QPRC 2016

A Multi-response Multilevel Model with Application in Nurse Care Coordination

Bing Si

Arizona State University 

Due to the aging of our society, hospitalized patients nowadays are likely to have multiple co-existing chronic illnesses. This has created a new challenge to patient care, because the care needs to be well coordinated within the health care team in order to effectively manage the overall health of a patient. Staff nurses, as the patient’s “ever-present” health care team members, play a vital role in the care coordination. Abundant research in the nursing society has shown that a greater amount of care coordination improves patient outcomes. However, little research is available to reveal the relationship between the care coordination activities conducted by nurses and their demographics and workload as well as the characteristics of their practice environment. Such research is important for nursing process improvement and designing of the best practices. The recently developed Nurse Care Coordination Instrument (NCCI) is the first of its kind that enables quantitative data to be collected to measure various aspects of nurse care coordination. Driven by this new development, we propose a multi-response multilevel model with joint fixed effect selection and joint random effect selection across multiple responses. This model is particularly suitable for modeling the unique data structure of the NCCI due to its ability of jointly modeling of multilevel predictors, including demographic and workload variables at the individual/nurse level and characteristics of the practice environment at the unit level, and multiple response variables that measure the key components of nurse care coordination. We develop a Block Coordinate Descent (BCD) algorithm integrated with an Expectation-Maximization (EM) framework for model estimation, and perform theoretical analysis to reveal the reason why the proposed model is able to outperform the existing multilevel method that models each response variable in separation. Asymptotic properties of the proposed model are derived. Simulation studies are performed. Finally, we present an application to a dataset collected across four U.S. hospitals using the NCCI and discuss implications of the findings.