Fits a weighted GLM (quasibinomial) to the trend in ever-smoking at age 25-34. This provides the "target" level for the Holford adjustment.
Usage
ever_smoke(
data,
time_horizon = 2100,
num_bins = 7,
model = "model2",
min_age = 15,
min_year = 2003,
age_cats = c("25-34"),
auto_holdout_bins = 2,
auto_tie_margin = 2,
auto_floor = 0.02,
auto_ceiling = 0.98,
auto_max_slope_mult = 2
)Arguments
- data
Data table of individual characteristics.
- time_horizon
Integer - the last year for projection.
- num_bins
Integer - bins for the period trend to reduce noise.
- model
Character - Model specification (interaction terms), or "auto" to select the structure as described above.
- min_age
Integer - youngest age for prediction.
- min_year
Integer - first year of survey data.
- age_cats
Character vector - age category for reference (e.g., "25-34").
- auto_holdout_bins
Integer - how many of the most recent year bins to hold out when scoring candidates (model = "auto" only).
- auto_tie_margin
Numeric - QAIC margin within which a simpler model is preferred to the best-scoring one (model = "auto" only). 2 is the conventional "no real difference" threshold.
- auto_floor, auto_ceiling
Numeric - the projected proportion for every stratum must stay inside this range over the whole projection.
- auto_max_slope_mult
Numeric - cap on any stratum's logit slope as a multiple of the main-effects model's common slope.
Details
The candidate models differ only in their interactions. The ones involving year_bin are the ones to be careful with: year_bin enters linearly on the logit scale, and the predictions run from before the data starts out to time_horizon, which is a twenty-odd year extrapolation from about sixteen years of observation. A sex or IMD difference in *level* is safe under that extrapolation; a difference in *slope* compounds with every projected year. So the risk in a richer model is not the extra parameters, it is what those parameters do a long way outside the data.
Setting model = "auto" picks the structure from the data, conservatively. Three stages:
1. Score every candidate on held-out time. The last auto_holdout_bins of the binned years are set aside, each model is fitted to the earlier bins and scored by weighted deviance on the held-out people. That scores the thing the projection actually does - predict forward along the fitted trend - rather than in-sample fit, which interactions can always improve.
2. Prefer simplicity. Candidates are compared on a holdout QAIC: the held-out deviance divided by the dispersion (estimated once, from the richest model on the training years, so every candidate is scaled the same), plus two per parameter. Models within auto_tie_margin of the best are treated as ties and the one with the fewest coefficients wins - the usual reading of differences under 2 on that scale. A relative tolerance does not work here: the deviance is dominated by irreducible person-level noise, so even a real slope difference only moves it by a fraction of a percent, and any percentage margin hands the choice to the plain model every time regardless of the data.
3. Guard the projection. The winner is refitted to all the data and its projection to time_horizon is checked: every stratum must stay inside [auto_floor, auto_ceiling], and no stratum's logit slope may exceed auto_max_slope_mult times the common slope from the main-effects model. A model that fails falls back to the next admissible candidate, and it says so. If nothing survives, the main-effects model is used; if even that fails the range check, that is a data problem and the function stops rather than projecting it.
Passing an explicit "model1" to "model8" behaves exactly as it always has.