Predicted response across a range of adstocked spend, derived from the
fitted saturation parameters. spend_for() is the inverse: the adstocked
spend that achieves a target response.
Usage
response_curve(spend, coefficient, type = "hill", ...)
spend_for(target, coefficient, type = "hill", max_spend = NULL, ...)Arguments
- spend
Numeric vector of adstocked spend levels to evaluate, in the units the saturation curve was fitted on (see Details), finite and non-negative.
response_curve()only;spend_for()has no such argument.- coefficient
The channel's fitted coefficient.
- type
Saturation curve type, as in
saturate(), or"none"for a linear response.- ...
Passed to
saturate(), for examplehalf_maxandshape.- target
Target response level.
spend_for()only. It must lie on the side of the response at zero spend that spending moves towards: above it for a positive coefficient, below it for a negative one.- max_spend
Upper bound of the search interval for
spend_for(). Defaults to 1000 times the curve's spend-scaled parameter (half_max,km, orlog(2)/rate).saturate_power()and"none"have no such parameter, somax_spendis required there.
Value
response_curve() returns a data frame with columns spend (the
levels given), response (the coefficient times the curve at each level)
and marginal (the slope of response there, as mroi() computes it),
one row per element of spend and in the same order. It has class
mm_response_curve, with a plot() method.
spend_for() returns a single adstocked spend level, or NA_real_ with a
warning when the target is not reachable within max_spend. A target
that no amount of spend can reach, such as a negative response from a
channel with a positive coefficient, is an error.
Details
spend is the saturation curve's input. When the regressor came from
media_transform(), that is adstocked spend: the media pressure in a
period, carryover included, rather than the spend of that period alone.
Under a normalised kernel, the default, a steady spend per period settles
at the same adstocked level, so a level can be read as the steady spend
per period that holds the channel there.
Both functions describe the response curve the model fitted, which is not
the same thing as what would happen if you actually spent that much. The
curve is identified only over the range of spend the data contains; asking
a saturated Hill curve what happens at ten times the observed maximum
returns a number, and that number is extrapolation. spend_for() returns
NA rather than a fabricated answer when the target lies beyond
max_spend, but it cannot tell you that a reachable target is outside the
data's support. Check the range of the adstocked series yourself;
plot() shades the levels outside it when given
the series as observed.
See also
plot.mm_response_curve(), saturate(), mroi(), and
marginal_roi() for the marginal return on raw spend through the full
transform
Examples
# The curve mm_weekly's television was simulated with: coefficient 5200,
# half_max 2250 and shape 1.6, on adstocked weekly spend.
data(mm_weekly)
truth <- attr(mm_weekly, "truth")
c(truth$beta[["tv"]], truth$half_max[["tv"]], truth$shape[["tv"]])
#> [1] 5200.0 2250.0 1.6
# A concave curve: marginal return falls from the first pound onward.
concave <- response_curve(
spend = seq(0, 5000, by = 1000),
coefficient = 5200, half_max = 2250, shape = 1
)
concave
#> spend response marginal
#> 1 0 0.000 2.3111041
#> 2 1000 1600.000 1.1076949
#> 3 2000 2447.059 0.6477545
#> 4 3000 2971.429 0.4244933
#> 5 4000 3328.000 0.2995231
#> 6 5000 3586.207 0.2225948
all(diff(concave$marginal) < 0)
#> [1] TRUE
# An S-curve: marginal return rises up to the curve's steepest point, near
# 900 here, and falls after it. Below that point each extra pound returns
# more than the one before, so the curve is not yet saturated there; whether
# the channel deserves more budget depends on how its marginal return
# compares with other channels' and with break-even.
s_curve <- response_curve(
spend = seq(0, 5000, by = 500),
coefficient = 5200, half_max = 2250, shape = 1.6
)
s_curve
#> spend response marginal
#> 1 0 0.0000 0.001128472
#> 2 500 429.9182 1.261993634
#> 3 1000 1115.8577 1.402248359
#> 4 1500 1785.0175 1.250419105
#> 5 2000 2355.7338 1.030821918
#> 6 2500 2818.6324 0.826120871
#> 7 3000 3188.0331 0.657873910
#> 8 3500 3482.5595 0.525816116
#> 9 4000 3718.8362 0.423714294
#> 10 4500 3910.1361 0.344863577
#> 11 5000 4066.6239 0.283637139
s_curve$spend[which.max(s_curve$marginal)]
#> [1] 1000
# What adstocked spend achieves a response of 2000?
spend_for(target = 2000, coefficient = 5200, half_max = 2250, shape = 1.6)
#> [1] 1677.288
# An unbounded curve needs an explicit search range
spend_for(target = 200000, coefficient = 5200, type = "power",
exponent = 0.5, max_spend = 1e6)
#> [1] 1479.29