pestim: pestim: a hierarchical model to estimate population counts...
In MobilePhoneESSnetBigData/pestim: Population Estimations Using Mobile Phone Data

Description Context The hierarchical model in a nutshell Computational paradigm

This package provides an implementation for a hierarchical model to combine both aggregated mobile phone data and external official (administrative or survey) data to produce estimates of population counts in each cell of a division of a territory.

This package has been developed in the context of a European research project within the European Statistical System called ESSnet on Big Data. More specifically this work corresponds to the work package on mobile phone data by which we assess the use of this data source in the production of official statistics. The goals of the project is many-fold. Firstly, the issue of accessing these data for the production of official statistics initially for research and then for standard production has been investigated. Secondly, in a hands-on bottom-up approach, we make some initial methodological proposals to produce concrete statistical output using those data sets compiled in the preceding phase. Thirdly, in parallel, IT tools, architecture and software development are assessed especially in contrast to traditional computer frameworks. Finally, quality is appraised especially in the context of the European Statistics Code of Practice and ESS Quality Assurance Framework. This package provides a first-step implementation of software routines to present a proof of concept about a methodological proposal (see below) to make inferences about a target population from a mobile phone dataset.

The methodological proposal giving rise to this package focuses on the inference exercise connecting aggregated mobile phone data with a target population under analysis. In concrete, the goal is to provide estimates of population counts in each cell in which we have divided the territory for which the telecommunication network provides count data. The estimation is assisted with official data at a larger time scale (either from a population register or from a survey).

The model rests on two working assumptions:

Given that mobile phone data and official data operate at different time scales, we assume that there exists an initial time instant in which we can equate population figures from both sources.
The mobility patterns of individuals do not depend on the mobile network operator which they are subscribed to.

The model works in two stages. Firstly at the initial time instant, we use data from both sources to make the inference for the actual population counts in each cell. Secondly, the time evolution of these counts are produced using the transition matrices from cell to cell of individuals provided by the mobile network operator.

The essence of the model is to emulate the ecological sampling setting in which the number of detected individuals in each cell follows a binomial distribution Bin(N_{i}, p_{i}) whose parameter N_{i} is the target of the model and is assigned a weakly informative prior and the detection probability is also assigned a weakly informative prior based upon both data sources.

Computations are conducted following the Bayesian paradigm. In this sense the generation of simulated populations according to different probability distributions is at the core of the package. In this sense the package contains basically three types of functions:

Auxiliary functions, providing computation of mathematical functions such as the ratio of two beta functions, the confluent hypergeometric function, an optimization routine for a concrete probability distribution, etc. Examples of these functions are ratioBeta, kummer, Phi, modeLambda.
Distribution-relation functions, providing computation regarding the generation of random deviates according to different probability distributions comprising both priors, posteriors, and the generation of parameter specifications for these distributions. Examples of these functions are dtriang, rtriang, ptriang, qtriang, dlambda, rlambda, rmatProb, rN0, rNt, rNtcondN0, rg, rp, alphaPrior, genAlpha, genUV.
Estimation-relation functions, providing computation of estimates based upon the populations generated with the preceding functions. Examples of these functions are postN0, postNt, postNtcondN0.