Hospital admissions for asthma, diabetes and COPD: is there an association with practice nurse staffing? A cross sectional study using routinely collected data

Ethical approval

Ethical approval was not required for this study.

Sample

We used the same sample that was used previously when studying the relationship between nurse staffing and quality of care measured by QOF scores [1]. In 2005/2006 QOF data were collected from 8409 practices. A number of practices were excluded from this study because they were small (< 1000 patients), were lacking in condition specific registers, registers indicated no patients, there was at least a 50% mismatch between actual registers and the number of patients reported for individual indicators or where it was not possible to estimate the number of FTE practice nurses. This reduced the number of practices remaining in the analysis down to 7456.

Data sources

Data were acquired from a number of sources. Data on admissions by practice were obtained from Dr. Foster Intelligence. This database contains records of all inpatient and day case care provided by the National Health Service (NHS) in England derived from the Commissioning Data Set. Admissions are coded according to the primary diagnoses associated to it, using the International Classification of Disease (ICD) version 10. We selected all non-elective admissions where the primary diagnoses was asthma, COPD or diabetes as classified by the Agency for Healthcare Research and Quality, Clinical Classification Software System for ICD10 [8]. We selected these three conditions because they represent QOF conditions where there are areas of significant activity by nurses in General Practice where the classifications to identify admissions were available. We extracted non-elective admissions data by each GP practice operating in England for the period April 2005 until end of March 2006 (the QOF reporting period). This allowed direct comparison with research on QOF outcomes for the same period [1].

Population and practice data were obtained from the Office of National Statistics and other sources were linked to each practice [6]. The number of FTE practice nurses employed by each practice was estimated using the method described by Griffiths and colleagues [1]. Data on the number of nurses employed by each practice was obtained from the healthcare specialist Binleys and the number of FTE practice nurses employed by PCTs from the NHS Workforce Projects Benchmarking database [9]. These data were used to calculate the FTE practice nurse estimate for each practice. A validation survey was conducted to check the accuracy of these estimates. There was good concordance between these estimates and FTE practice nurses actually employed by the practices sampled in the survey [1]. The estimate was sub-divided into quintiles ≤3038.01 patients per FTE practice nurse (n = 1421), 3038.02-3901.48 (n = 1421), 3901.49-4823.44 (n = 1422), 4823.45-6210.68 (n = 1420), ≥6210.69 (n = 1422) for the analysis purposes with a sixth category (n = 350) representing those practices that did not have a practice nurse.

Analysis

A two-level multilevel model (practices nested within PCTs) was estimated using MLwiN; a widely used multilevel modelling software package [10]. A similar set of independent variables to those used in an earlier paper [1] was used. These included variables related to geographical area (population density and the Index of Multiple Deprivation [11], which considers a range of factors in the area where the practice is located: income, employment, health and disability, education, skills and training, barriers to housing and services, living environment, and crime) practice patients (% aged ≥ 65, % from a racial or ethnic minority), practice profile (list size per FTE GP, singled handed practice, Primary Medical Services Contract) and general practitioner profile (% ≥ 45 years of age, % female, % UK qualified). A bias adjustment variable was also used to identify practices where recorded register size differed from the number of patients contributing to QOF indicators [1].

We modelled three dependent variables: the number of patients experiencing one or more non elective admission, the number of patients with two or more admissions and the ratio of actual to expected number of people admitted. In the first two cases data are modelled at the practice level and adjustment for patient characteristics are based on aggregated data. Patient gender at the practice level was excluded from the models because it has high colinearity with other patient profile explanatory variables [1].

To adjust for patient characteristics at the individual level we used the indirect standardisation method [12] and modelled a Standardised Admission Ratio (SAR). For the SAR, patient level hospital episode data for gender, age (five year bands) and deprivation are used to estimate the expected number of non-elective admissions. Deprivation was estimated using the Carstairs index for the patient's area of residence, which is based on small local area census results considering factors such as low socioeconomic status, car ownership, overcrowding and unemployment [13]. In this model, risk adjustment for age, gender and deprivation takes place at the patient level and for this reason we did not adjust for age, gender and deprivation at the practice level. The SAR models were compared against the other model for similarities and differences to determine whether our conclusions were sensitive to the approach taken to risk adjustment.

The dependent variable for all models is a count (admissions) that has a Poisson distribution. A Poisson model with a log link function was therefore fitted to the data. To incorporate the fact that we were modelling the number of patients with one or more admissions, which is a function of the number of patients at risk, an offset was required [10]. In the case of all patients admitted, and patients with two or more admissions, the register size was used as the offset and the expected number of patients admitted was used in the third (SAR) model.

The register size is a product of practice prevalence (of the condition of interest) and the practice list size. Therefore neither prevalence nor practice list size were included in models using register size as the offset. Practice list size also has a very strong correlation with the expected admissions and for this reason was dropped from the set of independent variables fitted in the SAR model.

Main findings

We found evidence of an association between nurse staffing levels in English general practices and non-elective hospital admissions for asthma, COPD and diabetes. The relationships seem to be associated with the skill mix (ratio of nurses to doctors) rather than the size of the clinical workforce relative to the number of patients.

Patients registered with practices employing more nurses were less likely to have a non-elective admission related to their asthma and COPD. For COPD, the significance of the association varied from one model to another. Where significant, (2/3 models) the association was negative with higher nurse staffing level associated with lower admissions for COPD. For diabetes, however, the association was positive and showed significance in two of the three models fitted. This suggests that although higher nurse staffing levels have been shown to be associated with better compliance with processes of care and better intermediate clinical outcomes resulting in achieving higher QOF scores [1], its association with non-elective hospital admissions is likely to be dependent on other factors including the specific disease, service configurations and patient related factors not included in our models.

