Replication code for "Latent Dirichlet Analysis of Categorical Survey Responses"

Author: Evan Munro and Serena Ng

This repository contains data and code for our paper, which uses hierarchical bayesian latent class models to summarize and interpret heterogeneity in categorical subjective expectations data.

• The main repository folder contains the replication scripts
• data contains:
  1. Raw data and associated description files in text format for Michigan data
  2. Card data in text and RData format and description files in text format
  3. README with sources and data guide
• posteriors Contains saved model posterior means that are used to create the tables for the paper

Replication Instructions

Setup

The replication code requires both an R and a Julia installation. Julia is required only for the plot in Figure 2.

1. Clone our Github repository and change working directory to the repository.

git clone https://github.com/evanmunro/lda-survey-exp
cd lda-survey-exp

2. Install the R package written for the models in the paper dhlvm from Github using the R package devtools:

devtools::install_github("evanmunro/dhlvm")

3. Make sure the following required R packages are installed through CRAN

reshape2, ggplot2, future.apply, kableExtra, stargazer

4. Make sure the following Julia packages are installed

Plots, Distributions

Reproducing Tables and Figures

Tables and Figures are saved in exhibits/. To reproduce Figure 2, run:

julia exhibits.jl

which is saved as dirichlet_density.pdf.

To reproduce Table 1 and Figures 3-5, run:

Rscript exhibits.R

Figure 3 combines mich1_pi_ev.pdf, mich3_pi_ev.pdf, and mich4_pi_ev.pdf. Figure 4 combines beta_PAGO.pdf,beta_PEXP.pdf, beta_BAGO.pdf, beta_BUS12.pdf, beta_DUR.pdf, and beta_UNEMP.pdf
Figure 5 combines beta_rotter_G.pdf, beta_rotter_H.pdf, and beta_rotter_K.pdf.

Table 1 is saved in TeX format as table1.txt.

Reproducing Data Cleaning, Model Estimation and Simulations

1. Rscript mich_analysis.R re-estimates LDA-DS on the Michigan data and saves the model posterior in posteriors/mich_estimate.RData
2. Rscript dhs_analysis.R re-estimates LDA-S on the Card data and saves the model posterior in posteriors/card_estimate.RData
3. Rscript simulations.R runs the simulations described verbally in Section 4.3 in the paper