Multi-state model with two uncertain states

[sstates]

I have a 22-year human “mark-recapture” dataset in which people are tested for the presence of antibodies for two tick-borne pathogens though time. Each individual can be positive, negative, or ‘uncertain’ for each pathogen. The ‘uncertain’ state refers to a test result that is borderline. My goal is to understand whether the presence of one pathogen may influence the probability of infection or recovery of the other. I have considered a couple approaches using multi-state models with state uncertainty (Conn and Cooch 2009; Kendall et al. 2012). I do not have robust design data, so I’m not sure how well my data fit these approaches, including the fact that I have two separate uncertain states. I would like some advice or leads to examples of two uncertain states, or if I should explore treating this as a case of 9 states in which each state includes the status of both pathogens are grouped together (0/0, 1/0, 1/equiv, etc). I admit that I would prefer to run this in Mark as I have experience with this program. This is a large dataset – I have 1800 encounter histories – but I do recognize that having 9 states is likely to be very computationally intensive. I appreciate any and all suggestions.

[Eurycea]

Do you have perfect detection of individuals? Are you interested in transitions between states?

[sstates]

Individuals didn't attend every year, but they are always identified correctly (at their first attendance they are given a unique ID number that is checked every subsequent time they attend). Yes, I am interested in state transitions.

[jlaake]

MARK contains this model RDMS2MisClass: Closed Robust Design Multi-state with 2 States Uncertain

I've not used it, so I don't know whether it will meet your needs. Even if your data are not RD, you can add a dummy secondary occasion to each primary occasion to get it to work. Simply set p=0 for the dummy occasion.

If that model doesn't work then you can look at the pre-print of the paper in the link below. It allows multi-state models to be defined based on a multivariate vector of any length and each variable (e.g., tick1 and tick2) can be uncertain. It is an open model and can handle both non-RD and RD data. The analysis code is in the R package called marked.

http://biorxiv.org/content/biorxiv/early/2015/08/28/025569.full.pdf

[cooch]

MARK contains this model RDMS2MisClass: Closed Robust Design Multi-state with 2 States Uncertain

I've not used it, so I don't know whether it will meet your needs. Even if your data are not RD, you can add a dummy secondary occasion to each primary occasion to get it to work. Simply set p=0 for the dummy occasion.

If that model doesn't work then you can look at the pre-print of the paper in the link below. It allows multi-state models to be defined based on a multivariate vector of any length and each variable (e.g., tick1 and tick2) can be uncertain. It is an open model and can handle both non-RD and RD data. The analysis code is in the R package called marked.

http://biorxiv.org/content/biorxiv/early/2015/08/28/025569.full.pdf

Hi Sarah -- following Jeff's note. Indeed, MARK has the model Jeff refers to, although it isn't document in 'the book'. While I appreciate the desire to 'stay with MARK', for these sorts of 'hidden Markov/state uncertianty' models, using MARK might require more than a few contortians 9Bill kendall being the expert in that regard). It might ultimately save you time to take the plunge and consider either some of the software Jeff is working up, or E-SURGE, which is what Paul Conn and I use in the paper your refer to. The learning curve isn't too bad.

[sstates]

Thanks, Jeff and Evan!
I had looked at the help file for MSRD with 2 uncertain states, but I got the sense that it might not be appropriate for my data - it seems to accept only the certain and corresponding uncertain state, whereas for each pathogen I have positive, negative, or uncertain.
Jeff, I've been reading through your paper this evening, and I think it will be really helpful. I work primarily in R, so I should be able to make it work.
Thanks again for the suggestions thus far!