Demographic recodes across coding eras

NCHS changed how it codes age, sex, and race several times over the decades. narcan’s recoders normalize each raw field to one stable scheme, dispatching on the data year. The recurring hazard is that codes which look continuous across a revision are not, and chaining them produces artifactual trends. This vignette walks the three families – age, sex, and race – and shows where each one breaks.

Every input below is a small illustrative vector, not an NCHS record, and every chunk runs.

library(narcan)

Age: two recodes, two outputs

narcan carries two age recoders because MCOD files carry two different age fields.

Detail age (age) to completed years. remap_age() reads the unit-coded detail-age field and returns completed years in a new age_years column. The encoding changed at 2003 (a 3-digit code before, 4-digit after), so pass year. Sub-year ages (months, weeks, days, hours, minutes) collapse to 0; not-stated codes become NA.

remap_age(data.frame(year = 2000, age = c(37, 206, 999)))$age_years    # pre-2003
#> [1] 37  0 NA
remap_age(data.frame(year = 2019, age = c(1037, 2006, 1999, 9999)))$age_years  # 2003+
#> [1] 37  0 NA NA

Pre-2003, a years-unit code is the age itself (37 -> 37) and 206 is 6 months -> 0. From 2003 a leading 1 marks years, so 1037 -> 37 and 2006 (months) -> 0; both 1999 and 9999 are not-stated -> NA.

AGER27 recode (ager27) to 5-year bins. convert_ager27() maps the NCHS 27-level age recode (codes 1-26 are age groups, 27 is “not stated” -> NA) to 5-year-bin starts (0, 5, …, 85). The finer under-5 categories collapse into the 0-4 bin, and everything 85 and older into 85.

convert_ager27(data.frame(ager27 = c(1, 7, 10, 23, 27)))
#>   age
#> 1   0
#> 2   5
#> 3  20
#> 4  85
#> 5  NA

Label the bins. categorize_age_5() turns 5-year-bin starts into an ordered factor with readable labels – the form tidyr::complete() and ggplot2 expect.

categorize_age_5(seq(0, 85, 5))
#>  [1] 0-4   5-9   10-14 15-19 20-24 25-29 30-34 35-39 40-44 45-49 50-54 55-59
#> [13] 60-64 65-69 70-74 75-79 80-84 85+  
#> 18 Levels: 0-4 < 5-9 < 10-14 < 15-19 < 20-24 < 25-29 < 30-34 < ... < 85+

The two age outputs are not interchangeable. age_years is single completed years; the AGER27 path – and the standard populations add_std_pop() uses – are 5-year bins. To bin age_years yourself:

age_years <- c(0, 3, 37, 88)
pmin(floor(age_years / 5) * 5, 85)     # -> 0, 0, 35, 85
#> [1]  0  0 35 85

Sex: numeric before 2003, character after

categorize_sex() maps either coding – numeric 1/2 (1979-2002) or character "M"/"F" (2003+) – to "male"/"female", the labels add_pop_counts() joins on. Pass year; it is authoritative.

categorize_sex(c(1, 2, 9), year = 2000)       # pre-2003 numeric
#> [1] "male"   "female" NA
categorize_sex(c("M", "F", "U"), year = 2019) # 2003+ character
#> [1] "male"   "female" NA
categorize_female(c(1, 2), year = 2000)       # -> 0 1 (male = 0, female = 1)
#> [1] 0 1

The trap: omit year and the era is guessed from the column type. A pre-2003 sex column re-read from a CSV as the characters "1"/"2" is guessed to be the modern "M"/"F" scheme and maps to all-NA – loudly, with warnings.

categorize_sex(c("1", "2"), year = NULL)     # guesses modern -> NA (with warnings)
#> Warning: categorize_sex(): `year` not supplied; inferring the coding era from
#> the column type (2003+ 'M'/'F'). Pass `year` to be explicit.
#> Warning: categorize_sex(): every value mapped to NA -- this usually means the
#> `year`/era does not match the column's coding (numeric 1/2 vs character M/F).
#> [1] NA NA
categorize_sex(c("1", "2"), year = 2000)     # year fixes it -> male, female
#> [1] "male"   "female"

Race: the most-revised field

Race coding changed repeatedly, and remap_race() dispatches on year to a standardized code that categorize_race() then labels. These eras are genuinely different schemes, not one evolving list.

