Enabling Data Disaggregation of Asian American Subgroups: A Dataset of Wikidata Names for Race Imputation

This repository contains code and data related to our work on using Wikidata to derive name-race distributions for race imputation of Asian American subgroups. Please reach out to Derek Ouyang at [email protected] with any questions or feedback.

Quick Start Guide

If you would like perform predictions on your own data immediately, the best place to start is notebooks/0_starter_code.Rmd.

The two key data files we generated through this work are:

• data/name_race_data/disagg_name_priors_main.rda: contains multiple dataframes like p_race_given_surname.
• data/geography/geo_race_table.rds: a list object containing multiple dataframes like p_county_given_race.

These are also available at Harvard Dataverse: https://doi.org/10.7910/DVN/LEOECM

Full File Structure

notebooks

• 0_starter_code.Rmd: A template for users to prepare their own data and perform predictions.
• 1_run_wqs.Rmd: Generates the raw Wikidata queries. See scripts/subgroup_queries.json, scripts/wikidata_query_helper_functions.R, and data/raw_wikidata/.
• 2_process_wiki_result.Rmd: Processes the raw Wikidata queries into first name and surname lists. See scripts/name_cleaning_helper_functions.R, external_data/, and data/intermediate_data/.
• 3_process_ipums.Rmd: Processes the raw IPUMS extracts into first name and surname lists. See external_data/ and data/intermediate_data/.
• 4_create_disagg_geo_tables.Rmd: Processes raw American Community Survey data into geography-race tables. See external_data/ and data/geography/.
• 5_create_name_prior_tables.Rmd: Processes first name and surname lists into name-race tables. See scripts/name_table_helper_functions.R and data/name_race_data/.
• 6_validate_afc.Rmd: Conducts the validation on EHR data presented in the paper. Note that the EHR data itself is not publicly available. See scripts/imputation_helper_functions.R and output/results/.
• 7_create_results.Rmd: Produces the figures and tables presented in the paper. See output/figures/.

scripts

• subgroup_queries.json
• wikidata_query_helper_functions.R
• name_table_helper_functions.R
• name_cleaning_helper_functions.R
• imputation_helper_functions.R: Contains the key function predict_disagg_race().

external_data

• census_surnames.rds: Surnames from the Census 2010 Surname list.
• surname_census_L2.rda: Surnames from Rosenman et al., 2022: https://doi.org/10.7910/DVN/SGKW0K
• ipums_race.xml and ipums_race.dat.gz: Extract from IPUMS of Asian Americans from historical censuses with detailed Asian subgroup information.
• ipums_bpl.xml and ipums_bpl.dat.gz: Extract from IPUMS of Asian Americans from historical censuses with detailed birthplace information.
• ipums_fbpl.xml and ipums_fbpl.dat.gz: Extract from IPUMS of Asian Americans from historical censuses with detailed father's birthplace information.

There are two additional SSA name list files that can be added to external_data/, with the filenames below to be correctly loaded in notebooks/6_validate_afc.Rmd and notebooks/7_create_results.Rmd. In particular, these files will enable you to generate the hybrid approach described in the paper. These SAS name list files are not available for direct public download, but can be requested from the original author at [email protected].

• external_data/SSA_Givennames.csv
• external_data/SSA_Surnames.csv

data

• raw_wikidata/: The rawest form of Wikidata query outputs as separate .rds files for each subgroup.
• geography/: Geography-race information from ACS, in raw count form and as distribution tables.
• intermediate_data/: Various intermediate files in the processing pipeline from raw Wikidata queries to name-race tables.
• name_race_data/: The final name-race tables, including alternative versions.

output

• results/: Various .rds files holding the input data necessary to produce the figures in the paper.
• figures/: PNG files of figures in the paper.

Future Work

We outline four potential directions of future work below.

First, while our approach could be extended to smaller Asian subgroups, we are limited by the sample sizes we can obtain from Wikidata. Names from a selection of smaller Asian subgroups could be grouped together as an “Other Asian” category, and the probability distributions for name and geography could be appended with this category, accordingly. Future work could explore the performance of such an augmented dataset in settings that include other Asian subgroups.

Second, future work could explore prediction performance in settings that include all race groups, where researchers have no race information at all. A potential solution is to deploy conventional BISG to first determine the likelihood that individuals identify as API, and then perform our disaggregated imputation, propagating uncertainties through the two stages. Those who have a high API likelihood may, of course, include NHPI individuals; prediction performance impacts are likely similar in nature to those driven by members of less common Asian subgroups. We expect Wikidata can be used to generate reasonable name-race distributions that include NHPI as an aggregated category or disaggregated categories. More generally, our methodology can likely be applied across all disaggregated subgroups required in the 2024 revised federal standards, but predictive performance is likely to be extremely heterogeneous, given sparse Wikidata coverage of individuals with structured name, birthplace, and citizenship information from less populous countries and regions, as well as the fundamental variation in distinctiveness of names across certain subgroups (i.e., ethnic subgroups that share a common written language).

Third, future work could explore whether performance could be improved by incorporating these probabilistic inputs into name algorithms that go beyond current Bayesian methods, such as machine learning models that have been shown to yield accuracy improvements. As before, validation with a subset of ground truth labels will provide the most valuable information about whether such augmented methods are advantageous in particular settings.

Fourth, this domain of work would fundamentally benefit from the release of disaggregated name-race distribution information sourced directly from large U.S. administrative datasets, such as SSA files or Census Bureau Decennial surveys. Should such data become available, future work could conduct a more thorough assessment of Wikidata's accuracy, and determine whether there are still use cases for Wikidata in subgroup race imputation.