Information Time Versus Calendar Time

Two Clocks in Sequential Monitoring

A group sequential study is observed at successive interim looks, but those looks can be described using two different clocks. Calendar time records when a look occurs: for example, month 4 or month 12 after a study begins. Information time records how much statistical information has accumulated by that look. The clocks need not advance at the same rate. Enrollment, outcome occurrence, follow-up, and data availability can all make information accumulate unevenly over calendar time.

The canonical joint distribution is organized by information levels, not by the number of weeks or months between looks. Thus, two monitoring schedules with the same ordered information levels have the same canonical joint law under the same model and parameter, even if the looks occur on different dates. This is a statement about the statistical distribution under specified assumptions; it does not mean that calendar time is unimportant for trial conduct, data quality, or interpretation.

Normalized information time is dimensionless, whereas calendar time has units such as months. It describes progress toward a specified final information level, not progress through the calendar. In particular, \(t_k\) is not generally equal to the fraction of the planned calendar duration that has elapsed.

The Canonical Model Uses Information Levels

For the canonical model, write \(S_k\) for the cumulative score at look \(k\), and \(Z_k=S_k/\sqrt{I_k}\) for its standardized statistic. In the usual canonical normal formulation, cumulative score increments over disjoint information intervals are independent and normally distributed. Under a fixed parameter value represented by canonical drift \(\theta\), this gives the representation

where \(B(I)\) is standard Brownian motion indexed by information \(I\). Equivalently, the cumulative scores have means \(\theta I_k\) and covariance \(\operatorname{Cov}(S_i,S_j)=I_i\) for \(i\le j\). The standardized statistics therefore satisfy

These expressions contain the information levels and the drift, but not the calendar dates \(c_k\). The calendar schedule can affect the law only insofar as it changes the information levels or breaks assumptions that support the canonical model. Under \(H_0\), where \(\theta=0\), the mean of each \(Z_k\) is zero; under the fixed alternative represented by \(\theta\), the mean at a look depends on the square root of its information.

Why a Change of Clock Preserves the Canonical Law

The Brownian representation gives a direct justification. Suppose the information levels \(I_1,\ldots,I_K\) are held fixed, while the calendar dates are replaced by another increasing sequence \(c'_1,\ldots,c'_K\). At each look, the canonical statistic is formed from the score accumulated up to information \(I_k\), not from the elapsed duration \(c_k\). Since the joint distribution of the Brownian values \(B(I_1),\ldots,B(I_K)\) is determined by their information-time arguments, changing the date labels leaves that distribution unchanged.

For \(i\le j\), dividing by \(\sqrt{I_iI_j}\) gives the correlation of the standardized statistics:

No calendar interval appears in this calculation. The same reasoning applies to the joint normal distribution at all looks: its means and covariance matrix are determined by \(\theta\) and the information levels. Consequently, any sequential rule that uses only these statistics at those looks has the same canonical probability law when only the calendar labels change. Its actual calendar duration, however, can change.

Information and Calendar Progress Need Not Match

The rate at which information accumulates can vary across calendar intervals. A long interval may add little information, while a shorter interval may add much more. Equal spacing of looks in calendar time therefore does not imply equal increments in information, and equal increments in information do not imply equal calendar spacing.

A useful way to describe a study is to report both clocks: when each look occurred and what information level was available. Calendar time is essential for operational questions, such as how long recruitment took or whether follow-up was complete. Information time is the relevant coordinate for the canonical joint distribution.

Assumptions, Diagnostics, and Limits

Interpreting the canonical model by information time requires more than simply recording \(I_k\). The group sequential approximation relies on a suitable score-based model, a valid information measure, and approximately independent score increments over disjoint information intervals. The joint normal approximation should be reasonable at the looks being considered. The information levels should be ordered and refer to the same parameter and statistical model throughout monitoring.

The simple invariance to calendar relabeling is clearest when the look schedule is fixed independently of the accumulating outcomes and the canonical information levels are the same. If a look is triggered by a data-dependent event, if information is itself random in a way related to the observed score, or if outcome availability depends on the outcomes, replacing dates while treating information as fixed may not preserve the joint law without additional assumptions. Calendar-driven changes in the population, treatment delivery, measurement, or follow-up can also undermine the model even when a numerical information estimate appears unchanged.

Practical checks include confirming that the reported information is cumulative and comparable across looks, checking that increments are positive, and comparing intended with observed information progress. Large changes in information accrual across calendar intervals are not by themselves a violation: they are precisely why the clocks must be distinguished. More concerning are inconsistent definitions of information, unexpected dependence between increments, or evidence that the statistical model changes over time. In those settings the canonical law may be a poor description.

This is different from an analysis that models the process directly as a function of calendar time. Such a model may be appropriate when the scientific question concerns when effects emerge, when hazards vary with time, or when calendar-time trends are central. The canonical information-time representation instead describes the joint behavior of sequential score statistics as information accumulates. Neither scale is universally preferable; the scale must match the statistical model and question.