SpendProbe.assert_increment_is_complete

SpendProbe.assert_increment_is_complete(*, effects, non_date_dims)

Check the evaluated nodes account for the whole move in the predictor.

Everything downstream rests on one identity:

\[\Delta \mu_t = \sum_c \Delta v_{t,c} + \sum_j \Delta e_{t,j}\]

– spend moves the linear predictor through channel_contribution and the resolved effects, and through nothing else. Up to here that is an assumption, and three separate mistakes break it silently: an effect that reports the wrong contribution variable, an effect that cannot be attributed at all, and a model-level node that reads channel_data outside any effect. Each drops a real part of the increment and reports the remainder as if it were the whole.

So it is checked rather than assumed, against the probe evaluation measure() already paid for. A structural check – does spend reach \(\mu\) other than through the accounted nodes – would miss the misreporting case, because a variable upstream of the true contribution blocks the same paths while entering \(\mu\) through a nonlinearity that makes the sum above false.

Parameters:

effectssequence of ChannelDependentEffect

The effects being evaluated alongside channel_contribution.

non_date_dimsmapping

Per node, the dimensions its evaluation carries after sample and date. The terms are added by name: an effect may legitimately drop a dimension the predictor has, and a panel model’s predictor need not order the ones it keeps the way spend does.

Raises:

NotImplementedError

If the accounted nodes do not reproduce the predictor’s move.