MSETools

This repository contains all of the code and analyses for the paper titled “On the Reliability of Multiple Systems Estimation for the Quantification of Modern Slavery” (Binette and Steorts, 2021). It is structured as follows:

• The R package MSETools provides a unified interface to multiple systems estimation (MSE) software. It implements best usage practices and computational speedups. Data from multiple system estimation studies of human trafficking has been reproduced for illustrations and analyses.

• The analyses folder contains the analyses and figures for Binette and Steorts (2021). Each analysis is provided as an Rmd document which can be knitted on any platform using the included cache. Figures are saved to png and pdf format into subfolders. Cache can be regenerated by knitting the Rmd documents on computing clusters using SLURM. To run the entire analysis from scratch, make sure that the MSETools package is installed (see instructions below) and use the command cd analyses && make clear_cache && srun make. Long-running programs on a cluster can be run within a detachable terminal (e.g. tmux) to avoid connection issues.

Installation

You can install the development version of MSETools from GitHub:

# install.packages("devtools")
devtools::install_github("OlivierBinette/MSETools")

Summary of MSETools

Implemented models

• dga: The dga() function provides an interface to the dga package of Lum, Johndrow and Ball (2015). The package implements decomposable graphical models with hyper-Dirichlet priors and Bayesian model averaging. Here it has been re-implemented in Rcpp and extended to allow more flexible prior distributions.
• LCMCR: The lcmcr() function provides an interface to the LCMCR package of Manrique-Vallier (2020), which implements the latent class model of Manrique-Vallier (2016). By default, MSETools initializes 200 parallel MCMC chains to provide cross-replication stability – this is necessary since the LCMCR Gibbs sampler fails to converge in some cases. Convergence diagnostics are available through MSETools::diagnostics().
• SparseMSE: The sparsemse() function provides an interface to the SparseMSE package of Chan, Silverman and Vincent (2019), which implements a Poisson log-linear approach with stepwise model selection and bootstrap confidence intervals.

Helper functions

• estimates() computes point estimates and confidence intervals for a list of models.
• batch.estimates() compute estimates as a SLURM job array for use in a cluster.
• diagnostics() computes convergence diagnostics for lcmcr objects.

Datasets

See Binette and Steorts (2021) for a description of the datasets reproduced herein.

Examples

library(MSETools)

Define a list of models fitted to the UK dataset:

models = list(lcmcr(UK), sparsemse(UK), dga(UK), independence(UK))

Compute estimates:

estimates(models)

Parallelize the computation of estimates on a computing cluster:

batch.estimates(models, njobs=4)

References

• Binette, O. and Steorts, Rebecca C. (2021) On the Reliability of Multiple Systems Estimation for the Quantification of Modern Slavery.
• Lax Chan, Bernard Silverman and Kyle Vincent (2019). SparseMSE: ‘Multiple Systems Estimation for Sparse Capture Data’. R package version 2.0.1. https://CRAN.R-project.org/package=SparseMSE
• James Johndrow, Kristian Lum and Patrick Ball (2021). dga: Capture-Recapture Estimation using Bayesian Model Averaging. R package version 2.0.1. https://CRAN.R-project.org/package=dga
• Daniel Manrique-Vallier (2020). LCMCR: Bayesian Non-Parametric Latent-Class Capture-Recapture. R package version 0.4.11. https://CRAN.R-project.org/package=LCMCR