Statistical approaches for monitoring early cancer diagnosis in England

Background: Increasing early-stage diagnosis is a priority of health policy in England. Numerous interventions to effect improvements were implemented during 2008-2013, including symptom awareness campaigns and bowel cancer screening. There were also structural changes to health services, with smaller increases in health spending from 2010 and a reorganisation in 2013. An analysis of early diagnosis trends and geographic inequalities is needed to assess the impact of these changes. There are challenges to monitoring early diagnosis trends during 2008-2013, and afterwards, however. Disease stage was not recorded in the cancer registrations of many patients. These patients have poorer outcomes, and assuming early diagnosis was as common for them as patients whose stage was recorded may introduce bias. Case-mix factors may bias comparisons of health services performance, and comparisons between local areas can be limited by sparse data. Finally, other indicators alongside stage can be used in monitoring. Consideration of the added value and interpretation of these is merited. These challenges have not previously been addressed in a national analysis of early diagnosis trends. The aim of this thesis is to address them, then apply the findings to evaluate trends and geographic inequalities during 2008-2013 for colorectal cancer, non-small cell lung cancer, and ovarian cancer. The implications for monitoring of early diagnosis after 2013 are then assessed. Methods: Different early diagnosis indicators were described, and reasons for missing data were considered, with reference to the literature and data analysis. A conceptual framework for determinants of early diagnosis was created, and used to identify case-mix factors to be adjusted for in the substantive analysis. The association between different indicators and survival was evaluated through data analysis and a systematic literature review. Methods for case-mix adjustment, geographic comparisons, and handling missing data were surveyed. Finally, an analysis of early diagnosis trends and geographic inequalities during 2008-2013 was conducted using multilevel logistic regression, with multiple imputation to reduce bias from missing stage data. Sensitivity and simulation analyses were performed to assess the treatment of missing data. Results: Stage was typically unreported for administrative reasons, but occasionally because the patient was frail; or died; or was treated privately. Improvements in staging data collection resulted in stage being less commonly missing for administrative reasons in 2013 than in 2008. Age, sex, comorbidities, and tumour morphology and topography were identified as key case-mix factors. Stage had clearest and most consistent interpretation of the indicators assessed. Multiple imputation was identified as the optimal approach to reduce bias from missing stage data. There was evidence for an increase in the percentage of patients diagnosed at stages I or II for each of the three cancers analysed, including a step-change improvement for colorectal cancer (from 32% in 2008-09 to 44% in 2012-13). Geographic inequalities reduced. For ovarian cancer, estimated trends were different between analyses which did and didn’t use multiple imputation. Sensitivity analyses indicated that the multiple imputation model was specified correctly, and that results were robust to some residual bias. It was found to be necessary to use information on one year’s survival time to impute stage accurately. Interpretation: Completeness of stage recording improved during 2008-2013, and afterwards to over 85% by 2018. Analysts may choose to disregard data on patients whose stage was not recorded, as they compromise a small proportion of the total. However, these patients have poorer outcomes, and the disparity between their outcomes and outcomes of patients with stage recorded has increased as stage recording has improved. Therefore, material bias may still be introduced from excluding them in analyses after 2013. Future evaluations, including of the overall impact of COVID-19 on early diagnosis, should use multiple imputation to account for patients whose stage is not recorded. This will give a more accurate picture of trends in early diagnosis of cancer in the whole population.