Statistical methods for survival analysis in large-scale electronic health records research

The relative survival framework is a popular method for the estimation of a subject's survival, corrected for the effect of non-disease related causes of death. A comparison is made between the observed all-cause survival and the expected survival, derived from published population mortality rates known as life tables, often stratified by age, sex, and calendar year. Under certain assumptions, relative survival provides an estimate of net survival, survival in a hypothetical world where subjects can only die due to their disease. In order to interpret relative survival as net survival, other-cause mortality rates for subjects with the disease of interest must be the same as expected mortality rates. When interest lies in the relative survival of diseases with multiple shared risk factors, for example lung cancer, the use of standard life tables is unsuitable, requiring additional stratification by these risk factors. The primary aim of this research is use a control population taken from large-scale linked electronic health records to adjust published life tables by comorbidity, and to investigate the impact of these and standard life tables on relative survival estimates. To achieve these research aims, bespoke software is developed to aid the management of large-scale health data, while investigations into mortality rates in the control population data is undertaken, showing biased results when follow-up requirements form part of patient selection. Comorbidity adjusted life tables are estimated using time-constant and time-updated exposures, and applied in a relative survival analysis in colorectal cancer, comparing groups defined by cardiovascular comorbidity status. This research extends concepts and methods previously developed to form novel approaches to the adjustment of background mortality data, taking into account the induced bias in the control population, and showing the importance of the use of correctly stratified life tables, with key implications for future studies investigating differential mortality rates.