Free-response text data contain valuable information that may not be adequately captured via conventional qualitative methods, but incorporating such unstructured data into statistical models remains challenging. Topic modeling provides a framework for extracting latent themes from text based on word distributions, enabling text to be incorporated as informative covariates. This paper develops a framework for integrating topic information extracted from free-response text into small area estimation models. Specifically, we use the posterior distribution from a Hierarchical Mixture of Unigrams model that accounts for informative sampling as covariates in a Bayesian Fay-Herriot model to improve estimation of survey responses collected from separate data sources. By combining information across surveys, our approach improves estimation precision while accounting for domain-level variability and uncertainty. We evaluate the proposed method through simulation-based metrics using data-thinning procedures and compare performance against a standard intercept-only Fay-Herriot model. Results demonstrate that incorporating posterior topic information improves estimation performance, and models that account for nonlinear relationships between topics and the response further improve upon linear specifications. We apply our method to publicly available free-response data from the American National Election Studies (ANES) describing the most important problems facing the country and estimate anxiety levels from the Household Pulse Survey (HPS). Results reveal meaningful nonlinear relationships between text-derived topics and anxiety, demonstrating the value of combining free-response text with traditional survey data. Overall, this framework provides a flexible approach for incorporating information from unstructured text into small area estimation and improving inference across multiple survey sources.

Preterm birth is the leading cause of perinatal mortality, increasing the risk of chronic medical conditions and developmental delays in children. In 2020, 13.4 million preterm births caused around 900,000 infant deaths globally, and in the U.S. alone, the preterm birth rate stood at 10.4%, leading to higher neonatal mortality, significant morbidity, and a minimum $25.2 billion economic cost. Subsequently, there is growing literature demonstrating health disparities in the U.S. in conjunction with pregnancy and its direct and mediated relationship with anxiety [1], [2]. Current efforts in creating large-scale, public health oriented data repositories, including NIH's AllofUs, have further energized research related to women's health, across many ethnic and socio-economic groups. Using the data extracted from AllofUS, we examine our hypothesis that having anxiety during pregnancy disproportionately affects pregnant women from minority groups and lower socio-economic backgrounds. This, in turn, leads to substantial health disparities in pregnancy outcomes, particularly premature delivery. In this work, we highlight two parallel modeling approaches for studying the dynamics of women's pregnancy, anxiety, and premature delivery, particularly among minorities and economically disadvantaged groups. Specifically, we we use ordinal logistic regression and cox proportional hazards models to identify parameters contributing to anxiety among pregnant women.

References: 

[1] Dario Elias, Hebe Campaña, Fernando A Poletta, Silvina L Heisecke, Juan A Gili, Julia Ratowiecki, Mariela Pawluk, Maria R Santos, Viviana Cosentino, Rocio Uranga, et al. Preterm birth etiological pathways: a bayesian networks and mediation analysis approach. Pediatric research, 91(7):1882–1889, 2022.

[2] Lianne Tomfohr-Madsen, Emily E Cameron, Christine Dunkel Schetter, Tavis Campbell, Maeve O’Beirne, Nicole Letourneau, and Gerald F Giesbrecht. Pregnancy anxiety and preterm birth: The moderating role of sleep. Health psychology, 38(11):1025, 2019.

Joint Statistical Meeting at Boston, MA with Dr. Paul A. Parker and Dr. Scott H. Holan

Title: Integration of ANES Free-Response Text Data to Improve Small Area Estimates of Anxiety Using the Household Pulse Survey 

Link to Slides

International Conference Health Policy Statistics at San Diego under PI, Dr. Sam Behseta

Title: Exploring Preterm Birth Health Disparities: Insights into Maternal Anxiety Across Demographic & Socio-Economic Groups