We estimate the relationship between measures of levels of nurse-to-patient staffing, organizational aspects of the nurse work environment, and nurse education-- three hospital-level measures derived from the nurse surveys-- and risk-adjusted 30-day inpatient mortality and failure to rescue across 665 adult acute care general hospitals in California (n = 271), Pennsylvania (n = 153), Florida (n = 168) and New Jersey (n = 73). These are four of the nation’s largest states, and account for over 20% of annual hospitalizations. The nurse survey data were collected in California, Pennsylvania and New Jersey between September 2005 and August 2006 and in Florida between November 2007 and April 2008. Patient discharge data from 2005–2006 and American Hospital Association (AHA) Annual Survey data from 2005 were linked to the nurse survey data for CA, NJ, and PA and patient discharge data from 2006–2007 and AHA data from 2007 were linked to the nurse survey data for Florida, using common hospital identifiers. Patient discharge data were obtained from the Office of Statewide Healthcare Planning and Development in CA, The Agency for Health Care Administration in FL, the Department of Health and Senior Services in NJ, and the Pennsylvania Health Care Cost Containment Council in PA.
The units of analysis in the study are 665 hospitals, but the units of observation are variously hospitals, patients, and nurses; and the statistical modeling is with reference to a hierarchical model in which patients are nested within hospitals. The hospitals included in our sample represent 86% of all general acute hospitals in the four states and account for over 90% of all adult general, vascular, and orthopedic surgical patient discharges in those states. Hospitals not analyzed are primarily small hospitals with fewer than 10 nurse respondents, which we regard as too few to provide reliable estimates of the hospital-level nursing characteristics of interest.
Adjustments in our models for differences in patient outcomes due to hospital characteristics not related to nursing are made using measures of hospital size, teaching status, and technology from the 2006 American Hospital Association Annual Survey. Three size categories (≤ 100 beds, 101–250 beds, ≥ 251 beds) were used. Teaching status was defined by number of medical residents and fellows (non-teaching without any residents/fellows; minor teaching with 1:4 or smaller trainee to bed ratios; major teaching with ratios higher than 1:4). Hospitals were designated as high technology if they had facilities for open-heart surgery, major organ transplants, or both.
The three key predictor variables — nurse staffing, nurse education and the nurse work environment — are hospital-level measures. They are averages of reports from large random samples of registered nurses from state licensure lists who identified themselves as working in one of our study hospitals. The strategy of measuring organizational features of hospitals by aggregating nurse-specific reports is derived from the sociology of organizations research literature20–29
and has been widely used in research reports on nursing outcomes.14,15,19,30,31
This method of measuring organizational features of hospitals is at least as accurate, and probably considerably more accurate, than reports by a single “key informant” within a hospital,32,33
and avoids the problem of hospital-level non-response bias where hospital officials may choose not to participate on the basis of the nursing features being studied.34
We obtained mail responses from over 100,000 nurses in the four states, including 39,038 staff nurses working in our study hospitals, for an average of approximately 60 nurse reports per hospital (the other nurse respondents were not working in hospitals).
The large mail survey undertaken in the study-- surveys were mailed to 272,783 nurses in the four states: 106,532 in CA, 49,385 in FL, 52,545 in NJ, and 64,321 in PA—had a response rate of 39% at the nurse level, owing to the impossibility of targeting the mailings to hospital staff nurses, providing monetary incentives, or undertaking extensive follow-ups with such a large sample. However, a high response rate from nurses is of secondary importance to having a high response at the hospital level and reliable reports from a representative sample of nurses in a large and unbiased sample of hospitals, covering a broad range of important issues. The survey included items that assessed, in addition to nurse workloads, nurse education, and the nurse work environment, nurse demographics, burnout, job dissatisfaction, intent to leave, and the quality of care, patient safety indicators, and frequency of adverse events on their unit. We have information from 9 out of every 10 hospitals in all four states. We also have evidence, from an intense re-survey of 1300 original non-respondents with a 91% response rate and a rigorous evaluation of possible bias, that there were no significant differences in responders and non-responders in reports of hospital-level organizational features of nursing.31,34
Hospital nurse staffing was calculated from nurse survey data by dividing the average number of patients reported by nurses on their units on their last shift by the average number of nurses on the unit. Nurses’ educational composition was the percentage of staff nurses in each hospital holding baccalaureate degrees in nursing or higher. The nurse practice environment was derived from the Practice Environment Scale of the Nursing Work Index-Revised (PES-NWI), an extensively-validated survey measure.4,30, 35–37
The 31 item Likert-type scale indicates the degree (1=strongly disagree to 4=strongly agree) to which various organizational features are present in the practice setting. In prior analyses we employed 5 subscales that were validated and shown to be strong predictors of patient and nurse outcomes: nurse participation in hospital affairs (9 items), nursing foundations for quality care (10 items), nurse manager ability, leadership, and support of nurses (5 items), staffing/resource adequacy (4 items), and nurse-physician relations (3 items). Published internal consistency coefficients (Cronbach’s alphas) for the five subscales range from .71 to .84. In the analyses reported here four of five PES-NWI subscales were used. The staffing/resource adequacy subscale was excluded because it empirically overlaps our direct measure of nurse staffing. Subscale measures were calculated for each hospital by averaging the values of all items on each of the subscales for all nurses in the hospital. These four aggregated subscales were then averaged to produce a single composite measure of the practice environment. PES-NWI subscales and the composite scale range in value from 1 to 4 and in the regression models were standardized to have a mean of 0 and standard deviation of 1.
Patients aged 19 – 89 years with a diagnosis related group (DRG) classification of general, orthopedic, or vascular surgery were included for a total of 1,262,120 patients. Measures included 30-day inpatient mortality and failure-to-rescue (defined as deaths for the subset of patients who experienced complications). International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM)
codes in the secondary diagnosis and procedure fields of discharge abstracts were scanned for evidence of 39 clinical events suggestive of complications.38
risk adjustment approach was used consisting of 27 comorbidities (excluding fluid and electrolyte disorders and coagulopathy40,41
). Additional adjustments included gender, age, transfer status and a series of 61 dummy variables indicating the type of surgery. Risk adjustment was enhanced by a 180-day look back to previous hospitalizations to distinguish between complications and comorbidities. C statistics (area under the receiver operating characteristics curve42
) for the risk adjusted mortality and failure-to-rescue models were 0.89 and 0.82, respectively.
Descriptive statistics are provided to show characteristics of the study hospitals as well as characteristics of surgical patients discharged from and the nurses who were surveyed in the different hospitals. Logistic regression models were used to estimate the effects of nurse staffing, nurse work environment, and nurse education on patient mortality and failure-to-rescue, before and after controlling for other patient and hospital characteristics. Our final model is one which includes an interaction effect involving nurse staffing and the nurse work environment. We use the estimated odds ratios from this final model to show the differing effect of staffing in hospitals with different environments and, alternatively, the different effects of work environments at different staffing levels. To account for the clustering of patients and nurses within study hospitals, all model estimates were computed using Huber-White (robust) procedures to adjust the standard errors of the estimated parameters. All analyses were conducted using STATA version 10.1 (STATA Corp, College Station, TX).