Fits a multinomial logistic regression "response surface" to estimate Current/Former/Never status. The model includes high-order polynomials for Age and Year, and interactions with Sex and IMD.
Two changes were made when the smoking trends were brought into the bootstrap:
1. The prediction grid is now pinned by the caller rather than read off the data. Under resampling the observed min/max age and year, and the set of IMD quintiles present, can all shift between iterations. Left unpinned that silently produces replicates with different numbers of rows, which then cannot be stacked and cannot form a covariance matrix.
2. Predicting outside the observed range of age or year now has to be asked for. See the note on `allow_extrapolation` below.
Usage
trend_fit(
data,
max_iterations = 1000,
age_var = "age",
year_var = "year",
sex_var = "sex",
smoker_state_var = "smk.state",
imd_var = "imd_quintile",
weight_var = "wt_int",
grid_ages = NULL,
grid_years = NULL,
grid_sex = c("Male", "Female"),
grid_imd = NULL,
expected_states = c("current", "former", "never"),
allow_extrapolation = FALSE,
boot_id = NULL,
tol = 1e-08
)Arguments
- data
Data table of survey data.
- max_iterations
Integer, passed to nnet::multinom.
- age_var, year_var, sex_var, smoker_state_var, imd_var, weight_var
Column names.
- grid_ages
Integer vector of ages to predict for. Defaults to the range observed in `data`, which is only safe outside bootstrap mode.
- grid_years
Integer vector of years to predict for. Defaults as above.
- grid_sex
Character vector of sex levels.
- grid_imd
Character vector of IMD quintile levels.
- expected_states
The smoking states the model must return a column for.
- allow_extrapolation
Logical. If FALSE (the default) it is an error for `grid_ages` or `grid_years` to reach beyond the range present in `data`.
- boot_id
Optional scalar written to a `boot_id` column on the output. Also suppresses the progress messages.
- tol
Tolerance for the "probabilities sum to one" check.