1992-2020 (bridged race). Detailed codes 1-7, plus a family of “other” codes that all collapse to 99.

r <- remap_race(data.frame(year = 2004, race = c(1, 2, 3, 6, 18)), year = 2004)
r$race                          # 1 2 3 6 99
#> [1]  1  2  3  6 99
categorize_race(r$race)         # white, black, american_indian, hawaiian, other
#> [1] white           black           american_indian hawaiian       
#> [5] other          
#> 9 Levels: total < white < black < american_indian < chinese < ... < other

The Asian and Pacific Islander subgroups (Chinese, Japanese, Hawaiian, Filipino) are resolved separately in this era but have no matching bridged population denominator, so collapse them to a single API group before computing rates.

Earlier eras differ. The same raw code means different things before 1992 – and there are two distinct pre-1992 schemes (1979-1988 and 1989-1991), not one. Code 8, for instance, is Filipino here (1979-1988, mapped to 7) but “other” (99) in 1989-1991. This is exactly why remap_race() refuses to run without a year.

remap_race(data.frame(year = 1985, race = c(0, 1, 7, 8)), year = 1985)$race  # 99 1 99 7
#> [1] 99  1 99  7

2021 is a gap. The bridged race fields were dropped and the single-race recodes were not yet populated, so remap_race() sets race to NA (with a warning).

remap_race(data.frame(year = 2021, race = c(1, 2)), year = 2021)$race
#> Warning in remap_race(data.frame(year = 2021, race = c(1, 2)), year = 2021):
#> race is retired in 2021 (bridged race dropped; single-race recodes not
#> populated until 2022); setting race to NA.
#> [1] NA NA

2022+ (single race). From 2022 the bridged column is gone; remap_race() reads the single-race Race Recode (racer5) and maps it into a non-colliding code space (101-106) so the two schemes cannot be silently mixed. categorize_race() labels these with an _only suffix (except code 106, multiracial).

r22 <- remap_race(data.frame(year = 2022, racer5 = c(1, 2, 4, 6)), year = 2022)
#> Warning in remap_race(data.frame(year = 2022, racer5 = c(1, 2, 4, 6)), year =
#> 2022): 2022+ race uses the single-race Race Recode 6 mapped to codes 101-106;
#> these are NOT comparable to the bridged race scheme (2020 and earlier).
r22$race                                # 101 102 104 106
#> [1] 101 102 104 106
categorize_race(c(101, 102, 104, 106))  # white_only, black_only, asian_only, multiracial
#> Warning in categorize_race(c(101, 102, 104, 106)): single-race codes (101-106)
#> are not comparable to the bridged race scheme (2020 and earlier); do not
#> combine the two into a single trend.
#> [1] white_only  black_only  asian_only  multiracial
#> 6 Levels: white_only < black_only < american_indian_only < ... < multiracial

Both remap_race() and categorize_race() warn on the single-race codes, because bridged (2020 and earlier) and single-race (2022+) categories are not comparable. Do not chain them into one trend.

Why the year matters

Two coding boundaries hide in these recoders, and neither is the ICD boundary:

Age and sex change their coding at 2003 (the death-certificate revision).
Race changes repeatedly, with the sharpest break at 2021/2022 (bridged to single race).

Neither lines up with the ICD-9-to-ICD-10 boundary at 1999 that the flag functions dispatch on. A long multi-year analysis therefore carries several independent comparability breaks at once, so each recoder reads the data year to apply the right map, and not-stated codes always become NA rather than a guessed value.

These recoders produce exactly the age, sex, and race columns that add_pop_counts() joins on, with labels that match the population data.

Age: two recodes, two outputs

Sex: numeric before 2003, character after

Race: the most-revised field

Why the year matters

See also