fbarraquand / predatorpreydynamics_wfrdata Goto Github PK

• We have looked at the likelihood surface for pairs of parameters in a case where we knew that there was a maximum (quasi-cycles, with explicit FR data).
• Would be interesting to redo this in cases where the maximum may not be so well-defined. First with quasi-cycles but without explicit FR data. Are there multiple models, including one at (0,0) for (C,D)?
• Would also be interesting to compare to the noisy LC simulation, for T=1000.
• We can also / directly do T= 100 if the likelihood surfaces are similar. Bolker suggested in EMDB that large samples & less noise = correlation in parameters more obvious (they get very precise, so their correlations too?). May be seen for both shorter TS and increased noise (e.g., sigma = 0.25 too?).

These T=100 (years) files were removed because the code was updated on the T=1000 files (up-to-date). We need to redo those from the T=1000 codes. Then perhaps get down to T=50 or T=25. We could also do functions of course...

initial value in 'vmmin' is not finite

-> work on starting values

• missing proper bibliographic references, including on identifiability and profile likelihoods
• references to equations, even with $\eqref{}$ seem somewhat problematic for HTML output
• include also an expected FIM for the "without FR data" case? -> allows to check identifiability irrespective of the optimization algorithm, around the true parameter value

Perhaps we need this combination to fully assess identifiability