The information fraction is the proportion of the total planned statistical information that has been observed so far in a sequential experiment. In the simplest case, it equals the current sample size divided by the maximum planned sample size. If you've planned an experiment for 100,000 users and 50,000 have been observed, your information fraction is 0.5.
Information fractions are the scaffolding that group sequential tests are built on. Every interim analysis in a GST occurs at a pre-specified information fraction, and the alpha spending function uses these fractions to determine how much of the significance budget to allocate at each look. Getting the information fraction wrong means the spending function allocates alpha incorrectly, which can inflate false positive rates or waste statistical power.
Why isn't information fraction always just sample size divided by max sample size?
In simple settings with a single metric and no variance reduction, the information fraction is exactly the ratio of current to planned sample size. But several common experimental features change the relationship.
Variance reduction. When you apply CUPED (a technique that uses pre-experiment data to reduce metric variance), the effective information per observation increases. After variance reduction, 50,000 observations might carry as much statistical information as 80,000 unreduced observations. The information fraction should reflect the reduced variance, not the raw sample count. Confidence accounts for this by computing information fractions based on the actual precision of the estimator, not just the number of users enrolled.
Unequal group sizes. If the experiment allocates 80% of traffic to control and 20% to treatment (as in some rollout-style experiments), each observation contributes less information than in a balanced 50/50 split. The information fraction needs to reflect the effective sample size for comparing the two groups, which depends on the allocation ratio.
Longitudinal data. When users contribute repeated observations over time, the information per user depends on how many observations each user has contributed and the correlation structure within users. The raw user count understates the complexity.
How do information fractions interact with alpha spending?
An alpha spending function takes the information fraction as its input and returns the cumulative amount of alpha (significance budget) that should be spent by that point. For example, an O'Brien-Fleming spending function at information fraction 0.5 might allocate only 0.5% of the total 5% alpha, reserving the rest for later looks. A Pocock spending function at the same fraction might allocate 2.5%.
The critical property: the spending function is defined as a continuous function of the information fraction, not as a function of calendar time or sample count. This means the actual timing of interim analyses can shift (more users arrived than expected, or fewer) and the spending function still produces correct boundaries as long as the information fraction is calculated correctly.
At Spotify, where experiments run across 300+ teams with varying traffic levels, this flexibility matters. An experiment might reach its planned information fractions faster or slower than anticipated. The GST framework handles this gracefully because the spending function tracks information, not time.
What happens when the information fraction is misspecified?
If the platform computes the information fraction using raw sample counts but the experiment uses CUPED variance reduction, the interim analyses occur at the wrong points on the spending function. The boundaries are either too strict (wasting power) or too lenient (inflating false positives).
This is one of the capability-matrix cells that separates experimentation platforms in practice. Supporting sequential testing and supporting variance reduction are both common. Supporting them together, with information fractions that correctly account for the variance reduction, is less common. Confidence computes information fractions that reflect the actual statistical precision at each analysis point, including the effect of variance reduction, unequal allocation, and metric type.