Converting a PartSA model to a Markov model

38.12 Converting a PartSA model to a Markov model

While PartSA models have their strengths, you may find that this technique cannot fully handle the complexity of disease progression. You may instead need to expand to separate health states and events supported by Markov models. If so, you can convert a PartSA model into a Markov model at the click of a button.

The process converts the PartSA structure into a tradition Markov model with Health States and Transitions. Based on the PartSA model, with healthier "states" at the top, the Markov model is constructed with progression from more healthy to less healthy/dead states.

We will use the Partitioned Survival Analysis Example Model PartSA-Example.trex to demonstrate.

To convert the model:

Right click on the PartSA node and select "Convert to Markov".
Choose either complex or simple transition probability formulas. For this example, choose complex.

The figure below shows the converted model.

The Markov model contains the following elements:

A straightforward Markov structure with the original states from the PartSA model as well as the explicit Dead state.
The original variables and distributions from the originating PartSA model.
Complex transition probabilities mimicking the PartSA state membership behavior. The transition probabilities are based off the underlying survival functions. The complex conversion front-loads transitions to better mimic the PartSA results. The simplier transitions just use the survival functions.
Continuous rewards from the PartSA model (note that other reward types are not converted, so the Chemo cost is lost and would have to be added manually).

The figure below compares the PartSA model's survivor curve to the Markov model's survivor curves generated from the Markov Cohort report.

Note that the converted Markov model should not be considered complete. It is a starting point for a representative Markov model.

Complex transition probabilites attempt to replicate state membership from the original PartSA model. That does not mean that those transition probabilities are correct. You need to examine the transition probabities to fit appropriately within the new Markov context. For example, the transitions to states further downstream (PFS to Dead) are set to zero by default. This is probably not appropriate.

The simpler transition probabilities will not match the original PartSA results as well, but they may still be a better place to start since transition probabilities will need adjustment.

Generating PartSA Survival Curve from a Markov Model

You can generate Survival Curve outputs from a Markov model for validating back to survival data. This is explained in detail in the section Partitioned Survival Curves from a Markov Model.