Skip to contents

Changelog

Source: NEWS.md

brfssdata 0.1.0

Initial CRAN release.

Correctness

  • The missing-code matcher now recognizes CDC’s calculated-variable buckets whose labels carry a trailing noun (“Don’t know, refused or missing values” and kin, plus the RACE2 family’s “component question” wordings). Before this, na = TRUE (the brfss_design() default) silently left genuine code-9 don’t-know/refused answers in 18 variable/code combinations, including _FRTLT1A (51,087 uncleared rows in 2021 alone), _HLTHPLN, _VEGLT1A, and the _LMT* family, overstating those denominators by several points. The extension was validated against every distinct catalog label: it adds exactly those buckets and loses none, and “Doctor refused when asked”-style substantive answers still never match.
  • Six variables that CDC stored as a number in some years and text in others (SEQNO, _RECORD, MRACEORG, WINDDOWN, _MSACODE, RCVFVCH4) are now written with one type across every hosted year, values unchanged. Multi-year reads previously promoted the numeric years to text, so the same MSA appeared as both "1120" and "1120.0" in one column, silently splitting groups and defeating missing-code matching. read_brfss() additionally refuses to combine files whose stored types conflict (brfssdata_type_conflict), so a stale cache mixed with current releases can never reproduce the bug silently; brfss_cache_clear() is the named remedy.
  • A user-supplied weight in brfss_design() now subsets to the rows the weight covers, with a brfssdata_weight_subset_note message. The documented child-module call, brfss_design(2023, weight = "_CLLCPWT"), previously always failed on real data, because a module weight is missing outside its module’s records (383,782 of 433,323 rows in 2023) and the design constructor refused missing weights. The automatic era weight still aborts on missing values, which there indicate a damaged file, and an explicitly named full-sample weight (_LLCPWT, _FINALWT) follows that same rule instead of silently subsetting.
  • The matcher also recognizes the abbreviations CDC’s 1998 to 2001 format libraries use for don’t-know and refused (“UNK/REF”, “UNK”, “REF”, “UNKNOWN”), together with the bare “N/A” and “N/A,REF” placeholders from the same years. Validated against every row of the shipped catalog, this matches 786 rows across 370 variables, all in 1998 to 2001, and loses none. Without it, na = TRUE left those codes in the data for years the catalog nominally covers: 145,594 cells in the 1998 file (_DRNKMO code 9999 alone on 117,351 rows) and 380,693 in 2001, so any mean or proportion from an affected variable was wrong. Spelled-out “not applicable” is deliberately not a token, because in later years it names ordinal scale positions such as GETHIV code 5.
  • na = TRUE says so when requested years have no value-label catalog at all (1985-1997, a classed brfssdata_na_coverage_warning) or mostly lack it (1998 covers under a quarter of that file’s variables, a brfssdata_na_coverage_note), instead of silently changing nothing.
  • labels = TRUE refuses to convert a variable whose label wording changed meaning across the requested years, keeping CDC’s numeric codes and naming it in a brfssdata_label_drift_warning. Applying the newest year’s labels to every row would restate what earlier respondents answered: CDC reused COLNTES1 and SIGMTES1 codes 3 to 5 for different colorectal screening intervals from 2022 on, so read_brfss(2021:2024, vars = "COLNTES1", labels = TRUE) would otherwise return a wrong interval distribution. This matches how every other ambiguity gate (incomplete formats, mismatched code sets, duplicate codes or labels, uncovered values) already behaves. The wording comparison folds < and > to words before stripping punctuation, so CDC’s house abbreviation (“anytime < 12 months ago” against “anytime less than 12 months ago”) counts as cosmetic and still converts.
  • The pooled state-participation warning (brfssdata_pooled_states_warning) counts participation over the rows a user-supplied weight covers, the population the design actually estimates, not over the whole annual file. Counting over the file names states the design never contained, and goes silent when the files agree on coverage but the weight’s domain does not, which is exactly the case for 2022 pooled with 2023 under _CLLCPWT.
  • Blank SAS character fields are stored as missing values, not "", and code matching is proof against R’s scientific notation, so a future round code at or above 100,000 cannot be skipped.
  • The label-catalog builder reads two SAS syntaxes it used to misread, so the shipped catalog gains rows it had been dropping and loses rows it had invented. A comma-separated code list (77,99 = 'UNK/REF', CDC’s style through 2001) previously kept only the last code, which is why the whole 7, 77, and 777 don’t-know family was missing from 1998 to 2001 and na = TRUE left those codes in place: a 2000 PHYSHLTH mean computed the documented way was 4.41 days against 3.34 once the don’t-know answers are cleared. A range written 244-<777 previously matched its bare endpoint, so the catalog claimed WEIGHT code 777 meant “244 +” when 777 is the don’t-know sentinel, and 2,736 respondents in 2000 alone carried it. Ranges now emit no code and mark their format incomplete, which keeps factor conversion refused where it always should have been.
  • Four label errors CDC published are corrected in the catalog, each recorded with the source that justifies it. CDC’s 2002 library assigns one generic format to 17 variables and labels its code 88 “Never smoked regularly”, so PHYSHLTH, MENTHLTH, POORHLTH and a dozen others carried a smoking label on a code that means None there; the smoking items keep CDC’s wording. DISPCODE 1999 code 2 read “REFUSED”, which the missing-code matcher took for a missing bucket although it names an interview disposition, and now reads “02-REFUSED” as CDC wrote it from 2000. _IMPNPH 2002 code 7 no longer inherits a don’t-know label from the raw question’s format. The builder now also reports every code whose missing or substantive status flips between adjacent years, so this class of error cannot ship unnoticed again.
