Data source and study date

This was a cross-sectional study using CPRD data linked to HES data and deprivation data (IMD 2010) in England. It investigated the historical recording of social factors among individuals aged ≥65 years, actively registered with a CPRD practice on a randomly chosen index date (1st January 2013), to allow assessment of both completeness and timeliness of recording of social factor data. Active registration on 01/01/2013 was determined by ensuring that patients’ start dates (the later of their registration date with the practice or the date the practice reached CPRD-defined quality criteria[21]) fell before the index date and their end dates (the earliest of their transfer out date, date of death or practice last collection date) were after the index date.

CPRD data are supplied in ten different files,[19] of which eight (patient, practice, clinical, consultation, additional clinical details, immunisation, referral and test files) were used for this study (S1 Table). These files include information about patients’ demography, lifestyle factors, clinical details, feedback from secondary care, therapy and laboratory results, stored in form of medical, therapy and other codes used by the GP practice staff.[21]

Social factors examined

In this study, social factors relevant at an individual level and informed by the conceptual framework of the WHO’s CSDH,[1] were examined in CPRD and included: religion, ethnicity, immigration status, deprivation based on LSOA of each individual’s residence,[23] living arrangements (living alone and cohabitation), residence (place of residence and homelessness) and marital status.

Lists of medical codes (S2 Table) for each factor were compiled by searching the CPRD’s Read code dictionary [21] for specific and broader text terms (using wild card searches) encompassing all social factors of interest. This was an iterative process that subsequently included a hierarchical search of the Read codes identified. The number of codes identified for these factors ranged from 86–465 (S2 Table). Further information (S1 Table) was accrued from other sources within the dataset as follows: the consultation files provided codes (‘consultation type’) on where the consultation took place and thus patients’ residence (for example in a care home), while the patient files provided information regarding patients’ marital status and their family number.[29] The latter variable can identify individuals sharing the same household and therefore can be used to get information for living arrangements (living alone, cohabitation), marital status and care home status. Similarly, the additional clinical details files provided coded information (‘entity type’) about residence, living alone and marital status. The linked hospitalisation data from HES provided additional information for ethnicity and residence, whilst the deprivation data provided deprivation scores for individuals’ LSOA as IMD quintile. The multiple code lists thus generated were discussed amongst the three of the authors (AJ, SLT and AJvH). These code lists were then utilised to systematically search for the Read codes in the clinical, immunisation, referral and test files. Additional information was sought from consultation type and entity types in the consultation and additional clinical details files respectively and also from the patient file and from linked HES and deprivation data. Some factors also provided information about another social factors: for example an individual coded as living alone was deemed not to be cohabiting, whereas an individual residing in a care home was considered not to be living alone.

The following example illustrates how information for social factors was assimilated. Type of residence (whether a patient lived in their own home, in sheltered accommodation, or in a care home) can be recorded in numerous way in both CPRD and HES. In CPRD, this information can be determined using the medical codes within multiple files as described above, using the entity type 132 for residence in the additional clinical details file, from the consultation file (e.g. “nursing-home visit”) and from the family number (as described below); residence data are also potentially available in HES by using information about individual’s location prior to hospital admission.

Exposure variables definition and categorisation

The code lists for the social factors of interest (religion, ethnicity, living arrangements (including living alone and cohabitation), immigration status, deprivation, residence (including place of residence and homelessness) and marital status) are presented in S2 Table. Ethnicity codes were those recommended for use by the Quality and Outcomes Framework, as used by Mathur et al.[24] Family number was used to derive additional information by modifying approaches used in previous studies,[27, 28, 30] as follows. Two adults, living in a household size of two or three, were identified as cohabiting (adults living in a couple) if the age difference between the couple was ≤15 years and age difference between the other household occupants and those living in a couple was >15 years. Couples identified as cohabiting were also allocated `partner-uncategorised’ category for marital status. Individuals from household size of two or more were identified as not living alone. Based on previous studies [26, 31, 32] care home was defined as a household with >3 individuals aged ≥65 years and if their total count was more than individuals aged <65 years. In sensitivity analyses households with >3 individuals aged ≥65 years and ≤3 individual aged ≤50 years were defined as a care home.

Religion was categorised into eight categories (Buddhists, Christians, Hindus, Jews, Muslims, Sikhs, Others and no religion (atheists)) to ensure comparability with Census data.[33] We hypothesised that certain minority religions might be more likely to be coded by GPs, and explored this by categorising one religion (Muslim) as a binary (yes/no) variable. Ethnicity was categorised in five groups: White, South Asian, Black, Others and Mixed as per the UK 2011 Census.[34]

Living alone and cohabitation were coded as binary variables (yes/no). Immigration status, a binary variable (immigrant/ not immigrant) was defined using: i) country of birth information and (to increase completeness of ascertainment) ii) codes for the first language spoken (S2 Table).

