This functionality is currently in limited testing.
The Experimentation panel lets analysts compare different user experience, marketing, or messaging variations to determine which is best at driving a specific outcome. You can evaluate the lift and confidence of any A/B experiment from any experimentation platform - online, offline, from Adobe solutions, Adobe Journey Optimizer, and even BYO (bring-your-own) data.
At this point, Adobe Analytics for Target (A4T) data brought into Adobe Experience Platform via the Analytics Source Connector cannot be analyzed in the Experimentation panel. We expect a resolution to this issue in 2023.
The Experimentation panel is available to use by all Customer Journey Analytics (CJA) users. No Admin rights or other permissions are required. However, the setup (steps 1 and 2 below) requires actions that only Admins can perform.
The recommended data schema is for the experiment data to be in an Object array that contains the experiment and variant data in two separate dimensions. If you have your experiment data in a single dimension with experiment and variant data in a delimited string, you can use the substring setting in data views to split them into two for use in the panel.
In CJA data views settings, admins can add context labels to a dimension or metric and CJA services like Experimentation panel can use these labels for their purposes. Two pre-defined labels are used for the Experimentation panel:
In your data view that contains experimentation data, pick two dimension, one with the experimentation data and one with the variant data. Then label those dimensions with the Experiment and the Variant labels.
Without these labels present, the Experiment panel does not work, since there are no experiments to work with.
If the necessary setup in CJA data views has not been completed, you will receive this message before you can proceed: “Please configure the experiment and variant dimensions in Data Views”.
Configure the panel input settings.
|Experiment||A set of variations on an experience that were exposed to end users in order to determine which is best to keep in perpetuity. An experiment is made up of two or more variants, one of which is considered the control variant. This setting is pre-populated with the dimensions that have been labeled with the Experiment label in data views, and the last 3 months’ worth of experiment data.|
|Control Variant||One of two or more alterations in an end user’s experience that are being compared for the purpose of identifying the better alternative. One variant must be selected as the control, and only one variant can be considered to be the control variant. This setting is pre-populated with the dimensions that have been labeled with the Variant label in data views. This setting pulls up the variant data that is associated with this experiment.|
|Success Metrics||The metric or metrics that a user is comparing variants with. The variant with the most desirable outcome for the conversion metric (whether highest or lowest) is declared the “best performing variant” of an experiment. You can add up to 5 metrics.|
|Normalizing Metric||The basis (People, Sessions, or Events) on which a test will be run. For example, a test may compare the conversion rates of several variations where Conversion rate is calculated as Conversions per session or Conversions per person.|
|Date Range||The date range is automatically set, based on the first hit received in CJA for the experiment selected. You can restrict or expand the date range to a more specific timeframe if needed.|
The Experimentation panel returns a rich set of data and visualizations to help you better understand how your experiments are performing. At the top of the panel, a summary line is provided to remind you of the panel settings you selected. At any time, you can edit the panel by clicking the edit pencil at the top right.
You also get a text summary that indicates whether the experiment is conclusive or not, and summarizes the outcome. Conclusiveness is based on statistical significance. (See “Statistical methodology” below.) You can see summary numbers for the best performing variant with the highest lift and confidence.
For each success metric you selected, one freeform table and one conversion rate trend will be shown.
The Line chart gives you the Control versus Control Variant performance:
This panel currently does not support analysis of A/A tests.
Experiment is Conclusive: Every time you view the experimentation report, Adobe analyzes the data that has accumulated in the experiment up to this point and will declare an experiment to be “Conclusive” when the anytime valid confidence crosses a threshold of 95% for at least one of the variants (with a Bonferonni correction applied when there are more than two arms, to correct for multiple hypothesis testing).
Best Performing Variant: When an experiment is declared to be conclusive, the variant with the highest conversion rate is labeled as the “best performing variant”. Note that this variant must either be the control or baseline variant, or one of the variants that crosses the 95% anytime valid confidence threshold (with Bonferonni corrections applied).
Conversion Rate: The conversion rate that is shown is a ratio of the success metric value, to the normalizing metric value. Note that this may sometimes be larger than 1, if the metric is not binary (1 or 0 for each unit in the experiment)
Lift: The Experiment report summary shows the Lift over Baseline, which is a measure of the percentage improvement in conversion rate of a given variant over the baseline. Defined precisely, it is the difference in performance between a given variant and the baseline, divided by the performance of the baseline, expressed as a percentage.
Confidence: The Anytime Valid Confidence that is shown, is a probabilistic measure of how much evidence there is that a given variant is the same as the control variant. A higher confidence indicates less evidence for the assumption that control and non-control variant have equal performance. More precisely, the confidence that is displayed is a probability (expressed as a percentage) that we would have observed a smaller difference in conversion rates between a given variant and the control, if in reality there is no difference in the true underlying conversion rates. In terms of p-values, the confidence displayed is 1 - p-value.
A full description of results should consider all available evidence (i.e. experiment design, sample sizes, conversion rates, confidence etc.), and not just the declaration of conclusive or not. Even when a result is not yet “conclusive”, there can still be compelling evidence for one variant being different from another (e.g. confidence intervals are nearly non-overlapping). Ideally, decision making should be informed by all statistical evidence, interpreted on a continuous spectrum.
To provide easily interpretable and safe statistical inference, Adobe has adopted a statistical methodology based on Anytime Valid Confidence Sequences.
A Confidence Sequence is a “sequential” analog of a Confidence Interval. To understand what a confidence sequence is, imagine repeating your experiments one hundred times, and calculating an estimate of the mean business metric (e.g. open rate of an email) and its associated 95%-Confidence Sequence for every new user that enters the experiment. A 95% Confidence Sequence will include the “true” value of the business metric in 95 out of the 100 experiments that you ran. (A 95% Confidence Interval could only be calculated once per experiment in order to give the same 95% coverage guarantee; not with every single new user). Confidence Sequences therefore allow you to continuously monitor experiments, without increasing False Positive error rates, i.e. they allow “peeking” at results.