Analysis methods

Eppo has four different methods for estimating the expected lift from experiment data and constructing a confidence interval around that estimate.

Overview

The details of each method, and the pros and cons are explained in detail below, but, in brief, the methods are:

Method	Description	Pros	Cons
Fixed-sample frequentist	Pick a target sample size, wait until you achieve that sample size, then make a decision	Maximizes ability to detect a treatment effect for a given sample size	Requires you to plan how long to run the experiment and then stick to the plan (even if the assumptions underlying the plan end up being incorrect)
Sequential frequentist (default)	Just start running your experiment, and make a decision when you want	Doesn't require a rigid plan and allows you to be flexible while still ensuring a certain false positive rate	Less power than fixed-sample, so need more sample to be able to reliably detect a lift
Sequential hybrid	This is a combination of the sequential and fixed sample methods: while the experiment is running we use sequential analysis, and once the experiment has finished, it switches to a fixed confidence interval. To maintain statistical guarantees, confidence intervals are slightly wider.	Combines the benefits of sequential and fixed sample testing: early stopping and high power at the end of an experiment	Requires to plan how long to run the experiment and slightly larger confidence intervals both during and at the end of the experiment
Bayesian	Combine data from the experiment with a prior belief about how likely different lifts are, and use the result to make a decision	Allows for making nuanced decisions based on the full distribution of expected lifts, which is particularly helpful when sample sizes are small	Aligning with stakeholders on how to interpret and use the results may require more effort, due to the very nuance and flexibility of the method (as well as lack of familiarity with Bayesian methods) Also, if prior beliefs are incorrect, and you don't have much data, then the lift estimates and confidence intervals will be incorrect

We chose to make sequential analysis the default because it allows for maximum flexibility while still making ship/no-ship decisions relatively straightforward: you can peek—and make a decision—at any time, without risking unacceptably high rates of detecting phantom effects.

While fixed-sample analysis might be more likely to detect an effect that is there if everything goes as planned, in practice experiments often don't go as planned, and that can invalidate the statistical guarantees of fixed-sample analyses.¹ In this way, sequential is the safest choice: it provides useful protections while maximizing the ability to adapt the experimentation process to a dynamic reality.

Sequential is not, however, always the best choice. Indeed, when sample size is hard to come by, it may not be a realistic option at all, which is why we've provided the others. We describe each method in more depth below, with a focus on the broad concepts, caveats to be aware of, and tips for how to best use and interpret the results; we provide all the Greek-letter-laden equations underlying each method on the statistical nitty-gritty page.

If you want to learn more about what you might want to consider when choosing an analysis method, you can find some discussion of that on the Analysis Plans page.

Fixed-sample analysis

A fixed-sample analysis is the most basic way to analyze the results from an experiment. First, you decide when you will look at the experiment results and make a decision: either after a fixed period of time, or once you have enough subjects in the experiment. Then, you start your experiment, wait until that predetermined decision point, and make a decision based on which, if any, metrics show significant movement compared to control.

The hard part with fixed-sample analysis is the first step: deciding when to look at the experiment results (and choose to ship or shut down). If you look at results too early, you might not have enough statistical power to detect a treatment effect even if it exists. If you look at results too late, then you've wasted time.

danger

Choosing when to make a decision is particularly fraught because, once you look at the results, you can only decide to ship, or to shut down—you can't decide to keep running and collect more data.

There are two primary ways this requirement gets violated, in practice:

1. You plan to run an experiment for a few weeks, but after a few days you notice that your metric has dropped by 10%. Faced with the prospect of weeks of lost engagement and revenue if you keep running, you decide to shut down the experiment.
2. You estimate how much sample size you need to be able to detect a lift, wait until you have collected that much data, then look at the results: the estimated lift is positive (or negative), but the confidence interval is too wide to exclude zero. So, you decide to continue running for a couple more weeks to "achieve significance"; if it then has an unambiguously positive lift, you ship it.

The problem in both cases stems from the fact that early on in an experiment, when you have less data, the lift estimate will fluctuate a lot purely by chance; if you collect more data, the fluctuations will tend to settle down, and the lift estimate will move closer to the true lift. The problem comes when you change how much data you collect based on the highly variable early results:

• If you stop an experiment if it has a randomly negative early lift, but continue it for an equally random positive lift, you don't give experiments that have bad luck early on a chance to recover; you will tend to shut down experiments that might, ultimately, have had positive lifts (thus missing out on potential improvements) and therefore bias your results downward (you will tend to think experiments have worse results than they actually do, and you'll miss). This is what would happen in scenario 1 above.

