Invisible Information, Unseen Connections: An Exploration of the Hidden Relationships that can Shape Data

Datasets often reflect complex and nuanced relationships which can be difficult to detect or fully represent with traditional epidemiological methods. This may be problematic as it can hinder further analyses, or give the investigator an incomplete picture of the outcome being studied. In this dissertation I explored three analytic contexts in which important relationships can go undetected and examined several methods that can be used to ascertain hidden or latent relationships in the data, drawing from meta-regression, latent class analysis, network analysis, and transmission modeling. In Aim 1, we used meta-regression to ascertain how the association between individual wealth, country level wealth, and Human Immunodeficiency Virus (HIV) burden has changed over time across a set of Sub-Saharan African (SSA) Countries. It has been assumed that like, The West, HIV is also a disease of poverty in SSA. However newer research suggests that this assumption may not be true. Here, we show that HIV may be positively associated with wealth in urban but not rural contexts, and that this association has waned over time. Aim 2 identifies patterns of sexual behavior and substance use across the life course, and examines the association between these patterns and sexually transmitted infection risk. Risk factors for sexually transmitted infections have proven challenging to study due to their tendency to be highly correlated or even collinear with one another. This collinearity is problematic because it inhibits the ability of statistical software to detect the effect of covariates in a regression model, rendering the coefficients of the variables uninformative. Consequently, alternative approaches are needed in order to identify behaviors that put individuals at risk for infection. This aim uses Latent Class Analysis which, unlike regression, uses collinearity to its advantage to identify response patterns. Our results reveal the existence of 5 archetypes that serve as the basis for the profiles present across our four age strata. However, the exact composition of each strata’s profiles varies in the magnitude that particular behaviors are endorsed, which we attribute to a combination of age, period, and cohort effects. Aim 3 constitutes the first part of a two-part analysis that uses network methodology to characterize and quantify patient movement and disease transmission. In this aim, a descriptive analysis of network structure was undertaken to describe the underlying interrelationship between hospital units and patient movement, using patient transfer data from the University Hospital at Michigan Medicine. We then characterized the resulting network to understand key structural features, including node centrality, graph centralization, degree distributions, and community structure. As a network, University Hospital is decentralized but highly transitivity . In Aim 4 we used an SEIR compartmental model to simulate COVID-19 in a hospital setting, to examine the relationship between the hospital network structure and disease transmission dynamics. The purpose of this analysis was to illustrate how the network relationship between locations can be an underlying structure that informs transmission dynamics within the hospital. In summary, the chapters of this dissertation illustrate contexts in which latent variable associations exist in data and provide tools researchers can use to extract them. It is our hope that this work provokes thought and sparks new lines of inquiry.