  • TYPEARTH code 88 is now read as CDC labeled it in 1998 and 1999, where their format libraries group it with 77 and 99 as unknown or refused; CDC splits it out as “NEVER SAW A DOCTOR” from 2000 on. Under na = TRUE this clears 11 rows in 1998 and 49 in 1999 that previously survived. The package reports each year’s published labels and does not reinterpret one year’s codebook with another’s.
  • The rename crosswalk covers the personal-doctor family (PERSDOC, PERSDOC2, PERSDOC3), one of CDC’s flagship healthcare-access measures. Reading PERSDOC3 across 2019 to 2023 returned 820,226 silently empty rows for 2019 and 2020 with no rename note, and brfss_crosswalk("PERSDOC3") reported that the variable had kept one name throughout, which was false. The zero-match message no longer makes that claim for any variable.
  • Pooled designs no longer dilute totals through a year that contributed nothing. brfss_design(2022:2023, states = "KY") built a design of 2022 rows alone, Kentucky having collected no 2023 data, yet divided every weight by the two requested years, so survey_total() reported half the state’s adult population. The participation check now treats a requested year with no in-scope rows as an empty year rather than skipping it. Means and proportions were never affected, because the rescaling cancels.
  • A cached year whose checksum fails verification can no longer be served silently for the rest of the session. The daily recheck marked every due file as checked before the repair was attempted, so a failed re-download (offline, or a 404) left the memo in place and the next call in the same session returned the unverified file with no message, while brfss_cache_info(verify = TRUE) reported it as failing. Only files that pass are marked.
  • read_brfss() checks that the years it read are the years asked for. A parquet file carrying the wrong year under the right filename, which a botched hand-copy of a cache produces, was served as a plausible tibble of the wrong survey year under download = FALSE.
  • na = TRUE warns, rather than merely noting, when the catalog covers none of the variables loaded for a year. The grade was keyed on whether the year had any catalog entries at all, so a 1998 request for six uncatalogued variables recoded nothing and said so only in a message, which message = FALSE in a report suppresses entirely.
  • Every TRUE/FALSE argument is now validated at entry with a classed error naming the argument and the value it received. Half of them reached base R’s && and failed with “invalid ‘y’ type in ‘x && y’”, which named nothing and carried no condition class, and verify and catalogs accepted anything at all: brfss_cache_info( verify = "yes") hashed nothing while reading, to the person who typed it, as a request to hash everything.
  • brfss_design() sets options(survey.lonely.psu = "adjust") only when the design it built actually contains a single-PSU stratum. It previously wrote the option on every call, including for years that have none (1995 and 2003 have no such stratum; 2023 has 101 of 2,146), so an unrelated survey analysis later in the session inherited the adjustment in place of survey’s fail-fast default. A pinned brfssdata.lonely_psu is still honored unconditionally.
  • A design variable missing from a damaged or foreign cached file now raises brfssdata_bad_design_var, the class the conditions page documents for it, and points at the cache remedy instead of at variable search. A weight the caller named is not treated as such: _LLCPWT2 exists in 2014 and 2016 but not 2015, and asking for it in 2015 reports a missing variable rather than accusing an intact file of corruption and advising its deletion.
  • _AGEG_ reads as CDC’s aggregate age groups in 1999 and 2000. CDC’s assignment files for those years point the column at AGEGFMT, whose six codes stop at 65+ and read 7 and 9 as unknown, while the column actually holds the twelve codes of _AGEGFMT, which both libraries define and which CDC assigns correctly from 2001. Codes 7 and 9 are the aggregates 18-34 and 55+, and 0, 8, 10, and 11 had no labels at all. Under the wrong format na = TRUE deleted the age group of 2,956 respondents whose age is known, every one of the 1,512 at code 7 in 2000 being aged 18 to 34.
  • A cached file holding another year’s data is caught even when that other year was also requested. Comparing the combined result against the whole request could not see the likeliest hand-copy error of all, one file copied over another’s name, because every row’s year was inside the request and the year simply counted twice; each file is now judged against its own name.
  • A failed catalog refresh no longer marks the catalog checked. The daily recheck memo was corrected on the data path in this release, and the value-label and crosswalk catalogs had the identical defect, where one failed refresh served a catalog known not to match the manifest for the rest of the session.
  • A cache directory that cannot be written to keeps its brfssdata_download_error class. The message naming the directory was assembled as a template and rendered in a frame that had no such directory to name, so the error died inside its own construction with a bare coercion failure carrying no class, and with it every fallback that subscribes to that class, including the bundled-snapshot degradation.
  • The stale-download sweep no longer matches files a user keeps. A dated snapshot such as manifest-2024.json is hex by accident, and the pattern that recognized this package’s own staged manifests recognized those too and deleted them silently.
  • na = TRUE says which named variables the catalog does not cover, rather than reporting a ratio. A 1999 read of PHYSHLTH beside GENHLTH sat at exactly one covered variable of two, just inside a threshold that required fewer than half, and said nothing at all while PHYSHLTH kept its 77s. A year that never asked a loaded question is no longer named either: an all-NA column carries no codes to leave behind.
  • download = FALSE no longer withholds the prefetch command for a year that is merely missing from this machine’s manifest. An air-gapped cache copied before the newest release is the documented workflow, and it made the package assert that a published year does not exist. Only a year in the future is refused outright.

