Cox regression lag sensitivity analysis

Runs Cox proportional hazards models at one or more lag periods to assess whether associations are robust to the exclusion of early events. For each lag, participants whose follow-up time is less than lag_years are removed from the analysis dataset; follow-up time is kept on its original scale (not shifted). This mirrors the approach used in UK Biobank sensitivity analyses to address reverse causation and detection bias.

Usage

assoc_lag(
  data,
  outcome_col,
  time_col,
  exposure_col,
  lag_years = c(1, 2),
  covariates = NULL,
  base = TRUE,
  strata = NULL,
  conf_level = 0.95
)

Arguments

data: (data.frame or data.table) Analysis dataset.
outcome_col: (character) Event indicator column (0/1 or logical).
time_col: (character) Follow-up time column (numeric, e.g. years).
exposure_col: (character) One or more exposure variable names.
lag_years: (numeric) One or more lag periods in the same units as time_col. Default: c(1, 2). Use 0 to include the unfiltered full-cohort result as a reference.
covariates: (character or NULL) Covariates for the Fully adjusted model. When NULL, only Unadjusted (and Age and sex adjusted if base = TRUE) are run.
base: (logical) Auto-detect age and sex and include an Age and sex adjusted model. Default: TRUE.
strata: (character or NULL) Optional stratification variable passed to survival::strata().
conf_level: (numeric) Confidence level for HR intervals. Default: 0.95.

Value

A data.table with one row per lag \(\times\) exposure \(\times\) term \(\times\) model combination, containing all columns produced by assoc_coxph plus:

lag_years: The lag period applied (numeric).
n_excluded: Number of participants excluded because their follow-up time was less than lag_years.

lag_years and n_excluded are placed immediately after model in the column order.

Details

Setting lag_years = 0 (or including 0 in the vector) runs the model on the full unfiltered cohort, providing a reference against which lagged results can be compared.

The same three adjustment models produced by assoc_coxph are available here (Unadjusted, Age and sex adjusted, Fully adjusted).

Examples

dt <- ops_toy(scenario = "association", n = 500)
#> ✔ ops_toy: 500 participants | 33 columns | scenario = "association" | seed = 42
dt <- dt[dm_timing != 1L]

res <- assoc_lag(
  data         = dt,
  outcome_col  = "dm_status",
  time_col     = "dm_followup_years",
  exposure_col = "p20116_i0",
  lag_years    = c(0, 1, 2),
  covariates   = c("bmi_cat", "tdi_cat"),
  base         = FALSE
)
#> ℹ outcome_col dm_status: logical detected, converting TRUE/FALSE -> 1/0
#> 
#> ── assoc_lag ───────────────────────────────────────────────────────────────────
#> ℹ 3 lag periods x 1 exposure x 1 model
#> 
#> ── Lag: 0 years ──
#> 
#> ℹ Excluded (time < 0 yr): 0 -- remaining: 463, events: 24
#> 
#> ── p20116_i0 ──
#> 
#> ── Lag: 1 year ──
#> 
#> ℹ Excluded (time < 1 yr): 3 -- remaining: 460, events: 21
#> 
#> ── p20116_i0 ──
#> 
#> ── Lag: 2 years ──
#> 
#> ℹ Excluded (time < 2 yr): 7 -- remaining: 456, events: 17
#> 
#> ── p20116_i0 ──
#> 
#> ✔ Done: 6 result rows across 3 lag periods, 1 exposure, and 1 model.
res[, .(lag_years, n, n_excluded, HR, CI_lower, CI_upper, p_value)]
#>    lag_years     n n_excluded        HR   CI_lower CI_upper   p_value
#>        <num> <int>      <int>     <num>      <num>    <num>     <num>
#> 1:         0   452          0 0.3927405 0.13251113 1.164016 0.0917984
#> 2:         0   452          0 0.4094191 0.09463282 1.771310 0.2321082
#> 3:         1   449          3 0.4610566 0.15228917 1.395853 0.1707241
#> 4:         1   449          3 0.4684395 0.10656446 2.059181 0.3154581
#> 5:         2   445          7 0.6225764 0.19727224 1.964804 0.4189975
#> 6:         2   445          7 0.6408615 0.14119296 2.908810 0.5642742