Skip to contents

Harmonizing geography with FIPS

Source: vignettes/geography-fips.Rmd
geography-fips.Rmd

County identifiers are not stable across data years. New counties appear, a few merge or are renamed, and – more disruptively – NCHS does not encode state the same way in every era. Public multiple-cause-of-death (MCOD) files use NCHS numeric state codes through 2002 and 2-letter postal abbreviations from 2003 on, and the numeric schemes overlap while meaning different states (NCHS "06" is Colorado; FIPS "06" is California). add_county_fips() translates whichever scheme a file uses into one 5-digit county_fips that is comparable across years.

Public-data caveat. Public-use MCOD carries state and county geography only through 2004. From 2005 on, any sub-national analysis requires the restricted NCHS All-County files. Every example below therefore uses data years 1999-2004, and nothing here works for later years without the restricted data.

All data here are small synthetic frames plus bundled narcan crosswalks (st_fips_map, ihme_fips); no chunk downloads anything and no restricted NCHS records are used.

state_abbrev_to_fips() – one column at a time

The simplest helper maps a postal abbreviation to its zero-padded 2-digit state FIPS. Pair the input with the output to see the mapping:

abbrevs <- c("CA", "NY", "TX", "PA")
data.frame(abbrev = abbrevs, state_fips = state_abbrev_to_fips(abbrevs))
#>   abbrev state_fips
#> 1     CA         06
#> 2     NY         36
#> 3     TX         48
#> 4     PA         42

"CA" becomes "06", not "6" – the zero-padding is what lets the result feed a FIPS-keyed join. Unrecognized abbreviations and the foreign/unknown code "ZZ" return NA (with a warning) rather than passing through unconverted.

add_county_fips() – the core call

Feed add_county_fips() the raw county column as character (countyrs for county of residence, countyoc for occurrence). A numeric column is refused, because numeric input has already dropped the leading zero (01001 reads as 1001, whose first two digits parse as a different state). Pass year – a scalar or a year column – so the coding scheme is chosen deterministically.

Both halves of the code change, not just the state

The state prefix is the obvious half. The county half is the one that bites.

Before 2003 the last three digits of countyrs are the county’s alphabetical rank within its state, not its FIPS code. NCHS 05019 is Los Angeles; FIPS 06019 is Fresno. Translating the state prefix and keeping the rank produces a code that looks like a county and is not the right one, for about 97% of pre-2003 records, with no error and no NA.

The raw record settles it. From data year 1982 the MCOD file carries a genuine FIPS geography block next to the NCHS one – fipsctyr for residence and fipsctyo for occurrence, each a complete 5-digit state + county – and import_mcod_fwf() reads both. When you pass countyrs for a pre-2003 year and fipsctyr is present, add_county_fips() takes the geography from it.

Start with a small synthetic frame that straddles the 2002/2003 boundary, shaped the way import_mcod_fwf() returns one. The 1999-2002 rows carry NCHS numeric state codes plus the record’s own FIPS county; the 2003-2004 rows carry postal abbreviations and a county half that is already FIPS, so they need no companion column:

deaths <- tibble::tibble(
    id       = 1:5,
    year     = c(1999L, 2001L, 2002L, 2003L, 2004L),
    countyrs = c("05019", "33031", "44057", "CA075", "TX201"),
    fipsctyr = c("06037", "36061", "48113", NA, NA)
)
deaths
#> # A tibble: 5 × 4
#>      id  year countyrs fipsctyr
#>   <int> <int> <chr>    <chr>   
#> 1     1  1999 05019    06037   
#> 2     2  2001 33031    36061   
#> 3     3  2002 44057    48113   
#> 4     4  2003 CA075    NA      
#> 5     5  2004 TX201    NA

Now harmonize. The call adds four columns – state_substr and county_substr (the two halves actually decoded), st_fips (2-digit state) and county_fips (5-digit state + county) – decoding each row under its own era’s scheme. It also reports, once per session, that it took the pre-2003 geography from fipsctyr; that notice appears in the output below:

harmonized <- add_county_fips(deaths, countyrs)
#> add_county_fips(): sourced pre-2003 county geography from `fipsctyr` for 3 row(s). The `countyrs` county digits are an alphabetical rank, not FIPS. This message is shown once per session; wrap the call in suppressMessages() to silence it.
#> This message is displayed once per session.
harmonized[c("id", "year", "countyrs", "st_fips", "county_fips")]
#> # A tibble: 5 × 5
#>      id  year countyrs st_fips county_fips
#>   <int> <int> <chr>    <chr>   <chr>      
#> 1     1  1999 05019    06      06037      
#> 2     2  2001 33031    36      36061      
#> 3     3  2002 44057    48      48113      
#> 4     4  2003 CA075    06      06075      
#> 5     5  2004 TX201    48      48201

