Setting and population
Our analysis examines the relative productivity of primary medical services in England. Our units of analysis are CCGs. Each general practice is a member of one CCG. Our analysis is conducted at this level, rather than at the level of general practices, since data on counts of appointments are aggregated at CCG level before publication. We excluded 6 (4.7%) CCGs representing 2.0% of registered patients due to incompleteness of appointment data (see Supplementary file
1).
Our analysis was conducted over two 9-month periods: April 2019 to December 2019, prior to the COVID-19 outbreak in the UK, and April 2020 to December 2020, the 9-month period following the outbreak. Some of the published datasets used to construct input and output variables were aggregated at the level of calendar quarters, constraining the selection of time-periods for analysis. We chose a 9-month rather than 12-month period because GP practice activity was substantially affected by the COVID-19 vaccination programme after December 2020.
Variables and data sources
Our data envelopment analysis used 4 input variables, the number of full-time equivalent (FTE) general practitioners (including partners, salaried GPs, trainees and locums), nurses, other clinical staff and administrative staff, and 5 output variables, the number of face-to-face and telephone appointments, home visits, secondary care referrals and prescriptions issued. Workforce, appointment and prescribing data was obtained from NHS Digital [
17‐
19]. Data on referrals to secondary care were obtained from NHS England [
20].
We regressed the resulting efficiency scores against 9 independent variables, selected to represent the size, and health needs of the registered populations, and the staffing skill-mix in each CCG and time period: (1) the number of registered patients, (2) the proportion of this population aged 65 years or more, (3) the mortality and (4) fertility rate, (5) the level of deprivation and (6) the prevalence of several long-term conditions, (7) the ratio of FTE GPs to nurses, (8) other direct patient contact staff and (9) administrative staff. Data on the number of registered patients and their age profile, the fertility and mortality rates were obtained from Public Health England [
21]. Deprivation was measured using the English Indices of Deprivation 2019 obtained from the Ministries of Housing, Communities and Local Government [
22]. Data on the reported prevalence of 20 conditions were obtained from NHS Digital: atrial fibrillation, asthma, cancer, coronary heart disease, chronic kidney disease, chronic obstructive pulmonary disease, dementia, depression, diabetes, epilepsy, heart failure, hypertension, learning disability, severe mental illness, obesity, osteoporosis, peripheral arterial disease, palliative care, stroke and transient ischaemic attack, and rheumatoid arthritis [
23].
Statistical methods
Data on staffing levels in GP practices are published on a quarterly basis. Over the 6 quarters of interest, a small proportion of practices failed to report the number of GPs (0.8%), nurses (2.6%), other direct patient contact staff (4.8%) and administrative staff (0.1%). These missing values were imputed by regressing the number of FTE staff by type against the registered population.
Data on the monthly count of appointments, published at CCG level, were adjusted to take account of the fact that not all practices reported data each month (3.6% missing), and a proportion (4.7%) of appointments were marked as having unknown appointment mode (face to face, telephone, video, or home visit) for each CCG. CCGs were excluded from the analysis if no information on the appointment mode was available, or if the appointment mode was unknown in more than 40% of appointments.
Data on staffing levels, prescriptions, and disease prevalence were sourced at the level of GP practices, and later aggregated to the level of CCGs using data on GP practice CCG membership from NHS Digital. Many CCGs underwent reconfigurations and mergers during the study period. All data were reframed into the latest CCG configuration using information on successor organisations from NHS Digital.
Data envelopment analysis (DEA) is a non-parametric, deterministic form of frontier analysis which can be used to estimate the relative efficiency of a set of decision-making units (DMUs), in our case CCGs [
24]. The method can accommodate multiple inputs and outputs and does not require prior knowledge about the relationship between these variables. DEA assigns an efficiency score between 0 and 1 to each of the DMUs, with a score of 1 meaning that the DMU is fully efficient, i.e. “none of its inputs or outputs can be improved without worsening some of its other inputs or outputs".
25 We use the output orientation of DEA, since we wished to estimate the additional outputs that could be delivered given the current input levels, and assumed variable returns to scale (VRS).
A window DEA technique was used to evaluate the DMUs over two time periods, by allowing the entities to be evaluated as different DMUs in each time period. Since data envelopment analysis is deterministic, it can be sensitive to measurement errors. A form of bootstrapping, described by Simar and Wilson, is used to find bias corrected efficiency scores [
25]. The change in a DMU’s productivity over time is found using the Malmquist productivity index [
26]. This is defined by distance functions which can be found using the calculated efficiency scores [
27].
Tobit regression was then used to regress the calculated efficiency scores against factors which may impact on efficiency. Tobit regression was used since our dependent variable, the bias corrected efficiency score, is right-censored. Independent variables were scaled to aid interpretation of the model coefficients. Our models were stratified by time period.
We used k-medoids clustering to assign a CCG to one of three groups based on its disease prevalence rates [
28]. These groups were labelled following a descriptive analysis of the results as (1) low disease prevalence, (2) high prevalence of strongly age-related conditions (e.g., chronic obstructive pulmonary disease, dementia, heart failure, osteoporosis) and (3) high prevalence of other conditions (e.g., obesity, severe mental illness, asthma, epilepsy)—see S1 table in the Supplementary file
1. The resulting assignment was used as a design (dummy) variable in our regression, along with the number of registered patients, proportion of patients aged over 65 years, deprivation, birth rate, death rate, FTE GP to nurse ratio, FTE GP to other clinical staff ratio and FTE GP to admin ratio.
All analyses were undertaken using R version 4.0.3 and the Benchmarking, Tidyverse, VGAM and Cluster packages [
29‐
33].