Place of residence had four categories: living in a care home, sheltered accommodation, other places of residence (e.g. prison, hospice, hostel, welfare home) and living in a household. Care home status was also considered as a binary (yes/no) variable, on the assumption that being in a care home might be more completely recorded by GPs than other places of residence (e.g living in a household). Homelessness was also a binary variable.

Relationship status was characterised by using following seven categories: single, married/civil partnership, widow/er, divorced, separated, partner-other (e.g. common-law husband/wife) and partner (uncategorised). As the last category was non-specific, an algorithm was developed to obtain more specific marital status information. If the ‘partner uncategorised’ status was preceded by any of the following three categories: 1) Single/engaged 2) Married/ civil partnership and 3) Partner-other category, the ‘partner uncategorised’ category was updated to that of the earlier observation.

Deprivation status is a composite score of 38 indicators for seven domains of deprivation (income, health and disability, employment, education and training, housing, living environment and crime).[23] These indices are available at the small area level (LSOA) as quintiles: quintile one representing the least deprived to quintile five representing the most deprived.[23]

Analysis

For the purposes of recording, the social factors that were likely to change with time (e.g. marital status, living alone status) were treated as time-varying exposure variables whereas ethnicity, religion and immigration status were deemed to be time-invariant.

In CPRD the event date (the date the event occurred as recorded by the GP) was used to ascertain when the factor was recorded in relation to the index date. If the event date was missing then the system date (the date when the event was recorded on the GP system) was used for these observations.[19] For information extracted from the patient files (such as marital status, family number), which does not include event dates, a conservative estimate of the date of recording was taken, using the date the patient registered with the practice[19], and the hospital admission date was utilised for HES data.[22]

All mentions of each factor of interest were identified within a patient’s linked records. Observations providing discordant information for a factor on the same date for a patient were excluded and the social factor recorded nearest the index date was used.

As family number provided information for social factors indirectly, and the date of recording this variable was unclear, information from family number was used only when data for a particular social factor were unavailable from other sources in CPRD or HES for that patient.

For each social factor, the following information was analysed:

Completeness of recording for individual social factors, and contribution from linkages

Completeness of recording for all social factors was better for females and amongst the oldest individuals (aged ≥85 years) for all factors except for religion, immigration status and IMD score (Table 1). Of the seven social factors ascertained, recording for deprivation data and ethnicity were the most complete, at ~82% (n = 486,426) and ~80% (n = 469,557) respectively (Table 1). The recording of ethnicity over time showed an increase in completeness in the year 2006 (when incentivisation was introduced) with a slight downward trend in 2011 and 2012 (Fig 1).

Download:

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

Fig 1. Patients with ethnicity records in Clinical Practice Research Datalink and Hospital Episode Statistics over time.

Abbreviations: CPRD Clinical Practice Research Datalink HES Hospital Episode Statistics.

https://doi.org/10.1371/journal.pone.0189038.g001

Download:

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

Table 1. Proportion of individuals with information on social factors available in Clinical Practice Research Datalink and Hospital Episodes Statistics: Age, sex and database distribution (N = 591037).

https://doi.org/10.1371/journal.pone.0189038.t001

The most incompletely recorded social factor was immigration status which available for only 4,187 (0.7%, data not shown) of the study population when country of birth codes were used alone. However, the additional use of ‘first language’ codes with country of birth codes more than doubled the information, to 1.6% (n = 9,713) of the study population (Table 1).

Religion was the second most poorly recorded factor, available for only 2.6% (n = 15,449) of study individuals (Table 1). Data on place of residence was recorded for 10.3% of the population, whereas living alone (yes/no) and marital status were recorded for nearly a third of the study population (29.2% and 27.2%, respectively).

The contribution of data from linked datasets to completeness of recording was particularly important for ethnicity, which showed a ~45% improvement (increasing from ~55% to ~80%, Table 1) after including linked hospitalisation data, and for IMD data (which was only available as linked data). For other social factors, there was hardly any evidence of improvement in completeness of recording from the linked data compared to using CPRD alone (Table 1).

The utilisation of family number in providing information for individuals who had no data for living alone, cohabitation, care home residence and marital status in either CPRD or HES showed that there was much higher completeness of recoding for living alone (70% versus 29%), cohabitation (60% versus 22%) and marital status (60% versus 27%) when information from family number was included (S3 Table, S4 Table and S5 Table). In contrast, family number contributed little to the completeness of recording of care home residence (11% versus 10%), irrespective of definitions used (S4 Table).

