In 1999 more than 90% of the 1.6 million people residing in the nation’s 18,000 nursing homes were older than 65 years, with a mean age of 85 years; most were single women, many had no living children, family, or friends, and nearly half (46%) were admitted from a hospital. 1–3 Many were cognitively impaired, and 75% required assistance with three or more activities of daily living (ADLs). 2–4
In 2001 nursing homes employed 150,680 RNs, 104,633 LPNs, and 575,981 certified nursing assistants (CNAs). 5 Public spending on nursing home care, mostly by state Medicaid programs, was $110.8 billion in 2003 and was projected to grow an average of 5.25% annually to $194.6 billion by 2014. 1, 6
Concern about the quality of care provided in hospitals and nursing homes has stimulated research and policy debate. 4, 7, 8 In hospitals, lower staffing levels have been associated with an increased risk of adverse outcomes, 9–11 including death. 12
But in nursing homes, where pain, pressure ulcers, malnutrition, and urinary incontinence are serious problems, evidence linking nurse staffing and outcomes is mixed. 4 Some studies report lower rates of adverse outcomes (for example, falls, pressure ulcers, and medication errors) associated with a higher percentage of RNs on staff and higher ratios of total nursing staff to residents. 4, 8, 13–16 Others have reported that the ratio of RNs to residents is not associated with adverse outcomes or mortality rates. 17, 18 These findings could indicate that the RN staffing levels and the amount of RN direct care time in the institutions studied were insufficient to make a difference in outcomes.
Despite these mixed findings, in 2004 the Institute of Medicine (IOM) recommended a staffing level in nursing homes of one RN for every 32 long-stay residents (45 minutes per resident per day, including time spent on administrative and managerial tasks and direct resident care). 19
Researchers and clinicians agree that pressure ulcer development is a nurse-sensitive outcome (a measurable outcome that is dependent on or effected by nursing interventions). 20 (For more on nurse-sensitive outcomes, see the National Database of Nursing Quality Indicators at www.nursingworld.org/quality/database.htm.) Despite guidelines designed to predict the likelihood and prevent the development of pressure ulcers, their incidence and prevalence remain high. 21, 22 They’re associated with higher rates of morbidity and mortality and poorer quality of life, 23 and they cost between $1.3 and $6.8 billion annually. Yet recent studies find that many nursing homes fail to provide within 48 hours of admission a plan of care or basic preventive measures that target pressure ulcers. 24
Most large epidemiologic studies have focused on resident characteristics, largely ignoring system factors such as an institution’s size, ownership (for profit or nonprofit), and staffing levels. 25 Spector, Horn, and others have considered the impact of nurse staffing on pressure ulcer prevention, but they didn’t explore this relationship in sufficient detail to determine the optimal amount of time RNs should spend in direct care to prevent adverse outcomes. 26–29 The purpose of the present study was to examine staffing levels and resident outcomes in sufficient detail to determine just that.
This study is a secondary analysis of a subset of data from the National Pressure Ulcer Long-Term Study (NPULS). Nurse staffing time was defined as the number of minutes (for RNs and LPNs) or hours (for CNAs) of direct care provided by RNs, LPNs, and CNAs per resident per day. The primary outcome variable was pressure ulcer development, but other adverse events in the database, including urinary tract infection (UTIs), weight loss, deterioration in ADLs, and hospitalization, were also assessed, as were two care processes (catheterization and administration of nutritional supplements) associated with pressure ulcer development. Details of the NPULS inclusion criteria, data collection methodology, and measurements of resident characteristics, treatment factors, and clinical outcomes are described in More on Methods and Statistics: National Pressure Ulcer Long-Term Care Study Database, page 66.
Resident selection in the study sample.
The existing NPULS database contained 1,524 residents from 95 facilities who didn’t have a pressure ulcer at study entry and for whom nurse staffing information was available. For a total of 148 residents at 13 facilities, RN direct care was 40 or more minutes (median, 81.3 minutes). We excluded these facilities and residents from our analyses because their characteristics differed significantly from long-stay nursing home residents who had less than 40 minutes per resident per day of RN direct care time. For example, the greater-than-or-equal-to-40-minute group remained in the NPULS study for significantly fewer days (65.5 versus 78.6 days, P < 0.001) and had significantly higher average scores on the Comprehensive Severity Index (CSI) in month 1 (81.8 versus 61.0, P < 0.001), significantly more enteral feeding (27.7% versus 15.1%, P < 0.001), significantly less hospitalization (2% versus 11.7%, P < 0.001), and a significantly higher death rate (10.1% versus 5%, P = 0.01). (For more on the CSI, see More on Methods and Statistics: National Pressure Ulcer Long-Term Care Study Database, page 66.) Therefore, to avoid skewing, the current study focuses on residents from 82 facilities in 19 states who received 40 minutes or less per resident per day of RN direct care and were thus more typical of long-stay nursing home residents. Data for the present analyses come from a subset (N = 1,376) of NPULS residents, all of whom had no pressure ulcer at study start. For details of statistical analyses see More on Methods and Statistics, page 66.
