Geographical variation in cardiovascular incidence: results from the British Women's Heart and Health Study

Cardiovascular disease (CVD) was defined as any of angina, unstable angina, myocardial infarction or stroke. Prevalent events were informed by either self-report at baseline or medical record review events dated prior to baseline. Incident events were informed by either self-report at the three- or seven-year follow-up, medical record review, or death certificate, using a cut-off of 30^th September 2007 for all sources. CVD deaths were indicated by ICD10 codes I200-259, I516, I600-679, I690-699, G460-469, and G450-3 (underlying or otherwise). Self-reported strokes were only included where symptoms were >24 hours (to exclude potential cases of transient ischaemic attack (TIA) mis-reported as stroke). For potential events identified by review of medical records, a number of criteria were applied in order to count an event as incident CVD for analysis. In doubtful cases a telephone discussion was held with the woman's general practitioner to confirm or reject CVD diagnoses. For MI, at least one of the following was required: (1) ECG evidence of MI, (2) raised cardiac enzyme (including troponin), (3) hospital letter confirming MI. For stroke, at least one of the following was required: (1) report of ischaemic/haemorrhagic stroke on CT or MRI scan, (2) hospital letter confirming stroke, (3) final diagnosis from GP notes is ischaemic/haemorrhagic/subarachnoid/other stroke (in absence of scan and letter). For unstable angina a hospital letter confirming the diagnosis was required.

Dates of CVD events were required for time-to-event analysis. Exact dates were available for events ascertained by record review or death certificates. In the case of events only ascertained by self-report (n = 67, 13% of all events), only a year was given; 31st December was therefore used for the purposes of analysis (those with missing year are excluded (n = 47 where no other incident event occurred)). For prevalent events, the date of the earliest pre-baseline event was used for analysis. Similarly for incident events, the date of the earliest post-baseline event was used.

In addition to age, data on a variety of known and potential risk factors for CVD were collected at baseline. Smoking status (current, ex or never smoker), family history of CVD (father, mother, brother or sister had a heart attack or stroke), alcohol intake (never, socially, or most days), limited fruit intake (both winter and summer fruit intake less than once a week/never), limited physical activity (never take regular exercise), socioeconomic position, and CVD medications (aspirin (British National Formulary code [26] 02.09), statins and other cholesterol-lowering drugs (code 02.12), beta-blockers (code 02.04), ACE inhibitors (code 02.02.05), and blood-pressure lowering drugs (codes 02.02.01, 02.02.08, 02.05.01-06, 02.06.02, 02.04)) were obtained through the questionnaire. Life-course socioeconomic position (SEP) was calculated as a score out of ten responses, with a high score representing greater deprivation [27]. Responses given as "don't know" to the individual components of the score are treated as missing. However, where the husband's social class was missing or not applicable, the responder's own social class was used where available (available for 392/816 (48%)).

Systolic blood pressure (SBP), waist circumference, obesity (body mass index >30 kg/m²) and leg length were obtained through the nurse examination. The presence of diabetes at baseline was identified from self-report (n = 31), record review (n = 19), both self-report and record review (n = 182), fasting glucose higher than 7 mmol/l only (n = 184), or diabetic medicines only (n = 1)). Blood samples were taken after a minimum eight hour fast and were used to obtain fasting glucose, total cholesterol, HDLc and triglycerides[11]. Low density lipoprotein cholesterol (LDLc) was calculated from total cholesterol, HDLc and triglycerides using Friedwald's equation [28]. The Carstairs index [29] of area deprivation (derived from the 1991 census, using data for Great Britain) was calculated using postcodes at baseline [30].

The 23 towns in the BWHHS were separated into four geographical regions used previously [11]: Scotland (comprising Ayr, Dumfermline, Falkirk), North England (comprising Burnley, Carlisle, Darlington, Grimsby, Harrogate, Hartlepool, Scunthorpe, Southport, Wigan), Midlands/Wales (comprising Mansfield, Merthyr Tydfil, Newcastle Under Lyme, Shrewsbury), and South England (comprising Bedford, Bristol, Exeter, Gloucester, Guildford, Ipswich, Lowestoft).

Regional differences in both fatal and non-fatal CVD incidence were analysed using Cox regression adjusted for age. Individuals were censored on the date of emigration, date of death, or 30^th September 2007. Loss to follow-up for non-fatal incident CVD following non-response to questionnaires (1091/3776 (29%) of those alive at seven-year follow-up) or non-return of record review (313/3730 (8%) of those alive at final record review) is not considered here. However, loss of information regarding non-fatal events is expected to be low since loss to follow-up for both sources is small (n = 180, 5% of those alive at seven-year follow-up).

