Is 2.413 always the Pocock boundary?

No. It is a familiar approximate value for a particular five-look, two-sided 5% design.

Can I inspect results every day anyway?

Not under a plan that only accounts for a different set of looks; additional monitoring needs an appropriate design.

Does failure to cross prove no effect?

No. It means the specified rejection criterion was not met, not that the null hypothesis has been established.

AI Glossary / Statistics and probability / Advanced

Pocock boundary

Definition

A Pocock boundary is a group sequential testing boundary that uses a constant critical value on the standardized test-statistic scale across a planned set of analyses. The boundary is chosen to control the overall probability of a false positive across the repeated looks at accumulating data. It is associated with Stuart Pocock's work on group sequential clinical trials. The constant value is a property of the specified design, not a universal number that can be reused for any study or any monitoring schedule.

What it means

Repeatedly testing new accumulated data at the usual fixed-sample significance threshold creates additional opportunities for a false positive. The analyses are correlated because later looks include observations from earlier looks, but this correlation does not make unrestricted repeated testing harmless. A Pocock design raises the common threshold to account for the planned sequence. Compared with a design that demands extremely strong early evidence, it makes early stopping more accessible, while generally requiring stronger evidence at the final analysis than a conventional single-look test.

How it works

Before data monitoring begins, specify the hypothesis, one-sided or two-sided error rate, number of analyses, information timing and stopping rules. Calculate the constant standardized boundary from the joint distribution of the sequential test statistics under the null hypothesis. At each planned look, compare the accumulated statistic with the relevant boundary. The classical construction uses planned information increments, often equally spaced. Information is a statistical precision measure and need not be identical to calendar time or participant count. If the schedule changes, use an appropriate planned adaptation or error-spending framework rather than assuming the original boundary still controls error.

Three practical examples

Clinical research: a trial can permit early stopping for evidence of a treatment difference while accounting for repeated analyses.
Online experimentation: a team may plan several information-based evaluations of a product comparison using a suitable sequential model.
AI evaluation: investigators can design staged data collection for a predeclared performance comparison, provided the statistic, dependence assumptions and monitoring procedure fit the sequential method.

Think of it this way

Think of agreeing on the rules for several checkpoints before starting a journey. Each checkpoint offers a chance to stop, so the stopping requirement must account for all those opportunities. Choosing an ordinary threshold at every checkpoint after seeing the results changes the rules and can make accidental success look convincing.

Worked example

Consider an illustrative two-sided design with five equally spaced information looks, an overall type I error target of 0.05 and the classical normal-theory assumptions. The documented Pocock critical value is approximately 2.413 at each look. Its ordinary two-sided normal-tail probability is about 0.0158, not 0.05. Suppose an original hypothetical sequence gives Z values of 1.6, 2.2 and 2.5 at the first three looks. The first two values do not cross the boundary; the third does. The value 2.2 would exceed the familiar fixed-look threshold of approximately 1.96, but it is insufficient under this planned sequential rule. This example applies the documented boundary rather than deriving it from those three values. The nominal 0.0158 at a look is not the overall sequential p-value, and multiplying it by five is not a valid calculation of the overall error because the cumulative looks are correlated.

Illustrative case study

In a fictional case, Beacon Experiments plans five evaluations of a new service workflow. The analysts specify the comparison, information targets and boundary before launching the study. After the second analysis, a manager notices that the ordinary p-value is below 0.05 and proposes declaring success. The team follows the planned sequential criterion and continues because the boundary has not been crossed. Its final report describes the monitoring design and uses inference appropriate to sequential observation rather than presenting an unadjusted fixed-sample result as if no interim analyses occurred.

Common misconceptions and mistakes

A constant standardized boundary does not mean a constant raw difference threshold: the standard error changes as information accumulates. The boundary is not simply 0.05 divided by the number of looks. A one-sided design requires its own specification, and a boundary for five looks does not authorize unlimited additional checking. Pocock-style error spending is related to the classical approach but is not automatically identical to its constant boundaries under every schedule. Futility rules, multiple endpoints and other adaptations introduce further design considerations. Crossing an efficacy boundary also does not settle safety, practical value or the validity of the underlying measurements.

Business use

The main business advantage is a transparent way to consider earlier decisions while protecting the planned false-positive rate. The trade-off should be evaluated against expected stopping time, maximum information and the consequences of errors. Keep a record of scheduled and unscheduled access to results, and separate the statistical stopping criterion from operational governance. Sequentially valid estimates and intervals may require special methods. The exact design should be calculated and reviewed for the actual application, especially where observations are delayed, correlated or collected at irregular information increments.

Check your understanding

The raw effect estimate stays constant while its standard error falls. Must its Z statistic stay constant?

Answer: No. Z divides the estimate by its standard error, so it can increase even though the estimated effect is unchanged.

Related terms

Type I error
O'Brien - Fleming boundary
Stopping rule
Group sequential design
interim analysis
alpha spending

Learn next

Learn how information fractions determine correlations between cumulative test statistics, then compare constant boundaries with error-spending designs and sequentially adjusted inference.

Sources

SAS, Summary of Methods in Group Sequential Designs: https://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/statug_seqdesign_sect030.htm Stata, Adaptive Designs manual, gsbounds Example 1, five-look Pocock boundaries: https://www.stata.com/manuals/adapt.pdf Penn State, STAT 509, Late Phase Studies: Interim Analyses: https://online.stat.psu.edu/stat509/Lesson10

Previous: Ply Next: Pod slurping