Atomic Bomb Survivor Mortality Data (Life Span Study Report 14)

A left-truncation example derived from the Radiation Effects Research Foundation's (RERF) Life Span Study (LSS) Report 14 cancer and noncancer disease mortality data, 1950-2003 (Ozasa et al. 2012). One row per subject, for all 86,611 subjects in the cohort.

Format

A data frame with 86,611 observations (one per subject) on 9 variables:

city: Factor: "Hiroshima" or "Nagasaki".
sex: Factor: "Male" or "Female".
gd3: Factor: ground distance, "<3km" or "3-10km".
agex: Numeric age at exposure in 1945, in years. Exactly entry - 5.15 for every subject (5.15 years separate the bombings from the start of follow-up), so agex and entry are collinear and must not both enter a model as covariates.
entry: Numeric left-truncation (delayed entry) time: attained age on 1950-10-01, in years.
time: Numeric event/censoring time: attained age at death, or at the 2003-12-31 end of follow-up for survivors, in years. Strictly greater than entry for every subject.
status: Integer event indicator: 1 if the subject died of any cause during follow-up, 0 if censored (50,620 deaths).
status_colon: Integer event indicator for the cause-specific colon cancer hazard: 1 if the subject died of colon cancer, 0 otherwise, with deaths from other causes treated as censored at their age at death (621 colon-cancer deaths).
dose: Numeric DS02-weighted colon dose (gamma + 10*neutron), in Gy. Note the unit: the RERF source column is in mGy.

Source

Radiation Effects Research Foundation (RERF) Life Span Study Report 14: Ozasa K, Shimizu Y, Suyama A, Kasagi F, Soda M, Grant EJ, Sakata R, Sugiyama H, Kodama K. Studies of the mortality of atomic bomb survivors, Report 14, 1950-2003: An overview of cancer and noncancer diseases. Radiat Res 2012 [March]; 177(3):229-43.

This report makes use of data obtained from the Radiation Effects Research Foundation (RERF), Hiroshima and Nagasaki, Japan. RERF is a private, non-profit foundation funded by the Japanese Ministry of Health, Labour and Welfare (MHLW) and the U.S. Department of Energy (DOE), the latter in part through DOE Award DE-HS0000031 to the National Academy of Sciences. The conclusions in this report are those of the authors and do not necessarily reflect the scientific judgment of RERF or its funding agencies.

Details

Left truncation. LSS follow-up begins on 1950-10-01, five years after the 1945 bombings, so a subject enters the risk set only at the attained age they had reached on that date: anyone exposed in 1945 who died before follow-up began is not in the cohort at all. Analyses on the age scale must therefore account for delayed entry via entry (or, in coxph, Surv(entry, time, status)). Entry ages range from 5.2 to 93.8 years, so the truncation is substantial and ignoring it biases the estimates - see the examples.

Reconstruction from grouped data. The RERF source file is a grouped person-time table stratified by city, sex, ground distance, Adult Health Study participation, age at exposure, attained age, calendar time and colon dose; it does not distribute individual subject records. Individual-level records are reconstructed from it, which is possible because the table's subject counts are recorded once per subject (in each subject's first calendar-time cell) and so sum to the cohort size. Each death in a cell becomes one record, with the attained age at death drawn uniformly within that cell's five-year attained-age band; the remaining subjects in each stratum are censored administratively at their attained age on 2003-12-31. Within-cell variation in age at exposure and in dose is collapsed to the cell mean, and the pairing of a particular death age to a particular subject within a stratum is arbitrary. The reconstruction reproduces the source table's 86,611 subjects, 50,620 deaths and 621 colon-cancer deaths exactly, and its 3,294,282 person-years to within 0.33%. See dev/data-raw/hiroshima.R to reproduce and dev/experiments/hiroshima_dataset/ for the original grouped table and its documentation.

Not a reanalysis of LSS Report 14. Ozasa et al. do not ungroup the table at all: they fit Poisson excess-relative-risk models (\(\lambda_0(\mathrm{strata})(1 + \beta d)\)) directly to the person-years, with the baseline stratified on 1200 cells. For analysing that table their approach is better; this reconstruction exists only because a proportional-hazards fit needs one row per subject.

How far the coefficients can be trusted has been checked against the grouped-data Poisson fit, which is the correct analysis of the source table (see dev/experiments/hiroshima_reconstruction_checks.R and dev/experiments/hiroshima_dataset/UNGROUPING.md). Conclusions about left truncation are robust: the difference between the truncated and naive fits is positive for every coefficient under all 12 reconstruction variants tried (RNG seed, within-band placement rule, dose level, entry offset, end of follow-up), and stable to four decimal places across seeds, because it depends only on the entry-age distribution, which the source table reports directly. Dose-response estimates are also recovered: fitted on this dataset with a matched stratification, the all-cause ERR per Gy is 0.145 for males and 0.276 for females, inside Ozasa et al.'s published intervals of 0.150 (0.104-0.199) and 0.298 (0.244-0.354). What the dataset does not support is quoting a coefficient to three decimals as an LSS14 estimate: the same 12 variants move the absolute coefficients by 1.8 to 4.1 standard errors, uncertainty that no reported standard error captures.

Examples

data(hiroshima)
if (FALSE) { # \dontrun{
# Delayed entry is handled by entry=, which currently requires the
# piecewise-constant baseline hazard, basis = "uniform".
fit <- coxph_mpl(
  Surv(time, status) ~ dose + sex + city,
  data   = hiroshima,
  entry  = entry,
  basis  = "uniform",
  smooth = 0
)
summary(fit)

# Cross-check: coxph()'s partial likelihood handles (start, stop] exactly.
coef(coxph(Surv(entry, time, status) ~ dose + sex + city, data = hiroshima))

# Ignoring the truncation biases the dose and city coefficients upwards
# (here by 1.3 and 2.1 standard errors respectively), because entry age is
# associated with both. The sex coefficient is essentially unaffected, entry
# age being unrelated to sex.
coef(coxph_mpl(
  Surv(time, status) ~ dose + sex + city,
  data = hiroshima, basis = "uniform", smooth = 0
), "Beta")
} # }