Excess mortality in the first COVID pandemic peak: cross-sectional analyses of the impact of age, sex, ethnicity, household size, and long-term conditions in people of known SARS-CoV-2 status in England

Study overview

This study used an observational cohort study design. Three main analyses are presented. First, the peak in mortality associated with the first SARS-CoV-2 peak is reported. The sentinel network mortality in the Oxford RCGP Research Surveillance Centre in 2020 is compared with 2019 for the same weeks reported by the UK Office of National Statistics (ONS).

Second, a relative survival analysis is conducted, comparing the mortality for 2020 with 2019, and estimating excess mortality across the whole population using ONS mortality rates for 2019. Finally, the association between SARS-CoV-2 status, demographic and clinical risk factors, and mortality is explored. Odds ratios (ORs) are estimated for all-cause mortality, and the modifying effect of age, sex, SARS-CoV-2 status, and household size examined. Models were further adjusted for ethnicity, socioeconomic status, smoking status, and underlying health conditions. The study data were collected from weeks 2–20 of 2020.

Setting and participants

The study population included all patients registered at general practices in the Oxford RCGP RSC network on 11 May 2020 and having ≥1 year of health records in the network (n = 4 413 734). The network extracts pseudonymised data from primary health care electronic records of member practices and is recruited to be nationally representative (see Supplementary Figures S1 and S2 for details).¹² Data include demographics, clinical conditions, medications, and laboratory results. The network also reports on mortality (Supplementary Figure S3).

Study variables

The main outcome of interest was all-cause mortality, obtained from primary care CMRs over the entire period of analysis: weeks 2–20 of 2019 and 2020 (7 January–19 May 2019, and 6 January–18 May 2020). Mortality data were obtained from a combination of coded data entered into the clinical record to indicate that the patient had died, and examination of patients who had been removed from the practice list by the national demographic service, which flags those who have died. Where available, the coded date of death was preferentially used.

The primary variables of interest included living in communal dwellings, SARS-CoV-2 exposure, socioeconomic and ethnic inequalities, and also learning disabilities. People were grouped into the same household based on having identical addresses, and households were divided into those with 1, 2–4, 5–8, and ≥9 residents. Residences with ≥9 people were described as communal establishments. This matching was done programmatically at data extraction without researchers having access to personal addresses. SARS-CoV-2 status was classified at four levels:

• definite case — supported by a positive virological test result;

• probable case — based on positive clinical code in the absence of a test;

• possible case — a code suggested testing for SARS-CoV-2 (but no result), clinical suspicion, or contact; and

• not a case — people with negative test results (see Supplementary Tables S1a and S1b for coding lists).

The SARS-CoV-2 status algorithm worked hierarchically, so if probable or possible cases subsequently had a negative test they were reclassified as not a case.

Other variables included:

• age;

• sex;

• socioeconomic status using the index of multiple deprivation (IMD), based on lower super output area (LSOA) — a geographical subunit with a minimum population of 1000 — divided into quintiles;¹³

• ethnicity divided into white, Asian, black, mixed, and others;¹⁴ and

• household size; determined using a pseudonymised household key based on identical address, this has been used in other studies.^10,15,16

• smoking status (comparing current, ex- and non-smokers);

• obesity (using the World Health Organization categorisation of overweight [body mass index {BMI} = 25–29 kg/m²]; obese [BMI = 30–34 kg/m²]; and severely obese [BMI≥35 kg/m²]; and

• population density (based on ONS locality data).¹⁷ The highest population density was in ‘conurbations’, medium levels in ‘city and town’, and lowest density in ‘rural areas’.

The following disease groups or clinical risk groups that might be associated with adverse outcomes were added in case codes as surrogates for SARS-CoV-2 exposure: upper and lower respiratory infections (URTI and LRTI, respectively), Type 1 and type 2 diabetes mellitus, hypertension, chronic kidney disease (CKD) defined as stage 3–5,¹⁸ heart disease (including myocardial infarction, other forms of coronary artery disease, and heart failure), chronic respiratory disease (asthma, chronic obstructive pulmonary disease, bronchiectasis, and other chronic lung conditions), people undergoing treatment for cancer or who may be immunocompromised due to taking medications for inflammatory conditions, and people with learning disability (Table 1).

Statistical analysis

Trends in mortality for weeks 2–20 of 2020 were analysed using descriptive statistics, and excess deaths due to COVID-19 modelled using a relative survival model.^19,20 In a relative survival model, the observed mortality rate in a cohort is compared to the age and sex specific mortality rates from a reference population. Assuming that the reference population mortality (that is, the background mortality from 2019) closely approximates what would have occurred in the absence of the disease, the difference between the observed and expected mortality will be close to cause-specific mortality.

The date of death as recorded in the primary care record was used. Each model was adjusted for age, sex, and household size. To ensure better model fit, age was examined as a continuous variable in the relative survival model, using a cubic polynomial term. The latest UK life expectancy tables were downloaded from the ONS via the Human Mortality Database maintained by the Max Planck Institute for Demographic Research.²¹ The life expectancy tables were imported as rate tables into the relative survival model. Current SARS-CoV-2 exposed (in weeks 2–20 of 2020) survival data were compared with these rates, enabling exposed survival to be measured relative to the counterfactual, exposure-free expected survival.²² Relative hazard ratios (RHR) and 95% confidence intervals (CI) were determined.

In patients with available SARS-CoV-2 status data, a multivariable logistic-regression model was fitted to examine the effect of age, sex, SARS-CoV-2 infection status, and household size on mortality. Further fully adjusted models examined these variables along with ethnicity, patient demographics (socioeconomic status and population density), smoking status, and underlying health conditions, including learning difficulties. Multiple imputation by the chained equations method was used (using all model covariates in the missingness model, including outcome but with no auxiliary variables) to impute missing data, imputing five datasets using predictive mean matching.²³ Each dataset was inverse probability weighted using an iterative proportional fitting algorithm to match the marginal covariate distributions of each imputed dataset to the full RCGP RSC population margins.²⁴ Outputs were employed in final multivariable, weighted logistic regressions.

Finally, each of the regression coefficient estimates, together with robust sandwich variance estimators, were pooled using Rubin’s rules.²⁵ All analyses were undertaken using R (version 3.5.3). A complete cases analysis was conducted as a sensitivity analysis. Both models are reported using ORs with 95% CIs.