Skip to contents

Forecast probabilities of smoking initiation, quitting and relapse

Source: R/quit_forecast.R
quit_forecast.Rd

Forecasts trends in transition probabilities using a Lee-Carter style Singular Value Decomposition (SVD) model.

Usage

quit_forecast(
  data,
  forecast_var,
  forecast_type = c("continuing", "stationary"),
  cont_limit = NULL,
  oldest_year = 2003,
  youngest_age = 11,
  oldest_age = 88,
  age_cont_limit = 88,
  first_year = 2010,
  jump_off_year = 2015,
  time_horizon = 2050,
  smooth_rate_dim = c(3, 3),
  k_smooth_age = 3,
  preserve_zeros = FALSE
)

Arguments

data

Data table with input probabilities.

forecast_var

Character - variable to forecast.

forecast_type

"continuing" (linear trend) or "stationary" (constant).

cont_limit

Integer - year where forecast becomes stationary.

oldest_year

Integer - start of historical data.

youngest_age

Integer - min age.

oldest_age

Integer - max age.

first_year

Integer - start year for trend fitting.

jump_off_year

Integer - end year of historical data.

time_horizon

Integer - end year of forecast.

smooth_rate_dim

Vector - dimensions for raster smoothing (c(3,3)).

k_smooth_age

Integer - knots for smoothing age component.

preserve_zeros

Logical - if TRUE, cells that are exactly zero in the input are kept out of the raster smoothing and put back at the floor value afterwards, instead of being clamped to 1e-6 and averaged in with their neighbours. This exists for initiation. Since the cumulative-curve fix in p_dense, a zero in the initiation surface is a real zero - nobody in that cohort starts at that age - not survey noise. Clamping it and letting the focal mean run over it drags mass down from the ages just below, which incorrectly increases the estimated values. Quitting and relapse keep the default FALSE: their zeros genuinely are sparse-cell noise and smoothing over them is the right treatment.

Details

The model assumes the logit of the probability can be decomposed into: Logit(P_xt) = Alpha_x + Beta_x * Kappa_t Where: - Alpha_x: Average age profile - Kappa_t: Time trend index - Beta_x: Sensitivity of each age to the time trend

Note that the output for the historical years is the reconstruction from this decomposition, not the input estimates: everything this function returns, past and future, has been through the smoothing and the rank-1 fit.