In tereom/quickcount: Estimate Election Results Mexico

knitr::opts_chunk$set(
    collapse = TRUE,
    comment = "#>", 
    cache = TRUE
)

To install the package use devtools (devtools is available on CRAN).

# devtools::install_github("tereom/quickcountmx")
library(quickcountmx)
library(tidyverse)

The package includes the results of the 2012 Guanajuato Governor election, which will be used to exemplify the functions.

data("gto_2012")
glimpse(gto_2012)

The variables are described in the package documentation ?gto_2012.

Random sampling

The functions select_sample_str and select_sample_prop allow for simple random sampling and stratified random sampling. For example to select a simple random sample with 2% of the data:

gto_srs <- select_sample_prop(gto_2012, frac = 0.02)

For stratified sample we can specify the desired sample size in each stratum or we can specify the percentage of the observations to be selected within each stratum (corresponding to stratified random sampling with proportional allocation).

If we are to specify sample size within strata we need to supply a data.frame with the sample sizes. For example, lets suppose we are stratifying by federal district and the data.frame allo_df specifies the allocation in each stratum, in this case we want to sample 3 polling stations in each stratum:

allo_df <- data.frame(distrito_fed_17 = 1:20, n = rep(3, 20))
allo_df

We use the function select_sample_str:

gto_equal <- select_sample_str(gto_2012, allo_df, n, distrito_fed_17)
table(gto_equal$distrito_fed_17)

And selecting with proportional allocation we can choose a stratified sample with 2% of the polling stations:

gto_str <- select_sample_prop(gto_2012, distrito_fed_17, 0.02, seed = 281982)
table(gto_str$distrito_fed_17)

Note that there is a parameter seed so we can replicate a sample.

Estimation

We begin exemplifying ratio estimation, to compute estimations using ratio estimator we need to know the size of each strata (in this case strata are the local districts):

# count number of polling stations per stratum
gto_stratum_sizes <- gto_2012 %>%
    dplyr::group_by(distrito_loc_17) %>%
    dplyr::summarise(n_stratum = n())

Now lets suppose we have a stratified random sample with 4% of the data (pps):

gto_sample <- select_sample_prop(gto_stratum_sizes, stratum = distrito_loc_17, 
    0.06, seed = 19291)

We then call the function ratio_estimation()

ratio_estimation(gto_sample, stratum = distrito_loc_17, n_stratum = n_stratum, 
    ... = pri_pvem:otros)

For Mr. P, we fit a model per candidate and then compute the proportions using the simulated counts for each polling station-candidate. The following function selects a sample of 6% percent of the data, fits the model using the sample and predicts for the population. Alternatively, one can select a sample before calling the function, and use the parameter frac_sample=1, however, for the estimation to take place the data must include all polling stations and NA whenever the polling station was not in the sample.

mrp_gto <- mrp_estimation(gto_2012, pri_pvem:otros, frac = 0.06, 
    stratum = distrito_loc_17, n_iter = 2000, n_burnin = 1500, 
    n_chains = 2, seed = 19291, parallel = TRUE, mc_cores = 8)

It is worth noting that parallelization is not available in Windows.

mrp_estimation returns a list with the fitted models, so we can evaluate examine the model and evaluate convergence:

mrp_gto$jags_fits$pri_pvem$model
plot(mrp_gto$jags_fits$pan_na)

It also includes a summary table with posterior estimations:

mrp_gto$post_summary

tereom/quickcount documentation built on Dec. 3, 2019, 5:36 a.m.