Cross-sectional study

We identified 2,236,001 patients from the primary care EHR who, on 21st March 2021, were currently fully registered with one of the GPs in the selected geographical areas. We excluded people who had died before or were more than 100 years old on 21^st March 2021, those residing outside the selected geographic area, and those without a high-quality pseudonymised unique address identifier RALF (Residential Anonymised Linkage Field) or residing in commercial properties. This resulted in a population estimate of 2,130,965 people in residential households or communal establishments for comparison with other data sources. This sample of 2,130,965 people was used to derive the population estimates and for all analyses comparing EHR population estimates with those from Census 2021 [1, 2] and Statistical Population Dataset [7] population estimates.

We created the second sample used for the analysis of households. This was done by excluding people living in communal establishments as those do not count as residential households according to the household definitions in all data sources.

We used Unique Property Reference Numbers (UPRN) [21], assigned to the GP-recorded alphanumerical addresses using the ASSIGN (AddreSS MatchInG to Unique Property Reference Numbers) address-matching algorithm [13] and pseudonymised with Open Pseudonymiser [23] into Residential Anonymised Linkage Fields (RALFs) [37] using a study-specific SALT or encryption key. We identified residential households by linking people with the same RALF derived from their registered patient address. We excluded people living in communal establishments. We included all households irrespective of their size. The final primary care residential households sample comprised 658,264 households containing 2,108,818 people (Figure 1). This sample was used to derive the household estimates and for all analyses comparing the estimates of the households obtained from EHR with those from Census 2021 [1, 2] and Admin-based occupied address dataset [21].

We categorised the ages of household members on 21^st March 2021 in whole years by ten-year intervals, and – for the Bland-Altman method [38, 39] – five-year intervals. We defined sex as a categorical binary variable (male, female) using the recorded value in the EHR. We categorised household sizes in two ways: firstly, to align with ONS Census 2021 definitions (sizes of one to seven with increments by one, and eight or more people), and secondly, to align with ONS ABOAD definitions (sizes one to four with increments by one, and five or more people). We used the 2019 Index of Multiple Deprivation (IMD) [40], the official measure of relative deprivation in England, produced for LSOA-level geographies [41] to categorise households into five quintiles, with quintile 1 being the most, and quintile 5 the least, deprived. We categorised people according to the presence of long-term conditions recorded in the primary care EHR (Supplementary Table S2). A borough/local authority is an administrative unit for local government with the following seven categories: Barking and Dagenham, Hackney, Havering, Newham, Redbridge, Tower Hamlets, and Waltham Forest.

We compared EHR estimates of the population of people in residential households and communal establishments, (n = 2,130,965 people; Figure 1), with Census 2021 population estimates by sex, age, local authority, and IMD quintile, and EHR household residential household estimates (n = 2,108,818 people in 658,264 households; Figure 1), with Census 2021 household estimates by number, size, and local authority. We defined an overcount and undercount in populations and households when the EHR, SPD or ABOAD estimates were respectively greater or less than the estimates derived from the Census. We estimated the 95% Limits of Agreement (LoA) between EHR and Census 2021 household and population estimates using the Bland and Altman method [38, 39]. This method is used to assess the level of agreement between two different methods of measurement, when the true values are unknown. It calculates the mean difference and standard deviations of the difference, which gives the range within which 95% of differences are expected to lie. We estimated the 95% LoA for a log-transformed number of people and a number of households by Middle layer Super Output Area (MSOA) – a geographical boundary representing between 2,000 and 6,000 households and a resident population between 5,000 and 15,000 persons [42] - and for log-transformed number of females and males by age group.

We performed sensitivity analyses by excluding people with no record of a recent clinical encounter with their GP 12, 36, or 60 months before 21^st March 2021, and examined the effect on levels of agreement in population estimates by age. Clinical encounters included GP consultations (in person, telephone, video and email), clinic consultations, administrative encounters (notes, letters, documents, reports and other), case reviews and repeat prescriptions. This included consultations during the COVID-19 pandemic. We tested the relationship between the clinical activity in the last 12, 36 and 60 months and a number of long-term conditions (LTCs).

