models.glm

Generalised linear model (gaussian, binomial, poisson or Gamma) on a table of any size: iteratively reweighted least squares with one streaming pass per iteration, memory independent of the number of rows. It follows R's glm.fit step for step -- starting values, link functions, convergence test, dispersion -- so it fits the model stats.glm fits (tested to agree to 1e-8, with the same iteration count), and writes the same glm model for models.predict.

Nesting (/, %in%) and cross-validation are not offered; use stats.glm for tables that fit in memory and need them. A model that does not converge within max_iter fails.

Kind: standard
Network: no

Inputs🔗

Name	Type	Description
`data`	file (Parquet or CSV)	Training table: one row per observation, with the target, the predictor columns, and any weights or id columns.
`formula`	string	Model formula, e.g. `ESTIMATE ~ B5 + log(B25) + FORTYPE + 0`. Use instead of `target` and `features`. Default empty.
`target`	string	The response column (when no formula is given). Default empty.
`features`	string	Column selector for the predictors (when no formula is given): names, globs and exclusions, e.g. `B,MB,-B0`. Default empty.
`intercept`	boolean	Include an intercept when building the model from `target` and `features`. With a formula, write `+ 0` in the formula instead. Default `true`.
`distribution`	string	Response distribution: `gaussian` (default), `binomial` (0/1, TRUE/FALSE or a proportion with the number of trials as weights), `poisson` (counts) or `Gamma` (positive values).
`link`	string	Link function: `canonical` (default; the distribution's own), `identity`, `log`, `logit` or `inverse`. Supported: gaussian identity/log/inverse, binomial logit, poisson log/identity, Gamma inverse/log/identity.
`id_columns`	string	Column selector for identifier columns, never used as predictors; kept as text when reading CSV. Default empty.
`weights`	string	Column of non-negative observation weights. Empty fits unweighted.
`categorical`	string	Column selector for predictors to treat as categorical even though they are stored as numbers. Default empty.
`na_action`	string	`fail` (default, naming the columns) or `drop`: what to do with a row that has a missing value in a model column.
`max_iter`	number	Maximum IRLS iterations; a model that has not converged fails. Default `25`.

Outputs🔗

Name	Type	Description
`model`	file (SpadeModel)	The fitted model (family `glm`), for `models.predict`.
`coefficients`	file (CSV)	One row per design column: `term`, `estimate`, `std_error`, `statistic`, `p_value`.
`fit`	json	Rows used and dropped, the arguments that shaped the fit, and `metrics`: `rmse` and `mae` on the response scale, `deviance`, `null_deviance`, `aic`, `bic`, `dispersion`, `df_residual`, `iterations`.

Example🔗

- id: "@counts"
  name: models.glm
  inputs:
    - "@plots"
  args:
    formula: "stem_count ~ elevation + forest_type"
    distribution: "poisson"