A control group is a set of users who were not exposed to the campaign, used as the counterfactual: what would have happened anyway, without the media. Subtracting the control group's behaviour from the exposed group's is what isolates the effect of the campaign, and it is the difference between reporting a lift and reporting an incremental lift.

Why the Control Is Built Afterwards

In digital and mobile advertising, a holdout group can be defined before a campaign runs - the ad server simply withholds the ad from some users and serves it to others. Out of home does not work that way. A billboard is visible to whoever walks past it, and nobody decides in advance who those people are.

So the control has to be constructed after the fact, from users we observed who were not exposed. The goal is a group that resembles the exposed group as closely as possible in every respect except having seen the media.

📘

This Is Why Exposure Comes First

The control is modelled on the exposed group, so it cannot be built until exposure has been computed. See Sources and Exposure.

What the Control Is Matched On

A control drawn carelessly would differ from the exposed group in ways that have nothing to do with the campaign, and those differences would show up in the results as though they were effects. Mira matches the control to the exposed group across four dimensions.

Geography

Control users are drawn from the areas around the campaign's own inventory, rather than from the country at large. Someone who lives and moves through the same neighbourhoods as the exposed group is a far better comparison than someone in a different market entirely - they shop at the same places, see the same local media, and experience the same local conditions.

Fidelity

Fidelity is how much we observe about a given device - how many distinct location signals it produces. It matters because it determines how much of a user's behaviour we can see at all.

A high-fidelity device is more likely to be seen converting than a low-fidelity one, regardless of whether it was exposed. If the exposed group were high-fidelity and the control low-fidelity, the exposed group would appear to convert more often purely as an artefact of observation. Mira matches the control to the distribution of fidelity in the exposed group so that both groups are observed to a comparable degree.

Demographics

The control is matched to the demographic composition of the exposed group. Age, income, and household characteristics all influence how likely someone is to convert, and a campaign's inventory placement means the exposed group is rarely demographically average.

Observability in the Conversion Environment

For web and app measurement, an exposure is connected to a conversion through a device graph - the dataset linking a mobile advertising ID to the other identifiers belonging to the same person. A device absent from that graph cannot be observed converting on a website, whatever it actually did.

Control candidates are therefore matched on their presence in the graph as well. Without this, the control would contain users who could not convert as far as our data is concerned, which would understate the baseline and inflate the measured lift.

What the Control Is Not

• Not a random sample of the population. A random sample would differ from the exposed group in geography, demographics, and observability all at once.
• Not a group we know to be unexposed. It is a group we did not observe being exposed. For a large campaign in a dense market these are close to the same thing, but they are not identical.
• Not fixed across studies. The control is specific to a campaign's exposed group, so two studies of different campaigns have different controls.

Checking the Control

The pre period is the test of whether the match worked. Before the campaign starts, neither group has been exposed, so both should be converting at a similar rate. When the pre-period rates line up, that is direct evidence the two groups are comparable; when they diverge sharply, the match is worth revisiting before the results are read.

This is one of the strongest arguments for running a pre period, beyond its role in the lift calculation itself: it is the only opportunity to verify the comparison rather than assume it.

📘

The Control Can Show a Lift, And That's Expected

It is entirely normal for the control group's conversion rate to rise from the pre period to the post period. Seasonality, other media, and events unrelated to the campaign all move both groups. That is precisely what the control is there to capture, and why net lift is the difference between the two lifts rather than the exposed group's lift alone. See the FAQ.

Availability

A control group is an option on a Study rather than a default, and it requires no advance configuration because it is built retroactively. A Study configured to use a control may report its exposed-group figures first and gain the control-based metrics once the control has been built and reviewed. See Incremental Lift.