Discovery and metadata

  • A brfss_vars() search that matches nothing says so (brfssdata_empty_result) and suggests near misses: variables whose name or label is a small edit away, variables matching every word of a multi-word pattern in any order, and, under a years filter, matches that exist only in other years. Unknown-variable errors from read_brfss() and brfss_codebook() likewise name the years that do carry the variable and offer a did-you-mean for close names.
  • brfss_crosswalk() reports CDC’s trailing-digit rename families (_DRNKWK1 to _DRNKWK3). A status column records how far human review of each family has gone (mechanically proposed families ship as "candidate"), and comparable/note record the reviewed verdict per generation pair. When a requested variable is empty in years a sibling generation covers, read_brfss() says so (brfssdata_rename_note); combining generations stays the analyst’s decision.
  • read_brfss() and brfss_design() gain states = (FIPS, postal abbreviation, or name), pushed into the DuckDB query so other states’ rows never reach R. For states this pre-filtering is variance-exact, because BRFSS strata nest within state. A requested state absent from a year warns (brfssdata_state_coverage_warning).
  • brfss_codebook() renders a per-variable card: label history, value labels with missing-type codes flagged, year availability, and the rename family. A card whose variable has no coded values says which of the two reasons applies, either that the requested years sit outside the catalog’s span or that CDC’s format for the variable is continuous or range-only (_BMI5), and a card built from a range format (PHYSHLTH, cataloged only as 77, 88, and 99) says that ordinary in-range values are valid and are not listed, so a truncated code list cannot read as the whole value set. Cards carry codes and labels only, with no units, scale factor, or valid range; those live in CDC’s codebook, named per year by brfss_year_info()$codebook_url. A cache-integrity failure while reading the rename crosswalk aborts, as it does everywhere else in the package, rather than quietly returning a card with no concept family. brfss_year_info() lists respondents, variables, states, hosted size, and CDC’s documentation page per year. brfss_citation() returns per-year bibentry citations, each with a stable BibTeX key (brfss2023 and the like, and brfssdata for the package), so toBibtex() output drops into a .bib file unedited.
  • brfss_labels() and brfss_crosswalk() say so when only some requested variables match (brfssdata_partial_match_note), naming the ones with no entries, instead of silently returning rows for the rest. A miss can be legitimate (continuous variables have no label entries; most variables belong to no rename family), which is exactly why it deserves a note rather than silence: without one, a legitimate absence and a typo look identical. brfss_missing_codes() inherits the note.
  • brfss_citation(integer(0)) errors with brfssdata_bad_years_arg instead of returning a malformed, data-free citation list with no signal that nothing was cited.
  • New package data: brfss_states (FIPS, names, abbreviations, Census regions for all 56 BRFSS jurisdictions) and brfss_std_pop_2000 (the 2000 projected U.S. standard population, all-ages and _AGE_G- matched adult groupings, for survey::svystandardize()).
  • The variable, label, and crosswalk catalogs ship as bundled snapshots, so metadata functions work on first use with no network; a snapshot is never served silently (brfssdata_bundled_fallback_note).
  • A brfss_vars() search that matches nothing keeps its single-word hint even when it has close matches to offer, and a search token now also reaches a variable through its name stem. Searching CDC’s own questionnaire wording, “personal doctor”, previously matched nothing and suggested five caregiving variables, none of them the personal-doctor question. The help now says that searches run over CDC’s 40-character SAS labels, whose wording differs from the questionnaire and is sometimes cut mid-word.
  • brfss_labels() returns codes in ascending order within a variable and year, the order a codebook reads in and the order brfss_codebook() already used.
  • Codebook cards say that a column may also be blank, from a question that was not asked or an interview that ended early, so the first mean() returning NA has an explanation on the card.

