Record linked retrospective cohort study of 4.6 million people exploring ethnic variations in disease: myocardial infarction in South Asians

Procedural issues

The procedural techniques took into account that the required data are held by two different bodies, with different responsibilities. The community health index (CHI) and hospitalisation data are collected and maintained by ISD, and the census and mortality data are collected and maintained by the General Register Office for Scotland (GROS). GROS performed a risk-assessment of the data security aspects of the proposed record linkage and produced recommendations, including that all the linkage work was performed on a stand-alone computer. Modifications were made to the computer operating system to monitor activity and an audit trail was implemented for all peripheral devices (floppy and CD-ROM drives, and external ports). The computer was in a locked room at GROS accessed with the agreement of both organisations. GROS maintained a register of visits. Neither organisation could view the other organisation's primary datasets (the census data and the hospital discharge/death linked data) in a form in which individuals were identifiable.

Use of record linkage and one-way encryption

The linkage methods used were identical to those developed in Scotland for administrative matching of health data. Date of birth, surname (using soundex codes to allow for variations in spelling), forename, address and full postcode were used to link. For many records there was an exact match. For the remaining records, a probability matching process was performed using methods pioneered by Newcombe [18]. Here, the rate of false positives is critical. Methods have been developed to identify how false positives occur and what kind of strategies a human checker employs to decide whether a pair match is good. These decision strategies were built into a partitioning computer algorithm, which groups (partitions) records according to the likelihood of correct matching. These partitions then allow the allocation of human checking to partitions, rather than individual records, so maximising linkage and minimising effort.

Prior to the linkage the project's steering group judged and specified that an effectiveness of 80% or more for all ethnic groups would be adequate for the purposes of this demonstration phase. This was a pragmatic decision that needed to be made (in the absence of data or theoretical precedent) to help assess the personnel time and resource to be directed to the matching process.

As both census and CHI datasets described essentially the same population, for every person on the CHI, there was a very high probability that the same individual was represented on the census. Where all linking variables matched, it was assumed that this represented the same individual. The various types of mismatches were addressed by the use of the partitioning algorithm described above. For example, all the instances of mismatch limited to one differing digit of year of birth were put into the same partition. A sample was then taken from each of the partitions, and the adequacy of each match was assessed by considering how much less likely the next best match was. This allowed quantification of the quality of the match for each partition.

Figure 1 illustrates in concept how record linkage was based on the use of subsets from three datasets: the SMR01-death linked dataset, extracted by coronary heart disease diagnostic codes on any record, and including personal identifiers and clinical information; the CHI dataset which contains personal identifiers and the CHI number; and the census file which contains personal identifiers and details of individuals' ethnicity.

The CHI dataset lists everyone registered with a General Practitioner or in receipt of screening services and forms a unique identifier for NHS use. More than 99% of the Scottish population is listed. A CHI number was attached to the SMR01-death linked dataset. A one-way cryptographic (hashing) algorithm was then used to encrypt the CHI number into a code such that it is currently impossible to reverse the process. Personal identifiers were removed and dates of birth converted to ages to create a file suitable for analysis.

A census extract with personal identifiers and a unique census identification number was taken from GROS's census file. The census number was encrypted using an algorithm developed by GROS. The file was then linked to the encrypted CHI dataset using the probability methods already described. The personal identifiers were then removed, leaving a file with an encrypted CHI number and its corresponding encrypted census number.

A new census extract containing ethnic code and an encrypted census number was joined to the above file by exact matching using the encrypted census number. The encrypted census numbers were now discarded leaving the ethnicity code and the encrypted CHI number. The subset of the SMR01-death linked file was exactly linked via the encrypted CHI numbers; The encrypted CHI was replaced with an unrelated serial number (to keep the hospital spells together), resulting in depersonalised clinical health records carrying census ethnicity codes. To reduce the risk of disclosure some variables were further aggregated and access to the analysis file was restricted to named researchers in the secure room described above.

Evaluation of the effectiveness of linkage between census data and CHI

The quality of the linkage depended on the adequacy of the linkage between the census data and the CHI. We tested the probability matching methods using data supplied by Scottish hospitals where the CHI number was present. We assessed whether the personal identifiers we used yielded the same CHI number.

Ethical approval, legal advice and public scrutiny

As no identified individual responses to the Census were linked to health records, and vice versa, we were following the principles of existing data protection legislation and guidance, including two National Statistics (NS) Protocols on Data Access and Confidentiality, and Data Matching. Legal advice (obtained by GROS) confirmed that the proposed use of census data was compliant with the Census Act 1920 and the Data Protection Act 1998. We obtained approval from the Scottish Multi-centre Research Ethics Committee and the Scottish Privacy Advisory Committee (PAC).

This work was supported by the Commission for Racial Equality, NHS Health Scotland (through The National Resource Centre for Ethnic Minority Health), ISD, the Scottish Executive Health Department and the Chief Medical Officer. We disseminated information widely to encourage people to discuss the proposal and the early findings.

