Novel Methods

Multilevel Spatiotemporal Models

Temporal patterns of hospitalizations over the life span on dialysis, where poor outcomes are typically exacerbated during the first year on dialysis, variations in hospitalizations among dialysis facilities across the U.S. contribute to spatial variation. Utilizing USRDS data, we propose a novel multilevel spatiotemporal functional model to study spatiotemporal patterns of hospitalization rates among dialysis facilities. Hospitalization rates of dialysis facilities are considered as spatially nested functional data with longitudinal hospitalizations nested in dialysis facilities and dialysis facilities nested in geographic regions. A multilevel Karhunen-Loeve expansion is utilized to model the two-level (facility and region) functional data, where spatial correlations are induced among region-specific principal component scores accounting for regional variation. Figure: Raw and predicted hospitalization rates at one month after transition to dialysis.

Publications

Profiling/Assessment of Dialysis Facilities

Profiling analysis aims to evaluate health care providers, such as hospitals, nursing homes, or dialysis facilities, with respect to a patient outcome. Various common patient outcomes of interest include hospital-wide or cause-specific in-hospital mortality and 30-day unplanned hospital readmission. For example, the Center for Medicare and Medicaid Services (CMS) has implement various measures of quality of care at Hospital Compare. Our research focuses on elucidating profiling models (e.g., fixed and random effects models, time-dynamic profiling models) for various patient outcomes, with applications to assessment of US dialysis facilities. Fig. Time-dynamic profiling of dialysis facilities for 30-day hospital readmission using novel metric SDRR(t).

Publications

Models for Time-Varying Effects

Methods to assess the effects of risk factors associated with patient outcome, where the effects are not constant, but are dynamic and vary over  the duration of follow-up time or by patient characteristics (e.g., age, disease severity) are important for longitudinal studies.  For example, understanding how cardiovascular (CV) risk evolves during the course of dialysis treatment and how it changes following critical  events like infection-related hospitalization may inform better patient care. These methods will also allow identification of the time periods  of increased outcome (e.g., CV) risk; this knowledge is potentially useful for formulation of CV risk reduction strategies.  Generalized varying coefficient models (GVCMs) depart from traditional simplistic modeling approaches that assume a static or constant effect size for risk factors,  e.g., "patients with baseline diabetes have 20% increased CV outcome risk." Clearly, although such a simplification is useful in some  studies, it cannot be used to quantify how the effects of individual risk factors vary depending on age at the start of dialysis, for instance. Our group and collaborators have developed novel GVCMs and extensions that allow for flexible models of time varying effects.

Publications

Multilevel Joint Modeling

We develop novel multilevel joint models (MJMs) that accounts for three-level hierarchical data, with longitudinal measurements, hospitalizations over time, nested within subjects and subjects further nested within dialysis facilities where they receive regular care. MJM accommodates the hierarchical structure of the data from the USRDS, through multilevel random effects and multilevel risk factors affecting both survival and longitudinal hospitalization outcomes. At the subject level, these include patient demographics and baseline comorbidities. At the facility level, facility staffing, such as the ratio of nurse to patients, may impact patient outcomes.

Publications

Self-Controlled Case Series Method, Exposure Onset Error

The self-controlled case series (SCCS) method is an approach to study the relationship between time-varying exposures and adverse events (AEs), such as AEs following vaccination or other acute exposures.  The SCCS design requires only subjects with one or more events. This aspect of the SCCS design is particularly useful for large longitudinal database applications. Another major advantage of the SCCS method is that it controls for all measured and unmeasured baseline confounders and is self-matched. Thus, the SCCS estimate of the relative incidence of events is not confounded by baseline differences in individual factors, such as socioeconomic status, underlying genetics, and baseline health status or comorbidities, which are difficult to accurately ascertain between exposure groups (e.g., vaccinated and unvaccinated individuals; patients on dialysis who do and do not acquire infections). Our work in this areas currently focuses on  extending the SCCS method to studies where the exposure onset time (e.g., infection time) is not known precisely. We refer to this as "exposure onset  measurement error."

 Publications

Development and Validation of Patient Outcome Risk Prediction

Risk prediction tools to inform patients, heath care providers, and researchers on the success of a treatment (e.g., surgery, kidney transplant, extra-corporeal membrane oxygenation [ECMO]) is an important component of the planning and initiation of treatment. Risk prediction tools, rigorously developed and validated, allow for more informed decision making at the individual patient level. Our group and collaborators have developed prediction tools for diverse patient populations, including patients on dialysis and neonates with congenital diaphragmatic hernia (CDH).

Publications

Regression Models for Joint Modeling of Disease Onset and Recurrence, Zero-Inflated Count Data

Cardiovascular disease remains one of the leading causes of hospitalization and death in the population of patients on dialysis. Our aim here is to develop methods to jointly model the relationship/association between covariates and (a) the probability of cardiovascular events (onset), a binary process and (b) the rate of events (recurrence) once the realization is positive - when the ‘hurdle’ is crossed - using a zero-truncated Poisson distribution. When the observation period or follow-up time, from the start of dialysis, varies among individuals the estimated probability of positive cardiovascular events during the study period will be biased. We develop strategies to eliminate this bias. In the context of zero-inflated count data, we are also developing functional linear model to be able to model functional predictors X(t) measured over time t, for instance.

Publications