Data sources: registered prevalence and primary care supply

Counts of general practice-registered, i.e. observed prevalence of CHD, stroke and hypertension in April 2007 were obtained from English general practice disease registers, produced for the purposes of incentivizing practices for achievement of QOF treatment targets. These patients are clinically confirmed cases of CVD who are receiving regular follow-up for their disease. QOF prevalence rates are based on total populations registered with practices, but to compare observed prevalence with expected prevalence estimated from resident population-based contextual data from the Census and other sources e.g. deprivation and proportion ethnic minority population, we derived residence-based QOF prevalence estimates for LAs using a lookup table - a pooled extract of England practice registers - from the National Strategic Tracing Service (now Personal Demographics Service) which apportioned practice populations to LA areas as at January 2006, by providing the exact number of practice population resident in each LA [14].

We apportioned counts of CVD patients registered by practices to LAs, in accordance with the proportion of each practice population resident in that area, which assumed that CVD prevalence was geographically uniform across a practice population (the average practice population is only about 6,400). We divided the aggregated count of CVD patients in each LA by total mid-2006 LA population estimates to give estimated crude prevalence, as used for QOF prevalence. Where less than 50 patients fell into an LA, the numbers were excluded from the look-up process. Three LAs could not be mapped due to discrepancies between QOF and NSTS datasets. In order to investigate the effect of healthcare supply upon diagnosis levels we included a measure of general practitioner availability in the form of the number of general practitioners (GPs) per thousand LA population, calculated in the same way [15].

Data sources: expected/estimated prevalence

Expected prevalence for each LA was obtained from the APHO epidemiologic models, which are based on the socio-demographic and behavioural characteristics of respondents with the respective conditions in the Health Survey for England (HSfE). To produce the models, HSfE data for the years 2003-4 was pooled (sample size 21,233) to increase the cases of diseases. Surveys over this period included boosts for ethnic minority and older people, and focused on CVD and its risk factors. Of the respondents included, 14,574 (68.6%) were of White ethnicity, 308 were Mixed (1.5%), 1,991 (9.4%) were Black or Black British, and 3,725 (17.5%) Asian or Asian British. The outcome variables were patient-reported doctor-diagnosed CHD and stroke, and for hypertension, normotensive-treated, hypertensive-treated but uncontrolled, and hypertensive-untreated groups; i.e. a combination of patient-reported and objectively-measured variables. Patient reports of doctor-diagnosed CHD and stroke have been extensively validated elsewhere [16–20]. For example, in the British Regional Heart Study, 80 per cent of men with a GP record of angina reported their diagnosis, and 70 per cent of men who reported an angina diagnosis had confirmation of this from the record review. The prevalence of diagnosed angina in these older men was 10.1 per cent according to self-reported history and 8.9 per cent according to GP record review [16]. At that time (1999) some cases, such as newly-registered patients, may not have had their diagnosis clearly recorded in GP records.

Ordinary least-squares (OLS) logistic regression models were fitted and explanatory variables for each disease outcome identified by reverse stepwise selection.

The baseline odds of each disease were obtained directly from the HSfE dataset. The strength of association between each explanatory variable and disease caseness was then used to calculate the relative odds, which were applied to the baseline odds to derive the prevalence estimates for each sub-group of risk factors. The variables which can be included in each local model are limited by the availability of local data for them from Census and other national sources. The core model variables are ten-year age band, gender (male and female), ethnicity (Asian/Asian British, Black/Black British, White, Mixed and Other including Chinese) and deprivation (based on Index of Multiple Deprivation 2004 scores) [21]. In the case of CHD and stroke, smoking prevalence is also included, and the stroke model does not include ethnicity. The models use 2006 mid-year quinary age-band population estimates by ethnic group from the UK Office for National Statistics (ONS), which were summed to 10-year age bands to match the model [22]. LAs are stratified into deprivation score bands based on cut-offs of Lower Super Output Area quintiles - the ONS categories used in the HSfE. Internal validation included using the models to predict the response for each subject in the source data, and area under the receiver operating characteristics (AUROC) curve. AUROC curve values were 0.834, 0.844, and 0.807 for the stroke, CHD and hypertension models respectively. External validation showed that prevalence gradients derived from the models - for example with age and smoking status - agree well with published results from population-based studies.

