53 252 patients from 1868 general practices who were subsequently diagnosed with cancer had complete information on the usage of investigation. The median number of patients with cancer per practice was 23 (IQR 11–40). There were more males in the sample than females (55% vs 45%) (see online supplemental appendix table 4). The majority of the patients (91%) were white. Nearly 77% of patients had at least one chronic disease, whereas 23% had three or more. Almost three-quarters of patients were 60 years or older at diagnosis. As previously reported, the patient population included in the NCDA was representative of those diagnosed with cancer in England with regard to age, gender, deprivation, ethnicity and cancer site.16

Overall, 56% (29 932/53 252) of patients in the analysis sample underwent at least one investigation before being diagnosed with cancer, but this percentage varied between practices (IQR of practice-level percentage 47%–68%). There was also substantial variation by cancer type and patient demographics (see online supplemental appendix table 4). When considering specific types of investigations, 44% (23 422/53 252) had blood test, 23% (12 368/53 252) underwent imaging and only 2% (845/53 252) had endoscopies. The use of these types of tests varied between practices with the practice-level IQRs being 33%–53%, 15%–31% and 0%–2%, respectively.

Due to the low number of endoscopies in the dataset, it was difficult to achieve convergence of all models for this outcome; therefore, only results for any investigation, blood test and imaging use are reported. In model 1 (including random intercept for general practice only to quantify practice variation after accounting for chance), the estimated SD of between-practice use on the log-odds scale was 0.36 (95% CI 0.33 to 0.38) for any investigation, 0.37 (0.35 to 0.40) for blood tests and 0.29 (0.26 to 0.33) for imaging, respectively. These estimates translate to ORs of 4.02, 4.33 and 3.12, respectively covering 95% midrange (ie, the 2.5th to 97.5th centiles of practice distribution). This means that there was greater than fourfold variation for use of any investigation and blood tests, and over a threefold variation in imaging, between the practices that use them the most and those that use them the least (excluding 2.5% of practices at each extreme).4

Considering the outputs of model 3 that included both a practice random effect and patient-level variables, the estimated SD of between-practice use increased to 0.44 (95% CI 0.41 to 0.47, p<0.001) for any test, 0.47 (0.44 to 0.51, p<0.001) for blood tests and 0.32 (0.29 to 0.36, p<0.001) for imaging. The corresponding ORs covering the 95% mid-range of practices were 5.61, 6.30 and 3.60, respectively. This means that patient case-mix was hiding variation between practices, suggesting that those with a higher proportion of patients in groups where tests were used most often were generally less likely to use tests across all their patients, and vice versa. Additional analyses showed that the increased between-practice variation in the adjusted model (model 3) was mostly driven by cancer site, meaning that patients with cancers which were overall associated with higher use of tests were more likely to belong to practices with lower than average use of investigation for any cancer site (online supplemental appendix table 5).

Considering the outputs of model 2 that examined patient-level factors without adjustment for practice variation, the odds of any investigation and imaging was lower in the oldest age group (80+ years) compared with the 60–69 years age group (see online supplemental appendix table 2). Female patients had slightly higher imaging than males (OR 1.08, 95% CI 1.02 to 1.14, p=0.007). Compared with white, non-white patients had fewer blood tests (OR 0.91, 95% CI 0.85 to 0.98, p=0.012). There was no significant association between the level of deprivation and the use of any type of investigation except blood tests where the most deprived group had fewer blood tests than the least deprived. There was decreased use of all types of investigations with increased comorbidities. Including the random effect for general practice in the case-mix (model 3) had no impact on person-level investigations; however, variation in the use of blood tests became non-significant for ethnicity (table 2).

The inclusion of practice factors in model 4 resulted in a very small change in the odds associated with patient factors for all types of investigations, with no material change in the SD of between-practice variation compared with model 3. However, we found (see online supplemental appendix table 3) that patients registered at rural practices had lower odds of undergoing any investigation (OR 0.86, 95% CI 0.76 to 0.96, p=0.010) compared with urban practices. Additionally, patients registered at practices with a higher proportion of patients aged 65 years or older had significantly higher odds of undergoing all kinds of investigations. Furthermore, patients registered at practices with larger list sizes had higher odds of blood test use (size 2: OR 1.11, 95% CI 1.02 to 1.21; size 3: OR 1.14, 95% CI 1.03 to 1.27; global p=0.024) but no such association was observed for other investigations.

Finally, patients registered at practices with higher TWW referral ratio had higher odds of imaging. There was no evidence that practice-level deprivation, patients per practice and variables representing key aspects of patients’ experience (access, continuity, satisfaction and doctor communication) had any impact on practice test use of any type.