how to choose the way of binarization, independently, concatenated or stacking?

[JhinAir]

how to choose the way of binarization, independently, concatenated or stacking?

[jernst98]

Dear Jing,

Here are five options I can think of when you do not have a full grid of mark/tissue types:

1. Run a stacked model.

2. Run a concatenated model restricting to a subset of stages and/or marks where you have a full grid and then run a concatenated model.

3. Run a concatenated model but treating as missing datasets for which you do not have data ('2' encoding in ChromHMM).

4. Run a concatenated model using impute data from a method such as ChromImpute where you have missing observed data.

5. Run independent models for each stage.

Option 1 has the advantage of using all the data directly and can highly different dynamic patterns so could be a good place to start, but does not provide per-stage annotations if that is what you are interested.
Option 2 has the advantage of being straightforward and allowing well controlled per-stage comparisons but the disadvantage of not using all the data
Option 3 relative to Option 2 uses more data but state usage comparisons will be confounded by the marks available in the stage
Option 4 relative to Option 2 uses more data and relative to Option 3 will be less confounded by the available marks at a stage, but if you have limited data for imputation it might not capture some of the stage specific differences you are interested in capturing
Option 5 is fine if you are interested in studying each stage independently but since you have different state definitions for each stage, stage comparisons are confounded

[JhinAir]

I see. Thank you very much for your very detailed explanations!