Where the accumulator starts, as a distribution over the rescaled state
at time zero. bias is the relative start point in (0, 1), the
fraction of the boundary separation above the lower boundary, exactly as
in wiener(); 0.5 is unbiased.
Details
gddm_start_point()All mass at
bias. Free parameterbias, logit link.gddm_start_uniform()Uniform on an interval of half-width
szaroundbias, withszmeasured as a fraction of the half separation. Free parametersbiasandsz, both logit link. This is the tape-safe counterpart of the Ratcliff start-point variability, which the analytic family can only reach by quadrature.
A point start is a delta function, which no fixed grid can hold. Both
components spread the initial mass with the same cubic B-spline used for
the non-decision-time shift, so the starting distribution is smooth in
bias and no grid index depends on a parameter. Both are then
normalized to unit mass on the grid, which also keeps the family honest
when bias and sz together push mass past a boundary.
Writing your own
A start term is the value of gddm_start_term(). Its fn is called as
fn(y, h, p), where y are the rescaled grid nodes on (-1, 1) and
h is the node spacing, and it returns a density at those nodes.