Access

  • read_brfss() returns respondent-level BRFSS microdata for any of the 40 published survey years (1985-2024) as a tibble. Each year is downloaded once from the package’s data releases, verified against a published sha256 checksum, and cached under tools::R_user_dir(); queries run locally through DuckDB, so selecting a handful of variables from a 300-plus column survey stays fast and repeat use works offline.
  • read_brfss() with no vars selection says it is loading every column before anything is read (brfssdata_full_load_note): a bare read_brfss(2024) materializes 302 columns and about 1.1 GB where a one-column projection is about 5 MB. The note names vars = c(...) as the remedy; it previously fired only from brfss_design(), which now inherits it from read_brfss().
  • brfss_design() builds a srvyr survey-design object with the era-correct final weight (_FINALWT through 2010, _LLCPWT from 2011), strata, and primary sampling units. Requests that pool years across the 2011 redesign fail unless allow_break = TRUE is set, because CDC states estimates are not comparable across that boundary. Files through 2000 carry genuine multi-respondent PSUs and keep the clustered variance estimator; from 2001 on each respondent is their own PSU, and the design drops the nominal cluster term for identical estimates at a fraction of the cost. A weight argument selects another final weight, such as the child weight _CLLCPWT, when CDC’s documentation calls for it. Module analyses that require CDC’s questionnaire-version datasets and their _LCPWTV1 to _LCPWTV3 final weights are not supported by the hosted annual files.
  • By default, brfss_design() sets the codes CDC uses for don’t-know, refused, and missing-type answers to NA (na = TRUE), so means and proportions cover substantive answers; read_brfss() defaults to na = FALSE and returns the file as published. The exported brfss_missing_codes() lists exactly which codes are affected.
  • brfss_design(weight = ) accepts CDC’s final analysis weights only: _FINALWT (1985-2010) and _LLCPWT (2011 on) for the full sample, the domain weights _CLLCPWT, _CHILDWT, and _HOUSEWT, and the 2007 questionnaire-version weights _FINALQ1, _FINALQ2, _CHILDQ1, and _CHILDQ2. Anything else, an intermediate pipeline stage or an arbitrary column, is a classed error (brfssdata_unrecognized_weight) unless unsafe_weight = TRUE says it is deliberate, and the override still warns. Previously only five modern intermediates drew a warning and everything else passed silently: the review built designs with weight = "GENHLTH" (female share shifted 1.8 points) and the 1985-2000 design weight _WT1 (fair/poor health shifted 0.95 points) without a signal. Weight values must now be positive and finite, a final weight requested outside its published span fails before anything downloads, and an explicitly named full-sample weight obeys the same completeness rule as the automatic path: missing values abort as a damaged file instead of silently subsetting the design. All weight columns, final and intermediate, are excluded from labeling and na recoding on every path.
  • brfss_design() detects the reverse weight mistake too: a requested variable whose answers sit almost entirely inside a module weight’s records (95% or more in every requested year) while the design uses a full-sample weight draws a brfssdata_module_weight_warning naming the weight to consider. The live case: 2023 child asthma (CASTHDX2, 99.7% confined to _CLLCPWT’s records) under the default _LLCPWT estimated 10.86% against the correct 10.36%, with no signal. It warns rather than fails because state-optional modules that CDC assigns to the core weight produce the same confinement shape; options(brfssdata.module_weight_check = FALSE) disables it.
