Deriving O'Brien-Fleming Information Scaling

Why Information Scaling Matters

An O’Brien-Fleming boundary is high at early analyses and approaches its terminal reference threshold as information accumulates. The shape has a direct explanation: the score is standardized by the square root of information, so a fixed threshold on the accumulated score becomes a larger threshold on the standardized scale when information is small. This tutorial derives that relationship and clarifies exactly what it does—and does not—say about a calibrated group sequential design.

The argument uses the information-standardized score and the information fraction. The earlier material on the O’Brien-Fleming boundary shape establishes the qualitative pattern; here the aim is to explain its approximate scaling algebraically. The scaling is a useful description of the boundary’s shape, not a substitute for calibrating the planned analyses to control the design’s overall Type I error.

For clarity, take positive values of \(\theta\) to favor the treatment and consider an upper efficacy boundary. A two-sided design applies the corresponding argument to \(|Z_k|\); a lower one-sided boundary follows by reversing the sign. Write \(c_*\) for the selected terminal reference critical value. For example, the single-analysis two-sided normal reference at significance level \(\alpha=0.05\) is approximately \(1.96\). In a group sequential design, however, the actual terminal critical value may be chosen as part of a joint calibration and need not equal that single-analysis reference.

Deriving the Inverse-Square-Root Shape

The O’Brien-Fleming scaling approximation sets the standardized threshold at look \(k\) to \(c_*/\sqrt{t_k}\). Substituting \(t_k=I_k/I_K\) and using \(Z_k=U_k/\sqrt{I_k}\) shows what this approximation means on the unstandardized score scale.

The last expression is the key justification: under this shape approximation, the efficacy threshold on the cumulative score scale is constant across looks. Dividing that fixed score threshold by \(\sqrt{I_k}\) to obtain a standardized threshold produces the factor \(1/\sqrt{I_k}\), or equivalently \(1/\sqrt{t_k}\) after expressing information relative to its final value. At the final analysis, \(t_K=1\), so the approximation gives \(c(1)=c_*\).

This derivation is algebraically exact once the approximate boundary \(c_*/\sqrt{t_k}\) has been posited. It does not prove that these values are the exact critical values that give a chosen familywise Type I error for every set of planned looks. The standardized scores across looks are correlated, and the probability of crossing at any look depends on their joint distribution. A design must calibrate its actual boundaries accordingly.

Worked Calculations

Assumptions and What the Approximation Leaves Out

The calculation rests on a particular statistical structure. The \(U_k\) values must be cumulative scores for the same treatment-effect parameter, and \(I_k\) must represent their corresponding cumulative information. Under the canonical model, the score has mean approximately \(\theta I_k\) and variance approximately \(I_k\), so \(Z_k=U_k/\sqrt{I_k}\) has approximately unit variance under \(H_0\). The planned information fractions must be meaningful for the analysis sequence; in particular, the simple expression treats \(I_k\) and \(I_K\) as specified information quantities rather than arbitrary sample-size percentages.

For a standard group sequential calibration, the joint behavior across looks also matters. In the canonical model, earlier and later scores are dependent; under the null, their correlation is determined by the ratio of the earlier to later information. The probability of crossing at least one boundary is therefore not obtained by treating the analyses as independent or by assigning the full \(\alpha\) separately at every look. The inverse-square-root rule explains the O’Brien-Fleming shape, while a joint calibration determines the actual cutoffs needed for a specified overall error rate.

Nor does the algebra make an approximate boundary robust to a misspecified information model. If the actual variance structure differs from the one used to calculate information, then \(Z_k\) may not be standardized as assumed. Likewise, if the information fractions are inaccurately planned or estimated, substituting them into \(c_*/\sqrt{t_k}\) can yield a boundary that does not match the intended design. The formula is a guide to shape under the stated model, not a repair for invalid scores or information estimates.

Comparison, Diagnostics, and Pitfalls

A flat, Pocock-style boundary keeps the standardized critical value roughly constant across looks. O’Brien-Fleming scaling instead demands a much larger standardized score early and moves toward the terminal reference later. Both shapes require a design-level calibration when used for group sequential testing; neither comparison changes the need to account for repeated looks. A nominal single-look cutoff applied independently at every analysis is not generally a valid substitute for that calibration.

Useful checks follow directly from the scaling argument:

• At \(t=1\), the formula must return \(c_*\).
• For \(0<t<1\), \(c_*/\sqrt{t}\) must exceed \(c_*\) when \(c_*>0\).
• Halving the information fraction multiplies the approximate threshold by \(\sqrt{2}\), not by \(2\).
• Multiplying the standardized boundary by \(\sqrt{I_k}\) should recover the same score threshold \(c_*\sqrt{I_K}\) at each look.

A common mistake is to say that the required standardized threshold is inversely proportional to the information fraction itself. That would give \(c_*/t\), making the threshold too high. The score standardization uses \(\sqrt{I_k}\), so the correct relationship is \(c_*/\sqrt{t_k}\). A second mistake is to interpret this formula as an exact boundary that automatically spends the desired Type I error. The correction is to separate the shape approximation from joint calibration: use the scaling to understand the geometry, then use a valid group sequential calibration for the specified design and error rate.

The approximation is most useful for understanding why early O’Brien-Fleming thresholds are stringent and for checking whether a proposed boundary has the expected scale. It does not provide exact calibrated values for every finite set of looks, replace the assumed score model, or determine how a design should be adapted when information changes. Those questions require methods beyond this scaling derivation.