Resident demographics and outcomes.
The mean age of the 1,376 study residents was 81.6 years (SD ± 12.7 years, range = 19 to 104 years); 75.3% were female. The mean CSI score in month 1 was 61 (SD ± 37). The mean study duration for the whole resident population was 78.6 days (SD ± 15 days, median = 84 days [12 weeks]). See Figure 1 , page 60 , for the frequencies of selected resident outcomes.
Facility variables and nurse staffing.
Across 82 facilities, the mean RN time per resident per day was 16 minutes (median = 15.1, range = 0.7 to 36.9), the mean LPN time per resident per day was 30.6 minutes (median = 31.5, range = 0 to 122.0), and the mean CNA time per resident per day was 1.7 hours (median = 1.9, range = 0.04 to 6.1). Facility ownership status was not significantly associated with nurse staffing levels. Nursing time per resident per day by for-profit and nonprofit status, respectively, was 15.1 versus 19 minutes for RNs (P = 0.08), 31.9 versus 26.6 minutes for LPNs (P = 0.26), and 1.7 versus 2 hours for CNAs (P = 0.14).
Nurse staffing and resident outcomes.
For each outcome we found either a trend of decreasing percentage of residents having complications, such as pressure ulcer development, weight loss, and hospitalization, with each 10-minute increase in RN direct care time—with the lowest complication rates for 30 to 40 minutes per resident per day—or a threshold effect, with better outcomes with 30 to 40 minutes per resident per day, as compared with less care. Thirty to 40 minutes of direct care per resident per day was associated with fewer UTIs and catheterizations, less deterioration in the ability to perform ADLs, and more use of nutritional supplements. (Figure 2, at right).
We examined numerous possible ranges for CNA time per resident per day. Only one outcome, the development of pressure ulcers, was significantly associated with CNA time (χ2 = 17.8, P < 0.001): 16% of the residents who had a CNA time of 2.25 hours or more per day (median = 3.1 hours) developed at least one pressure ulcer, as did 23% of those who had a CNA time of 2 to 2.24 hours per day (median = 2.1 hours) and 32% of those who had less than 2 hours per day (median = 1.8 hours).
Two time ranges for LPN time per resident per day produced the only difference in outcomes: less than 45 minutes (median = 29.3 minutes) and 45 minutes or more (median = 51.8 minutes). More LPN time was associated with fewer pressure ulcers (χ2 = 7.22, P = 0.007). However, more LPN time was also associated with more deterioration in the ability to perform ADLs (χ2 = 22.6, P < 0.001) and greater catheter use (χ2 = 7.61, P = 0.006) in bivariate analyses, but did not remain in multivariate analyses when other variables were controlled for.
Correlations among variables.
RN, LPN, and CNA times were not highly correlated with study duration, CSI score in month 1, or resident age (−0.18 < r < 0.17), even though a few correlations were statistically significant. The correlation of RN time with CSI score was r = −0.04 (P = 0.10); the correlation of RN time with study duration was r = 0.10 (P < 0.001). The correlation of RN time with CNA time was r = 0.238 (P = 0.033), the correlation of RN time with LPN time was r = −0.037 (P = 0.745), and the correlation of LPN time with CNA time was r = 0.563 (P < 0.001). Interaction terms between staff time variables and also between staff time and other resident, treatment, and facility variables were allowed to enter models but none was significant.
Figure 3 (page 64) presents significant predictors of pressure ulcer development identified through logistic regression. The largest negative coefficient and smallest odds ratio (OR) of 0.16 was for RN direct care of 30 to 40 minutes per resident per day. More CNA and LPN time was significantly associated with less likelihood of developing pressure ulcers, but RN time was a stronger predictor.
