Skip to contents

A bare intercept-and-slope least squares fit of y on transformed media, supplied so that tune_carryover() has a sensible default and the simple case stays one line. It exists to be replaced: pass your own fit_fn and predict_fn to tune carryover against the model you intend to fit.

Usage

fit_ols(x, y)

# S3 method for class 'mm_ols'
predict(object, newdata = NULL, ...)

# S3 method for class 'mm_ols'
coef(object, ...)

# S3 method for class 'mm_ols'
fitted(object, ...)

# S3 method for class 'mm_ols'
print(x, ...)

Arguments

x

For fit_ols(), a numeric vector of transformed (adstocked) media. For print(), an mm_ols object.

y

Numeric vector of the response, the same length as x.

object

An mm_ols object.

newdata

Numeric vector of transformed media to predict from. When omitted, predict() returns the fitted values.

...

Ignored, present for generic consistency.

Value

fit_ols() returns an object of class mm_ols: a list with elements intercept, slope, n and fitted.values, the fitted values of the n finite pairs used. predict() returns a numeric vector the same length as newdata. coef() returns the intercept and slope, named (Intercept) and media, so a fit on a column named media works with contributions(). fitted() returns fitted.values.

Details

This is deliberately the simplest possible model. It has no trend, no seasonality, no price term and no other controls, so carryover parameters chosen against it absorb whatever those omitted terms would have explained. That is fine for a first pass and wrong for a deliverable.

To tune against a model with controls, pass them to tune_carryover()'s controls argument, which switches the default model to least squares on the adstocked media plus the controls and keeps every row aligned with every split. For several channels at once use tune_carryover_joint(), and for carryover tuned jointly with saturation and a model penalty, step_adstock() in a recipes pipeline. The examples of tune_carryover() show the first; vignette("carryover") and vignette("tidymodels") work through the other two.

Pairs where x or y is missing or non-finite are dropped before fitting. With fewer than two pairs left the slope is zero and the intercept is the one remaining y, or NA when none remains. Otherwise the slope is set to zero when x has no spread relative to its own scale, so the answer does not depend on the units x is measured in.

See also

tune_carryover(), which uses fit_ols() and rmse() as its defaults.

Examples

set.seed(1)
spend <- adstock_geometric(c(100, 50, 0, 0, 200, 100, 0, 50), decay = 0.5)
kpi <- 10 + 0.4 * spend + rnorm(8, sd = 0.1)

m <- fit_ols(spend, kpi)
m
#> <mm_ols> fitted on 8 observations
#> response = 10.06 + 0.3991 * media
coef(m)
#> (Intercept)       media 
#>  10.0634323   0.3991045 
predict(m, spend)
#> [1] 30.01866 30.01866 20.04104 15.05224 52.46828 51.22108 30.64226 30.33046