Representativeness

Amongst those with ethnicity data available, (Table 2), White ethnicity was recorded for the majority (~95%) and the ethnic composition of the study population was comparable to the English Census data[34] (Fig 2). In contrast, amongst the small number of individuals with available data on religion, 85% (n = 13,074) were recorded as Christians (Table 2), with an over-representation of the minority religion categories in CPRD (Table 2) compared to Census data[33], for example Muslim (3.1% in CPRD versus 1.3% in the Census), Hindu (2.5% versus 0.8%), Jewish (1.6% versus 0.7%) and Sikh (1.1% versus 0.4%). When Muslim religion was considered as a binary variable, using the entire study population as the denominator and assuming those without a code were non-Muslim, there was appreciable under-recording of Muslim status (n = 481, 0.1%) compared to English Census (1.3%).[33]

Download:

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

Fig 2. Comparing ethnicity recording (denominator: Those with available data) in electronic health records with English Census 2011.

Abbreviations: CPRD Clinical Practice Research Datalink HES Hospital Episode Statistics.

https://doi.org/10.1371/journal.pone.0189038.g002

Download:

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

Table 2. Social factors in Clinical Practice Research Datalink: Recording and categorisation.

https://doi.org/10.1371/journal.pone.0189038.t002

Similarly, among those with data on immigrant status, there was marked over-representation of immigrants (n = 7,866, ~81% of the total) among those with recorded data (Table 2), but under-representation when immigrant status was considered as a binary variable (1.3% of the total study population (Fig 3) compared to 9.9% non-UK born individuals in the English Census).[35]

Download:

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

Fig 3. Comparing recording of immigration status, care home residence, not living alone, cohabitation and religion in electronic health records and English Census 2011.

Abbreviations CPRD Clinical Practice Research Datalink HES Hospital Episode Statistics FN family number.

https://doi.org/10.1371/journal.pone.0189038.g003

For living arrangements, amongst those with available data, the proportion of individuals recorded as living in a household (~50%, Table 2) was under-reported in CPRD compared to English Census data (in which ~96% of people aged ≥65 years were recorded as living in household) and living in a care home was over-reported (~48%) compared to Census (3.2%).[36, 37] However, once care home residence was categorised as a binary yes/no variable, representativeness improved markedly; in the total study population, 4.9% of individuals were categorised care home residents compared to 3.2% in the English Census data (Fig 3).[37]

The data from EHR for marital status amongst those with non-missing data were also not comparable to the Census data, [38–40] with 68% being recorded as married or in a civil partnership, compared to 55.9% in the Census data. Data were comparable for the sub-categories of: ‘single’ (4.5% versus 5.5% in the Census) and ‘separated’ (1.3% versus 1.2%), but there were small number of individuals in both these sub-categories, making it difficult to draw any conclusions.

The number of individuals with a code indicating that they were homeless was also very small, representing just 0.02% (n = 92) of total study population. There were no corresponding data in the 2011 Census, but data for statutory homelessness and homelessness prevention and relief data (2013) from local authorities in England for individuals aged ≥65 years showed that the proportion of homeless individuals accepted for assistance was 0.01%,[41] providing a minimum estimate of the true proportion of homeless individuals (as not all would have been accepted for assistance).

Amongst those with available data, individuals categorised as those not living alone and as cohabiting were both over-represented in the data (96.7% and 88.6% respectively, Table 2). When considered as a binary variable using the entire study population as denominator, these factors were under-represented (28% and 19% respectively) compared to the Census data (68.5% and 58.6%, respectively).[39, 40] However, when information from family number was added, the percentage of those not living alone (68.9%) or cohabiting (52.2%) were fairly comparable (68.5% and 58.6% respectively) to the Census data (Fig 3).[39, 40]

For deprivation (Table 2), the data showed a slightly lower proportion of study population from the two most deprived quintiles of IMD status, suggesting that older patients in the practices consenting for linkage with deprivation data tended to be from more affluent areas. This is in contrast to a previous study which suggested that overall, including patients of all ages, those in linked CPRD IMD data are comparable to the UK population.[42]

Timeliness

The recording of time-varying social factors in relation to the index date varied considerably (Fig 4). Amongst those who had information available, 34.7% of individuals had data on whether they lived alone recorded within 5 years of the index date if data from family number was not included, but this decreased to about 20% if family number data were also considered (Fig 4 and S1 Fig). The equivalent percentages for marital status, without and with family number data were 19.5% and 13.5%, respectively. (Fig 4 and S2 Fig). Little difference (58.8% versus 59.5%) was observed for recording of residence within this defined period for analyses including and excluding family number (Fig 4 and S1 Fig).

Download:

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

Fig 4. Timeliness of recording of living alone, cohabitation, marital status and place of residence.

*data from family number analysis excluded

https://doi.org/10.1371/journal.pone.0189038.g004

The equivalent figures for timeliness when the entire study population (n = 591037) was considered, varied from 3.7% for cohabitation status to 14.2% for living alone data (including use of family number, S3 Fig).