When ownership status was allowed to enter this model, it was significant (χ2 = 13.05, P < 0.001, OR = 0.54, confidence interval [CI] = 0.39 to 0.76). Nonprofit status was associated with less likelihood of developing pressure ulcers. When ownership status entered the model, several variables in Figure 3 became nonsignificant (P > 0.05); they were high-calorie and high-protein enteral nutrition, disease-specific enteral nutrition, fluid order, RN direct care time of 10 to 20 minutes, and all of the CNA timing variables.
Deterioration of the ability to perform ADLs (Figure 4, page 65) was positively associated with dehydration, severity of illness in month 1, new admission, and age older than 85 years. RN direct care time of 30 to 40 minutes per resident per day (P = 0.046, OR = 0.58, CI = 0.34 to 0.99) was the strongest predictor of less deterioration in the ability to perform ADLs after forcing all other nurse staffing variables into the model. When only significant nurse staffing variables were in the most parsimonious model, RN direct care time of 30 to 40 minutes per resident per day was significant (P = 0.005, OR = 0.52, CI = 0.32 to 0.82). When ownership status was allowed to enter this model, it was significant (P = 0.02, OR = 0.67, CI = 0.47 to 0.94). Nonprofit status was associated with less likelihood of decline in the ability to perform ADLs. When ownership status entered, no variables became nonsignificant and no new ones entered the model.
The strongest predictor of hospitalization was higher severity of illness in month 1 (Wald χ2 = 57.32). Catheter use and longer study duration also were significantly associated with increased likelihood of hospitalization. Again, RN direct care time of 30 to 40 minutes per resident per day was associated with the least likelihood of hospitalization, and ownership status was not significant.
The strongest predictor of UTI was catheter use (P < 0.001, OR = 2.45, CI = 1.69 to 3.53). In Figure 2, page 61, RN direct care time of 30 to 40 minutes was significantly associated with less likelihood of UTI (P = 0.009); however, it was also associated with significantly fewer catheterizations (P = 0.011). After adjusting for these and other variables, RN direct care time of 30 to 40 minutes per resident per day was nearly but not quite significantly associated with less likelihood of developing UTI (P = 0.055, OR = 0.60, CI = 0.36 to 1.01).
All variables found to be significant in logistic regressions remained significant when either hierarchical or Cox models were applied. RN direct care time of 30 to 40 minutes per resident per day was always among the most significant predictors of better outcomes.
This study provides a comprehensive assessment of associations of average nursing time with five adverse clinical outcomes (development of pressure ulcers and UTIs, deterioration of ability to perform ADLs, weight loss, and hospitalization) and two care processes (catheterization and administration of nutritional supplements) after controlling for physical limitations; medical diagnosis; severity of illness; and other resident, treatment, and facility factors for long-stay, at-risk nursing home residents. In each analysis, RN staffing of 30 to 40 minutes per resident per day was strongly associated with better outcomes, and that association was much stronger than those for LPN and CNA times.
We first put the time per resident per day into the models as a continuous variable. For the outcome of pressure ulcer development, RN, LPN, and CNA time was significant. For the other outcomes, only RN time was significant. However, knowing only that more time is better wouldn’t be that helpful to nursing homes and policy makers. Therefore, we looked at 5-minute increments for each staff type and rates of outcomes. The outcomes were flat, having little variation, except where we ultimately chose to break the data as described: 10-minute increments for RNs, one break at 45 minutes for LPNs, and two breaks for CNAs.
Severity of illness was strongly associated with all outcomes: sicker residents were more likely to develop pressure ulcers, experience decline in the ability to perform ADLs, be hospitalized, have weight loss, and die. Another strong resident factor was being newly admitted. Newly admitted residents remained in the study for fewer days (71.2 versus 80.8 days, P < 0.001), which may partially account for why fewer newly admitted residents developed pressure ulcers. Although severity of illness and new admission status are resident factors and not directly controllable by nursing home management, nurse staffing is directly controllable by management. After taking severity of illness; new admission status; study duration; and many other patient, facility, and treatment variables into account, we found strong and consistent associations between the average time RNs provided direct nursing care and pressure ulcer development, weight loss, deterioration in the ability to perform ADLs, and hospitalization, and weaker association with the development of UTI. More RN time was also associated with reduced use of catheters and increased resident use of standard medical nutritional supplements, which we’ve found in prior analyses to be associated with less likelihood of developing pressure ulcers. 28 The greatest reductions in adverse outcomes and improvements in care processes resulted when RNs spent 30 to 40 minutes per resident per day on direct resident care.
