R/get.par.start.R

In gompertztrunc: Conducting Maximum Likelihood Estimation with Truncated Mortality Data

#' Get starting values for parameters
#'
#' Uses linear modeling to compute initial values for MLE optimizer
#'
#'
#' @param formula the estimation formula
#' @param data data matrix with y, u, l, and covariates, including cohort
#'
#' @return Named vector of initial parameter estimates
#'
#' @export


## get parameter start values
get.par.start <- function(formula, data) {
  ##     da = mydata
  ##     formula = formula(y ~ -1 + treat)
  ## we start with lm with intercept

  newform <- stats::update.formula(formula, . ~ . + 1)
  m <- stats::lm(newform, data)

  ## (Intercept)        treat
  ##      56.004       -1.494
  ## then let m.start be the intercept
  M.start <- stats::coef(m)["(Intercept)"]
  names(M.start) <- ""

  ## let b.start = 1/10
  b.start <- 1 / 10
  a.start <- bM2a(b = b.start, M = M.start)
  b0.start <- log(a.start)

  ## then use entropy-based ball park. Say entropy is 0.1
  ## say 1 year of change is 1%, so 1 year is like 10% decrease
  ## -perc*H*e0 = change = -(-.1) * .1 * 100 = +1 year
  coef.vec <- stats::coef(m)[names(stats::coef(m)) != "(Intercept)"]
  b.vec.start <- -1 * coef.vec * .1

  ## rename
  par.start <- c(
    log.b.start = log(b.start), b0.start = b0.start,
    b.vec.start
  )

  return(par.start)
}