First, thank you for sharing this nice example in R code. Your team is to be commended for producing extremely clear and well commented code. I have two comments/suggestions with respect to this example.
1) The probability sensitivity analysis is incomplete
This comment/suggestion relates to the probability sensitivity analysis of the model for which there are at least 3 sources of variability that should have been considered (and possibly others):
a) The transition probabilities
b) The costs for each state
c) The utilities for each state
On lines 216 and 217, random samples from the OS and PFS survival curves are generated and on lines 227-230 the Markov trace is updated with these random sample values, effectively addressing point (a) above. However, in lines 236-37 the costs and QALYs are calculated as fixed values, previously specified in lines 96-102. I think that the analysis should sample these quantities from appropriate distributions as well as they are very unlikely to be known with absolute certainty (excepting perhaps the utility of death).
2) Extending the model to allow a decreasing utility within a health state
It is common for the utility of a person to decline over time, even within a single health state. For example, if this partitioned survival analysis was comparing two policies using two different chemotherapy regimens for cancer, it would be very usually the case that quality of life (and the utility) of the person declines over time while under therapy, even if the cancer is not progressing. This is due to side-effects of treatment. This could be very important when comparing novel targeted therapies or immunotherapies that do not tend to have this "declining utility" property with existing treatments that do. In such cases we are typically trading off expensive and better tolerated drugs with less expensive and worse tolerated drugs, and the implications of this trade-off for costs and QALYs is important to model. It would be nice to demonstrate how your model would handle such a situation.
First, thank you for sharing this nice example in R code. Your team is to be commended for producing extremely clear and well commented code. I have two comments/suggestions with respect to this example.
1) The probability sensitivity analysis is incomplete
This comment/suggestion relates to the probability sensitivity analysis of the model for which there are at least 3 sources of variability that should have been considered (and possibly others):
a) The transition probabilities
b) The costs for each state
c) The utilities for each state
On lines 216 and 217, random samples from the OS and PFS survival curves are generated and on lines 227-230 the Markov trace is updated with these random sample values, effectively addressing point (a) above. However, in lines 236-37 the costs and QALYs are calculated as fixed values, previously specified in lines 96-102. I think that the analysis should sample these quantities from appropriate distributions as well as they are very unlikely to be known with absolute certainty (excepting perhaps the utility of death).
2) Extending the model to allow a decreasing utility within a health state
It is common for the utility of a person to decline over time, even within a single health state. For example, if this partitioned survival analysis was comparing two policies using two different chemotherapy regimens for cancer, it would be very usually the case that quality of life (and the utility) of the person declines over time while under therapy, even if the cancer is not progressing. This is due to side-effects of treatment. This could be very important when comparing novel targeted therapies or immunotherapies that do not tend to have this "declining utility" property with existing treatments that do. In such cases we are typically trading off expensive and better tolerated drugs with less expensive and worse tolerated drugs, and the implications of this trade-off for costs and QALYs is important to model. It would be nice to demonstrate how your model would handle such a situation.