netest_refit_abc: Refit TERGM with Biased Edges Dissolution Approximation

In EpiModel/EasyABCMPI: Approximate Bayesian Computation for EpiModel on High-Performance Computing Clusters

Description

Refits a TERGM estimated with netest using the edges dissolution approximation that had residual biased estimates.

Usage

netest_refit_abc(
  est,
  nsims,
  ncores,
  nsteps,
  coefs.vec,
  targets.vec,
  prior.min = 0,
  prior.max = 0,
  p_acc_min = 0.1
)

Arguments

est	An object of class netest, from the netest function.
nsims	Number of simulations per sequential ABC wave (more = fewer waves required for convergence).
ncores	Number of cores for each wave of simulations.
nsteps	Number of time steps to simulate the TERGM for each refitting.
coefs.vec	Integer vector of coefficient position to refit. Implicit default is the full coefficient vector, but use this to subset fitting to selected coefficients.
targets.vec	Integer vector of target statistic position to fit to. Implicit default is the full target stats vector, but use this to subset fitting to selected targets.
prior.min	Absolute lower bound of adjustment to be made across all model coefficients.
prior.max	Absolute upper bound of adjustment to be made across all model coefficients.
p_acc_min	Convergence threshold for ABC waves, as a proportion of simulations accepted (range from 0 to 1, with lower = greater precision).

Details

Fitting an TERGM in netest can be accelerated with the edges-dissolution approximation method (edapprox = TRUE), but this sometimes results in bias (poor fit to the target statistics) when the mean network degree is high, mean duration is short, or other combinations. All of this poor fit is revealed with model diagnostics in netdx. If the approximation method does not work, then the immediate next step is to try fitting a full TERGM without the approximation (edapprox = FALSE). This is often successful must has challenges with larger and more complex network models, particularly those with longer durations. Therefore, this function provides an alternative approach using approximate Bayesian computation (ABC) methods. This function simply wraps a commonly used sequential ABC approach and does the additional coefficient adjustment automatically. It may not work for all models, and there is no guarantee of model convergence or even improvement from the original fit; one should evaluate that fit with an additional round of model diagnostics after the ABC algorithm completes.

Examples

## Not run: 
# Original fit with edges dissolution approximation: High mean degree and
# short duration often leads to poor performance of the approximation method
nw <- network_initialize(1000)
formation <- ~edges + concurrent
target.stats <- c(1*(1000/2), 250)
coef.diss <- dissolution_coefs(dissolution = ~offset(edges), duration = 10)
est <- netest(nw, formation, target.stats, coef.diss)

# Diagnostics will show bias here
dx <- netdx(est, nsims = 10, nsteps = 300, ncores = 4)
print(dx)

# Refit the model with Bayesian ABC, with uniform prior adjustments applying
# to all model coefficients
est_new <- netest_refit_abc(est, nsims = 10, ncores = 5, nsteps = 300,
                            prior.min = -0.5, prior.max = 0)

# Raw output from EasyABC
est_new$refit

# Compare old and new coefficient values
est$coef.form
est_new$coef.form

# Rerun diagnostics to evaluate fit (no guarantee that ABC works)
dx_new <- netdx(est_new, nsims = 10, nsteps = 300, ncores = 4)
dx_new

## End(Not run)

EpiModel/EasyABCMPI documentation built on Dec. 5, 2020, 8:43 p.m.