https://github.com/alexgao09/structuredpriorsmrp_public
Repository for reproducing results in structured priors MRP paper
https://github.com/alexgao09/structuredpriorsmrp_public
Last synced: 4 months ago
JSON representation
Repository for reproducing results in structured priors MRP paper
- Host: GitHub
- URL: https://github.com/alexgao09/structuredpriorsmrp_public
- Owner: alexgao09
- License: mit
- Created: 2019-09-08T00:17:53.000Z (over 5 years ago)
- Default Branch: master
- Last Pushed: 2020-06-27T16:16:16.000Z (almost 5 years ago)
- Last Synced: 2024-06-05T02:31:49.169Z (10 months ago)
- Language: R
- Homepage:
- Size: 196 KB
- Stars: 17
- Watchers: 3
- Forks: 4
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- jimsghstars - alexgao09/structuredpriorsmrp_public - Repository for reproducing results in structured priors MRP paper (R)
README
# Code for reproducing results in "Improving multilevel regression and poststratification with structured priors"
## Authors
- Yuxiang Gao
- Lauren Kennedy
- Daniel Simpson
- Andrew Gelman
Accepted for publication in _Bayesian Analysis_. arXiv preprint: https://arxiv.org/abs/1908.06716
## Steps to run simulation pipeline in simulation directory (directed structured priors example)
1. The two .dta files in poststrat_pipeline_testing_age3_v3_posteriorvariance.R are retrieved from the paper, Estimating State Public Opinion with Multi-level Regression and Poststratification using R. URL: https://scholar.princeton.edu/jkastellec/publications. Download these two files before running the simulation pipeline, and put them in the simulation directory.
2. Choose age preference curve with coef_age in poststrat_pipeline_testing_age3_v3_posteriorvariance.R
3. Choose sample size, simulation runs, probability indices and number of age categories with sample_size, runs, r, age_grouping_multiplier in poststrat_pipeline_testing_age3_v3_posteriorvariance.R
4. Run poststrat_pipeline_testing_age3_v3_posteriorvariance.R
5. Run threemodelwriteup_v3_posteriorvariance.R with the same configurations. This will produce the bias plots shown in the paper for a given configuration.
6. Run proportion_agecat.R to visualize the share of simulations that the RW and AR priors outperformed the IID priors
## Steps to run simulation pipeline in simulation_spatialmrp directory (undirected structured priors example)
1. Choose sample_size in icar_mrp_simulation_bym2.R and then run the script. This conducts spatial MRP simulations, based off binomial regression
2. Run icar_mrp_viz.R with the same configurations to visualize the spatial MRP simulations
3. Run proportion_puma.R to visualize share of simulations that the BYM2 priors outperformed the IID priors
## Steps to run data analysis on 2008 Annenberg phone survey in realdata_annenberg directory
1. Request access for the 2008 National Annenberg Election Survey (NAES) telephone from the Annenberg Public Policy Center and put the phone survey text file into the realdata_annenberg directory. Rename it to annenbergphone2008.txt
2. acs_ps.rds is the 2006-2010 5-year American Community Survey (ACS) in a cleaned-up format. Create a new path ~/Desktop/annenbergdata/fiveyearacs/ and then download the 5-year ACS and unzip in this path. Finally, run ACS_PS_annenberg_fiveyear.R in the preprocess directory to get acs_ps.rds in ~/Desktop/annenbergdata/fiveyearacs/
3. Make sure state_level_update.dta, gay_marriage_megapoll.dta, acs_ps.rds, annenbergphone2008.txt are in realdata_annenberg directory. The first two files are from Step 1 in the simulation pipeline.
4. Run fitmodel_annenberg_v3.R for age_grouping_multiplier = 12,48,72 and for all three models in the variables m_file,m_name
5. Run poststrat_annenberg_v3.R
## Notes:
1. To download the 2006-2010 5-year ACS, download both United States Population Records and United States Housing Unit Records in the data set 2006-2010 ACS 5-year Public Use Microdata Samples (PUMS) - CSV format, which can be found on https://factfinder.census.gov/faces/nav/jsf/pages/searchresults.xhtml?refresh=t#
2. The US heatmap and proportion tables in the Annenberg data analysis are derived from annenbergphone2008.txt and acs_ps.rds.
The covariate N in acs_ps.rds, which corresponds to a person's weight, is summed when calculating heatmap and proportion values. As an example, to get the proportions for the education covariate in the 2006-2010 ACS, we use:
```R
acs_ps %>% group_by(education) %>%
summarise(percentage = sum(N)/sum(acs_ps$N)) %>%
dplyr::mutate(percentage = round(100 * percentage, 2)) %>% data.frame()
```
3. For a working simulation example that doesn't require downloading data files, see my StanCon 2019 repo.
4. Noncentered parameterizations were used for the simulation studies whereas centered parameterizations were used for the real data analysis. NUTS sampling in Stan struggled with the noncentered parameterizations for the real data analysis but centering fixed this.
5. In the simulation_spatialmrp directory: poverty_poststrat.rds contains the poststratification matrix that's derived from the 2013-2017 5-year ACS, ma_adj_matrix_sparse.rds contains the adjacency matrix of the 52 PUMAs in the state MA, smooth_x.rds contains a sample from a Gaussian Markov random field over the space of 52 PUMAs. ma_adj_matrix_sparse.rds is derived from the tigris package.