TODO.md

In bbolker/fitsir: SIR fitting tools

Overall to-do list

Package stuff

• self-starting
  • prevalence is OK, but a little hacky (not enabled by default)
  • incidence is not so good
• work on neg binomial fit ... quasi-Poisson
• plotting methods for data, fits?

Optimization-related

• larger random/Latin hypercube sim; k-NN clustering, basins of attraction, etc.
• stochastic global opt? (Mullen JSS paper)

Other data sets

• 1918 flu

Longer-term/lower priority

• priors

• missing data

• think a lot about the interface: how to optimize efficiency, flexibility, breadth, transparency, simplicity ... ? Want simple model objects (S3 class) that allow simple summaries, diagnostics, fits ... extend mle2 class?

• should/must mle2 class be extended via S4? or export coef?
• (summary.pars gets Roxygenized documentation confused)
• more general parameter transformation machinery -- from constrained to unconstrained scale to epidemiological parameters (R0, r, rho, etc.) and reverse
• flexibility in models?
• throw in old BIRS erlangSEIR stuff?
• gradient matching?
• state-space models: pomp, JAGS, etc. ?
• tie-in with Tiff Bogich's "Public Library of Models" (plom.io) (related to Ballesteros's epi-models)?
• tie-in with other epi R packages: EpiModel, EpiDynamics, pomp ...

exploration to-do list

• Figure 1: prevalence?? or preferably re-do with incidence if not too hard
• (a) cumulative distributions of attained log-likelihood for N-M and BFGS, showing three levels (pick an example with three)
• (b) points and trajectory for each plateau (i.e., flat, jump-up plus decay, good fit)
• (c) ??? basins of attraction for difference plateaus: i.e. plot the starting points from the LHS coloured accord to which plateau they end up in. Summaries???

• check Hessians (analytical + finite-diff of grad, which you can get via numDeriv)

• do this for real epidemic curves

• re-run resList.grad (BFGS) with negative log-likelihood outcomes
• big LHS experiment:
  • sensitivity equations vs. no explicit gradient
  • different optimizers (via optimx?)
    • no point in running derivative-free methods with sens equations
    • in optimx, can't run no-gradient with derivative-ful methods
    • ?check latest optim (optimrx from r-forge?) install.packages("optimrx",repos="http://r-forge.r-project.org")
  • NLL vs least squares?
  • rk4 vs lsoda

• compare:

  • simulated data
  • real data
    • Bombay
    • Harbin
    • Phila 1918 flu
    • Gani & Leach smallpox?
    • ??? i.e., is it just that the Bombay data are weird?

• visualization

• try with 2 parameters (fix gamma, N, show likelihood surface for beta, I0); 3 parameters?

• reparameterization for most independent results?

• understand why NLL vs least-squares should matter.

  • just numerical sensitivity/"noise"?
  • should get same best fit value (extreme points of surface should agree ...)

bbolker/fitsir documentation built on June 4, 2019, 8:28 a.m.