• If you stop an experiment if it has a randomly large early lift, but continue it if the lift is mediocre, you don't give experiments with outlier lifts (positive or negative) the chance to come back to earth. You will tend to conclude that experiments had a (positive or negative) effect that was actually just random noise. Since you only ship the positive experiments, while shutting down both negative and neutral ones, in practice the result is that you will overestimate how much positive impact you got from experiments. This is what would happen in scenario 2 above.²

In order to avoid pitfalls like the two above scenarios, it's important, when using fixed-sample analysis, to set a decision-making process (for example, when you will make a decision, and what criteria you will use to make the decision) before starting the experiment, and then stick to it.

There are primarily two ways to pick a decision point before running the experiment:

1. Pick a set period of time to run an experiment.

Sometimes organizational or business constraints determine how long an experiment can be run. This is fine as long as the duration is long enough to actually be able to detect a difference, if there is one, but there's no way to know how long is "long enough" without...

2. Do a power analysis to determine target sample size.

A power analysis looks at the historical data for a given metric, and tries to predict how many subjects will be needed in order to be able to detect a lift of a given size. First, you determine the minimum lift you care about (the minimum detectable effect, or MDE) for each metric you're going to observe in the experiment. Then, you can use a tool like Eppo's Sample Size Calculator to get the estimated sample size needed to detect that MDE.

One danger with a power analysis is that historical behavior is not always a perfect predictor of future behavior, which can leave you unable to detect a true effect even after achieving the predetermined sample size. In addition, some metrics and populations might not have historical data at all—as with new users, or recently created metrics.

A more pernicious danger is that your power analysis is based on some fundamental assumptions about the experiment—such as which metrics you care about, which metrics are likely to move, and how you'll balance tradeoffs between different metrics—and if those assumptions are wrong it can force you into suboptimal decisions in order to ensure statistical validity. For example, sometimes an experiment uncovers unexpected patterns of behavior or even potential bugs, and you might want to be able to gather more data in order to confirm and explore such anomalies, without putting the whole experiment at risk; a fixed-sample analysis precludes that option.

Pros of fixed-sample analysis

• Minimizes false-negative rate for a given sample size. If you have small sample sizes and struggle to get enough data to detect treatment effects, a fixed-sample experiment will give you the best shot at detecting an effect.

Cons of fixed-sample analysis

• You have to select your sample size or duration ahead of time. Eppo's Sample Size Calculator makes this significantly easier, but if the assumptions you put into it end up not holding (for example, if the variance in your metrics is higher than it historically has been), then you can still end up underpowered at the predetermined sample size.

• You cannot make decisions before the sample size is achieved. (AKA: No peeking! 🙈) Shipping/shutting down experiments with big gains/drops early will violate the statistical guarantees of fixed-sample analysis³

• If you end up not seeing a treatment effect, you cannot decide to continue running the experiment to collect more data. The flip side of the peeking problem, continuing to run an experiment because of a lack of significant metric movement will also violate the guarantees on false positive rate. This also means that, if, after looking at the results, you decide you want to understand how the experiment affected different segments or subpopulations, you can't just keep running to be able to slice and dice accordingly. Instead, you need to start a new experiment, and run it longer before making a decision.⁴

• You cannot adapt when you decide to ship or shut down based on experiment results. If, for example, you see hints of a surprising trend, you cannot decide to continue running to be able to confirm or disprove that trend. Or, if the business needs change—for example, a metric that you didn't care that much about when you planned the experiment has since become much more important—you can't adapt to those changes.

Sequential analysis (the default)

Sequential analysis allows you to run your experiment without predetermining the duration or sample size, and allows you to make a ship or shutdown decision at any time—while still guaranteeing the specified false positive rate. To do this, we give up some power compared to the fixed-sample method. In exchange, you get much more flexibility in when and how you make decisions; you can monitor experiment results continuously without causing problems; and you can avoid the time and error-prone choices (such as what MDE is acceptable for each metric) required to plan an analysis ahead of time.