  • quiet = TRUE governs progress and housekeeping output only: download progress, cache notes, the full-load hint, and the recode tally. Signals about what the data mean fire regardless of quiet: the rename note, the missing-code coverage signals, and the weight-domain subset note. Previously all three were suppressed by quiet = TRUE, which every article passed, so a brfss_design(weight = "_CLLCPWT", quiet = TRUE) call could silently estimate a different population. Requesting na = TRUE for years before 1998, where no value-label catalog exists, is now a classed warning (brfssdata_na_coverage_warning) rather than a message: nothing was cleared there, and a 1993 PHYSHLTH mean with the 77/99 codes left in is materially wrong. Silence any of these by class, e.g. suppressMessages(..., classes = "brfssdata_rename_note").
  • brfss_labels() exposes CDC’s value-label catalog (1998-2024), and labels = TRUE converts variables with safe one-to-one maps to factors; ambiguous maps keep their numeric codes. labels = "both" keeps the code in the level text so it survives conversion.
  • brfss_vars() searches variable names and labels across years; brfss_years() lists the published years. Variable names match case-insensitively everywhere, and returned columns keep CDC’s canonical spelling.
  • brfss_download() prefetches years and the metadata catalogs for offline use; brfss_cache_dir(), brfss_cache_info(), and brfss_cache_clear() manage the cache. A cached file that fails integrity checks is re-downloaded automatically or named in a classed error with its remedy.
  • Downloads fail instead of hanging on a dead network: the transfer aborts if a connection takes more than a minute to establish or an established transfer sits below 100 bytes/s for five minutes, since libcurl sets no ceiling of its own. When curl reports a classed transport error, the brfssdata_download_error message names the likely cause, distinguishing an unreachable GitHub (offline or blocked network) from a proxy or TLS-interception failure, a stalled connection, and a rejected request, each with its remedy. Failures the transport cannot classify keep the generic wording.
  • Cached survey years are re-verified against the manifest’s checksums at most once a day per session, the same cadence the metadata catalogs already use. A cached file whose hash no longer matches is treated like a damaged one: announced (brfssdata_cache_note) and re-downloaded verified, with the old file kept until a verified replacement lands. Previously only file size was compared after the initial download, so a same-size corrupted copy was accepted without a condition. On that same path, a fully cached read_brfss() with download = TRUE now lets the manifest refresh on its daily cadence, so a user who cached a year before a corrected republish is told within a day; when the refresh fails (offline), the cached manifest is used with a brfssdata_manifest_note and the read proceeds as before. download = FALSE checks nothing, downloads nothing, and never deletes anything.
  • brfss_cache_info(verify = TRUE) adds a verified column: TRUE when a cached file’s sha256 matches the manifest entry, FALSE on a mismatch, NA where the manifest has no entry to compare against. Off by default because hashing reads every byte, roughly two seconds for the full 40-year, 737 MB cache, and brfss_download() lists the cache on every call.
  • A cached manifest that is present but unreadable does not count as fresh. The manifest is the one asset fetched without an expected hash, and any non-empty payload is accepted, so a captive-portal or proxy error page served with HTTP 200 could otherwise sit in the cache looking new. For a day after that, brfss_years() would return integer(0) and every checksum lookup would return NULL, so downloads would proceed unverified. Freshness now reads the content, the bundled copy wins over an unusable cache with a brfssdata_manifest_note naming brfss_years(refresh = TRUE) as the repair, and a downloaded manifest reaches the cache only once its payload parses.
  • A request whose states filter leaves no rows aborts with a classed brfssdata_no_eligible_rows error naming the states, years, and weight, instead of failing inside the survey package with “group length is 0 but data length > 0”. Kentucky and Pennsylvania in 2023 are the live cases. read_brfss() still returns the zero-row tibble, which is a usable answer where a zero-row design is not.
  • Every error, warning, and message carries a documented condition class; see ?brfssdata-conditions.

