cpsvote: A Social Science Toolbox for Using the Current Population Survey’s Voting and Registration Supplement
cpsvote helps you work with data from the Current Population Survey’s (CPS) Voting and Registration Supplement (VRS), published by the U.S. Census Bureau and Bureau of Labor Statistics. This high-quality, large-sample survey has been conducted after every federal election (in November of even years) since 1964, surveying Americans about their voting practices and registration. The raw data, archived by the National Bureau of Economic Research, is spread across several fixed-width files with different question locations and formats.
This package consolidates common questions and provides the data in a structure that is much easier to work with and interpret, since much of the basic factor recoding has already been done. We also calculate alternative sample weights based on demonstrated changes in non-response bias over the decades, recommended by several elections researchers as a best practice. Documentation of this reweighting is provided in
We have provided access to VRS data from 1994 to 2018, and anticipate updating the package when 2020 data becomes available.
Version 0.1 is on CRAN!
You can also install the development version from our GitHub repository.
We have written several functions to transform the VRS from its original format into a more workable structure. The easiest way to access the data is with the
This will load the prepared VRS data into your environment as a tibble called
cps. The first time you try to load a given year of data, the raw data file will be downloaded to your computer (defaulting to the relative path “./cps_data”). This can take some time depending on your internet speeds. In future instances, R will just read from the data files that have already been downloaded (defaulting to the same “cps_data” folder), as long as you correctly specify where these are stored. See
?cps_allyears_10k for a description of the columns and fields that
We recommend using a single R project for your CPS analysis where these files can be stored (this will work with the default options), or storing one set of CPS files in a steady location and specifying this absolute file path each time you load in the data. If you specify a location that does not have the correct files, these functions will attempt to re-download the data from NBER, which can take up noticeable time and storage space.
We have also included a 10,000 row sample of the full VRS data, which comes with the package as
cps_allyears_10k. This is particularly useful for planning out a given analysis before you download the full data sets.
The CPS has survey weights that are necessary to calculate accurate estimates about the US population. Two R packages that work with survey weighting are
srvyr (a tidyverse-compatible wrapper for
survey). You can see more examples and details on weighting in
vignette("voting"), but here is one example of using
srvyr to calculate state-level voter turnout among eligible voters in 2018.
These estimates follow closely Dr. Michael McDonald’s estimates of turnout among eligible voters in the November 2018 General Election. For a detailed examination of how non-response bias has affected the use of CPS for estimating turnout, see
vignette("voting"). We thank the U.S. Elections Project at the University of Florida for the turnout estimates.
In addition to the basic function listed above, you can customize several steps in the process of reading in the VRS data. If you’ve worked with the CPS before, you may already have some code to read in analyze this survey data. We still hope that this package can help you organize your workflow or ease some of the more tedious steps necessary to work with the CPS.
Be sure to refer to the CPS documentation files when working with alternative versions of the VRS data. We have included the function
cps_download_docs() to provide the documentation versions that match this data. These are all in PDF format (and several are not text-based), so they are not easy to search through.
cps_load_basic() is a wrapper for several constituent steps that have their own parameters and assumptions. We’ve detailed the changes made to get from the raw data file to the cleaned file in
cps_download_data(path = "cps_data", years = seq(1994, 2018, 2)) cps_download_docs(path = "cps_data", years = seq(1994, 2018, 2)) cps_read(years = seq(1994, 2018, 2), dir = "cps_data", cols = cpsvote::cps_cols, names_col = "new_name", join_dfs = TRUE) %>% cps_label(factors = cpsvote::cps_factors, names_col = "new_name", na_vals = c("-1", "BLANK", "NOT IN UNIVERSE"), expand_year = TRUE, rescale_weight = TRUE, toupper = TRUE) %>% cps_refactor(move_levels = TRUE) %>% cps_recode_vote(vote_col = "VRS_VOTE", items = c("DON'T KNOW", "REFUSED", "NO RESPONSE")) %>% cps_reweight_turnout()
cps_download_data()will download the data files from NBER according to
yearsinto the folder at
path. This is automatically called by
cps_read()when the CPS data files are not found in the provided
dir- it will search for files with the 4-digit year associated with their data.