Pros of sequential analysis

• You don't have to predetermine your sample size. Unlike with fixed-sample analysis, sequential analysis does not require you to do a power analysis beforehand.⁵ More importantly, it also does not require you to restart the experiment if any of the parameters of that power analysis (such as the expected metric values and variances) end up being incorrect. Note that it can still be insightful to use Eppo's Sample Size Calculator to understand, operationally, when you can expect to be able to detect an effect if it exists.

• You can make any decision at any time, safely. Go ahead, peek away 👀! If you want, you can check the results every day and ship as soon as you detect a nonzero treatment effect: you don't have to worry about it being a false positive just because you haven't reached a predetermined sample size.[^earlystop] Also, you can decide to keep running the experiment to collect more sample, and, unlike with fixed-sample analysis, doing so will not invalidate the statistical guarantees.

Cons of sequential analysis

• You will have less power, which means experiments might take longer. For a given sample size, sequential analysis is less likely to detect a true effect than is fixed-sample analysis. In other words, to have the same statistical power as fixed-sample analysis, you'll need to run your experiment longer if you use sequential analysis. However, if you see strong effects,a sequential analysis you can ship early, which you usually cannot do with fixed-sample analysis. The ultimate result is that, in general, sequential can allow you to make decisions faster when the lifts are large (positive or negative):

  Lift is...	Decision	Faster method
  Substantially bigger than MDE	Ship	Sequential
  Moderately larger or equal to MDE	Ship	Fixed-sample
  Small or nonexistent	Shut down	Fixed-sample
  Moderately negative	Shut down	Fixed-sample
  Very negative	Shut down	Sequential

  Another caveat is that, if you end up having to restart a fixed-sample experiment because the initial power analysis was wrong, or because you just got unlucky, it's possible that you could still make a decision faster with sequential analysis.

Sequential hybrid analysis

The sequential analysis method is attractive because it allows us to continuously monitor experiment results. However, the downside is that the confidence intervals necessarily have to be more conservative. On average, it may or may not be faster than the fixed sample analysis depending on assumptions, but in the worst case it definitely requires much more samples. This can make for a difficult decision between the two options.

The sequential hybrid option aims to take the best of both worlds: continuous monitoring as well as tight confidence intervals at the end of the experiment. We achieve this by combining (slightly more conservative versions of) the sequential approach while the experiment is running, and then switching to the fixed sample analysis once the end date has been reached.

Of course, there is no free lunch; there is a price to pay: first, the confidence intervals during each of the two phases are slightly wider (about 10-15%), and second it requires setting an end date of the experiment ahead of time. However, we believe this makes for an attractive trade-off.

Sequential hybrid as two one-sided tests

Another way to use sequential hybrid to take the best of both worlds is to stop early for degradations only but wait until the pre-planned end date to declare winning variants. The inflexibility of the fixed-sample methodology is often most apparent when the test is doing poorly; if a variant is significantly degrading metrics, you will likely want to pull the plug instead of fulfilling your promise not to peek. Significant degradations due to poor user experiences also often have large effect sizes that offset the loss of power from the sequential methodology. In these cases, the point estimates for the lift are also of less interest compared to experiments with "winning" variants.

Conversely, for detecting improvements, it is often helpful to have additional power and to have more reliable estimates of the treatment effect, which are both advantages of the fixed-sample approach. As a result, a sensible approach is to use sequential hybrid's sequential test for early detection of poorly performing variants and its fixed-sample approach for detecting improvements.

This approach is effectively two one-sided tests: a sequential test with a significance level $\frac{\alpha}{4}$ is performed continuously on the degradation tail and a fixed-sample test with a significance level $\frac{\alpha}{4}$ is performed on the experiment's end date on the improvement tail. To understand where the $\frac{\alpha}{4}$ comes from, first recognize that the core idea of the sequential hybrid methodology is that we allocate half of the "$\alpha$ budget" to the sequential test and half to the fixed sample test. This means that the two-tailed sequential test has a significance level $\frac{\alpha}{2}$ and half of that $\alpha$ is allocated to each tail, leaving $\frac{\alpha}{4}$ for each tail. If we stop the test early only for degradations, we only reject the null hypothesis on the degradation tail, which means that this test is effectively a one-sided test with significance level $\frac{\alpha}{4}$. Similarly, when we run the fixed-sample test at the end of the experiment, the two-sided test has significance level $\frac{\alpha}{2}$ and the one-sided test (improvements only) has significance level $\frac{\alpha}{4}$. Note that if you use this approach, you may want to consider setting the confidence level to 90% ($\alpha = 0.1$) to follow the convention of allocating $\alpha$ = 0.025 to each tail, which normally would be achieved by setting the confidence level to 95%.

