Methods
Sampling
Hospital Units. This study was conducted in 22 acute medicine units in 11 hospitals in Quebec, Canada. Units were selected to generate a stratified sample covering a variety of organizational contexts of nursing care, based on predefined criteria and informed by a survey sent to all Quebec hospitals (50 out of 100 institutions responded). Diversity of institutions was sought on the following criteria: institutional teaching status (university and community), size, location (urban, suburban and rural), nursing workforce profiles (different proportions of nurses holding university degrees) and work reorganization track records (stable structure with no recent modification, recent work reorganization initiatives such as introducing new staff categories and enhancing nurses' roles).
Patients. Patients on the 22 units were selected based on 4 criteria: (i) hospitalizations of at least 48 h, (ii) age 18 years and older, (iii) admission diagnoses typical of care provided on medicine units and (iv) hospitalizations overlapping with a concurrent nurse survey to characterize nursing care delivery models on the units. The observation period was restricted to the first 30 days of the selected patients' hospitalization, to exclude long-stay patients and increase homogeneity of the patient sample.
Assuming a 3% potential rate of selected events based on the literature, we calculated that a sample of 2600 patients would be required to achieve a 2.5% margin of error in point estimates of risk, based on a 0.05 significance level. The final sample totaled 2699 patients, varying from 117 to 128 per unit.
Nursing Care Organization Model (Independent Variable)
The independent measure was a four-category variable representing the nursing care organization models. Cluster analysis of data from the 22 units elicited 4 nursing care organization models with considerable face validity (see Fig. 1). The unit types clustered along two axes, one related to overall staffing intensity and the second, to the proportion of more educated nurses and the quality of the professional practice environment. Two models were variations on a professional model and two others, on a functional model (see Table 1 for details).
(Enlarge Image)
Figure 1.
Four nursing care organization models
Professional Models of Nursing Care Organization. The two professional models reflect managerial decisions that recognize nursing as a professional discipline. These models employ more nursing workers with higher formal education and have professional governance structures supporting the efforts of these knowledge workers. As such, these models are characterized by a higher proportion of care hours provided by RNs and by nurses' perception of greater support for their professional practice. Table 1 describes the two professional models' distinctive features.
Functional Models of Nursing Care Organization. The functional models reflect a view of nursing as a broad set of tasks that can be carried out by a variety of workers, presumably in response to factors such as economic and labour-market constraints. As such, these models draw more on less educated staff, including licensed practical nurses (LPNs) and unregulated assistive staff, to deliver nursing services than do the professional models. They are characterized by a lower proportion of care hours provided by RNs, and by nurses' perception that the practice environment is less supportive of a 'professionalized' approach to RNs' work. Table 1 describes the two functional models' distinctive features. The term 'adaptive' refers to the use of both LPNs and RNs to wider scope of practice relative to other units.
Patient Safety Outcomes (Dependent Variables)
Based on the literature, we selected six patient outcomes (medication administration errors, falls, pneumonia, urinary tract infection, unjustified restraints and pressure ulcers) for study. They were identified from abstraction of each patient's medical record for the stay, using a standardized protocol adapted from earlier work. Each event's severity was rated according to its consequences for the patient, using a standardized algorithm. Events were sorted into those 'without' consequences (which had potential for harm and may or may not have required intervention or follow-up, but did not cause lasting clinically detectable harm) and those 'with' consequences (causing a temporary or permanent change in the patient's condition and requiring an intervention, treatment or extended hospitalization). These categories transcend more restrictive definitions of adverse events, taking into account situations at different points on the safety continuum in terms of impacts on patients.
Patient-level Control Variables
We used two approaches to address baseline differences in patient risk for events. First, we applied strict inclusion rules to increase homogeneity of the patient pool. Second, we included four indicators as control variables in our regression model to capture severity of conditions and presumed risk for negative outcomes: age-adjusted Charlson Comorbidity Index [(CCI), in which comorbid conditions are scored, weighted and totaled, with points added for age], number of risk factors (including alcoholism, smoking, drug addiction, obesity, cognitive disorders such as Alzheimer's disease, mental health problems such as depression or schizophrenia and illiteracy), length of stay and number of diagnoses at admission.
Data Collection
We screened patients' records retrospectively to detect occurrences, after admission, of the six safety-related events. Eligible patients were tracked for outcomes until transfer, hospital discharge, death or the end of a 30-day period. Using a template drawing upon previous studies, three experienced nurses screened records post-discharge. Reviewers underwent 4-h training sessions and received a training manual.
The protocol involved a two-stage review. First, each record was screened by one reviewer to check for the occurrence of at least one event by systematically examining incident reports, discharge summaries, medications, lab results, nursing and physician notes or comments. In this first-stage review, data were also collected on patients' demographics (age and sex), conditions (main diagnoses at admission, comorbidities and risk factors) and length of stay. To assess inter-rater reliability, the first seven records of each unit were independently examined by a second reviewer. This quality control was conducted on 6% of records that is in the general range of rates ranging from 1 to 5% in other studies from the literature. The Kappa coefficient for inter-rater agreement for this first stage of the review process was 0.98.
If one or more safety-related events were identified in the first-stage screening, the record was reviewed again, more thoroughly and independently. In that review, assessors also rated each event's severity. The Kappa coefficient for measurement of agreement for the second review was 0.97. In cases of disagreement and discrepancies, the two reviewers reached mutual agreement after discussion.
Statistical Analysis
The data were first summarized using descriptive statistics and frequency tables. Prior to logistic regression modeling, bivariate analyses were conducted to examine the associations among the variables, identify potential confounding variables and detect possible problems with multicollinearity (using a P-value threshold of 5%).
Two composite outcomes were constructed: a binary variable for each patient indicating occurrence of any of the six events, with or without consequences, and a second indicating whether the patient had experienced any events with consequences. Subsequently, regression modeling was used to assess associations between the four nursing care organization models (independent variables) and both dependent variables. We used binary logistic regression with adjustments for patients' characteristics, including age-adjusted CCI, number of risk factors, number of diagnoses at admission and length of stay. The Hosmer–Lemeshow test and DFBETAS analysis were used to assess goodness-of-fit of the final models.