The datasets

brfssdata publishes one parquet file per BRFSS survey year, 40 files covering 1985 through 2024. Each file is derived from the SAS Transport file CDC distributes for that year by the pipeline in data-raw/, which keeps every row and every column, recodes nothing, renames nothing, and publishes a SHA-256 checksum next to each year it uploads. Variable names keep CDC’s spelling, including the leading underscore on calculated variables such as _LLCPWT and _AGE_G, and values keep their original numeric codes. The only column brfssdata adds is year. What does change on the way in is housekeeping. SAS variable labels move out of the data and into the searchable catalog behind brfss_vars(); the Windows-1252 bytes older files carry in their character columns are re-encoded to UTF-8 so the parquet is valid; blank SAS character fields, which are SAS’s missing value for character data, are stored as nulls rather than empty strings; and the six variables CDC stored as a number in some years and text in others (SEQNO, _RECORD, MRACEORG, WINDDOWN, _MSACODE, RCVFVCH4) are written with one storage type across every year, values unchanged. Without that last step a multi-year read would silently promote the numeric years to text and split one code into 1120 and "1120.0"; with it, and a guard in the package that refuses to combine files whose stored types conflict, that failure mode is gone.

The files are hosted as GitHub release assets, one release per survey year, at https://github.com/muntasirmasum/brfssdata/releases/download/data-YYYY/brfss_YYYY.parquet. read_brfss() downloads a year once into the local cache and queries it with DuckDB afterwards, so a request for two variables reads those two columns, plus year, and leaves the rest of the file alone. Nothing here replaces CDC’s own distribution. The transport files, codebooks, and questionnaires on cdc.gov remain the authoritative source, and this collection is a convenience layer over them.

The catalog

The table below lists every published year with its respondent count, its column count, the size of the hosted parquet file, and the final weight brfss_design() selects for that year. That weight is _FINALWT through 2010 and _LLCPWT from 2011 on, which is the mechanical consequence of the 2011 redesign discussed below.

stats <- read.csv("brfss_year_stats.csv")

catalog <- data.frame(
  Year = stats$year,
  Respondents = format(stats$respondents, big.mark = ",", trim = TRUE),
  Columns = stats$variables,
  `Size (MB)` = stats$size_mb,
  Weight = ifelse(stats$year < 2011, "_FINALWT", "_LLCPWT"),
  check.names = FALSE
)

knitr::kable(catalog, align = "rrrrl")

Year	Respondents	Columns	Size (MB)	Weight
1985	25,221	101	0.8	_FINALWT
1986	34,395	109	1.2	_FINALWT
1987	50,081	117	1.9	_FINALWT
1988	56,448	153	2.3	_FINALWT
1989	66,867	163	2.8	_FINALWT
1990	81,557	194	3.7	_FINALWT
1991	87,846	181	3.8	_FINALWT
1992	96,213	182	4.1	_FINALWT
1993	102,263	198	4.1	_FINALWT
1994	105,853	210	4.9	_FINALWT
1995	113,934	213	4.8	_FINALWT
1996	124,085	264	6.2	_FINALWT
1997	135,582	269	6.2	_FINALWT
1998	149,342	327	8.2	_FINALWT
1999	159,989	282	8.1	_FINALWT
2000	184,450	290	10.5	_FINALWT
2001	212,510	292	11.7	_FINALWT
2002	247,964	311	15.7	_FINALWT
2003	264,684	295	17.9	_FINALWT
2004	303,822	294	16.6	_FINALWT
2005	356,112	326	24.2	_FINALWT
2006	355,710	303	23.4	_FINALWT
2007	430,912	363	36.9	_FINALWT
2008	414,509	292	27.6	_FINALWT
2009	432,607	406	36.6	_FINALWT
2010	451,075	398	32.4	_FINALWT
2011	506,467	455	44.4	_LLCPWT
2012	475,687	360	31.8	_LLCPWT
2013	491,773	337	41.3	_LLCPWT
2014	464,664	280	25.8	_LLCPWT
2015	441,456	331	33.3	_LLCPWT
2016	486,303	276	24.8	_LLCPWT
2017	450,016	359	35.3	_LLCPWT
2018	437,436	276	22.5	_LLCPWT
2019	418,268	343	32.6	_LLCPWT
2020	401,958	280	21.4	_LLCPWT
2021	438,693	304	26.0	_LLCPWT
2022	445,132	329	26.3	_LLCPWT
2023	433,323	351	29.1	_LLCPWT
2024	457,670	302	25.6	_LLCPWT

The whole collection holds about 11.4 million respondents in roughly 737 MB of parquet, which is small enough that a working cache of every year fits comfortably on a laptop. The column count is CDC’s variables for that year plus the added year, so it tracks the size of the core questionnaire together with whatever optional modules and calculated variables went into the public-use file. CDC’s 1985 page reports 100 variables, for instance, and the hosted 1985 file has 101 columns.