We compared EHR, SPD and Census population estimates by sex and age, and EHR, ABOAD and Census household estimates by local authority and household size.

Overall, the population estimate of people living in residential or communal establishments in NEL derived from primary care EHRs was 2,130,965, 7.1% greater than the Census 2021 estimate of 1,990,087 people (Figure 2). The difference in estimates varied by local authority, being lowest in Havering (2.2%) and greatest in Newham (13.2%).

Overall, agreement in population estimates was closer for the female (3.6%) than for the male population (10.7%) (Figure 3, Table S3, Appendix 1). This varied by age group, with people aged 30-39 years of either sex showing the greatest (female: 9.6%; male: 25.5%), and people aged 20-29 years the smallest, difference (female: 0.8%; male: -0.3%) (Figure 3).

Overall, the difference between EHR and Census 2021 estimates lay between -3% and 9% for 95% of the age groups for the female population and -8% and 27% for 95% of the age groups for the male population, using the Bland and Altman method (see Appendix Figures S1a and S1b).

Agreement between EHR and Census population estimates varied by deprivation quintile, with the greatest difference seen for the population living in the two most deprived quintiles, where the majority (68%) based on the EHR population estimate lived (Figure 4). In 95% of MSOAs, the difference between Census 2021 and EHR estimates lay between -9.3% and +24.0% (Figure S2).

EHR of residential households estimates were on average 9.1% lower than Census 2021 estimates, ranging from -6.3% in Barking and Dagenham to -10.9% in Hackney (Figure 5).

The majority (78%) of EHR-derived households comprised fewer than four members, with 3.0% and 4.7% respectively being households of seven, or eight or more, members. EHR estimates of household size were lower than Census 2021 estimates across all household sizes for households of four or less, and greater than Census estimates for households with five or more, members (Figure 6), with the greatest percentage overcount observed for household sizes of seven and eight or more members. This was consistent across all local authorities (see Appendix Figure S4).

The population estimate from EHR was on average 11.6% lower than the Census 2021 estimate when people with no record of a recent clinical encounter with their GP in the 12 months prior to 21^st March 2021 were excluded (Figure 7). This was most marked in the 10-19 and 20-29 year age groups. EHR population counts were slightly higher than Census 2021 estimates when excluding people with no clinical activity in the preceding 36 (1.3%) or 60 months (4.3%), with less marked variation by age (Figure 7, Appendix Table S4).

A record of a recent clinical encounter with a GP in the 12, 36, or 60 months prior to 21^st March 2021 was much higher for adult patients (≥18 years) with a long-term condition than those without, suggesting that exclusion of people without a record of a recent clinical encounter would exclude many patients without long term conditions. (Figure 8).

The population estimates derived from EHRs were similar to those derived from SPD and both were higher than the Census 2021 population estimates overall (Figures 9 and 10, Supplementary Table S3) and higher for males than for females (males: 10.7%, 7.9% respectively; females: 3.6%, 2.7% respectively).

Both EHR and ABOAD demonstrated similar levels of agreement with Census 2021 in the number of estimated households/occupied addresses (ABOAD -7.3%; EHR -9.1%) (Figure 11). Relative to the Census, both ABOAD and EHR household sizes showed a similar pattern of undercounting households of four people or less and overcounting households of five people or more (Figure 12, Supplementary Table S5).

