5 Estimation Strategies

For this chapter we need the following packages and the predefined objects.

library(tidyverse)
library(sf)
library(raster)
library(knitr)
library(kableExtra)
library(ggthemes)
library(DT)
library(transformr)
library(gganimate)
set.seed(762)

census.classified.final.sf <- readRDS("C:/Users/Marco/Vysoká škola ekonomická v Praze/Tony Wei Tse Hung - YAY/working objects/census.classified.final.sf.rds") 

coverage.areas.final <- readRDS("C:/Users/Marco/Vysoká škola ekonomická v Praze/Tony Wei Tse Hung - YAY/working objects/coverage.areas.final.rds") 

c.vec <- readRDS("C:/Users/Marco/Vysoká škola ekonomická v Praze/Tony Wei Tse Hung - YAY/working objects/C.vec.df.final.new.rds") 

census.geo.body <- readRDS("C:/Users/Marco/Vysoká škola ekonomická v Praze/Tony Wei Tse Hung - YAY/working objects/shape.focusarea.rds")

coverage.areas.comp <- coverage.areas.final %>%
  left_join(c.vec, by = "antenna.ID") %>%
  filter(!phones.sum == 0)

# select only the internal id and the mobile phone population per tile
census.de.100m.tile.pop <- census.classified.final.sf %>% 
  dplyr::select(internal.id, pop)

We have implemented two specific estimation strategies which each have two specifications:

  • Voronoi tessellation (tower locations OR coverage area centroids as seeds)

  • MLE Poisson

These four estimation techniques produce a tile specific estimate. The sum of these estimates always equals the sum of the c.vector, i.e. all registered mobile phones. This characteristic explains the approach of all of these estimators: Taking the complete number of registered phones in the c.vector and allocating them across all the tiles. There is however a fundamental difference between the Voronoi estimator and the MLE Poisson: The second one allows for the realistic setting of overlapping antenna coverage. As shown in the chapter before there are many tiles that are covered by multiple antennas. Current research suggests that this can result in more accurate estimations.

This chapter will focus on the two estimation strategies with their chosen specifications. In the next subchapter we will focus on the Voronoi estimator.