Contradictory evidence exists in the literature regarding whether the attainment of higher QOF scores is associated with positive effect on morbidity, mortality, hospital referrals and non-elective admissions [14–17]. Although the QOF might accurately assess the process of care, it is questionable whether the same is possible for clinical outcomes, despite some intermediate outcomes being included in the framework. This is in part due to the lack of adjustment for case mix when QOF rewards are calculated. Evidence from Canada also suggests that the relationship between performance and the actual health gains is questionable [18]. There is also controversy around the appropriateness and evidence-base of the targets relating to the control of acetylated haemoglobin (HbA1c) levels in diabetic patients that needs to be achieved in general practice [19]. We previously found that higher nurse staffing levels were associated with attainment of higher QOF scores in four out of eight clinical domains and with performance on specific indicators of intermediate clinical outcomes within the QOF clinical domains [1]. Using hospital non-elective admissions as an indicator of the quality of care delivery in general practice offers an external measure that can be useful in validating QOF performance measures. It can also be taken as a proxy measure of clinical outcomes.

We found partial support for the relationship between staffing and quality that we observed previously when using this external measure. In this study, we found that, for asthma, intermediate nurse staffing levels in the 2^nd and 3^rd quintiles (3038-3901 and 3901-4824 patients per FTE practice nurse) had the lowest admission rates in the crude and the SAR models respectively. This might indicate that an optimal mix of GPs and nurses, that does not require the highest levels of nurse staffing, could be the best strategy for delivering care to asthma patients in general practice.

However, simple changes in staff-mix are not sufficient without consideration of the context in which people work and the organisational factors related to that [20]. Hence, the variation in the relationship between higher nurse staffing levels and admission rates across the three clinical areas may relate to variations in the activity of nurses and/or how services are locally organised in those areas. This further confirms our earlier conclusion that there is a need to investigate the configuration of services and deployment of nurses more specifically [1].

While higher levels of nurse staffing were associated with better QOF performance for diabetes, including levels of HbA_1C, in our previous study[1] they were associated with higher levels of admissions here. Although we did not directly assess the association between QOF performance and admissions in this study, this finding does confirm that the relationship is not a straightforward one. There are known associations between several socioeconomic and patient related factors and developing diabetes and its complications that are independent of the access to and quality of care provided. The National Diabetes Audit revealed that "the most deprived in the UK are two and a half times more likely to have diabetes" and that "complications of diabetes such as heart disease, stroke and kidney damage are three and a half times higher in the lower socio-economic groups" [21]. Reid and colleagues also found that patient factors were the strongest predictor of the large variation in all admission rates between 120 London general practices, especially for emergency admissions [22]. We have attempted to control for the socioeconomic factors, including deprivation at the practice level, as well as patient related factors, in the SAR. However, there are other patient related factors, like complexity of the condition and patient adherence to medication that can contribute to the likelihood of a diabetes related admission, which we were unable to include in the model.

Strengths and limitations

This study includes data covering the vast majority of patients in English general practice. Although we excluded some practices, our study has examined evidence from the vast majority of English general practices providing care to 48 million patients. We have controlled for potential confounding variables but observational studies such as this cannot account for unmeasured factors. It may be that higher nurse staffing is associated with other unmeasured attributes of quality within the practice and if this is the case increasing nurse staffing will not bring benefits unless these factors are attended to. What remains unclear is whether there is a causal relationship at all.

We have used the rate of hospital non-elective admissions as an indicator of quality of primary care delivered to patients; however although rates of hospital non-elective admission for some chronic conditions are possible indicators of the quality of care, they should be interpreted in conjunction with measures of population composition and deprivation. Hospital non-elective admission rates are likely to be confounded by socioeconomic characteristics of the population, case-mix and secondary care characteristics (e.g. hospital policies) which are not under the control of the primary care providers. Lack of adjustment for this confounding can result in primary care trusts and general practitioners being penalised for patterns of service use that they cannot directly control [2]. This was echoed by Giuffrida et. al. [23] who argued that using admission rates for asthma, diabetes and epilepsy as an indicator of primary care quality can be misleading. In our analysis, we have not controlled for secondary care related factors that might impact on admissions.

The approaches used to model the data have their own advantages and disadvantages. The SAR model allows for some adjustment at the patient level albeit using a simple model whereas the admissions models are based on data collected solely on practices therefore only suffers from model misspecification at that level. There could be multiple reasons why some practices have more single admissions than others that go beyond the care they provide and there may be a stronger case for using subsequent rather than first admissions as a proxy measure of quality [14]. In our findings whereas the relationship between nurse staffing and diabetes non-elective admissions was significantly positive in the models for one or more admissions, the overall association was not statistically significant in the repeat admissions model although in all cases the relationships were broadly similar across all models. The lack of consistency in relationships seems to indicate that nonetheless there is residual confounding.

We did not explore the cost implications of the variations in staffing observed. The absolute numbers of admissions across these conditions is high and so the relatively small reductions in rates of admissions translate into large absolute numbers of people. However when considering the numbers of people admitted per practice it seems clear that, even if causality was assumed, employing nurses to reduce admissions may be an expensive solution and could only be justified if there was a wide benefit in quality associated with a sufficiently large group of patients or across several conditions. However, hospital admissions are expensive and so economic gains through prevented admissions could also be large.