In the case of the CHD and stroke models, smoking status is also included. Local smoking prevalence estimates are not available from the HSfE because of small sample sizes, so the CHD model uses synthetic estimates from the Neighbourhood Statistics website, which are for the period 2003-2005 [23]. Model assumptions include that the proportion of smokers, ex-smokers and never-smokers is uniform across ethnic categories and that the proportion of ex-smokers in each age-sex group is constant across areas. Sensitivity analysis has shown that varying the smoking prevalence has a very small effect on prevalence. Further technical details of the models are available in additional files 1 (CHD), 2 (hypertension) and 3 (stroke) in the web appendix, and also on the APHO website [13].

Spatial analyses

Observed: expected prevalence ratios for LAs were calculated in Excel 2007 and mapped using the geographic information systems package ArcGIS 9. Two exploratory spatial data analysis methods commonly used in geographical studies were used to investigate patterns in O:E relationships, Local Moran's I (LMI) analysis and geographically weighted regression (GWR). The LMI technique is used to identify geographic clusters and outliers in data by testing for randomness in spatial distribution across a dataset, localities with significance scores (Z scores) greater than two standard deviations being considered to be either clusters or outliers [24]. Strongly positive Z scores indicate statistically significant similar values in close geographic proximity hence the presence of a cluster; a strongly negative Z score demonstrates a locality with a significantly dissimilar value in relation to its neighbouring localities thus indicating an outlier.

GWR is a form of spatial statistics which disaggregates geographic data into spatial blocks using a probability distribution kernel, which moves from location to location across the dataset to test for geographic variation in regression relationships. In situations where there is geographic variation in the strength of a regression relationship, a phenomenon referred to as spatial non-stationarity, the use of GWR will improve model goodness-of-fit to data, expressed as the trade-off between statistical predictor bias (linked to R-squared values) and variance (linked to degrees of freedom). In comparison to a classical model GWR will produce higher correlation coefficients, lower residuals and higher degrees of freedom than traditional ordinary least squares (OLS) regression [25, 26].

In this paper, we used GWR to assess whether a linear regression relationship between observed and expected prevalence existed and if so whether it varied in strength over space, the purpose of which should be viewed as distinct from that of mapping observed to expected ratios, as the latter aims to measure equality between two variables rather than to assess predictability of an association. Both OLS and GWR models were run in the software package GWR 3 to test for spatial non-stationarity. The optimal bandwidth for the kernel was estimated using the Akaike Information Criterion [27]. Two rounds of regression were performed, the first a univariate regression involving expected prevalence as the independent variable and observed prevalence as the dependent variable; the second bivariate including whole-time equivalent GP supply as an additional variable.

Our research conformed to the Helsinki Declaration http://www.wma.net/en/30publications/10policies/b3/, and to local legislation. It did not require ethical approval or patient consent as it is a secondary analysis of publicly-available data.

Observed/diagnosed prevalence of CVD

Spatial analysis

Mapping of observed and expected prevalence demonstrated that observed and predicted prevalence for both diseases showed strong north-south and southwest-southeast gradients, with the southeast generally showing lowest disease levels, especially in the case of expected prevalence. However, within these geographical trends in both QOF and modelled prevalence, mapping of O:E ratios demonstrated spatial variation between neighbouring LAs. LMI Z scores indicated the presence of some statistically significant clusters and outliers in O:E ratios (Figure 1, Figure 2 and Figure 3).

For all three conditions, particularly CHD and stroke, London and much of its hinterland showed statistically significant discrepancies in O:E ratios, with observed prevalence much lower than the epidemiological models predicted. For CHD there were also significant clusters and outliers in parts of northern England, especially in Cumbria and Yorkshire, though unlike for the London area the clusters typically related to O:E ratios tending towards unity.

Further geographic analysis with GWR revealed significant spatial variation in the O:E relationship for CHD, hypertension and stroke, with GWR describing the dataset more accurately than traditional OLS regression. The linearity of the relationship between observed and expected prevalence was weakest for hypertension with the lowest overall regression coefficients and highest residual sum of squares for observed against expected prevalence; stronger associations were observed for CHD and stroke.