Read the before/after together. Row 1’s raw countyrs is "05019": NCHS state 05 is California, and 019 is the nineteenth California county alphabetically – Los Angeles. The answer is "06037", taken from fipsctyr. Had the rank been carried through, the result would have been "06019", which is a real county (Fresno) and the wrong one. In 2003 California enters instead through the abbreviation "CA", giving "06075", and no companion column is needed because the county half is already FIPS. Codes that cannot be placed – "ZZ" foreign residence, FIPS state "00", a county masked to "999" on the public file, or a territory code such as NCHS "62", which narcan does not cover – yield NA in st_fips/county_fips rather than a spurious string. The state_substr and county_substr columns keep the literal code, so you can still count how much of a public extract is masked.

Note also that county_substr reads "037", not the "019" in the input: on a row sourced this way, both halves come from fipsctyr. The same is true of the other two pre-2003 rows – in every one of them the county digits change, which is the point. A rank and a FIPS county code agree only when a county’s alphabetical position happens to equal its FIPS number, and since FIPS county codes step by two (001, 003, 005 …) while ranks step by one, that happens only at rank 001 – the first county alphabetically in each state. Roughly 97% of pre-2003 records got a wrong county from the old code.

If you are working with a frame that has no fipsctyr/fipsctyo – a hand-built extract, say – add_county_fips() reports the opposite thing, once per session: that the county half is a rank and the result is not usable for county-level work. use_fips_cols = "no" restores the old behavior but does not buy you silence; it swaps the first notice for the second, because the returned county is then the one the notice is warning about. Use suppressMessages() if you want a quiet call. Data years 1979-1981 are the one span with no remedy at all – no FIPS county was recorded, so the true county cannot be recovered from the file and only st_fips is usable. Those years report it too – but both notices share one once-per-session identifier, so if an earlier call in the same session already raised the rank notice, a later 1979-1981 frame will be silent. Do not rely on the message to tell you which years you are looking at; check the data years.

Neither notice depends on scheme. That argument selects how the state half is decoded; use_fips_cols selects where the county half comes from. Pinning scheme = "nchs" on 1999 data still sources the county from fipsctyr.

The bundled crosswalks

Both helpers read st_fips_map, which lines up state name, postal abbreviation, FIPS code, and NCHS code so the two numeric schemes can be told apart:

head(st_fips_map)
#> # A tibble: 6 × 4
#>   name       abbrev  fips  nchs
#>   <chr>      <chr>  <dbl> <dbl>
#> 1 Alabama    AL         1     1
#> 2 Alaska     AK         2     2
#> 3 Arizona    AZ         4     3
#> 4 Arkansas   AR         5     4
#> 5 California CA         6     5
#> 6 Colorado   CO         8     6

ihme_fips handles the other half of the problem – counties that split, merge, or are renamed over a long series. It collapses unstable orig_fips codes to a temporally stable ihme_fips (from Dwyer-Lindgren et al., JAMA 2016), so a county can be compared across periods:

head(ihme_fips)
#>    state group orig_fips ihme_fips
#> 1 Alaska     1     02158     02158
#> 2 Alaska     1     02270     02158
#> 3 Alaska     2     02140     02188
#> 4 Alaska     2     02188     02188
#> 5 Alaska     3     02010     02013
#> 6 Alaska     3     02013     02013

For a multi-year county analysis, remap county_fips through ihme_fips before comparing – several Alaska areas and Virginia independent cities are otherwise not comparable across the 1999-2004 window.

From FIPS to denominators

Once county_fips or a state_fips column exists, it is the join key for sub-national denominators. add_county_fips() names its state column st_fips; add_pop_counts() keys on state_fips, so rename st_fips to state_fips for a state join, or use the 5-digit county_fips directly for a county join (see the companion vignette("population-denominators")). As above, this path is only valid for 1999-2004 public data – 2005+ sub-national denominators require the restricted NCHS All-County files.

See also