Analysis of data on incidence of myocardial infarction and mortality

A diagnosis of AMI was recorded by a clinician at hospital discharge or recorded as the underlying cause of death on a death certificate. There is national guidance promoting uniformity of approach. The approach to the diagnosis would be similarly applied in different ethnic groups. Increasing use of laboratory tests makes the diagnosis less open to the subjective judgement of a clinician who might be influenced by the ethnicity of patients. The Scottish Morbidity Record database (SMR01) records inpatient and day case discharges with a diagnostic accuracy of around 90% for CHD [19]. Data on discharges with CHD diagnoses (ICD9 410–414 or ICD10 I20-I25 in any diagnostic position) between 1981 and December 2003 were extracted (the data from 1981 – 2001 being used to check for prior AMI). A first AMI was identified using ICD10 codes I21 and 122 and ICD9 code 410, but restricting to first admissions for AMI based on no hospital admission for AMI in the preceding 10 years. Deprivation categories were assigned using the postcode sector of residence in 2001, which provides the Scottish Index of Multiple Deprivation (SIMD). The SIMD scorers were divided into quintiles for analysis.

The South Asian population comprised people who were designated by the head of household at the 2001 census as Pakistani (23, 660 people, 86.9% matching of census and community health index), Indian (11, 690, 87.6% match), Bangladeshi (1463, 84.9% match), and other South Asian (4526, 82.2 % match).

The incidence of AMI was calculated using the data between April 2001 and December 2003 inclusive. Poisson regression modelling was used to calculate incidence rate ratios and 95% confidence intervals, by sex. Odds ratios for survival following AMI at fixed intervals from 30 to 720 days after admission were calculated using logistic regression. Kaplan-Meier plots were produced using an age adjusted survival function, setting age at 68 years. Adjusted hazard ratios were estimated using Cox regression. We included age (in five groups), sex and previous admission for diabetes in the Cox model. Graphical methods were used to confirm the proportional hazards assumption. Adjustment for area based social and economic deprivation based on the Scottish Index of Multiple Deprivation was done.

The checks included review of procedures and analysis programs created by M.S., first by C. F., and subsequently in consultation with an external adviser with experience of similar analyses, KMc (see acknowledgements).

Quality of linkage of the subset of the SMR01-death file to CHI

630,121 SMR01-death records came to ISD with a CHI number applied by the hospitals supplying the data. 629,493 yielded the same CHI number on matching to the CHI database using the fields in our linkage methodology. 628 records had differing CHI numbers, and are therefore potential false linkages from our type of probability linkage. 125 of the 628 had CHI which do not appear in the CHI data set, so these must be errors on the original SMR01 record. 503 (628-125 entries) false positives might have arisen in the probability linkage process. Applying the worst assumption that all the 503 CHIs derived from the linkage were incorrect and the CHIs as applied by the hospitals were correct, this implies an upper limit to the false positive linkage rate of 0.08 %.

Quality of overall linkage of census and CHI records

Myocardial infarction incidence and survival

An age-adjusted Kaplan-Meier plot of survival by sex and ethnic group is shown in figure 2. In a Cox regression model the hazard ratio among South Asians compared to non-South Asians was 0.50 (95% CI 0.37, 0.68). After adjustment for age, sex and any previous admission for diabetes the hazard ratio was 0.59 (95% CI 0.43, 0.81). With those over 75 years of age excluded (associations between disease and social economic inequalities are least in the elderly) the hazard ratio was 0.68 (95% CI 0.46, 1.01). These results were unaltered when subjected to checking procedures.

Ethics and methods

We surmounted formidable technical, ethical, administrative and legal challenges to link census, CHI, hospital and mortality data for analysis of the relationship between ethnicity and disease. It is vital to demonstrate confidentiality, and in particular to address the public's concerns. The encryption ensured records could be linked in an unidentifiable form, but our security procedures provided additional safeguards. It is feasible for someone to try out a large number of possible CHI numbers to look for the corresponding encrypted code. This can be avoided by using extra encryption [20].

There is an obligation to ensure that potential ethnic inequalities are highlighted so that they can be assessed and addressed [1, 17]. We had to balance individuals' right to data privacy (individual consent was impossible) and the potential benefits to society of producing information derived from potentially sensitive data [21]. We judge our methods struck the balance. The proportion of records ascribed an ethnicity code (94%) was slightly smaller than that normally achieved within ISD (typically around 98%). As the non-White ethnic minority population (about 2%) is comparatively small the false positive rate is critical. A rigorous matching methodology appropriate for administrative matching was, therefore, used. Census forms are completed by the public and processed electronically using optical recognition and keying from images. The success of these processes depends on legibility. The detailed spelling of a name, particularly if it is associated with a minority ethnic group, is more likely to be prone to error when transcribed by a third party onto NHS records, than in the census where the census informant writes it. Such errors could lead to varying accuracy of linkage by ethnic group. Nevertheless, we met our prior stated standard of 80% for every ethnic group. In future, we hope to examine matching rates at each stage of the linkage and by sex.

Acute myocardial infarction-strengths and weaknesses of the analysis

The major strength of this analysis is the population coverage with the inclusion of community and hospital deaths as the SMR 01 database contains over 99% of hospital admissions for AMI in Scotland (A Redpath, ISD, personal communication, 2006). Possible, as opposed to probable or definite, AMIs are coded as chest pain and are excluded from our figures. The quality of data on AMI in the database has been validated as reliable [19].

