Hospital admissions

To identify relevant admissions for our dependent variables, we used a four step approach. First, we mapped Read codes—the clinical codes used in UK primary care—for the QOF dementia register to the diagnostic codes used for hospital care (ICD10). This ensured that the criteria for identifying hospital admissions matched the eligibility criteria for the QOF dementia register. The Health and Social Care Information Centre (HSCIC) provided a mapping algorithm, and we checked this against a mapping based on a large UK primary care database, the Clinical Practice Research Datalink. The final list of ICD10 codes for dementia was: F00-F03, G300-G309, G310, G311, G318, F051 and F107. Second, we identified individuals in the Hospital Episode Statistics (HES) who had at least one of these codes as their primary or secondary diagnosis during the sample period 2006/07–2010/11. Third, we extracted three samples from HES based on three types of admission (i.e. our dependent variables). Fourth, we dropped all admissions that occurred before a patient’s first (inpatient) dementia diagnosis, and, to ensure consistency with QOF registers which are defined at the end of the financial year, admissions occurring during the year the patient died.

Emergency dementia admissions were defined as those where the patient had a primary diagnosis of dementia on admission. Emergency ambulatory care sensitive conditions were defined by specific ICD10 codes [15]. Elective conditions comprised cataract, hip replacement, hernia, prostate disease, or hearing loss, selected in consultation with our lay project advisors as those that might reasonably be identified at the QOF annual dementia review. For each admission type, we defined the dependent variable as the number of practice patients aged 18 or over with at least one admission to hospital within a year. Table 1 shows the three models that were tested with each of the three dependent variables (admission types), and Table 2 provides descriptive statistics.

Quality of care

The primary explanatory variable of interest was the quality of care provided by the practice, which was proxied by the QOF indicator scores for the dementia review. The quality indicators for dementia were introduced in April 2006, and we compiled a set of panel data covering the financial years 2006/7 to 2010/11. QOF indicator scores are freely available at practice level (http://qof.hscic.gov.uk/), but are not published at patient level. We measured achievement based on the total number of people in the practice for whom the indicator was achieved, divided by the total number of eligible people (Table 1). We did not use QOF points or thresholds in the achievement measure, as points-based measures understate between-practice variation in quality and may be subject to gaming behaviour [16]. GPs may ‘exception report’ individuals who are considered unsuitable for treatment, or who are newly registered with the practice or newly diagnosed, or who make an informed dissent. In the base case analysis, we included exception-reported individuals in the denominator for the achievement calculation. A typical practice excludes (exception reports) around 8% of eligible patients from the dementia annual review [17], but clearly these individuals are still at risk of hospitalisation. To test the impact of exception reporting, the exception rate was included as a separate covariate in all analyses (i.e. additional to the achievement measure). As a sensitivity analysis, we also tested a modified quality indicator that excluded exception-reported patients from the denominator (see Table 1; Model SA2); this higher QOF achievement rate is the measure used for reimbursing practice performance. The average annual review rates were relatively stable over our sample period, but performance was sufficiently variable to allow differences within and between practices to be tested.

Exposure terms

Our base case exposure term—the pool of individuals in a practice who were at risk of admission—was the reported QOF dementia register. However, less than half of those in England with dementia have a formal diagnosis [18], and the QOF dementia register is therefore highly likely to underestimate the ‘true’ disease prevalence. To test the robustness of findings, we used an alternative measure for our exposure term (Table 1; Model SA1). This alternative measure was derived from published estimates of ‘true’ population prevalence figures [19], officially considered to be “the most authoritative to date” [20]. We used age-gender band prevalence estimates for early and late onset dementia to predict the QOF register for each GP practice in each year, adjusting for the number of patients in nursing homes where prevalence is higher (details are provided in S1 Appendix). We tested the effect of this revised exposure term in our regressions. On average, the predicted register is more than twice the reported QOF dementia register (Table 2). In Fig. 1, the impact of this difference between the registers is mapped at local health authority (PCO, primary care organisation) level. The figure shows how PCO average register estimates vary when based on (a) the reported QOF dementia register and (b) the predicted register, with darker red indicating larger numbers. In map (c), darker blue indicates a larger change when moving from (a) to (b). These figures were generated using ArcGIS software and we used the same intervals in figures (a) and (b) to facilitate comparison.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. GP practice QOF dementia registers: average for English Local Health Authorities, 2010/11.

https://doi.org/10.1371/journal.pone.0121506.g001

Other covariates

Hospital admissions are influenced by factors other than the quality of primary care, and people with dementia often have complex health and social care needs. Our analyses therefore adjusted for an array of factors (Table 2). These were selected in consultation with lay advisors to the research project, who have experience of caring for family members with dementia.

To control for primary care practice characteristics, we used data from the General Medical Services (GMS) dataset, QOF dataset and Attribution Dataset (ADS). To control for local population characteristics, we used data from the ONS Neighbourhood Statistics which are available at Lower Super Output Area (LSOA) level. LSOAs are defined geographic units that cover an average population of 1,500 people. As defined by the 2001 Census, there were 32,482 LSOAs in England during the study period (2006/7 to 2010/11). The ADS provides a breakdown of the practice population by LSOA. We derived a weighted average of the local population characteristics of the practice (e.g. deprivation, ethnicity, informal care) and assigned this to the practice. ADS data are collected at the start of the financial year, whereas QOF data are collected at the end of the financial year. We therefore adjusted the estimates based on ADS by taking moving averages across two years of data.

