In one line
If CPA and ROAS are averages over everything you have spent, marginal CPA and marginal ROAS are the values for the next dollar you add — or the last dollar you remove. Scale-up and pull-back decisions belong to these, not the average.
Why the average alone is dangerous
"This campaign's CPA is $8, let's spend more" is a risky call. $8 is the average over everything spent, not the efficiency of the next dollar. Where the spend-conversion curve (the response curve) flattens, average CPA can still look fine while marginal CPA has already gone much worse. A channel with great average ROAS is a loss to feed once its marginal ROAS drops below your target line.
The decision rule
Marginal CPA ÷ average CPA above 1 means the next dollar is worse than the average (a saturation signal); near 1 means there's still headroom to scale. "Where should the next dollar go / where should it come from?" is answered by the marginal metric.
Equimarginal principle
Allocation is optimal when marginal efficiency is equal across channels. Pull from low and push to high, and you naturally converge there.
Caveat — the curve isn't symmetric
Cutting isn't a clean rewind of scaling. Because of learning-phase resets, a minimum viable spend floor, and non-reversibility (accumulated creative fatigue), a removed dollar comes back more expensive than its value on the curve.
Go deeper
How to judge increases, reallocations, and cuts with a single marginal metric is covered in Cut, scale, or reallocate budget with one metric — marginal ROAS.