The national average for U.S. nursing homes reported in the 1997 Online Survey, Certification and Reporting System was 48 minutes of RN time, but these data included director of nursing time, nurse manager time, and nurse administrator time, whether it was spent in direct care or not. 32 We excluded director of nursing and nurse manager time because we were interested in associations with direct resident care times. The Center for Medicare and Medicaid Services (CMS) conducted a study in 2001 that found that, for long-stay nursing home residents, total licensed nurse staffing time should be 1.3 hours for RNs and LPNs combined, 0.57 hours for RNs, and at least 2.8 hours for CNAs. 33 Our findings are close: we found better outcomes with RN time of 30 to 40 minutes and LPN time greater than 45 minutes, with median times of 34 minutes for RNs and 52 minutes for LPNs (for a total of 86 minutes, or 1.43 hours). We also found that a CNA time of 2.25 hours or longer per resident per day was associated with a lower incidence of pressure ulcers.
The lack of strong, consistent relationships between the direct care time provided by LPNs and CNAs and adverse outcomes highlights the crucial role RNs play in the quality of care in nursing homes and how important it is that the time they spend in direct care of residents be sufficient. A recent study of more than 5 million discharge abstracts of hospitalized patients also found that relationships between nurse staffing and adverse patient outcomes are much stronger for RNs than for other nursing personnel. 9 We hypothesize that nurses influence quality in a facility by providing expertise in direct care and evaluation. Although LPNs and CNAs are important members of the nursing team in hospitals and nursing homes, the convergence of the results of our study with those in the study by Needleman and colleagues is further evidence that workforce policy should focus on increasing the proportion of RNs providing direct resident care in nursing homes, a position advocated by IOM reports. 4, 8, 19 The 1996 IOM report Nurse Staffing in Hospitals and Nursing Homes: Is It Adequate? recommended that by 2000, 24-hour presence of RNs in nursing facilities be required (under current law, the Omnibus Reconciliation Act of 1987, of which the Nursing Home Reform Act is a part, only eight hours of RN coverage is required). 8 The 2001 IOM report Improving the Quality of Long-Term Care recommended that the CMS implement earlier IOM recommendations to increase RN coverage in nursing homes. 4 The 2004 IOM report Keeping Patients Safe: Transforming the Work Environment of Nurses recommended that RN time be increased to 45 minutes per resident per day, consistent with the threshold for better outcomes in this study. 19 Our findings demonstrate consistently that this level of direct care is related to improvement in many of the key safety-related outcomes, supporting the IOM staffing recommendation.
The goal of increasing the proportion of RNs in nursing homes will require significant financial commitment from payers. Grabowski reports that, contrary to prior studies, increasing Medicaid reimbursement improves nursing home quality, and the probable mechanism for achieving this improvement is better staffing, particularly RN staffing. 29 However, the current RN shortage and much larger shortage that is expected as RNs age and retire in the future makes the task of increasing RN staffing challenging. 34
Considering the large number of residents affected by these outcomes, estimates of decreased rates of adverse outcomes associated with more RN minutes are clinically important. While they were being studied, one out of three residents experienced a deterioration in the ability to perform ADLs; more than one-quarter developed a pressure ulcer or experienced weight loss. Between 10% and 20% were hospitalized, developed a UTI, were catheterized, or had some combination of these outcomes; 5% died. These outcomes increase facility expenditures, resource use, and spending on liability-related matters, and they cause residents to experience pain and suffering and loved ones to experience emotional distress. Our findings suggest a potentially greater risk of nursing home residents suffering avoidable adverse outcomes if nothing is done to improve RN staffing levels, and nursing homes and society incurring greater costs as a result of such outcomes. Dorr and colleagues found that there would be a savings to society of almost $3,200 per at-risk resident per year if RN direct care time were increased from less than 10 minutes per resident per day to 30 to 40 minutes per resident per day. 35
It’s important to note several limitations to this study. First, this was an observational study, not an interventional study. Noninterventional studies can only show associations that reflect assumptions about causality, unlike randomized, controlled trials, which can demonstrate causality with more certainty. Second, nurse staffing data were not resident specific; we didn’t conduct time and motion studies to determine the exact number of minutes each type of nurse spent performing specific interventions for each resident. Third, the time spent on specific nursing activities was not documented. Also, we didn’t include facilities with an average RN direct care time of 40 minutes or longer per resident per day, for the reasons specified above. Future studies should look at facilities with the highest levels of staffing to determine whether the relationships found in our study continue to hold. Although this study identifies associations between nurse staffing and patient outcomes, these associations provide directions for further research and practice changes.
