Put a non-decision-time bound on a family

Records bound on fam, sets the ndt link that makes the bound structural, and, for a per-group bound, arranges for the per-row floor to reach the family's likelihood as data. It does NOT touch the likelihood: what a consumer's density has to do is three lines, written below, and doing it there rather than through a wrapper is what keeps this seam free of every assumption about where that density lives.

Usage

ndt_bound_attach(fam, bound)

Arguments

fam: A frmtmb_family with a distributional parameter named ndt.
bound: The result of ndt_bound().

Value

fam, modified.

What the consumer's density must do

Call ndt_apply() wherever it would have read dpars[["ndt"]], in the density and in the mean and in the simulator alike. A family that attaches a bound here and then does NOT call it fits, converges and reports a non-decision time measured at 37.5 percent too small, with nothing refusing; ?ndt_apply has that measurement. Nothing on this side can check it. That is the whole of it, and it is one line rather than three, because the three lines of arithmetic come with a refusal that is longer than they are and that is the part that matters. An earlier version of this seam documented the arithmetic and left the refusal in prose; a consumer who copied the first and not the second got a fraction silently read as seconds, which is the defect the per-group bound exists to remove.

What it changes on the family

links$ndt, ndt_bound, and for a per-group bound aterm_data and the ndt starting value, which becomes the fraction one half. A scalar bound leaves the family's own starting value alone. It also records the family's pre-attach aterm_data and ndt starting value in a reserved slot named ndt_seam, which is what makes a second call replace the first bound instead of composing with it; a family with a slot of that name of its own would lose it. Nothing else moves, and st is not handled: a family with an across-trial range on the non-decision time sets that bound itself.

Calling it twice replaces the first bound rather than stacking on it, which matters because frmtmb::frm() runs family_finalize again on the influence(), frm_simulate() and prior-predictive paths.

Examples

# what a family_finalize() written against this seam looks like
finalize <- function(fam, y, aterms) {
  if (!is.null(ndt_bound_of(fam))) return(fam)
  ndt_bound_attach(fam, ndt_bound(y, aterms, what = "my_family"))
}