redist_smc: SMC Redistricting Sampler (McCartan and Imai 2023)

Description

redist_smc uses a Sequential Monte Carlo algorithm (McCartan and Imai 2023) to generate representative samples of congressional or legislative redistricting plans according to contiguity, population, compactness, and administrative boundary constraints.

Usage

redist_smc(
  map,
  nsims,
  counties = NULL,
  compactness = 1,
  constraints = list(),
  resample = TRUE,
  runs = 1L,
  ncores = 0L,
  init_particles = NULL,
  n_steps = NULL,
  adapt_k_thresh = 0.99,
  seq_alpha = 0.5,
  truncate = (compactness != 1),
  trunc_fn = redist_quantile_trunc,
  pop_temper = 0,
  final_infl = 1,
  est_label_mult = 1,
  ref_name = NULL,
  verbose = FALSE,
  silent = FALSE
)

Value

redist_smc returns a redist_plans object containing the simulated plans.

Arguments

map: A redist_map() object.
nsims: The number of samples to draw.
counties: A vector containing county (or other administrative or geographic unit) labels for each unit, which may be integers ranging from 1 to the number of counties, or a factor or character vector. If provided, the algorithm will only generate maps which split up to ndists-1 counties. Even there are fewer counties than ndists - 1, the spanning trees will change the results of the simulations. There is no strength parameter associated with this constraint. To adjust the number of county splits further, or to constrain a second type of administrative split, consider using add_constr_splits(), add_constr_multisplits(), and add_constr_total_splits().
compactness: Controls the compactness of the generated districts, with higher values preferring more compact districts. Must be nonnegative. See the 'Details' section for more information, and computational considerations.
constraints: A redist_constr() object or a list containing information on sampling constraints. See constraints for more information.
resample: Whether to perform a final resampling step so that the generated plans can be used immediately. Set this to FALSE to perform direct importance sampling estimates, or to adjust the weights manually.
runs: How many independent parallel runs to conduct. Each run will have nsims simulations. Multiple runs allows for estimation of simulation standard errors. Output will only be shown for the first run. For compatibility with MCMC methods, runs are identified with the chain column in the output.
ncores: How many cores to use to parallelize plan generation within each run. The default, 0, will use the number of available cores on the machine as long as nsims and the number of units is large enough. If runs>1 you will need to set this manually. If more than one core is used, the sampler output will not be fully reproducible with set.seed(). If full reproducibility is desired, set ncores=1.
init_particles: A matrix of partial plans to begin sampling from. For advanced use only. The matrix must have nsims columns and a row for every precinct. It is important to ensure that the existing districts meet contiguity and population constraints, or there may be major issues when sampling.
n_steps: How many steps to run the SMC algorithm for. Each step splits off a new district. Defaults to all remaining districts. If fewer than the number of remaining splits, reference plans are disabled.
adapt_k_thresh: The threshold value used in the heuristic to select a value k_i for each splitting iteration. Higher values are more accurate but may require more computation. Set to 1 for the most conservative sampling. Must be between 0 and 1.
seq_alpha: The amount to adjust the weights by at each resampling step; higher values prefer exploitation, while lower values prefer exploration. Must be between 0 and 1.
truncate: Whether to truncate the importance sampling weights at the final step by trunc_fn. Recommended if compactness is not 1. Truncation only applied if resample=TRUE.
trunc_fn: A function which takes in a vector of weights and returns a truncated vector. If the loo package is installed (strongly recommended), will default to Pareto-smoothed Importance Sampling (PSIS) rather than naive truncation.
pop_temper: The strength of the automatic population tempering. Try values of 0.01-0.05 to start if the algorithm gets stuck on the final few splits.
final_infl: A multiplier for the population constraint on the final iteration. Used to loosen the constraint when the sampler is getting stuck on the final split. pop_temper should be tried first, since using final_infl will actually change the target distribution.
est_label_mult: A multiplier for the number of importance samples to use in estimating the number of ways to sequentially label the districts. Lower values increase speed at the cost of accuracy. Only applied when there are more than 13 districts.
ref_name: a name for the existing plan, which will be added as a reference plan, or FALSE to not include the initial plan in the output. Defaults to the column name of the existing plan.
verbose: Whether to print out intermediate information while sampling. Recommended.
silent: Whether to suppress all diagnostic information.

Details

This function draws samples from a specific target measure controlled by the map, compactness, and constraints parameters.

Key to ensuring good performance is monitoring the efficiency of the resampling process at each SMC stage. Unless silent=FALSE, this function will print out the effective sample size of each resampling step to allow the user to monitor the efficiency. If verbose=TRUE the function will also print out information on the \(k_i\) values automatically chosen and the acceptance rate (based on the population constraint) at each step. Users should also check diagnostics of the sample by running summary.redist_plans().

Higher values of compactness sample more compact districts; setting this parameter to 1 is computationally efficient and generates nicely compact districts. Values of other than 1 may lead to highly variable importance sampling weights. In these cases, these weights are by default truncated using redist_quantile_trunc() to stabilize the resulting estimates, but if truncation is used, a specific truncation function should probably be chosen by the user.

References

McCartan, C., & Imai, K. (2023). Sequential Monte Carlo for Sampling Balanced and Compact Redistricting Plans. Annals of Applied Statistics 17(4). Available at tools:::Rd_expr_doi("10.1214/23-AOAS1763").

Examples

Run this code

# \donttest{
data(fl25)

fl_map <- redist_map(fl25, ndists = 3, pop_tol = 0.1)

sampled_basic <- redist_smc(fl_map, 5000)

constr <- redist_constr(fl_map)
constr <- add_constr_incumbency(constr, strength = 100, incumbents = c(3, 6, 25))
sampled_constr <- redist_smc(fl_map, 5000, constraints = constr)

# Multiple parallel independent runs
redist_smc(fl_map, 1000, runs = 2)

# One run with multiple cores
redist_smc(fl_map, 1000, ncores = 2)
# }

Run the code above in your browser using DataLab