Why Separate Repetition From Time?
In a replicate crossover design, a participant receives at least one formulation more than once. A later response to that formulation may differ from its earlier response, but responses can also vary with study period or calendar time. For example, a change across periods could reflect ordinary time-related variation rather than repeated exposure itself. A schedule should therefore be checked for whether its treatment order allows these explanations to be separated.
The earlier material on period balance and replicate sequences establishes how treatments are allocated across periods. Here the question is more specific: does the order of repeated treatments provide comparisons that distinguish repeat status from period or calendar time? This is a property of the planned observations and the model, not a guarantee that a particular biological explanation is correct.
This binary definition treats all later exposures as one category. A model could instead distinguish second, third, or later exposures, or use separate repeat-status indicators for test and reference. Each additional term requires its own estimability check. The schedule audit must use the same definition of repetition that the proposed model will use.
A Model-Based Check for Confounding
A simple fixed-effects model for a response \(Y_{ip}\), such as a log-transformed pharmacokinetic endpoint, is
Here \(\mu_i\) is a participant effect, \(T_{ip}\) indicates treatment (with a chosen reference coding), \(\pi_p\) is a categorical period effect, \(\rho\) is the repeat-status coefficient, and \(\varepsilon_{ip}\) is the remaining error. This model is a diagnostic framework, not a claim that these terms explain every source of variation. Its interpretation requires an appropriate response scale, a defensible error model, and no omitted design feature that makes the planned comparisons misleading. In particular, the coefficient \(\rho\) is a conditional model contrast; it does not by itself prove a biological mechanism.
To check whether \(\rho\) is estimable, collect the nuisance columns (participant, treatment, period, and any other included terms) in a design matrix \(Z\), and let \(q\) be the column of repeat indicators. The repeat coefficient is separately estimable precisely when \(q\) is not in the column space of \(Z\). Equivalently, there must be no representation of \(q\) using only the nuisance columns. In a numerical audit, this can be checked by evaluating the rank of \(Z\) and of the matrix formed by adjoining \(q\).
The check can also be understood by removing the part of \(q\) explained by the nuisance columns. If the residualized repeat column is zero, the model has no remaining repeat-status variation after accounting for those terms. If it is nonzero, the repeat coefficient is identifiable in the stated model. This is an algebraic condition; very small residual variation can still make an estimate imprecise.
Worked Schedule Audits
Worked Example: Repeat Status Confounded With Period
Consider two four-period sequences, TRTR and RTRT. Each participant receives each treatment twice. For each sequence, list the treatment and whether that period is a later exposure to that treatment.
| Sequence | Period 1 | Period 2 | Period 3 | Period 4 |
|---|---|---|---|---|
| TRTR | T, \(Q=0\) | R, \(Q=0\) | T, \(Q=1\) | R, \(Q=1\) |
| RTRT | R, \(Q=0\) | T, \(Q=0\) | R, \(Q=1\) | T, \(Q=1\) |
Across both sequences, every observation in periods 1 and 2 has \(Q=0\), and every observation in periods 3 and 4 has \(Q=1\). With categorical period effects, the repeat column is exactly the sum of the period-3 and period-4 indicator columns:
Therefore \(q\) is in the span of the period columns. Adding \(\rho Q_{ip}\) to a model that already has categorical period effects cannot separate a repeat-status effect from those period effects. The schedule still has treatment replication, but this particular ordering does not support estimating those two sets of effects separately under the stated model.
Worked Example: Repeat Status Not Confounded With Period in the Model
Now consider sequence A = TTRR and sequence B = TRRT. Their repeat-status patterns differ by period:
| Sequence | Period 1 | Period 2 | Period 3 | Period 4 |
|---|---|---|---|---|
| A: TTRR | T, \(Q=0\) | T, \(Q=1\) | R, \(Q=0\) | R, \(Q=1\) |
| B: TRRT | T, \(Q=0\) | R, \(Q=0\) | R, \(Q=1\) | T, \(Q=1\) |
Check the model with participant effects, treatment, and categorical period effects. Suppose, for contradiction, that \(q\) could be represented using only those nuisance columns. Write that representation as \(Q_{ip}=a_i+b_p+c_{T_{ip}}\), where \(a_i\) is the participant contribution, \(b_p\) the period contribution, and \(c_T\) the treatment contribution.
For sequence A, comparing periods 2 and 1 gives \(b_2-b_1=1\), because treatment is T in both periods and \(Q\) changes from 0 to 1. Comparing periods 3 and 2 gives \((b_3-b_2)+(c_R-c_T)=-1\). For sequence B, comparing periods 2 and 1 gives \((b_2-b_1)+(c_R-c_T)=0\). Since \(b_2-b_1=1\), the last equation implies \(c_R-c_T=-1\). The comparison of periods 3 and 2 for sequence A would then require \(b_3-b_2=0\). But for sequence B, periods 3 and 2 both have treatment R and \(Q\) changes from 0 to 1, so \(b_3-b_2=1\). This contradiction shows that \(q\) is not in the span of these nuisance columns. The repeat coefficient is therefore estimable under this model.