Although the original data from NPULS were collected in the mid 1990s, the results are relevant to today’s nursing home environment. We are implementing these findings in nursing homes today, and the same factors correlate comparably with outcomes.
Studies that determine the effectiveness of processes of care performed in response to specific risk factors are needed to inform practice. We also need to continue to develop models of care; better documentation systems that support practice decisions; and health information systems that guide, document, and evaluate practice. However, these analyses suggest that, in addition, we need adequate RN staffing if we are going to improve the nursing home experience and resident outcomes.
MORE ON METHODS AND STATISTICS
National Pressure Ulcer Long-Term Care Study Database
The National Pressure Ulcer Long-Term Care Study (NPULS) was an observational retrospective cohort study designed to provide data for improving practice. During the study period in 1996 and 1997, we collected comprehensive data from resident charts in 109 long-term care facilities in 23 states. 27, 28, 30 Facilities owned by six providers volunteered to participate; 76% (83 facilities) were for profit, with an average daily census of 120 residents (median = 114, range = 32 to 300) during the study.
NPULS inclusion criteria.
To qualify for the NPULS, residents had to be 18 years old or older, have been a resident of the facility for 14 days or longer, and be at risk for developing pressure ulcers (defined as a score of 17 or less on the Braden Scale for Predicting Pressure Sore Risk [Braden Scale] 31) or have had a pressure ulcer at the start of the study.
No resident identifiers were included in the database, except those permitted by the privacy rule established under the Health Insurance Portability and Accountability Act of 1996. This research complied with the rules for human experimentation as outlined in the Declaration of Helsinki and was approved as qualifying for expedited review status by the institutional review boards.
NPULS data collection.
We trained 19 data abstractors (employed by the study facilities) to use the data collection instrument and the Comprehensive Severity Index (CSI) software system (for more on CSI, see below) and measured their reliability by comparison with experienced trainers; this ensured the accuracy of data collection. The criterion for accuracy of the abstractor’s data collection (interrater reliability) was agreement with the trainer on at least 90% of the measurements for each resident; if this was not achieved, the abstractor received additional training until 90% agreement was achieved.
For each resident, we retrospectively reviewed 12 consecutive weeks of medical records, data from the long-term care Minimum Data Set (MDS; an assessment tool used by all Medicare-certified long-term care facilities), and other written records such as physician orders and medication logs. If the resident was discharged during the 12 weeks under review, records for the actual length of stay were included. For newly admitted residents, the review of records began on the resident’s admission date (the study start date) and continued for 12 weeks. For existing residents with a pressure ulcer, the review included the four weeks before its identification and extended for eight weeks afterward. For existing residents without a pressure ulcer, the review included records beginning when residents were identified as being at risk (having a Braden Scale score of 17 or less) and extended for 12 weeks afterward. All study periods fell between February 1996 and October 1997.
A multidisciplinary panel created a database designed to assess relationships among resident characteristics, resident treatment characteristics, facility variables, and clinical outcomes. The determination of which data elements were to be included in the study was made on the basis of clinical expertise, published literature, and pressure ulcer guidelines. 21, 22 (NPULS data collection instruments and study methodology are described in greater detail elsewhere. 27, 28)
Resident characteristics included demographics, medical history, severity of illness measured by CSI scores, Braden Scale scores, nutritional assessments, study duration (the number of days that the resident was in the facility during the respective 12-week study period), and activities of daily living (ADLs; defined using sections G.1 and G.2 from the MDS). Dependency was defined as being totally dependent or requiring extensive assistance in at least seven of the following eight ADLs: bed mobility, transfer, locomotion on the unit, dressing, eating, toilet use, personal hygiene, and bathing. The threshold of seven or more ADLs was found to be statistically significant in previous analyses we conducted. 28
Resident treatment characteristics included medications, nutritional interventions, pressure ulcer management strategies, and incontinence interventions. Treatment with “nutritional supplement” means the use of oral medical nutritional products at any time during the study period. Chart data were used to determine whether catheterization had occurred.