We derived two measures to reflect differences between practices in terms of access to care. First, the variable on 48-hour access to primary care came from the annual GP Patient Survey. Second, to adjust for patients’ geographical access to emergency care we estimated the straight line distance from the primary care practice to the nearest acute hospital, using postcode data for primary care practices and acute hospitals from the HSCIC. Calculations were based on the grid reference from the postcodes of all the practice branches and of acute hospitals [21]. Straight line distances are a good proxy for road distances, and are less computationally demanding [22, 23].

The level of social support provided in the community may help prevent acute hospital admissions [24]. We used three measures of social support: the proportion of practice patients aged 65 and over who were nursing homes residents, based on time series data from the HSCIC; and two variables based on Attendance Allowance (AA) data from the Department of Work and Pensions. AA is a social benefit for people aged 65 or over who have a physical or mental disability and need help with personal care. It is not means tested, is available at lower and higher rates depending on need, and can be used to pay for care from any provider. We derived two practice-level measures: the percentage of people claiming AA; and the percentage of AA claimants who were paid at the higher rate. Around 16 percent of practice patients aged 65 and over claim AA, with over half of those receiving the higher rate.

Our practice-level measure of deprivation was the percentage of people aged 60 or over living in income deprivation. Uptake of AA was positively correlated with this measure of deprivation, but the relationship was not linear. We therefore used a set of interaction terms to test whether the effect of AA varied depending on the level of deprivation. Specifically, we interacted the percentage of AA claimants in the practice with a categorical measure of deprivation that grouped practices according to whether the percentage of people aged 60 or over living in income deprivation was low (<20%; 50% of practices), medium (20% to 35%; 37% of practices) or high (> 35%; 13% of practices).

Model specification and statistical analysis

After excluding patients younger than 18 and those who changed practices within a year, we merged the HES patient-level data with the QOF practice-level data and linked these to the covariates derived from the other datasets. We dropped practices with a list size of fewer than 1,000 patients because these practices have relatively small numbers of patients with dementia and practice performance is therefore subject to large fluctuations over time. Our final dataset provided an unbalanced panel of between 7,697 and 7,965 GP practices per year, with most practices (82.5%) contributing five years of data.

We ran panel count data models to test the impact of primary care quality on hospital admissions. Our response variable was defined as the number of people admitted to hospital adm_it at least once in year t = 2006/07,…,2010/11 from GP practice i = 1,…,N. Around 11% of practices had no emergency admissions for dementia over the 5 year study period, and for elective admissions the figure was over 40%. As these practices could not contribute data to fixed effects models, we estimated random effects Poisson models for the three admission types specified as follows: (1)

In this equation, Q_it is the variable for GP practice quality as measured by the QOF; X_it is a vector of covariates that capture differences in the practice patient population and the supply of and access to other care resources; D_t is a vector of time dummy variables to control for temporal trends in hospital admission; risk_it is the number of patients at risk of admission which in our base case analysis (Model BC) is defined as the QOF dementia register (see subsection: Exposure terms).

The GP practice-specific effects α_i capture unobserved, time-invariant practice effects, in other words any practice behaviour that influences admission rates, including, but not limited to, decisions specifically on hospital care. Following the convention we assume a conjugate gamma density for the multiplicative GP practice-specific random effects [25].

Estimation of random effects is carried out by maximizing the likelihood that results after integrating out the conjugate random effects [26]. We calculated bootstrapped standard errors for the random effects model. Statistical significance was assessed at the 5%, 1% and 0.1% levels.

The coefficient estimates can be interpreted as the percentage change in admissions resulted from a partial change in a variable (semi-elasticity). The exponentiated coefficients or incidence rate ratios (IRRs) also facilitate a semi-elasticity interpretation with (IRR-1)*100 estimating the result of a discrete (one percentage point) change in the QOF rate on the percentage change in admissions. For instance, an IRR of 0.9975 implies a reduction in admissions of 0.25%. Details of the method used to calculate IRRs for the AA and deprivation variables which involve interaction terms are provided in S2 Appendix.

We ran two sensitivity analyses to test the robustness of findings (Table 1). First, we used an alternative measure for our exposure term (Model SA1; Sensitivity-1). Second, we tested a modified measure of achievement that excluded exception-reported patients (Model SA2; Sensitivity-2), as this measure is the basis on which practices are rewarded. Analyses were performed using Stata, version 12 (StataCorp LP, TX).

Ethics statement

This was a retrospective analysis of previously collected, non-identifiable information, and involved no change in the management of patients. Obtaining individual consent was not feasible so patient records were anonymized and de-identified prior to analysis. The Health and Social Care Information Centre (HSCIC) handles requests for de-identified data and has a legal responsibility to ensure there is an appropriate legal basis to permit the release and subsequent processing of data, that all necessary approvals are in place, and that organisations have appropriate arrangements and safeguards for secure data handling. The HSCIC approved the release of the Hospital Episode Statistics (http://www.hscic.gov.uk/hes) data to the University of York (Data Re-Use Agreements RU115; RU536; RU750).