Translating Hazard Ratios Into Sequential Drift

Why a Hazard Ratio Determines a Drift Path

A prespecified hazard ratio describes the planned treatment effect on the log-hazard scale. In a group sequential trial, however, monitoring uses standardized statistics at successive looks. The link between the two is the approximate mean of the logrank statistic: as information accumulates, the expected statistic moves away from zero in the direction implied by the treatment effect. This mean path is called the sequential drift.

The drift calculation is useful for planning and interpretation, but it is not a prediction of the exact statistic at any look. A realized statistic also includes sampling variation. The approximation depends on a proportional-hazards model and on information being correctly characterized at each look.

The information \(I_k\) measures the precision accumulated by look \(k\), on the scale of the log hazard ratio. It is not necessarily equal to the number of participants or to the number of events. Define the information fraction \(t_k=I_k/I_K\), where \(I_K\) is the planned information at the final look. Thus \(I_k=t_k I_K\). The information fraction is a design-scale quantity; it need not equal the fraction of calendar time elapsed.

From the Log Hazard Ratio to Sequential Drift

Under the proportional-hazards alternative, the expected score is approximately the information multiplied by the log hazard ratio. This is the score-scale approximation used to translate an assumed effect into a mean standardized statistic:

The null variance of the score is approximately \(I_k\). Standardizing the score by \(\sqrt{I_k}\) therefore gives the approximate mean of \(Z_k\):

This is the basic drift formula. It applies look by look: insert that look’s information, or its information fraction and the planned final information. If \(HR\) is below one, the mean is negative under the stated convention; if the statistic is instead defined with the opposite score sign, the signs reverse. The magnitude grows in proportion to the square root of information, not in proportion to information itself.

The square-root dependence follows directly from the score and variance scales. Information multiplies the log hazard ratio in the expected score, while the standard deviation grows as the square root of information. Dividing the former by the latter leaves a factor of \(\sqrt{I_k}\). Consequently, doubling information does not double the mean standardized statistic; it multiplies its magnitude by \(\sqrt{2}\).

Using Information Fractions Correctly

Often the planned schedule is given in information fractions rather than as an information value at every look. Because \(I_k=t_k I_K\), it is straightforward to calculate the expected drift at each fraction once the final information and hazard ratio are specified. Fractions must refer to information, rather than being substituted as if they were event fractions or time fractions without justification.

If the total information is not supplied directly, it may sometimes be approximated from the planned number of events and the treatment allocation. For a two-group logrank comparison with treatment risk-set fraction near \(q\), a common event-information approximation is \(I\approx Dq(1-q)\), where \(D\) is the event count. This requires the risk-set allocation to remain reasonably close to \(q\); allocation at randomization alone does not ensure that condition under differential censoring or changing risk sets. The earlier material on Schoenfeld approximation covers the use of event counts for planning; here the event approximation is only a route to the information used in the drift formula.

Diagnostics, Alternatives, and Limitations

Before relying on a drift calculation, check that the model and information scale match the question. The hazard ratio should be explicitly defined as treatment relative to control; reversing that comparison changes \(\log(HR)\) to its negative. Confirm the sign convention of the logrank score as well. A sign mismatch can make a beneficial assumed effect appear to drift in the wrong direction even when the arithmetic is correct.

Next, check the proportional-hazards assumption. If the treatment effect changes over time, one constant hazard ratio may not describe the score’s expected behavior across all looks. The accumulated logrank score then reflects effects across event times and risk sets; simply applying one prespecified value of \(\log(HR)\) at every look may be misleading. A single average hazard ratio also need not imply the same drift path as a constant hazard ratio. When non-proportional hazards are plausible, use a design-specific model or simulation that represents the time-varying effects and censoring pattern rather than treating this simple formula as exact.

Information also requires scrutiny. It is determined by the event and risk-set structure, not merely by enrollment or calendar time. Loss to follow-up, unequal allocation in the risk sets, and administrative censoring can make realized information differ from the planned value. If information fractions are miscalculated, the square-root scaling will still be applied correctly to the wrong inputs. Use the information appropriate to the analysis cutoff, and distinguish planned information from realized information.

The main alternatives address different needs. A model-based survival analysis can estimate a treatment effect from observed data, whereas this calculation starts with a prespecified effect and projects its approximate mean score. A design-specific simulation can represent non-proportional hazards, censoring, or changing allocation more flexibly, but requires a more detailed data-generating model. Neither alternative changes the meaning of the simple drift formula: it is a concise approximation, not a complete description of the joint sequential statistics or their boundary-crossing probabilities.

The practical value of the calculation is that it puts an assumed treatment effect on the same standardized scale as the sequential statistics. It also makes the role of information explicit: a fixed hazard ratio produces a mean path whose magnitude increases like the square root of information, provided the proportional-hazards and information assumptions remain credible.