Facility variables included nurse staffing time, ownership status, existence of skin-care teams, use of outside skin-care consultants, use of in-house nurses who floated to understaffed units, and hours of dietitian time per month. Nurse staffing time was reported by each facility as the number of hours of direct care provided to residents by RNs, LPNs, and CNAs per 24-hour period for each month of the study, as was resident census per month for the same time periods. Time spent on managerial and administrative tasks was not included in RN direct care time. For each facility we computed average monthly RN, LPN, and CNA minutes or hours per resident per day for each month during the whole study period. These numbers had almost no variation so we didn’t use the monthly averages but instead used the aver-age over the whole study period.
Resident outcomes included development of pressure ulcer (stages I through IV), urinary tract infection (UTI), or weight loss; deterioration in the ability to perform ADLs; and hospitalization. Pressure ulcer documentation was abstracted from medical records; ulcers smaller than 0.25 cm2 were eliminated. A pressure ulcer was considered new if the first pressure ulcer assessment or treatment occurred 15 days or longer after the study began.
UTI was defined according to the criteria of the International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM; code 599.0). Weight loss was defined according to MDS section K (oral–nutritional status), item 3a (“Weight loss 5% or more in last 30 days; or 10% or more in last 180 days”), or from recorded chart weights showing a 5% or greater weight loss in study month 1 ([weight obtained closest to the end of the first month minus the first weight] divided by the first weight) or a weight loss of 10 lbs. or more in three months (the last weight obtained minus the first weight).
Deterioration of the ability to perform ADLs was defined according to section G.9 of the MDS, in which such deterioration, as compared to the patient’s status 90 days previously or since the most recent assessment, if sooner, is documented. Hospitalization dates and deaths were abstracted from resident records.
Severity of illness. The CSI was used to adjust for each resident’s severity of illness. 27, 28, 30 CSI scores are based on the degree of abnormality of over 2,100 possible signs, symptoms, and physical findings (for example, blood pressure, pulse, temperature, confusion, dyspnea, dysphagia) specific to a resident’s disease or diseases, rather than diagnostic information (ICD-9-CM coding) alone. In general, scores of less than 16 are considered low severity; those greater than 60 are considered high severity. More abnormal findings produce higher CSI scores. Three monthly CSI scores—one for each month of the study—were computed, using the most abnormal value found for each criterion during each month. For example, a resident might have the highest temperature one day, the highest blood pressure on another day, and so on; each of these most-abnormal values were used to calculate the monthly severity score.
In our study of RN staffing time and outcomes of long-stay nursing home residents, we conducted bivariate analyses using χ2 tests to examine associations between increments of nursing time and resident outcomes and between increments of nursing time and the use of treatments associated with pressure ulcer prevention. Logistic regression analyses were used to determine associations among resident, treatment, and facility characteristics and outcomes. Multicollinearity and goodness of fit were assessed. The importance of each predictor was determined by coefficient magnitude, odds ratios, and 95% confidence intervals.
In a stepwise procedure, variables were allowed to enter the model if P < 0.1 and then were dropped from the model if P > 0.05, to adjust for potentially confounding variables. After automatic variable selection, all nurse staffing variables were forced into the model regardless of P value. Variables that were allowed to enter but were not significant (at P ≤ 0.05) are listed in the figure captions (see Figures 3 and 4, page 64 and page 65, respectively).
Discrimination c statistic was assessed using the area under the receiver-operating-characteristic curve. The Hosmer–Lemeshow goodness-of-fit test was used to evaluate calibration. A nonsignificant P value indicates that the estimated probabilities weren’t significantly different from actual rates of occurrence and is further indication that the model was a good fit.
Facility-level measures, including staff time, were the same for all residents in a facility. Hierarchical models to account for clustering by facility were constructed and compared to the nonhierarchical models. Based on hierarchical and nonhierarchical analyses having the same significant variables, we elected to present findings from nonhierarchical models with resident as the unit of analysis (for more, see Hierarchical Models, page 68). Variables such as staff time and other facility-level measures were allowed to enter models. We also computed Cox proportional hazards regressions to evaluate the robustness of the findings. All analyses were performed with SAS statistical software (version 8.2, SAS Institute).
The article by Horn and colleagues notes that hierarchical models were used to account for “clustering by facility” (See More on Methods: Statistical Analyses, page 67). What does this mean and why is it important in such studies?
