Holdouts

Eppo's Holdouts allows you to validate the aggregate impact of Experimentation.

This functionality offers the ability to set a holdout audience in feature flags that keeps an audience isolated from all active experiments, and measures metric changes for the holdout audience versus an audience that only experiences winning experiments.

The feature can be used with Eppo's SDK for an end-to-end managed experiment or in an Analysis-Only mode where you use an external randomization library.

Using Eppo SDKs

Begin by configuring your Assignment Source to annotate your holdout columns.

Journey of a Holdout user

Define one or more holdout groups with a relevant key and the desired active date range. You can have multiple holdout groups running at once (e.g one for each team that runs experiments).

The holdout is split into two groups: status quo who always see the control experience, and winning variants who see the treatment from experiments that are rolled out.

Assignment period

1. Create experiments on your own schedule.
2. When subjects are assigned to eligible experiments (more details below) with an active Holdout, Eppo's SDK may assign them to a holdout group. The assignments control and holdout should lead to the same end user experience. For this cohort the SDK will return the control variation.

Eligible experiments

• Any experiment allocation that starts after the start date will have a holdout applied.
• Any experiment allocations that starts after the end date will not have a holdout applied
• Experiments allocation that start during the assignment window but do not end in that time, will have the holdout applied but not included in the analysis. In this case some traffic is wasted.

3. Once an experiment is complete and a winning variation is selected, all non-holdout users (variations and control) will receive it. Users will now either get assigned the winning variant or holdout.
4. Run as many experiments as you want by repeating these steps.

Evaluation period

You may begin an analysis at the conclusion of your desired assignment period by creating a Holdout Experiment.

1. The evaluation is ongoing and newly shipped winning variations will be added to it automatically.
2. For that holdout group, after the evaluation period finishes, end the holdout group.
3. A report can now be generated reflecting data from the evaluation period showing how the holdout group did against users who saw all experiment winners.

SDK Behavior

Application developers should expect no changes to the SDKs.

For the correctness of Holdout Analysis to take place, the experience of subjects assigned to a Holdout should match those in the control of each experiment during the enrollment phase.

When invoking the get*Assignment methods with a subject and flag keys, subjects assigned to a Holdout will have the control variation returned. This makes it foolproof to design your User Experience without worrying about handling a special case.

Changes to assignment logging

The callback function is augmented with a holdout value if the subject was assigned to one. This should be transmitted to your Data Warehouse to facilitate filtering and computation of lifts.

An example in GoLang:

import (
  "gopkg.in/segmentio/analytics-go.v3"
)

...

type ExampleAssignmentLogger struct {}

func (al *ExampleAssignmentLogger) LogAssignment(event eppoclient.AssignmentEvent) {
    client.Enqueue(analytics.Track{
        UserId: event.Subject,
        Event:  "Eppo Randomization Event",
        HoldoutKey: event.Holdout
    })
}

Analysis-Only Holdouts

If your Holdouts are deployed by an external randomization library and you are not using Eppo's feature flags, please refer to this section.

Eppo enables you to bring your own Holdouts and provides a full-featured analysis of that Holdout on key metrics.

Assignment setup

While Eppo is flexible to how you configure Holdouts on your end, we recommend you setup your Holdouts in the following fashion:

• There should be two variants: status_quo and winning_variants
• Traffic should be evenly split between these variants within your withheld traffic group
• Users in status_quo should always see the control experience until the holdout is released
• Users in winning_variants should be exposed to the winning variant of each experiment after it has been concluded and rolled out

Ensure you are logging exposure to the holdout and each variant in an Assignment Source. Eppo requires the following information:

• An experiment key which is the name of the Holdout the user is enrolled in
• The variant name that the user is enrolled in for the applicable Holdout

With this setup, Holdouts will be logged with Experiments in AssignmentSQL that is setup in Eppo.

Configuring a Holdouts Analysis

Create a new Holdout experiment over your desired date range by clicking on the "+Create" button on the Analysis tab.

Configure the analysis with the variation names in your holdout, such as status_quo and winning_variants. This allows the Eppo generated SQL to correctly query the data in your warehouse.

You are able to link experiments to your holdout to provide a comprehensive report of how each experiment impacted the primary Holdout metric and show the experience provided by the winning variants. To do so, go to the Overview tab and add Experiments that are part of the Holdout. Note that these experiments must be configured in Eppo to be linked.

Holdout Analysis

Regardless of whether you use Eppo's SDK or your own, the artifact will be a report detailing impact of the holdout variation over the control, with annotations display the shipped experiments.

Lift on the primary metric will be computed across the winning variants comparing the holdout subjects against everyone else.