We developed a dynamic method of deriving population and household estimates from primary care EHRs in an urban, multi-ethnic and disadvantaged London region. Population estimates derived from EHRs on the Census date were 7.1% higher than those derived from the Census, and this was most marked for men aged 30-39 years. The highest agreement between estimates from EHRs and the Census was in the people aged 20-29 years, but this close agreement may be partially attributed to the overcount of people with incorrect GP registrations (e.g. those who left the area without registering), being balanced out but the undercount of local residents who did not register at GP as they did not need to use GP services. Overall, there was better agreement in the estimated female than male population. In this highly disadvantaged region, EHR population estimates were closer to Census estimates in the less deprived areas. Exclusion of people with no recorded clinical activity in the EHR did not reduce these observed differences. In contrast, household estimates derived from EHRs were around 9% lower than those derived from the Census, especially for households of four or fewer, which comprise the majority of households in this region.

There has been a considerable effort to evaluate the accuracy of population and household estimates derived from administrative data as a supplement for the ONS decennial Census which is less frequent and more costly. Our relatively simple method of estimating populations and households produced estimates similar to those derived from the SPD and ABOAD datasets both of which include linked data from multiple administrative datasets, including from the NHS PDS which records addresses of all NHS registered patients in England.

We compared the population counts obtained in this research to the estimates produced on 1^st April 2021 from the primary care registration database held by the NHS Application and Infrastructure Services system (NHAIS) [43]. This system, which is being decommissioned in 2024, holds a snapshot of patients registered with GPs for each NHS region in England and is used primarily to process payments to GPs based on list size. While quality-assured, it is recognised that its accuracy is impacted by many factors including NHS reorganisations, such as occurred in April 2021, and the impact of the recent COVID-19 pandemic on patient registrations and patient list updating by GPs [44]. In 2021 our EHR population estimate of 2,130,965 people (Figure 1), was overall 7.6% lower than that of 2,304,999 people derived from the NHAIS dataset provided in the Spotlight report [45] for the equivalent geography, but much closer (-3.0%) when comparing the number of currently fully registered patients by the local authority (2,236,001 patients compared to 2,304,999 respectively). The NHAIS database has also been compared to Census 2021 estimates, which suggested a much higher discrepancy of (-15.8%) between the Census 2021 and NHS list with respective derived populations of 1,990,200 and 2,304,999 for the NEL area [45]. Much of this discrepancy was attributed in this report to GP list inflation related to patients moving out of the area and failing to de-register or be de-registered by their GPs. We were able to examine the effect of recency of clinical activity on population estimates and did not find these improved levels of agreement with the Census estimates to any extent. Furthermore, we took care to define residential households and their populations as closely as possible to the definitions used in the Census extracts. This may account for differences between the EHR estimates and Census being smaller than those found by the Spotlight report (7.1% versus 15.8% overcount respectively).

There are several factors in addition to those mentioned in the previous section which could affect the extent to which our population and household estimates for this region agreed with those reported by the Census 2021. These arise from methodological differences in defining the eligible populations and households, as well as wider sociodemographic factors affecting population stability and inclusion in the respective datasets.

The level of the alignment of the GP-registered population with the residential population for a given geography depends on the proportions of people who do or do not live in that geography and are registered or not registered with a GP providing services in that geography. We were careful to include only patients who resided in the NEL region and excluded those with addresses outside of the region, though we did not have information on those living in the region and using GP services outside of NEL.

Differences may also arise due to the lack of comparable definitions of populations and households in the different data sources. Census 2021 data are collected primarily to obtain population statistics while patient registrations are recorded in the EHR to provide health care. Given these differences, the Census 2021 definitions of population and households cannot be directly applied to EHR data. While we were not able to test the intended length of stay at the address to confirm the usual residence rule within the EHR database, we excluded patients with temporary registrations. We were also unable to identify rough sleepers from the primary care data.

The Census 2021 definition of a household as “one person living alone, or a group of people (not necessarily related) living at the same address who share cooking facilities and share a living room or sitting room, or dining area” cannot be verified in the EHR. We identified households by assuming that people who share the same RALF at the same point in time comprise a household. This in turn depends on the ability to assign a UPRN to the recorded patient address. The primary care records in the Discovery Data Service (DDS) allocate UPRNs in real-time using the validated ASSIGN algorithm [46]. We excluded 2.4% of patients (53,757 out of 2,236,001) in our population due to no or a poor quality match of their address to a UPRN (Figure 1).

