ESTIMATION OF TREATMENT OUTCOME CUMULATIVE INCIDENCE (Cl) AND KM ANALYSIS
The 01 method has been promoted on the basis that lack of independence is a problem in actuarial analysis of data describing outcome with competing risks.
7.4 CONCLUSIONS
7.4.1 Modelling failure specific prognostic factors
Multivariate analysis using competing risks models as implemented in the BMDP software package allov^ exploring the relation between various prognostic factors and different types of first recurrence to be explored. With this approach prognostic factors may be used as predictive factors for individualising treatments based on the most likely failure types. This analysis shows the influence of a particular covariate on time and type of a specific failure by providing three hypothesis tests and has two potential fields of application.
Firstly, it enables us to predict the risk of specific types of failure with improved precision, which could be used to identify subgroups of patients with significantly different failure patterns. When this approach was used for locally advanced NSCLC the two common failure types (LRF and DF) were considered together and patients were divided into four prognostic groups with different risks for LRF and DF. These four prognostic groups defined by the model showed a close agreement when the model predicted failure rates over time were compared graphically with the KM estimates. This informal validation method used overall data and applied it to the strata. Even if the model fits the data properly in our study we are aware that the most appropriate way would be to use an independent data set to validate the method. The same approach was used in a slightly different way in HNSCC where the subgroups were formed to include patients with different risks for T, N and M failure separately. Three prognostic groups for each type of failure were defined, compared and found to be significantly different. The individual estimates could also be used for tailoring of
therapy based on individual risk profiling when they are not grouped as above. Using this approach, selection of primary therapeutic interventions could be based on the consideration of the contribution of relative risks of failures at local (T), nodal (N) and distant (M) positions. We have shown this approach for HNSCC analysis.
Secondly, the influence of each covariate on the time and the type of failure could be considered one by one in exploring the clinical significance of established clinico- pathological factors or new biological markers and stratification of patients into trials of new treatments.
In NSCLC model, the clinical stage was the only significant factor associated with both the time and the type of failure. Patients with stage III disease were prone to have a higher and earlier distant failure relative to local failure. Thus, intensified local treatment may not suffice to improve therapeutic outcome in this group of patients. The literature is rapidly expanding in the field of biological markers and their potential role as prognostic or predictive factors remains to be tested. It is possible that further biological characterisation of these tumours in a competing risk analysis would enable us to predict the risk of specific types of failure with improved precision.
In the HNSCC model, all other molecular biomarkers except for p53 were predictive for at least one of the types of failures. The predictive information of molecular markers on failure types for an individual patient at the time of presentation would provide the possibility of intensifying and thus selecting ideal patients for different treatment schedules. Moreover this analysis showed that this approach could also possibly be used to understand the relation between the CHART arm and various cell
systemic therapies, studies of failure-type specific predictive markers are of great interest and we feel that the approach taken here warrants further study.
The BMDP competing risks analysis is based on the KM principle and assumes that the failure times for the types of failure are independent. This assumption is untestable.
Competing risks analysis of failure-specific prognostic markers is a powerful tool for selection of patients with different risk profiles for specific types of failures, which eventually may lead to tailored therapies.
7.4.2 Estimation of treatment outcome
The underlying assumptions and the difficulties in interpretation of the three estimators compared here lead us to the conclusion that no single method is appropriate on its own when estimating treatment outcome. The assumption of independence of different event types in general may not be clinically meaningful where the lowest risk of LRF might be very much related to the highest risk of DF. A second assumption independence of the censoring mechanism is statistically untestable. In clinical research censoring could indicate an unfavourable prognosis (patients too sick, or bedridden), or conversely a favourable prognosis (drop out of the study once they are cured). This problem is common to all survival analysis.
The KM (any) method gives the upper limit of an event rate under study by ignoring all the other events, it does not address the question of which event is likely to occur first. The KM (any) estimate should be interpreted with caution when estimating late
side effects because this estimate could be misleading if any other treatments have been used for any types of failures.
The KM (1^) method considers only the first event type rather than any event type in the analysis. This estimate focuses on the first failure and is not influenced by other competing events, so it is always a higher figure compared to the Cl estimate where the competing events has a direct influence on the resulting estimate. The KM (1®*) estimate should be preferred to have an idea about the upper limit of an event of interest occurring first. The biases discussed for the KM (any) estimate are not relevant for KM (1®*) estimate so it is a reliable estimate of the highest probability of treatment failure and late side effects.
The 01 method has been promoted on the basis that lack of independence is a problem in actuarial analysis of data describing outcome with competing risks. However the 01 estimate is subject to the assumption of exclusivity of outcomes and gives the lowest possible figure, where should be interpreted with this in mind. Whenever late radiation morbidity is under study, the 01 estimate should be interpreted with great caution since the estimate could give misleading results in a group of patients with same treatment intensity but different prognosis.
The KM and the 01 methods should be used as complementary analyses whenever possible considering the various clinical situations. The underlying mechanisms and assumptions of statistical methods used in competing risks situations should be well understood by the clinicians who will interpret them in the clinical practice.
Publications from the material presented in this thesis
1. Ô Umk, SM Bentzen, Ml Saunders, M. Parmar. Pattem of failure after CHART or