Additional multistate model applications
Unobservable States single observable state, single unobservable state
Example Colonial seabird (survey colonies) 2 states: Breeder (B) Non-breeder (N) Only breeders are observable Non-breeder are temporarily emigrated
B N History B 0 B B N Unobservable What state at occasion 2? State B? (1-p B ) State N? (p N =0) or
History B 0 B: probability statement State B? (1-p B ) State N? (p N =0) or
Unobservable states in MARK Unobservable state If robust design, This case can also be treated as temporary emigration
Unobservable States multiple observable states, one or more unobservable states
A Observable Recapture areas A and B (e.g., 2 br. colonies that are surveyed) B U S i A i AU S i U i UA S i B i BU S i U i UA S i A i AB S i B i BA Unobservable (“rest of the world”) Temporary emigrated animals Multiple observable, 1 unobservable
Important assumption S i U cannot be estimated separately It must be assumed equal to S i A or S i B This assumption is not required when all states are observables
Unobservable states in MARK >1 observable / Unobservable states
Multiple observable and multiple unobservable states A Pre-Breed A Breed B Pre-breed B Breed A Young B Young Example: Roseate Tern (Lebreton et al. 2003)
History: a 0 0 B We could decompose transitions and survival
A Breeders B Breeders S i Ab i AbAn S i Bb i BbAb A Non-breed Unobservable B Non-breed S i Ab i AnAb S i Ab i AnBb S i Ab i AbAn S i Bb i BnAb S i Bb i BbBn S i Bb i BnBb Recapture Areas Paired observable, unobservable
Design Considerations Exploit opportunity for robust design Include telemetry (preferably with p =1 and survival monitored) Create buffer zone around study area ( search for marked animals, non-breeders? ) Collect behavioral cues each time an animal is seen (breeding behavior?)
Conclusions Multistate models for unobservable states Potential for including unobservable states Not all models identifiable Robust design can be used to improve estimates (more info on p)
Conclusions Existence of unobservable states should be minimized through design. Robust design, telemetry, and other sources of information provide means to adjust for unobservable or misclassified states.
Mis-classified or Unknown States
How do we adjust for fact that calves are not seen each time with mother? Photo credit: USGS - Sirenia Project 2 states: C = female with Calf N = female with No calf
Assumption: breeding status is known for each sighted female Problem: Sometimes a calf is present but not sighted. Implication: Calf might be missed each time female is sighted and hence misclassified as non-breeder. Solution: adjust for misclassification (i.e., estimate calf detection probability).
Multi-state Model for Misclassification (Kendall et al. 2003) CC CN NC
How do we estimate detectability for the calf? Two sighting occasions per year (robust design). Entire range of population is covered twice in a short period of time so that if female is breeder calf will be there both times. Each time individual is seen the date and presence/absence of first-year calf is noted. Calf detection probability is estimated from sighting history of female and calf combined.
Other examples of mis-classified states Disease status / dynamics Diseased animals show no symptoms Recovered animals still show symptoms Sex is mis-assigned Seabirds – behavioral assessment Sharks – claspers not always visible
Multi-event approach (quick overview) Program E-SURGE
E-SURGE Extension of multistate framework Hidden Markov Chains
E-SURGE Modeling done in 2 distinct steps: 1- State Transition Matrices (same as multistate modeling) 2- Observation Matrices (allows for misclassification)
Transition Matrices 3 alive states: P, B, N + “dead” PBN† P B N † 0001 True State at time t True State at time t+1
Observation Matrices 4 possible observations (“events”) PBN0 P B N † 0001 True State at time t Observation at time t