Multilevel and structural equation models in health services research: application to research on the organizational context of nursing care

The past two decades have witnessed a growing b ody of US research evidence supporting contentions that hospitals with better nurse staffing, h igher proportions of bachelor-educated nurses and superior nurse practice environments hav e better outcomes of patient mortality, patient sa tisfaction with care, and nurse wellbeing. Mo re recently, the Registered Nurse Forecasting (RN4 CAST) research consortium confirmed the direc tion and consistency of these associations across¨ European hospitals at a pan-European scale. This P hD study capitalizes on the RN4CAST study. It enco urages the use of contemporary statistical te chniques by refining our understanding of the mech anisms underlying established associations, b y studying phenomena at multiple organizational le vels, and by promoting an evaluation of similarity ¨of concepts across groups. Two important issues a re tackled. A first issue is a lack of¨ alignment between statistical techniques and¨ the many comprehensive theories that have been pro posed to explain why inputs of nursing are associa ted with nurse and patient outcomes. A first challenge in the statistical analysis is to a ccount for the complex hierarchical nature of the¨ data and theories. A hierarchical data struct ure is also called a multilevel data structur e. In the RN4CAST study, a complex four-level ¨data structure occurred because the study design¨ corresponds with the clustering of nurses and ¨patients (level 1) within nursing units (level 2) ¨within hospitals (level 3) within countries¨ (level 4). This complex data structure requires th e use of multilevel modeling techniques. A second¨ complexity arises from the theoretical specificati on of mediator and moderator variables. Mediators¨ illuminate the mechanisms that explain why as sociations can be observed. Moderators are context ual factors that define the conditions under which ¨such associations can be predicted. A third chall enge relates to understanding underlying dime nsions, also called latent variables, from a set o f survey items. Hypotheses involving latent variables require a statistical technique called st ructural equation modeling. The analysis of m ediators, moderators and latent variables must be performed in a multilevel framework to comply ¨with suggested theoretical models. A s econd issue relates to a lack of measurement invariance evaluation in the cross-group compariso ns that are so important to this field of res earch. Measurement invariance pertains to whether measurement instruments and the constructs be ing measured operate in the same way across popula tions of interest. Lack of measurement invari ance can result in the incorrect interpretation of ¨what the researcher believed were true¨ cross-group differences. Measurement invariance is ¨studied in a latent variable framework and,¨ like mediation and moderation analysis, relie s on the use of structural equation models. In a first introductory chapter, we provide an ove rview of the patient safety literature and th eoretical models and landmark studies in this fiel d of research. It concludes with the research ¨objectives of this PhD study. A second introducto ry chapter provides an overview of the R N4CAST study design and gives a concise intro duction to the fundamental principles of the two m ain statistical techniques that are used in the Ph D study: multilevel modeling and structural equati on modeling. The next part of th is PhD study is organized into four thematic¨ chapters that all extend on RN4CAST data or s tudy concepts. The first core chapter shows that a ¨high proportion of nurses working in Europea n hospitals reported having performed tasks below¨ their skill level during their last shift, which i s a newly theorized concept in the evaluation of nurses&##39; work structure. In line with¨ anecdotal evidence, we confirm a pronounced overal l effect of being a foreign trained nurse from a d eveloping country and an increase in reports of tasks below skill level. In the second core cha pter, we hypothesize and confirm that nursing work environment components exhibit different effects on nurse outcomes at different systems l evels. These new linkages have rich potential for¨ policy makers and hospital human resources manager s, as it may provide them with empirically-ba sed strategies in designing level-specific interve ntions to improve nurse wellbeing and patient ¨safety. In a third core chapter, we show that nec essary nursing care left undone, another only¨recently theorized concept, is associated wi th a poor nursing work environment, performing mor e tasks below skill level, and a complex inte rplay between nurse staffing and nurses&##39; education levels. We also showed that nursing ¨care left undone explains the association be tween these components of nursing organization of¨ care and patient satisfaction. This had recen tly been theorized, but was not evaluated statisti cally. In the fourth and last thematic chapter, we ¨evaluate measurement invariance of our work¨ environment survey items in Belgium. This study ev idenced that some items are not invariant across D utch and French speaking nurses. This study also a llowed us to conclude that nursing unit manag ers, when compared to staff nurses, hold more positive views of several work environment dimens ions. In a concluding chapter, we argue that the mounting evidence linking inputs of ¨nursing to quality and patient safety allows huma n resources managers and policy makers to identify ¨the impact of different systems policies of¨ nursing organization on outcomes. It should at the ¨very least trigger a critical reflection on good practices and on the transferability, impleme ntation and evaluation of investments in nurse edu cation, work environments and staffing levels.status: publishe