Hi,
I have wondered about the rationale behind the direct contribution plot in general and the connection to fitted saturation curves in general.
My understanding is that the scatter plot created by plot_direct_contribution_curves links spend in a given period to the channel’s estimated contribution in that period. The contribution effectively can come from past spend or from spend in the current period. The plot does not account for delayed effects of the spend in the current period (as is stated in the docstring).
I feel the resulting plot is not very informative, maybe even misleading, exemplified by the fact that we often have points with non-zero contribution at zero spend (where the entire impact comes from past time points). In summary, interpreting this visualization naively will overestimate the impact of spend on time points before which there was already some spend and underestimate the impact of spend on time points before which there was no spend (provided there is some adstock effect).
Moreover, plotting the fitted saturation curve over that plot is even more misleading, as it suggests a correspondence between the points and the curve which is not there given non-zero adstock. To my understanding, the meaning of the y-axis is really different for points versus curve. For the former it’s “return in spend period” and for the latter it’s “total return of spend (taking into account delayed effects)”.
Is my understanding correct, and if so, wouldn’t it be a good idea to (1) disentangle the plots and (2) be clear in the documentation about the restricted interpretability of the scatter plots?