cps_download_docs()will downlaod the pdf documentation into
pathfor each year supplied in
years.The documentation here is aligned with the NBER data, and other data sources (such as ICPSR) may have edited the data such that their data or documentation does not line up with the NBER data and documentation. By using the NBER data through
cps_download_docs(), you can make sure that the fields you look up in documentation are the proper fields referenced in the data.
cps_read()is the function that actually loads in the original, (mostly) numeric data from files defined by the arguments
dir. Since the raw data is in fixed-width files, you have to define the range of characters that are read. You can see the default set of columns in the included data set
cps_cols, or supply
colswith your own specifications of columns (for details on adding other columns, see
names_colargument details which variable in
colswill become the column names for the output; we have provided the original CPS names as
cps_name, but recommend using
new_nameas it is more informative and accounts for questions changing names (“PES5”, “PES6”, etc.) across multiple years.
join_dfslets you join multiple years into one
tibble, and should only be used if you’re sure that a column name (like “PES5”) refers to the same question across all years you read in.
cps_label()replaces the numeric entries from the raw data with appropriate factor levels (as given by the data documentation; see
cps_download_docs()). We have taken the factor levels as written from the PDFs, including capitalization, typos, and differences across years. This is provided in the included
cps_factorsdataset, but you can supply the
factorsargument with your own coding (for details on changing factor levels or adding them for a new column, see
names_colargument defines which column of
factorscontains the column names that match the incoming data set to be labelled. Further:
na_valsdefines which factor levels should be marked as
expand_yearturns the two-digit years in some files into four-digit years (e.g. “94” becomes “1994”), and
rescale_weightdivides the given weight by 10,000 (as noted by the data documentation) to ensure accurate population sums.
toupperwill make all the factor levels upper case, which is useful because as-is the factors are a mix of sentence case and upper case.
cps_refactordeals with all of the typos, capitalization, and shifting questions across years. We have attempted here to consolidate factor levels and variables in a way that makes sense. For example, one common method of assessing vote mode (in-person on Election Day, early in-person, or by mail) has been split between two separate questions from 2004 onwards, and this function consolidates those two questions (and the one question of previous surveys) into one
VRS_VOTEMETHOD_CONvariable. Note that this function will only work with certain column names in the data; see
?cps_refactorfor more details.
cps_recode_vote()recodes the variable
VRS_VOTEaccording to two different assessments of voter turnout. The new variable
cps_turnoutwill calculate turnout the same way that the Census does, while another new variable
hurachen_turnoutwill calculate turnout according to Hur & Achen (2013). These two methods differ in how they count responses of “Don’t know”, “Refused”, and “No response”; see
vignette("background")for more details.
cps_reweight_turnout()adds a new variable,
turnout_weight, that reweights the original
WEIGHTaccording to Hur & Achen (2013) to account for the adjusted turnout measure. This corrects for increased nonresponse to the VRS over time, as well as a general pattern of respondents overreporting their personal voting history (though the CPS sees less overreporting than other surveys). See
You can use different combinations of these functions to customize which CPS data is read in. For example, this code would load the 2014 VRS data with the original column names and numeric data.
You can then apply factor labels to this data.
Note that some features (like
cps_refactor()) won’t work on certain customized versions of the data, because they are relatively hard-coded based on specific column names. For example, correcting “HIPSANIC” to “HISPANIC” only works if you know which column represents the Hispanic flag. Feel free to take the code from functions like this and adapt based on your own column names.
vignette("basics")provides an intro to the package with some basic instructions for use, and mirrors our GitHub README
vignette("background")describes our intellectual rationale for creating this package
vignette("add-variables")describes how additional variables from the CPS can be merged with the default dataset
vignette("voting")does a deep dive into how to use the CPS and the default datasets from
cpsvoteto look at voter turnout and mode of voting