Respondents over time

library(ggplot2)

ggplot(stats, aes(year, respondents)) +
  geom_vline(xintercept = 2011, linetype = "dashed", color = "grey55") +
  annotate(
    "text",
    x = 2010.4,
    y = max(stats$respondents),
    label = "2011 redesign",
    hjust = 1,
    size = 3.4,
    color = "grey35"
  ) +
  geom_line(color = "#1f4e79", linewidth = 0.7) +
  geom_point(color = "#1f4e79", size = 1.3) +
  scale_y_continuous(
    labels = function(x) format(x, big.mark = ",", scientific = FALSE)
  ) +
  labs(x = NULL, y = "Respondents") +
  theme_minimal(base_size = 12)

The shape of that curve is mostly the history of which states were in the survey and how many interviews they funded. Sample size climbs as states join and expand their samples, peaks in 2011, the year cell-phone respondents entered the combined public-use file, and has since varied between roughly 400,000 and 490,000 completed interviews a year, consistent with the “more than 400,000 adult interviews each year” CDC describes as the system’s current volume.

Milestones

CDC initiated BRFSS in 1984 with 15 states collecting risk-behavior data through monthly telephone interviews. Participation grew from there, and the survey now collects data in all 50 states, the District of Columbia, and participating US territories, completing more than 400,000 adult interviews a year. A public-use file for 1984 exists too, 12,258 records from those first 15 states, but its documentation now sits in CDC’s web archive instead of the current annual-data index, and it is not part of these releases, which begin at 1985.

The 2011 survey year is the one methodological boundary that matters for almost every analysis. CDC added a cell-phone sampling frame and replaced post-stratification with iterative proportional fitting, or raking, which allowed cell-phone data to be incorporated and brought additional demographic characteristics such as education, marital status, and home ownership into the weighting. The annual data pages since then carry the same caution, that the year’s data are not directly comparable to BRFSS years before 2011 because of the change in weighting methodology and the addition of the cell phone sampling frame. brfss_design() treats that statement as a hard boundary and refuses to pool across it without allow_break = TRUE. Another consequence of 2011 is visible in the catalog above. The file grows to 455 columns that year, the widest in the series, and the respondent count reaches its maximum.

Coverage is not fixed after 2011 either, and the reporting areas in a public-use file change from year to year. For 2020 the data cover 50 states, the District of Columbia, Guam, and Puerto Rico. A year later the US Virgin Islands are included and the state count is 49. By 2023 it is 48, because Kentucky and Pennsylvania did not collect enough data that year to meet CDC’s minimum requirements for the public file, with the District of Columbia and the same three territories still in. State-level or region-level work should check the reporting areas for each year instead of assuming a constant panel of states.

Per-year CDC documentation

Each survey year has its own documentation page on cdc.gov, all of them reachable from the annual survey data index. Recent years follow the pattern https://www.cdc.gov/brfss/annual_data/annual_YYYY.html, for example annual_2023.html. Years through 2011 use an .htm extension instead, and the pages for 1984 through 1988 have moved to CDC’s web archive.

Those pages are where to answer the questions this package deliberately leaves alone. A recent year carries the annual overview (background, design, data collection and processing, and statistical and analytical issues), the codebook with frequencies for every variable, the calculated-variable documentation, the weighting formula, the summary data quality report with response rates, the comparability notes, a listing of which optional modules each state used, and the questionnaire itself. Early years are much thinner. The 1985 page offers the data files and a SAS conversion program and little else.

Every page also states the record count for the year, and that is the number reproduced in the catalog table above. CDC reports 433,323 records for 2023 and 401,958 for 2020, and the hosted files hold exactly that many rows.

