrences in an inhomogeneous Poisson process, while our attribution algorithm is based on iteratively removing the last ad in the path.

Research has continued to shed light on the extent and significance of gender disparity in social, cultural and economic spheres. More recently, computational tools from the data science and Natural Language Processing (NLP) communities have been proposed for measuring such disparity at scale using empirically rigorous methodologies. In this article, we contribute to this line of research by studying gender disparity in 2,443 copyright-expired literary texts published in the pre-modern period, defined in this work as the period ranging from the beginning of the nineteenth through the early twentieth century. Using a replicable data science methodology relying on publicly available and established NLP components, we extract three different gendered character prevalence measures within these texts. We use an extensive set of statistical tests to robustly demonstrate a significant disparity between the prevalence of female characters and male characters in pre-modern literature. We also show that the proportion of female characters in literary texts significantly increases in female-authored texts compared to the same proportion in male-authored texts. However, regression-based analysis shows that, over the 120 year period covered by the corpus, female character prevalence does not change significantly over time, and remains below the parity level of 50%, regardless of the gender of the author. Qualitative analyses further show that descriptions associated with female characters across the corpus are markedly different (and stereotypical) from the descriptions associated with male characters.