Study population
We obtained electronic health record data from patients registered with general practices contributing to the Clinical Practice Research Datalink (CPRD) during April 1, 2015, to March 31, 2016. The CPRD, a large database of anonymised electronic health records from UK primary care, contains patient-level data covering approximately 7% of the UK population [
14]. CPRD data have been validated extensively and are representative of the UK population in terms of age, sex [
14], and ethnic background [
15]. We included patients of any age if their records were acceptable for research purposes (a data quality indicator provided by CPRD) and were registered at practices with continuous high-quality data reporting (CPRD defined up-to-standard) [
16] at any time during the study period. We grouped patient data into their respective general practices.
The protocol was approved by the Independent Scientific Advisory Committee (ISAC) of the MHRA (ISAC protocol number 17_06R). Ethics approval for observational research using the CPRD with approval from ISAC was granted by a National Research Ethics Service committee (Trent MultiResearch Ethics Committee, REC reference number 05/MRE04/87).
Included tests
We examined 44 specific tests (28 laboratory, 11 imaging and 5 other miscellaneous tests). The tests were chosen because they are commonly used tests or included in the guidelines or in the Quality Outcomes Framework (QOF) (Additional file
1: "Extended included tests" section).
We grouped tests into their respective general practices, via their practice identification number. To avoid double counting, if the same code was recorded multiple times for the same patient on the same day, it was counted as only one test. Similarly, codes likely referring to the same test, or separate components of a single test (e.g. individual components of a full blood count), were grouped and counted as one test.
Statistical analysis
To identify which test was subject to the greatest between-practice variation in its use, we calculated, for each 44 tests, an unadjusted coefficient of variation (CoV) and then an adjusted CoV.
To calculate the unadjusted coefficient of variation, we did the following: we initially determined the number of tests ordered from each practice from April 1, 2015, to March 31, 2016. We then calculated the total person-years of observation for each general practice. Patients alive and registered for the entire year contributed 1 person-year of observation to the total. Patients who were born, died, registered, or deregistered during the year were included, but their contribution to the person-year calculation was adjusted proportionately (e.g. a patient who was registered and alive for only 6 months contributed 0.5 person-years).
We then calculated the mean unadjusted rate of use for each specific test across all 444 general practices; we also calculated the corresponding standard deviation. We used these two numbers to calculate the unadjusted CoV (standard deviation/mean × 100) [
17]. The use of CoV facilitates a direct comparison of the variation in use between tests controlling for differences in sample size. It is expressed as a percentage (the ratio of the standard deviation to the mean), with larger percentages reflecting greater variation.
To calculate the adjusted CoV, we constructed a generalised linear model with Poisson errors to estimate the number of tests ordered from each general practice adjusted for practice differences in patient age, sex and deprivation. We constructed 44 Poisson models for each test. The age covariate represents the median age of each general practice, the sex covariate represents the proportion of female patients in each practice and the deprivation covariate represents the practice-level Index of Multiple Deprivation (IMD) deciles. We constructed Poisson models to adjust for differences in patient demographics (age, sex and IMD) between general practices. We did not construct Poisson models to compare the predictive ability of patient demographics on the rates of test use.
We then calculated the adjusted rate of use for each test in every general practice by dividing the adjusted number of tests by the person-years for each general practice (the same process we followed to calculate the unadjusted rate of use). The adjusted rates were used to calculate the adjusted CoV for each test, as described above [
17]. We ranked tests according to their CoV. We present both the unadjusted and adjusted CoV in Additional file
1, but only the adjusted CoV in the main manuscript.
To identify the tests that had both a high rate of use and a high CoV, we calculated the overall median rate of test use and the overall median adjusted CoV. We then classified tests into four categories: (1) high variability, low rate, (2) high variability, high rate, (3) low variability, low rate or (4) low variability, high rate. These categories reflect a test’s measure in relation to the median value, e.g. high variability, low rate refers to tests with a coefficient of variation above the median coefficient of variation, but a rate of test use below the median rate of test use.
Role of the funding source
This study was funded by an independent grant from the National Institute for Health Research (NIHR) School of Primary Care Research (Grant reference number 386) and the Primary Care Research Trust. Independent expert peer reviewers provided feedback on the grant application underpinning this study but had no further role in study design, data collection, analysis, interpretation or drafting of the manuscript.