In nursing and health care research, it can be difficult to isolate a few key variables in order to understand their effects on patient outcomes. For example, a researcher may ask “Why do people in nursing homes get pressure ulcers?” As Horn points out, many factors contribute to pressure ulcer development. Multivariate analysis is a statistical method for assessing which factors (independent variables) influence specific resident outcomes (dependent variables). In this case, the authors examined several independent variables (staffing time, facility characteristics, and history of pressure ulcers, among others) and several dependent variables (development of pressure ulcers, weight loss, and ability to perform activities of daily living, among others).
Hierarchical analysis is a type of multivariate analysis that addresses two additional issues. One is the clustering mentioned by Horn and colleagues. Some observations are clustered—another term is nested—within others, creating a possible correlation among participants. Consider an example from educational research, a field in which hierarchical analysis has a longer history than it does in health care research (this example is taken from the Raudenbush text, below). In any school system, many individual students are “nested” within a single class, and many classes are “nested” within a single school. In addition to their individual differences, students in the same class may share other factors in common—for example, the influence of a particular teacher—that affects their test scores. Therefore, when comparing test scores across classes, we want to know the degree of this within-class correlation. This correlation within a group of study subjects is tested using the intraclass correlation coefficient (ICC), a measure of the degree of resemblance of smaller units nested within larger ones. The magnitude of the ICC influences estimates of variance and tests of significance. (Other statistical effects of this nesting aren’t explored here, but readers who are interested in a more thorough statistical discussion can refer to the texts listed below.)
The second issue is, what do the additional layers—or levels, in hierarchical terminology—contribute to our understanding of the research question? We have to make at least two assumptions about these “higher” levels (in our example, the class and school levels, as opposed to the individual student). First, results across classes or schools aren’t necessarily a simple aggregate of individual students’ results. Second, class- and school-level factors may contribute to an explanation of the outcome of interest—for example, variations in grade-point averages in a particular class or school, above what individual student characteristics may contribute (as would be found, for example, if teachers assigned grades using different standards).
As Raudenbush suggests, once you become aware of the issue, you notice that hierarchical (multilevel) problems are everywhere. In health care research, we often encounter multilevel problems: patients nested within clinicians nested within primary care practices or, as another example, residents nested within units nested within nursing homes. Hierarchical analyses and models are used because nonhierachical techniques, such as logistic regression, don’t take nesting or ICC into account.
Consequently, Horn and colleagues employed a hierarchical analysis to “take into account” the possible ICC among nursing home patients. Facility variables, such as nurse staffing time and ownership status, were considered as second-level variables to see whether they contributed to resident outcomes over and above the effect of the first-level variables. They compared these results to those of (nonhierarchical) logistic regression analysis. By indicating that the hierarchical analysis yielded no statistically significant differences from the nonhierarchical analysis, the authors suggest that the ICC is negligible, that there’s not necessarily a second-level contribution, and that the study findings could be accurately presented using logistic regression.
Because the dependent variable in the logistic regression is binary—for example, residents either did or did not develop pressure ulcers—the coefficient and, more importantly, the odds ratio tells the reader whether, after adjusting for the presence of other variable s, a given independent variable significantly increases or decreases the odds of the dependent variable occurring—for example, whether RN direct care time increases or decreases the odds of pressure ulcer development, as shown in Figure 3 (page 64). A statistically signficant odds ratio greater than 1 indicates an increase in the odds of the dependent variable occurring, while a statistically significant odds ratio less than 1 indicates a decrease in those odds. —Arthur E. Blank, PhD, codirector, Division of Research, and director, Center for the Evaluation of Health Programs, Department of Family Medicine and Social Medicine, Albert Einstein Medical College–Montefiore Medical Center, Bronx, NY
Goldstein H. Multilevel statistical models. 3rd ed. London: E. Arnold; 2003.
Hox JJ. Multilevel analysis: techniques and applications. Mahwah, NJ: Lawrence Erlbaum Associates; 2002.
Kreft IGG, de Leeuw J. Introducing multilevel modeling. Thousand Oaks, CA: Sage Publications; 1998.
Raudenbush SW, Bryk AS. Hierarchical linear models: applications and data analysis methods. 2nd ed. Thousand Oaks, CA: Sage Publications; 2002.
Snijders TAB, Bosker RJ. Multilevel analysis: an introduction to basic and advanced multilevel modeling. Thousand Oaks, CA: Sage Publications; 1999.