Performance and ergonomics

  • Metadata reads are served from a session memo. The label catalog was previously read from parquet, through its own DuckDB connection, on every call that used it, twice per brfss_design() with defaults; the manifest JSON was re-parsed up to three times per read and about fifteen times per brfss_download(catalogs = TRUE). Each is now parsed once per file state (the memo invalidates when the on-disk file changes), and the missing-code matcher runs only on the requested years’ catalog rows. Results are byte-identical; repeat calls in a session skip the parquet and JSON work entirely.
  • A vars typo no longer costs a download: when a requested year is not yet cached, read_brfss() and brfss_design() first check the requested variables against the variable catalog (cached or bundled, never fetched for this purpose) and abort with the same brfssdata_bad_var error the query would raise, before any 20-30 MB year file transfers. Years the catalog does not cover skip the check, so a stale catalog can never block a valid read.
  • The “Did you mean years = ...?” hint for a year-shaped first argument, previously only in brfss_labels(), now also fires in brfss_crosswalk(), brfss_codebook(), and brfss_missing_codes().
  • print() on a brfss_codebook caps at 10 cards and says how many more there are; print(x, n = Inf) renders everything. A codebook of the full catalog previously printed thousands of cards.
  • brfss_cache_clear() called with no years argument in an interactive session now asks before deleting every cached year. Scripts, tests, and rendered documents are never prompted, and an explicit years = NULL keeps the old clear-everything behavior without a question.
  • Every session option the package reads is documented in one place, ?brfssdata-options, including the previously undocumented brfssdata.repo.
  • A cache directory that cannot be created or written to is reported as the permission problem it is. The failure previously arrived as a download error suggesting the user might be offline, and a failed directory creation was followed by a note saying where downloads would be cached, which is the one population the package documents restricted-network workarounds for.
  • brfssdata.cache_dir is validated where it is read, so a mistyped path in .Rprofile fails immediately instead of reporting an empty cache. Partial downloads left behind by a killed session are swept, while files the package did not write are still never touched. brfss_cache_info() lists regular files only.
  • quiet = TRUE suppresses the first-run cache-directory note, which reached the console through the manifest refresh whatever the caller asked for. A re-download caused by a failed integrity check now reports itself even under quiet = TRUE, being an integrity event rather than progress narration.
  • The duckdb (>= 1.5.5) requirement is checked when a connection is opened, not only at install time. duckdb loads lazily and is not in the package’s imports for version purposes, so a downgrade or a shadowing site library silently removed the guard that keeps duckdb from writing to the user’s home directory.
  • brfss_years() takes download and quiet, the vocabulary the other exports use, so a strictly offline call is expressible.
  • srvyr’s survey_mean(), survey_prop(), survey_total(), unweighted(), and as_survey_design() are re-exported, so the README’s three-line estimate runs with library(brfssdata) and dplyr alone. Forgetting library(srvyr) previously let the design build and group_by() succeed, then failed at the last verb with a bare “could not find function” and no mention of srvyr.