This algebra establishes estimability, not precision or adequacy for a real study. With only one participant per sequence, the data would not support reliable population-level inference. Additional sequences and participants are needed for balanced allocation and useful precision.
Worked Example: Period Number and Calendar Time
Suppose two participants follow the same four-period timing, with observations on days 0, 7, 14, and 21 for one participant and days 2, 9, 16, and 23 for the other. Let \(D_{ip}\) be elapsed calendar day from a common origin. For every observation, \(D_{ip}=s_i+7(p-1)\), where \(s_i\) is the participant’s start day. The participant component absorbs \(s_i\), and period effects absorb \(7(p-1)\). Thus calendar day is exactly representable by participant and period columns; a separate calendar-day coefficient cannot be distinguished from those effects in this schedule and model.
Now suppose the second participant’s observations are on days 2, 10, 17, and 26. The day differences between participant 2 and participant 1 by period are 2, 3, 3, and 5 days. If \(D_{ip}\) were expressible as a participant contribution plus a common categorical period contribution, that between-participant difference would be constant across periods. It is not. Calendar day is therefore not an exact linear combination of participant and period columns in this schedule. A specific calendar-time term may be estimable, subject to the other model columns and sufficient data.
This check concerns the actual scheduled times, not merely the labels “period 1” through “period 4.” A study with identical intervals can have calendar time structurally aligned with period even when participants start on different dates. Conversely, variable timing can create some separation, but does not automatically ensure useful precision or remove other biases.
How to Audit a Proposed Schedule
State whether repetition means any later exposure, a particular exposure number, or repeat exposure to a specified treatment. Specify whether the time adjustment is categorical period, a linear trend, or actual calendar time.
For every participant-period record, enter treatment, repeat status, period, and the planned timing variables. Include the nuisance terms that will actually appear in the analysis model.
Compare the span of the nuisance columns with the span after adding the repeat column. If the repeat column adds no independent dimension, the coefficient is not separately estimable under that model. A rank deficiency elsewhere is not enough to reach that conclusion.
Tabulate repeat status by period, treatment, and sequence. Look for periods or treatments where both repeat and first-exposure observations occur, and examine how much information remains after adjusting for the nuisance terms.
Categorical period effects are a flexible way to adjust for differences among periods, but they consume more degrees of freedom than a prespecified linear period trend. A linear trend is more parsimonious, but it assumes that the period contribution changes linearly on the model’s response scale. Calendar time is another possible term; it answers a different question from period number and must be checked against participant, period, and other included columns. These approaches are not interchangeable merely because all describe “time.”
Diagnostics, Limits, and a Common Pitfall
A schedule audit should use the actual allocation, including planned sequence counts and timing. Useful warning signs include a repeat indicator that is identical to a period grouping, a treatment that is repeated only in one period, or calendar dates that are a fixed function of participant and period. Design-matrix rank and singular values can reveal exact or near dependencies. Cross-tabulations and plots of repeat status by period and sequence help explain what the algebra means. Near-dependence is not exact confounding, but it can produce a large standard error and unstable estimates.
Even when the repeat coefficient is estimable, interpretation has limits. The coefficient is conditional on the chosen model and captures the average contrast represented by its design column. It does not automatically distinguish pharmacological adaptation from other changes associated with order, nor does it remedy carryover, nonadherence, or informative dropout. Model assumptions and the study’s timing and conduct still matter.
Common mistake: “The full design matrix has deficient rank, so the repeat coefficient cannot be estimated.” This conclusion is too broad. A dependency may involve only nuisance columns, while the repeat column remains independent of their span. The correction is to test the repeat coefficient’s estimability directly: check whether \(q\) is in the span of the other model columns, or equivalently whether every null-space dependency leaves its coefficient unchanged.
The central design lesson is that replication alone is not enough to separate repeat exposure from time. Sequence order must produce variation in repeat status that remains after the intended time and treatment adjustments. If it does not, changing the model to omit important time effects may yield a number, but it does not create the missing design information.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a schedule, all first exposures occur in periods 1–2 and all later exposures occur in periods 3–4. Can repeat status be separated from categorical period effects?
- What does rank deficiency of a full design matrix establish, and what additional check is needed for the repeat coefficient?
- Why can different participant start dates fail to separate calendar time from period?
- If the repeat column is independent of all nuisance columns, what does that establish, and what does it not establish?
Solutions. (1) No. The repeat column is the sum of the period-3 and period-4 indicator columns, so it is confounded with categorical period effects. (2) Rank deficiency establishes that at least one parameter or contrast is not estimable. To assess the repeat coefficient, check whether its column is in the span of the nuisance columns, or use a coefficient-specific estimability check. (3) If every participant follows the same spacing, calendar day can equal a participant-specific start contribution plus a common period contribution; those terms absorb the calendar-day column. (4) It establishes algebraic estimability under the stated model. It does not establish adequate precision, model validity, absence of bias, or a causal biological explanation.