Main findings

The findings presented here indicate that despite almost universal access to free primary healthcare services, and a significant financial incentive for these services to find and register new cases on their computer systems, there may be significant under-diagnosis of all three CVD conditions examined in this study in many areas in England. The difference between expected and recorded disease is most marked for hypertension, with strikingly wide variation in diagnosis levels between areas. Hypertension and stroke showed similar geographic variation to CHD, but without the clusters/outliers for northern England; additionally the contrast between the ratios for London and LAs immediately to the north was more pronounced for stroke.

An obvious question is: why was there such a discrepancy between expected and practice-registered disease prevalence, when the model-based CHD and stroke prevalence estimates were based on patient reports of doctor-diagnosed disease? Reasons why practice computer systems may under-record cases are likely to include inadequate searching of practice records (which are now in the UK mainly electronic) for previous diagnoses or CVD-related prescriptions [28, 29], lack of linkage of practice and hospital records (practice still have to enter codes for hospital admissions manually), and high population mobility in urban areas: data for new patients is often still entered into practice systems manually and previous diagnoses may not be added for some time [30]. The QOF does not appear to have resulted in substantial improvement in recording. The crude QOF CHD prevalence was 3.44% in 2006-7, the year after QOF was implemented, but was 3.54% in 2009/10 - a minimal increase. In the case of hypertension, for which the modelled estimates used a combination of doctor-diagnosed disease and blood pressure measurements, the gap is greatest in younger and middle age groups, when males in particular are less likely to be seen by their practices [31]. In contrast, diagnosis appears to be more complete in the north of England.

From a spatial analytic perspective, there are significant geographic discrepancies between QOF prevalence and modelled prevalence for all three conditions in much of London and its hinterland. The inclusion of a measure of GP supply in the GWR analyses suggests that, despite needs-based healthcare resource allocation in the UK, persistent differences in availability of primary care services is an important limiting factor in diagnosis. An analysis of Gini coefficients to measure geographical equity in GPs per 100,000 population in England and Scotland showed that equity in England rose between 1974 and 1994, but then decreased, and in 2006 it was below the 1974 level [32].

The results also provide some contrast to a study investigating revascularisation rates in males, which found that even after adjusting for the higher CHD burden in the north of the country related to socio-demographic composition, the likelihood of receiving surgery during the 1990s decreased outside of southern England, suggesting a geographic imbalance in the provision of tertiary cardiology care [33]. Considerable efforts have been made subsequently to ensure that access to tertiary care is more equitable. However our analysis suggests that from a primary care perspective, it is London and the surrounding area which would benefit most from increased resource allocation, although GP supply itself seems to be only one influential factor affecting diagnosis levels. Further work will need to investigate the effectiveness and yield of strategies for CVD case-finding by practices, and to validate the predictions of the prevalence models at practice level.

Strengths and limitations of the study

Strengths of the study include the use of new data from practice disease registers and recently-developed geographical analytical techniques and software. Limitations include the fact that modelling was carried out at LA level, which may conceal much wider variation at lower levels. The spatial scale of the analyses includes many LAs which are highly heterogeneous in socio-demographic composition, so the study may have missed significant small area variation; this is especially likely in northern areas where regression associations were weakest. There is imprecision in the model prevalence estimates, for example, because of the size of the Health Survey sample we used, especially in subgroups with small numbers of cases - the precision of the estimates is also affected by the prevalence of the disease outcomes. Model restrictions include that the proportion of smokers, ex-smokers and never-smokers is uniform across ethnic categories, when, for example, smoking prevalence varied from 16.4% in the Black/Black British sample population to 24.1% in the White sample population. Smoking prevalence is itself a modelled estimate, and ex-smoking prevalence is assumed to be the same nationwide. In addition the number of variables in the local prevalence models is constrained by the availability of local data on risk factors, so that it is not possible to include other well-established CVD risk factors as variables. Some of the local risk factor data relies on estimates based on 2001 Census data, although this will be improved by, for example, more accurate and timely data on smoking prevalence from the new ONS Integrated Household Survey. However we were reassured by the results of model validation, e.g. ROC curves. In addition, comparing model predictions to QOF registrations is also a form of external validation. For example, QOF registered counts of CHD were greater than expected counts in only three of 352 LAs, and in these only by small amounts. If the models were very imprecise, much more under-estimation would be expected. Further model validation is needed, and is occurring as the models are being used to guide local case-finding initiatives.