Bayesian analysis

Both fixed-sample and sequential methods described above use a frequentist approach, where we test how likely it would be to see the observed data under the assumption that the treatment and the control were identical; that is, they take it as given that the lift will be zero on average—but with random fluctuations—and ask how often those fluctuations would produce the kind of results observed during the experiment.⁶ In contrast, the Bayesian approach takes the data as given, as well as what we believed before we collected data, and asks what distribution of lifts are most compatible with the combination of the two.

More formally, the Bayesian method starts with a prior distribution for the lift, which describes our beliefs about what the lift might be before we run the experiment. Then, we use the data gathered in the experiment to update that prior and produce a posterior distribution (so called because it comes after the data), which describes our new beliefs about what the lift might be, given both the prior we started with and the data we've observed.

caution

One important difference between Bayesian and the frequentist methods described above is that the center of the confidence interval is not the lift measured from the data, even when CUPED is disabled.⁷ Instead, the lift measured from the data is used to update our prior, and the resulting posterior distribution determines both the center and bounds of the confidence interval.

The prior we use is described specifically on the Statistical nitty-gritty page, but in essence we set our pre-experiment belief to be that the lift on any given metric will be, on average, zero, and that there will be random fluctuations around that average such that for 50% of experiments the lift will fall between -21% and +21%, and for 95% of experiments the lift will fall between -62% and +62%; if your experiments tend to show bigger lifts in either direction, then our Bayesian confidence intervals⁸ might be too narrow (biasing toward showing an effect when there is none).

Many ways to be Bayesian

Being "Bayesian" simply means that you start with a prior belief, update it with data, and make decisions using the resulting posterior—it doesn't dictate how to set your prior. You might use a prior on the distribution of a metric at the per-subject level, or you might set your prior on the distribution of the mean of the metric, across subjects. If you have a deep understanding of each metric and of the patterns in your data, you can establish a complex prior that captures all the dynamics of your product and user base; using a correct prior can allow you to make correct decisions with less data than with frequentist methods.

However, developing such a deep understanding of a complicated system requires a lot of research and specific knowledge, and using a complex prior also requires using computational methods that do not scale well.⁹ In general, there is a tradeoff between doing a very specific analysis that requires less data but more time and expertise; vs. doing an analysis that requires more data but less time and expertise—and that is more generalizable to different metrics and different contexts.

We may not have a deep understanding of your particular product, but we do have a deep understanding of experiments and the lifts that are typical across many kinds of experiment. So, we establish our priors on the lift itself, rather than on the aggregation of each metric or the behavior of individual subjects, which allows us to take advantage of our prior knowledge and provide experiment results in a way that Bayesians can use to make nuanced decisions.

Wait, what is the N you're using to update your prior?

The N is 1!

In general, compared to frequentist methods, Bayesian methods can be more straightforward to communicate about, but also provide fewer guarantees. In many ways, they hide fewer things from you than frequentist methods do, but as a result they force you to confront assumptions and choices that frequentist methods can sweep under the rug. They require more care and attention but are less rigid. This means that they also allow you to adapt the level of rigor with which you make decisions to the business and data context; you can set a very high bar for mission-critical decisions where data is readily available, or you can relax requirements if you need to move quickly despite low sample sizes. This is of course both a strength and weakness: it allows for greater flexibility in the decision-making process but requires more decisions about that process.

Some summary statistics that can help make decisions surrounding Bayesian experiments are described in the documentation on statistical details.

You can, of course, decide to use Bayesian results like you would frequentist results: if the confidence interval is above zero, there was a positive lift, and the experiment gets shipped. In some ways, that negates key benefits of a Bayesian approach, and since that approach depends so much on a choice of prior it's often prudent to think more about what happens if I'm wrong than with frequentist methods (luckily, Bayesian methods make it much easier to think about exactly that question). However, even if you apply a simple decision rule to Bayesian experiment results, there are a number of ways where Bayesian analysis can better allow you to make experiment decisions quickly and rigorously.