When a result looks surprising, the codebook for that year is usually the fastest explanation, because a question’s wording, its answer categories, or the set of states that asked it may have changed.

Variables across years

The hosted files are the survey as CDC released it, which means variable sets drift. The catalog that ships alongside the data holds 2,128 distinct variable names across the 40 years. brfss_vars() searches it by name and label and reports the years each variable appears in. A search for smok pulls in most of the smoking battery at once, so the six names below are picked out of a much longer result.

library(brfssdata)

smoking <- brfss_vars("smok")
nrow(smoking)
#> [1] 96

smoking[smoking$variable %in% c(
  "SMOKE100", "SMOKENOW", "SMOKEDAY", "SMOKDAY2", "STOPSMOK", "STOPSMK2"
), ]
#> # A tibble: 6 × 3
#>   variable label                                   years    
#>   <chr>    <chr>                                   <chr>    
#> 1 SMOKDAY2 FREQUENCY OF DAYS NOW SMOKING           2005-2024
#> 2 SMOKE100 SMOKED AT LEAST 100 CIGARETTES          1985-2024
#> 3 SMOKEDAY FREQUENCY OF DAYS NOW SMOKING           1996-2004
#> 4 SMOKENOW CURRENTLY SMOKE                         1985-1995
#> 5 STOPSMK2 STOPPED SMOKING IN PAST 12 MONTHS       2001-2024
#> 6 STOPSMOK QUIT SMOKING A DAY OR MORE IN PAST YEAR 1990-2000

The pattern in those rows is typical. One anchor question survives the whole series, SMOKE100 running from 1985 to 2024, while the question next to it is renamed as its wording is revised, so current smoking frequency is SMOKENOW through 1995, SMOKEDAY from 1996 to 2004, and SMOKDAY2 from 2005 on. Quit attempts follow the same path from STOPSMOK to STOPSMK2. An analysis that spans those breaks has to harmonize the variants deliberately, and the year ranges are what tell you where the seams are.

Because variable sets differ, combining years fills a variable that a year does not carry with NA instead of failing. That is convenient but easy to misread, so confirm with brfss_vars() that a variable actually exists in every year requested before treating a column of missing values as a real zero-prevalence result.

brfss_vars("SMOKDAY2", years = c(2003, 2005))
#> # A tibble: 1 × 3
#>   variable label                         years
#>   <chr>    <chr>                         <chr>
#> 1 SMOKDAY2 FREQUENCY OF DAYS NOW SMOKING 2005

The reply names 2005 only, so in a read spanning both years the 2003 rows would come back all NA for that column.

Known limitations

The collection has a few known gaps.

The files carry no upstream provenance yet. brfss_year_info() reports what the hosted copy contains, but not the CDC URL it was built from, when it was retrieved, or the hash of the upstream Transport file, so there is no way from inside the package to tell which CDC revision of a year you are holding. Those columns are planned for the next data republish.

Calculated variables keep CDC’s stored scaling. _BMI5 and _DRNKWK2 carry two implied decimal places, so a stored 2704 is a BMI of 27.04, and nothing in the files or catalogs records that factor, nor units or valid ranges. The per-year CDC codebook, linked from brfss_year_info()$codebook_url, is the source for all three.

Value labels stop at 1998, because CDC distributes no usable format libraries for earlier years. na = TRUE therefore has no catalog to consult before 1998 and warns rather than guessing; special codes in those years must be recoded by hand from CDC’s codebooks.

Module analyses that require CDC’s questionnaire-version datasets and their _LCPWTV1 to _LCPWTV3 weights are not supported, because the hosted annual files do not include those weights. For the module weights the files do carry, the confinement check in brfss_design() warns when a module variable is analyzed under a full-sample weight.

Releases are not immutable. A corrected year replaces the published bytes under the same tag; the manifest checksum and the daily cache recheck notice the change, but there is no way yet to pin an analysis to an exact prior snapshot.

Citing the data

All data originate from the CDC BRFSS annual survey files, which are in the public domain. Suggested citation for the data:

Centers for Disease Control and Prevention (CDC). Behavioral Risk Factor Surveillance System Survey Data. Atlanta, Georgia: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, [appropriate year].

The hosted parquet files are built directly from CDC’s published SAS Transport files. That pipeline lives in data-raw/, and each published year ships with a .sha256 file, so a downloaded parquet can be checked against the release it came from.