Skip to contents

Sweep Candidate Knots of a Piecewise-Linear Mixed Model

Source: R/piecewise_lmm.R
tune_breakpoint_lmm.Rd

Fits `y ~ t + pmax(t - k, 0) + (1 + t | id)` by maximum likelihood for each candidate change point `k` and reports AIC, BIC and the in-sample RMSE (observed minus the conditional fit). A knot whose model fails to fit gets `NA` metrics and a `status`.

Usage

tune_breakpoint_lmm(
  data,
  response,
  time_col = "time",
  id_col = "arl_number",
  knots = NULL,
  metric = c("RMSE", "AIC", "BIC")
)

Arguments

data

A data frame in long format.

response

Name of the numeric response.

time_col, id_col

Names of the time and subject columns.

knots

Candidate change points. `NULL` uses the observed times strictly between the smallest and the largest.

metric

Criterion that picks the selected knot (lowest wins): `"RMSE"`, `"AIC"` or `"BIC"`.

Value

A tibble with `knot`, `AIC`, `BIC`, `RMSE`, `singular`, `status` and `selected` (`TRUE` for the best knot by `metric`; none if every fit failed).

Details

Needs `lme4`. Rows with a missing response, time or id are not used. Note that the RMSE favours flexible fits; prefer AIC or BIC when the knot should generalise.

See also

[impute_piecewise_lmm()]

Examples

if (FALSE) { # \dontrun{
d <- lme4::sleepstudy
tune_breakpoint_lmm(d, "Reaction", "Days", "Subject", knots = 2:7, metric = "AIC")
} # }