The diagnosis of AMI was based on criteria used by clinicians to make the diagnosis, however, which would be inconsistent over time. This should not, however, affect ethnic groups differently. If misclassification of diagnosis is non-differential then the differences would probably be underestimated, but if they were differential differences would be exaggerated. We do not have data to assess these options, and the anonymised methods precluded validity studies at this stage, but are a high priority for future work.

Linkage rates were slightly lower for South Asians than non-South Asians. If non-linkage occurred at random, as is likely, this would reduce the power of the study, but not bias the results. If those not matched were at different risk of AMI or death following AMI, from those who were matched, this would bias the results. We intend to explore this in future research. The limited information available to allow adjustment for co-morbidity did not explain differences in survival, and in future analyses, assuming ethical permission can be obtained, we would include more explanatory variables from the Census.

We did not separate Indians and Pakistanis (the dominant South Asian populations in Scotland), as desirable [11], because of lack of numbers, but this is technically feasible, and is planned for the future as events accrue. The pattern of cardiovascular risk factors varies among different South Asian populations [11], but the excess mortality is shared by all [15].

Findings on incidence of and survival from acute myocardial infarction in relation to previous knowledge

Few studies have reported the incidence of CHD among South Asians in the UK [15]. Our findings corroborate those of Tunstall-Pedoe and colleagues who reported a 30% (not statistically significant) increased rate of first and recurrent AMI among Asian (predominantly Bangladeshi) men in London [16]. The data accord with high CHD mortality in South Asian-born populations, compared to White populations [15, 22].

Studies of hospital admission rates for AMI in the UK have reported a similar [15, 23] or higher [24, 25] risk for South Asians, but these did not include community deaths. Wilkinson [26] reported a two-fold and Hughes a 4.9 fold increase in the incidence of AMI [27] – the latter result being an outlier. Britton et al reported a doubling of the incidence of fatal CHD or MI in 560 South Asian civil servants compared to White ones in the Whitehall 2 cohort study [28]. Our findings are consistent with the relatively high CHD mortality in those born in the Indian subcontinent around the 1971–2001 Censuses [1, 10, 29–31] and the relatively high CHD prevalence [11, 12, 32].

As survival from AMI was comparatively good, our data indicate that the high rates of CHD mortality arise from higher incidence rather than from higher case fatality. Previous studies of fatality after AMI in South Asians were on hospital cases [33]. Muhktar found no significant difference in survival up to 4 years after discharge in 111 South Asians and matched non-South Asian controls [34]. Studies agreeing with this conclusion include those by Lowry et al in Birmingham [35], Hughes et al in London [27], Lear et al in Leicester [36], Liew et al in London [37], Gupta et al in Toronto, Canada [38], and Mak et al in Singapore [33]. Wilkinson et al's work is unusual in showing a substantially poorer survival in South Asians [26]. Lawrence and Littler's data showed better survival in 124 South Asians in Birmingham compared with matched White people [39]. Unpublished data from the UK national audit (MINAP) show 'Asians' were substantially less likely to die from AMI than 'Caucasians' (Zaman, M. J., lecture, London School of Hygiene and Tropical Medicine, 12/1/07).

Our results of better survival are unlikely to be chance or analytic errors. How might we explain them? Were South Asians diagnosed at an earlier stage of their disease than the rest of the population? With the recent emphasis on heart disease in South Asians this is possible. South Asians may have been quicker to seek medical attention as suggested by Chaturvedi [40]. Although South Asians may present promptly to hospital they are more likely to encounter delays between arrival and treatment as shown in some (but not all) studies in the UK [41] and Canada [38]. South Asians may have quicker access to emergency treatment in Scotland, particularly in Glasgow and Edinburgh (where most of them live), because of the proximity of their inner-city residences to major hospitals. South Asians are less likely to get revascularisation as shown by Feder et al in London [42].

Alternatively, non-South Asians, who are mainly White Scottish people, may have particularly poor survival in Scotland while South Asians in Scotland have average survival. There is evidence for this. Our mortality patterns in South Asians are similar to those reported by Mak et al in Singapore [33] and compatible with the in-hospital mortality figures in studies by Mukhtar et al [34], Lowry et al [35], Lear et al [36], and Gupta et al [38]. Macintyre et al [43] have shown an association between social and economic deprivation and survival in Scotland, which might account for Scotland's relatively high mortality. This still leaves the question of why this might not apply to South Asians. In our results, adjustment for deprivation accounted for some but not all of the difference in survival between South Asians and non-South Asians.

The greater risk of AMI among South Asians but better prognosis needs further research including on the possible occurrence of AMI at an early stage in the progression of atheroma [44]. Anatomical differences in the coronary artery vasculature in South Asians [45] disappear when populations are matched closely, and are unlikely to explain the high risk[46].

What was known before

What is known now?

Ethical and legal obstacles to linking, retrospectively, ethnic codes in the UK census to NHS health and mortality databases are surmountable

The method has international implications and mandates review of current policies on prospective coding by ethnic group