Differences may also arise as a consequence of non-participation in the Census and statistical assumptions and methods used to adjust for non-response. Almost all local authorities in NEL had Census 2021 response rates lower than the average for England and Wales (97%) and three had response rates lower than the average for London (95%) including Hackney (93%), Newham (94%) and Tower Hamlets (94%), with levels of response especially low in younger people aged 20-24, and in Tower Hamlets (90%), in Redbridge (93%) and in Hackney (91%) [47, 48]. There was also uncertainty about the number of households estimated from the Census 2021 in Tower Hamlets, Barking and Dagenham and Waltham Forest, with higher estimates than expected when compared to alternative sources [49]. Residential mobility might explain some of the variation in levels of agreement by the local authority observed in our study. Data from the Trust for London suggests that the inner east London boroughs had among the highest levels of residential mobility in London between 2015 and 2020 [50].

Our findings when we compared EHR estimates to those derived from multisource ONS administrative datasets showed similar levels of agreement with Census 2021. Earlier work combining local administrative datasets with EHRs suggests that multiple administrative data sources may improve enumeration [14]. However, this requires complex linkage which may introduce errors or biases due to linkage and is more expensive than a single source. These biases may arise from issues in the data or the matching algorithms and may result in missed or false matches, or over- and under-representation of some groups in linked data. These biases may be hard to measure and, even when quantified, may not provide information on how linkage bias affects the analyses using linked data [51, 52]. Deriving the estimates from the single EHR source eliminates the risk of linkage bias. A further advantage of the single EHR source is that only pseudonymised identifiers are used to identify individuals and households.

Local authorities need to have up-to-date estimates of their populations and households for planning purposes. For example, between the Census 2011 and Census 2021, the population of Tower Hamlets increased by 56,200, an increase of 22.1% [53]. Primary care EHRs provide a suitable, simple and dynamic method to obtain these estimates between decennial Census dates with a high level of geographic and demographic granularity. As EHRs also include health information, estimates derived from EHRs can provide detailed information on the health status of populations for local authorities and researchers. The method is cost-effective as it utilises already collected pseudonymised data, is simple to use, and can be used to produce estimates in near-real time, including information on household age, sex and ethnic composition. Regional primary care datasets such as the DDS have the potential to be used to create such estimates for local authorities and researchers.

Strengths of our study include the use of a high-quality, validated algorithm [15] to allocate UPRNs to patient addresses, enabling encryption and safe use of address data for research without compromising patient privacy and confidentiality. We used a published and transparent method to identify household members at a point in time through the shared address identifier RALF [54]. We compared derived population and household estimates in an ethnically diverse mobile population with high levels of deprivation and lower than average Census participation which may limit generalisability to other geographies. We tested the effects of list inflation through sensitivity analyses. Our assessment of patients’ engagement with general practice included administrative encounters (for example letters) which could be associated with patients who are still registered with the GP but who no longer reside in the area. However, we were unable to disaggregate these from face-to-face phone or video consultations. While we could not create definitions of populations and households that were entirely comparable to those used in the Census, we excluded patients who were temporarily registered and excluded patients with low quality or no UPRN matches. We were only able to compare our estimates with Census estimates using aggregated numbers rather than on an individual person-level basis, limiting a more granular analysis of the differences between the two sources of data.

Further analyses to compare the Census and EHR estimates using individual-level data are planned as part of the ADRUK-funded Healthy Households project [55]. These will overcome the limitations of using aggregate data and will include methods that estimate the household population size from independent sources. The generalisability of our findings in NEL using the Discovery Data Service could be tested in other regions with similar regional primary care datasets. This would enable lower cost dynamic population and household estimations for planning purposes.