Skip to contents

Rolling-origin validation for ordered repeated measures

Source: R/stats_rolling.R
fit_rolling.Rd

Instead of one hold-out block per person, the origin walks forward: train on each person's first initial rows and predict the next assess, then move the origin on by step and repeat. Every fold trains only on the past, so the result is a forecast, not a fit.

Usage

fit_rolling(
  data,
  y,
  x,
  id,
  method = c("lm", "glm", "ml"),
  time = NULL,
  scope = "both",
  subgroup = NULL,
  initial = NULL,
  assess = 1L,
  step = NULL,
  folds = 5L,
  min_train = 10L,
  tune = FALSE,
  valid_prop = 0.2,
  ...
)

Arguments

data

Data frame.

y

Outcome column name.

x

Predictors: names, numeric positions, a:b range, formula, or data frame.

id

Person/unit ID column.

method

Which family to fit each fold with: "lm", "glm", or "ml".

time

Optional ordering column.

scope

"both" (pooled + individual), "pooled", "individual", "subgroup", or "all" (pooled + subgroup + individual).

subgroup

Optional subgroup mapping: an find_subgroups() result, a grouping column in data, or a named vector of labels per person.

initial

Rows each person trains on in the first fold. Defaults to whatever leaves room for folds folds.

assess

Rows predicted per fold.

step

How far the origin moves between folds. Defaults to assess (contiguous, non-overlapping test blocks).

folds

Number of folds.

min_train

Minimum complete training rows per person.

tune

Logical. Tune each fold on a validation block carved from that fold's training rows, so no fold's test rows influence its own settings.

valid_prop

Proportion of each fold's training rows used for tuning.

...

Passed to the underlying fitter, e.g. model or family.

Value

An idiographic_fit whose predictions carry a fold column.

Details

Predictions gain a fold column. Metrics pool across folds, so metrics() answers "how well does this model forecast this person over time" rather than "how well did it do on one arbitrary split".

Examples

fit <- fit_rolling(srl, y = "effort", x = "efficacy:monitoring", id = "name",
                   time = "day", method = "ml", model = "ridge", folds = 3)
metrics(fit, overall = TRUE)
#>        scope model estimator  subject subgroup   n     rmse      mae     bias
#> 1     pooled ridge    native .overall     .all 108 20.18199 15.97641 1.847201
#> 2 individual ridge    native .overall     .all 108 18.40502 13.93209 2.781897
#>   r_squared
#> 1 0.4599228
#> 2 0.5508409