Pros of Bayesian analysis

• You don't have to be quite so careful with statistical terminology. A frequentist confidence interval is much easier to use than it is to explain, in a precise way. But Bayesian credible intervals allow you to describe experiment results in a way that is often more natural, especially for non-technical stakeholders. So, you can freely say things like "There is an 85% chance the treatment is better than the control," which is a no-no for frequentist confidence intervals.¹⁰

• You can make a decision based on a nuanced understanding of the probabilities. Frequentist methods for experimentation inevitably boil down to: Are we confident that the lift is above (or below) zero? With Bayesian results, you have the whole posterior distribution, not just a binary "yes" or "no". For example: even if the lift is very likely to be positive, how likely is it to be big enough to actually matter? When an experiment can move metrics in different ways, and different movements in different metrics can have very different costs to the business, this level of nuance allows you to make the best decision for each particular situation.

• You can make decisions even with little data. In the context of small samples, it can often be very difficult to be confident that the treatment is actually moving a metric. In essence, frequentist methods start "from scratch": even if you know a lot about your system outside of the data you've collected, there's no way to use that information—the data stands on its own. Bayesian methods allow you to discount very unlikely theories (like having a lift be 1,000%), and therefore the data doesn't need to do as much work all on its own to narrow down the possible explanations for what the treatment is doing.

• You can make a decision at any time. Bayesian analysis is not immune to peeking, but it does avoid the issue simply by making no promises about the false positive rate. The idea of there being a "true effect" or "no effect" doesn't really make sense in a Bayesian paradigm: the lift is not simply zero vs. non-zero, but rather has a continuous distribution of values. Since there is no "positive result" or "negative result", there can't be a false positive or false negative, which means there are no statistical guarantees to violate by making a decision too early.

Cons of Bayesian analysis

• You have to trust the prior. With enough data, having an incorrect prior won't significantly bias your results. But since one of the benefits of Bayesian approaches is that they are easier to make decisions with when sample sizes are small, and the prior plays a larger role in just these cases, the choice of prior can sometimes determine whether a variant gets shipped or not. Even if you establish that the prior reflects the historical patterns in your pre-experiment data, you still have to assume that the experiment you're running is similar to those you've run in the past, and (particularly when your product is growing rapidly) this might not be true.

• Stakeholders might not be used to the Bayesian way of thinking. Although some Bayesian concepts are more intuitive than their frequentist counterparts, frequentism is still the more common paradigm people learn for doing statistical inference, so switching from "This metric had a statistically significant increase" to "The treatment is 89% likely to have increased this metric" might require educating stakeholders on how to use Bayesian results.

• You have to decide how to make decisions.

  The output of Bayesian inference is a statistical distribution, rather than a "yes"/"no" answer to the question "Is the treatment better than control?" This means that decision-makers need to align on how to translate that distribution into a decision on whether to ship a particular treatment. For example, do you care more about risk, or probability the treatment beats control? What probability that the lift exceeds the MDE is considered "good enough to ship"? (The documentation on Bayesian summary statistics might provide some useful decision-making procedures.)

Footnotes

1. Problems arise with fixed-sample tests whenever how long an experiment is run is influenced by the early results. We describe some common but problematic practices here. ↩

2. Although scenario 2 was described as reaching the decision point and then continuing, it is identical to one where you peek early and then ship if you detect an effect or else let the experiment run to the end. ↩

3. Even if you do restart the experiment, you may be unable to measure the intended treatment effect, as some portion of the new control group will have been in the old treatment group, and so don't represent a true control. ↩

4. Specifically, that confidence interval correctly limits the false positive rate in a way that reflects the confidence level you've set. ↩

5. Technically, fixed-sample methods do not require a power analysis, but without one there is no way to know when is an appropriate to look at the results. ↩

6. That is,**** they use Null Hypothesis Significance Testing. ↩

7. When using CUPED, the frequentist methods display the predicted lift after controlling for pre-experiment differences between variants. ↩

8. Technically, credible intervals. ↩

9. In particular, Markov chain Monte Carlo methods. ↩

10. Technically, and pedantically, a frequentist confidence interval allows you to say "If we ran this experiment many times, 95% of the confidence intervals would contain the true lift value," but not "The true lift is 95% likely to be in this interval" or even "If we ran this experiment many times, our estimate of the lift would fall within this interval 95% of the time." ↩