Browsing by Author "Kotze, Loamie"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- ItemMarkov modelling of disease progression in the presence of missing covariates(Stellenbosch : Stellenbosch University, 2019-04) Kotze, Loamie; Mostert, P. J.; Stellenbosch University. Faculty of Economic and Management Sciences. Dept. of Statistics and Actuarial Science.ENGLISH SUMMARY : Breast cancer is a very prevalent cancer amongst women. The stages of breast cancer are influenced by characteristics such as age, hormone receptor statuses, HER2 status and staging information (TNM staging). This study aims to model the progression of breast cancer using a multi-state model which evaluates three pre-defined stages of the disease. A secondary aim is to determine an appropriate technique to impute missing data in the covariates. The disease progression can be modelled by using multi-state models and it is of interest to analyse the effect of different risk factors on the transitions between the states. The variable of interest can be seen as the state of the individual at that time point. The transition intensities of the multi-state model provides the hazards of moving from one state to another and can be used to calculate the mean sojourn time in any given state. A combination of claims data and authorisation treatment request data were obtained from Isimo Health for 393 breast cancer patients. Based on this, a dataset was simulated using the TPmsm package in R statistical programming. The simulated data were used to test two imputation techniques, one based on chained equations and one based on random forests, for the missing data present in the covariates. The latter technique performed the best based on several performance measures, and was used to impute the dataset from Isimo Health. Thereafter, a multi-state Markov model was fitted to the imputed data with three pre-defined states including curative (receive treatment with the intent to cure), non-curative (receive treatment with the intent to provide improved survival or symptom control) and death. It was observed that the Markov assumption does not hold and, therefore a semi-Markov model was fitted to the data. The findings showed that only one of the covariates, namely staging, had a significant effect on the transition probabilities. This is only the case for the transition between the non-curative and death state. Covariates as a whole, did have a significant effect on the transitions from curative to non-curative and non-curative to death. However, there was no significant effect on the transition from curative to death. It can be concluded, based on statistical measures, that the missForest package efficiently imputes missing covariates before modelling disease progression with multi-state models using the p3state.msm package.