Flexible parametric models for cancer patient survival : loss in expectation of life and further developments

Abstract

Population-based cancer studies can contribute to a better understanding of cancer patient survival. The general aim of this thesis was to develop and apply statistical methods for population-based cancer studies to ensure understanding in this area. In any analysis setting, it is important that the statistical methods are appropriate. Since nonproportional hazards are common in population-based data where follow-up is long, Study I assessed the ability of flexible parametric models (FPMs) in capturing time-dependent effects in a simulation setting. This study also attempted to determine whether the Akaike Information Criterion (AIC) or the Bayesian Information Criterion (BIC) were more appropriate in selecting a good-fitting model. Results indicated that bias was small for estimated survival proportions, hazard rates and log hazard ratios when degrees of freedom were selected using guidance from the AIC or BIC. Neither the AIC nor the BIC constantly outperformed the other. We concluded that FPMs accurately capture time-dependent effects but that users should perform appropriate sensitivity analyses.

There have been large changes in treatments for chronic myeloid leukaemia (CML) patients over time; most notably is the introduction of imatinib mesylate, a targeted therapy, in 2001. Studies have shown that relative survival of CML patients greatly improved after the introduction of this treatment. Since CML is a chronic disease, quantifying survival in terms of life expectancy is highly relevant. In Study II we aimed to quantify life expectancy improvements using the loss in expectation of life measure. Results indicated that patients diagnosed in 2013 would be expected to lose less than three of their remaining life years due to their CML diagnosis. We concluded that the life expectancy among CML patients is approaching that seen in the general population and improvements over time are largely due to imatinib mesylate and allogeneic stem cell transplantation.

Population-based mortality rates are often used as a proxy for the mortality rate a diseased population would have experienced had they been disease-free; these are used in relative survival analysis and when calculating standardised mortality ratios. Population-based mortality rates are commonly available by age, calendar year and sex which might limit analyses to these factors. In Study III we described methodology to adjust population mortality rates by additional variables, such as socioeconomic status, that are available in a control population. We presented both Poisson and flexible parametric methods and found that both methods estimate similar mortality rates by socioeconomic status. Adjusting for the additional uncertainty associated with the methodology presented made little difference to five-year relative survival of breast cancer patients.

Socioeconomic status is known to affect the survival of breast cancer patients. Stage at diagnosis of breast cancer is also known to have a strong effect on survival and the distribution of stage at diagnosis often differs by socioeconomic status. In Study IV we aimed to quantify survival differences between education groups as a proxy for socioeconomic status. We estimated that 572 life years could be saved, and 25 deaths could be postponed five years beyond diagnosis, if differences between education group in the breast cancer stage-distribution could be removed in three regions of Sweden. If differences between education groups in stage-specific breast cancer survival could be removed, we estimated that 692 life years could be saved, and 27 deaths postponed five years beyond diagnosis.

In conclusion, this thesis reassures users of FPMs of their performance under non-proportional hazards. Secondly, methodology is described to further adjust expected mortality rates which could aid researchers in estimating survival measures by additional, non-standard variables. Finally, this thesis provides a greater understanding of the probable prognosis and the burden of CML and breast cancer diagnoses via the loss in expectation of life and other similar measures.