Further adjustment was made for known and potential risk factors (see Table 1, section A). Corresponding Kaplan-Meier survival functions were also produced for both incident CVD and prevalent/incident CVD combined.

The impact of missing values in covariates was investigated using multiple imputed datasets, generated using chained equations (with 10 cycles of regression switching). Ten datasets were imputed using all other factors (including age and region) to impute missing values. Log survival time and binary indicator for event were also included as predictors [31]. Linear regression was used to impute missing values for continuous variables (SBP, LDLc, waist circumference, leg length, Carstairs index, age at baseline), multinomial logistic regression for categorical variables (alcohol intake, smoking status, region), and logistic regression for binary variables (diabetes, obesity, family history, limited physical activity, limited fruit intake, CVD medication, life-course socioeconomic position components). For the life-course socioeconomic position measure, imputations were made for the ten components separately. Clustering by town is accounted for in all analyses through the use of robust standard errors.

Survivorship may also play a role. Although prevalent non-fatal CVD events are included in analysis, fatal CVD events (and other cause mortality) occurring prior to baseline are not, since by definition an individual must be alive to participate in the study. Those still alive at the start of the study will thus be a healthier group. The impact of this survivorship may vary between regions since mortality (CVD and other causes) may differ between regions prior to baseline. This would imply that regions exhibiting lower CVD prevalence in fact represent a group from which more of the least healthy individuals have already died.

It may also be possible that a different type of individual is at risk of early CVD (for example, before age 60). If this early onset CVD differed by region, but later onset CVD was more similar between regions, this would be consistent with the observations described here. Lastly, it is possible that there are regional differences in response to recent guidance on the management of cardiovascular risk factors and both primary and secondary prevention.

Regional differences in cardiovascular disease incidence and mortality have been reported for a number of other countries, including the United Kingdom [32], the United States [19, 33], Canada [14], Finland [20], Sweden [10], France [13, 16], and across Europe [15]. Some of these studies have sought to simply identify and describe regional differences. Others have attempted to use known risk factors to explain the regional variations in CVD [19, 33], although in some cases analysis was performed on aggregated data (i.e. individual patient data were not available) [14]. A study of regional differences in 12-year stroke incidence in the United States [33] showed that the differences could largely, but not entirely, be explained by known risk factors in a cohort of 7000 men and women aged 45-74. More prominent regional differences remained after adjustment in the women. This supports the results from the BWHHS cohort, which suggest that known risk factors do not account for all observed geographical variation in CVD. A further study in the United States has reported regional differences in stroke and fatal CVD, but not MI or non-fatal/fatal CVD combined [19]. Adjustment for known risk factors in this study further exaggerated regional differences. However, restriction of participants to physicians and exclusion of 16% of subjects with missing baseline data raises questions regarding the generalisability of these results. Finally, the BRHS has provided much of the UK evidence for regional variations in CVD in men. This work has provided evidence that for men aged 40-59 at baseline, around 75% of the between-town variation in 5-, 10-, and 15-year incidence of coronary heart disease can be explained by smoking, systolic blood pressure, exercise, social class, and height [18]. This again suggests that most - but not all - observed regional variations can be explained by classical risk factors.

One of the potential weaknesses of this analysis is missing data. This problem has two components, since there are missing data in the CVD risk factors included as covariates in the Cox models, and also non-response at each of the three follow-up time-points. The former is addressed through the use of multiple imputation, which assumes that the missing covariate values are missing at random (MAR) - that the missing values depend on observed values of other variables, but not on the missing values themselves. Given the well-documented strong relationships between many of the covariates of interest here, this would seem a generally reasonable assumption. Furthermore, imputation based on non-missing covariates is reasonable since individuals do not generally have many missing covariates (of those with at least one missing value at baseline, 1839/1976 (93%) have at least 9/15 non-missing values). The results from analysis of the multiply imputed datasets suggest that accounting for the missing data explains some of the observed differences between regions.

The second component of the missing data issue is non-response to either the baseline or follow-up questionnaires. Non-response to the follow-up questionnaires is not addressed here, since it is anticipated that the majority of self-report CVD events missed as a result of this non-response would be picked up by record review (only 5% of individuals are lost to follow-up from both sources). Furthermore, there is little evidence of a difference in response by region at either three-year (χ² p = 0.09) or seven-year follow-up (χ² p = 0.18). Non-response at baseline however is more pronounced than at subsequent waves (59% response across all regions) and there is strong evidence for a difference in response at baseline by region (χ² p < 0.0005). Since non-responders at baseline are excluded from all analyses, this has implications for bias and generalisability of the results presented here. If regions with higher non-response at baseline actually have higher CVD incidence, the large number of non-responders (those at likely to be at highest risk of CVD) would render estimates of regional differences conservative.