Method comparison

Three ways to measure incrementality — holdout, on/off, ITS

How do I measure the incremental effect of advertising?

A randomized holdout is the most accurate. Otherwise use pre/post or ITS.

Incrementality is the outcome advertising actually created. There are three main designs. A randomized holdout, a pre/post comparison around turning a campaign on or off, and interrupted time series (ITS). The accuracy ranking is clear, and so is the difficulty ranking.

Side by side

Incrementality measurement methods
MethodDesignCausal claimRequirementsCaution
Control holdoutSplit exposed and unexposed groups at random, at the same timeSupportedPlatform holdout feature or geo splitSmall samples give intervals too wide to conclude
Turn on / turn offCompare before and after a campaign starts or stopsLimitedA clear intervention date and enough pre/post dataSeasonality and promotions are mixed in
ITS (interrupted time series)Extend the pre-intervention trend to estimate the counterfactualLimitedMany pre-period observations and a stable trendIgnoring autocorrelation understates the interval

Which one to use

If your platform supports holdouts
Use a holdout. Randomization is what allows a causal claim.
If the campaign cannot be paused, such as brand
Use ITS to extend the pre-period trend, with autocorrelation correction.
If the budget is too small for a control group
Start with pre/post, but add a control series and read it as difference-in-differences.
If the result is non-significant
Do not conclude no effect. Check whether power was sufficient first.

Frequently asked questions

Can pre/post alone prove incrementality?

It is not recommended. Seasonality, promotions, and competitive shifts land in the same window. Adding a control series and reading it as difference-in-differences removes much of that bias.

If a holdout is non-significant, does advertising not work?

No. Small samples or small effects go undetected even when a real effect exists. Read the width of the confidence interval and check whether the test had enough power.

How many pre-period points does ITS need?

There is no fixed minimum, but you need enough to separate trend from seasonality. A short pre-period makes the counterfactual unstable and widens the interval.

Tools for this

← All method comparisons