markovchainSequence, markovchainListFit and markovchainFit to handle NA

[spedygiorgio]

It could be extremely useful markovchainSequence, markovchainListFit and markovchainFit to consistently handle NA (remove them from the compute of transitions).

[ncordon]

What do you mean by removing NAs from the list of transitions? For example, for input (1 NA 2), should we consider no transition at all, or should we have an unique (1 2) transition?
Thanks

[ysaibhargav]

@ncordon, the former may be one simple way to resolve it. But it may also be possible to estimate parameters without completely ignoring such occurrences in the input data (perhaps by treating missing states as latent variables).

[ysaibhargav]

So after a discussion with @spedygiorgio, we arrived at the conclusion that it may be useful to resolve this by both methods - the argument na.behaviour = "remove" would treat ("a", NA, "b") as no transition at all, while "latent"/"infer" would model the missing states as latent variables to compute the chain parameters.

[ncordon]

I think I've proposed a suitable solution with the first approach (no transition at all). I will try to improve it with the flag na.behaviour .

[spedygiorgio]

Regarding the flag na.behaviour='infer' a possible solution is to obtain the best "guess" of the unknown state from the given data.

[spedygiorgio]

@ncordon and @vandit15 thanks for your commits. I am not sure that we will merge your commit in the imminent release but they will become part of the next one certainly. Btw could you also please update the help file and the package vignette of your commits adding some words on your changes?