Split-half null distribution of profile turnover — compute_turnover

Estimates, for each profession, the distribution of turnover_netto when there is no real change, by comparing two random halves of the postings of the same window.

Usage

compute_turnover_null(
  incidence,
  prof_col,
  skill_col = "skill_id",
  id_col = "general_id",
  time_col = "mese_idx",
  base,
  target,
  n_rep = 20L,
  prob = 0.95,
  seed = 1L
)

Arguments

incidence: Posting-level data.frame with id_col, time_col, prof_col and skill_col, one row per posting and skill.
prof_col: Profession column.
skill_col: Skill column. Default "skill_id".
id_col: Posting identifier. Default "general_id".
time_col: Period column. Default "mese_idx".
base, target: Numeric vectors c(from, to).
n_rep: Number of random splits per window. Default 20.
prob: Quantile of the null distribution. Default 0.95.
seed: Integer seed. Default 1.

Value

A data.table with the profession column, n_nullo (null values), media_nullo and q_nullo (the prob quantile).

Details

For each replicate and each window (base and target), the postings of the window are split at random into two halves; the two halves play the role of base and target in compute_profile_turnover(). The null sample of a profession collects 2 * n_rep values of turnover_netto; its prob quantile is the threshold above which observed turnover exceeds the variation expected from sampling alone. Each posting must belong to one profession. The global RNG state is restored.

Examples

set.seed(6)
inc <- data.table::data.table(general_id = rep(1:200, each = 2),
  mese_idx = rep(rep(1:2, each = 100), each = 2), cp4 = "p1",
  skill_id = sample(letters[1:5], 400, replace = TRUE))
compute_turnover_null(inc, "cp4", base = c(1, 1), target = c(2, 2),
  n_rep = 5)
#>       cp4 n_nullo media_nullo    q_nullo
#>    <char>   <int>       <num>      <num>
#> 1:     p1      10  0.02599631 0.07890718