Documentation

  • A Value labels and missing codes article walks the labels and na workflow end to end: what converts and what deliberately does not, the split defaults, auditing with brfss_missing_codes(), the 88-means-“None” recode, why %in% misreads NA-cleared data, and silencing signals by condition class.
  • The Age-adjusted prevalence article passes the outcome to survey::svystandardize(), as excluding.missing = ~ fair_poor + age_group. svystandardize() filters on excluding.missing before it calibrates, so naming only the age variable leaves the age weights calibrated over respondents that na.rm = TRUE discards afterwards, returning the age-group means weighted by standard share times weighted response rate instead of by standard share alone. The two agree only when item nonresponse is flat across age, which it is not. The recipe also labels the age factor from brfss_std_pop_2000 and asserts the level order, because svystandardize() matches population to the levels of by by position without checking names. brfss_std_pop_2000 now documents that ordering contract.
  • The articles no longer call a whole-file estimate national. The 2023 public-use file covers 52 reporting areas, 48 states plus the District of Columbia, Guam, Puerto Rico, and the U.S. Virgin Islands, with Kentucky and Pennsylvania absent.
  • The age-adjustment article states which standard reproduces CDC’s published tables. CDC age-adjusts most BRFSS questions with three age groups, 18 to 44, 45 to 64, and 65 and over, and the article previously presented its six-group result as the number CDC’s tables report, which it is not: 2023 Florida came out 0.26 points below CDC’s published figure. The three-group recipe, obtainable by summing the shipped standard-population rows, matches to rounding, and six groups are shown as the finer alternative. ?brfss_std_pop_2000 no longer claims the six-group set is what CDC uses.
  • The Stata section no longer promises that value labels carry CDC’s codes through haven::write_dta(). Factor export renumbers values to level positions, so code 7 arrives as 6 and a “Never” answer coded 8 arrives as 5, and a code-based recode in Stata silently misses the answers it names. The article gives routes that keep the codes.
  • The Korn-Graubard interval in the validation article returns a number. It was demonstrated without na.rm = TRUE on a design that recodes missing values, so the published page showed NA NA NA for the method the section recommends.
  • Corrected doc claims: a modern survey year runs to 45 MB, not 35; DuckDB accepts unquoted underscore-prefixed identifiers, so quoting them is portability advice and not a syntax requirement; the data-meta tag hosts nine assets, including the crosswalk and year inventory a mirror needs; srvyr’s filter() on a design deletes rows and estimates the domain through each stratum’s recorded PSU count, rather than keeping respondents at zero weight; labels = TRUE alone signals nothing on pre-1998 years and the recode tally is governed by quiet. The articles also note how a degenerate subgroup behaves, which variables are stored as text, and how to keep don’t-know and refused apart, the distinction R’s single NA cannot hold.
  • ?brfss_design and the vignette no longer say flatly that primary sampling units are applied. From 2001 on they are not, deliberately, and a design built for those years prints ids: 1, which reads as a dropped design feature to anyone auditing against a Stata svyset line. The era rule was documented only under the Choosing a weight heading, where nobody looking for it would find it; it now has its own section, Why some years have no PSU term, which states the numbers: on 2023 the estimate, the standard error, and the degrees of freedom are identical to the last bit of a double with and without the cluster term, while on 1995, where PSUs are shared, the two specifications genuinely differ.

Known limitations

  • Upstream provenance is not recorded yet: brfss_year_info() carries no CDC source URL, retrieval date, or hash of the upstream SAS Transport file, so the package cannot say which CDC revision of a year the hosted copy was built from. The published checksums cover the hosted copies. Provenance columns are planned for 0.1.1, which requires a data republish.
  • Calculated variables keep CDC’s stored scaling. _BMI5 and _DRNKWK2 carry two implied decimals (a stored 2704 in _BMI5 is a BMI of 27.04), and neither the files nor the catalogs record scale factors, units, or valid ranges. Codebook cards say so and point to CDC’s per-year codebook (brfss_year_info()$codebook_url). Curated scale metadata is 0.2.0 work.
  • No value-label catalog exists before 1998, so na = TRUE has nothing to consult there and warns instead of guessing; special codes in 1985 to 1997 must be recoded by hand from CDC’s codebooks. This is a deliberate policy of signaling over hand-curation.
  • Coverage inside 1998 to 2001 is uneven for the same reason. CDC’s 1999 library defines no format for the variables its assignment file points at, PHYSHLTH and MENTHLTH among them, so those columns have no catalog entries that year and na = TRUE cannot clear their 77s. A read that touches them says how many of the loaded variables the catalog covered.
  • Module analyses that require CDC’s questionnaire-version datasets and their _LCPWTV1 to _LCPWTV3 final weights are unsupported; the hosted annual files do not include those weights. The module-weight confinement check covers the reachable cases; curated module-to-weight metadata is 0.2.0 work.
  • Data releases are not immutable snapshots. A corrected year replaces the published bytes under the same tag; the manifest checksum and the daily cache recheck notice the change, but an analysis cannot yet pin itself to an exact prior snapshot. Versioned snapshots are 0.2.0 work.