Why the Sequence of Occasions Matters
Repeated-dose data can show that a PK parameter differs across occasions, but the differences do not all have the same explanation. A parameter may fluctuate around a stable subject-specific level from one occasion to the next. Alternatively, it may move systematically with elapsed time or occasion number, for example because the underlying PK process changes during follow-up. A sequence can also contain both a trend and occasion-to-occasion deviations.
The distinction matters because an occasion random effect describes unsystematic variation around a mean pattern, whereas a trend describes a directional change in that mean pattern. Treating a trend as random occasion variability can inflate the estimated occasion variance and obscure clinically relevant change. Treating random fluctuations as a trend can instead impose a directional pattern that is not supported by the data.
A Model That Separates the Components
Let \(\theta_{ij}\) be a positive PK parameter for subject \(i\) on occasion \(j\). Let \(u_{ij}\) be a time variable or occasion index, centered at a meaningful reference point. A basic log-scale model is
Here, \(\theta_{\mathrm{pop}}\) is the typical parameter value at the reference time for the reference covariate values, \(\mathbf{x}_i\) contains subject-level covariates, and \(\boldsymbol{\beta}\) contains their coefficients. The subject effect \(\eta_i\) represents persistent between-subject deviation. The coefficient \(\gamma\) represents a common linear trend on the log-parameter scale. The occasion effect \(\kappa_{ij}\) represents a deviation specific to subject \(i\) and occasion \(j\).
A common starting assumption is \(\eta_i\sim N(0,\omega^2)\) and \(\kappa_{ij}\sim N(0,\pi^2)\), with occasion effects independent across occasions and independent of \(\eta_i\). These assumptions say that the random deviations have mean zero, constant variances, and no residual temporal pattern after the trend is included. If dependence between interindividual and interoccasion effects is part of the model, the covariance structure must be specified rather than silently assuming independence. The meaning of \(\gamma\) depends on the time scale: it is a change in log parameter per unit of \(u\), not a change per occasion unless \(u\) is coded as occasion number.
Centering \(u_{ij}\) makes the typical value interpretable at the chosen reference time. With \(u\) centered at the first occasion, for example, \(\theta_{\mathrm{pop}}\) refers to that occasion for reference covariates. The log scale keeps \(\theta_{ij}\) positive after transformation: a change of \(\Delta u\) at fixed random effects multiplies the parameter by \(\exp(\gamma\Delta u)\).
Worked Example: Calculating a Linear Trend
Suppose hypothetical occasion-level summaries of the log clearance deviation from a reference value are \(0.00,\ 0.03,\ 0.04,\ 0.09\) at occasion indices \(u=0,1,2,3\). For a simple descriptive calculation, fit a straight line by least squares to these four summaries. This calculation illustrates the pattern; a population PK analysis should account for the uncertainty in the underlying parameter estimates.
The mean index is \(\bar{u}=(0+1+2+3)/4=1.5\), and the mean log deviation is \(\bar{y}=(0.00+0.03+0.04+0.09)/4=0.04\). The centered values are \(u-\bar{u}=(-1.5,-0.5,0.5,1.5)\) and \(y-\bar{y}=(-0.04,-0.01,0.00,0.05)\).
The least-squares slope is
The fitted intercept at \(u=0\) is \(\bar{y}-\hat{\gamma}\bar{u}=0.04-(0.0280)(1.5)=-0.0020\). Thus the fitted log deviations are \(-0.002,\ 0.026,\ 0.054,\ 0.082\). The slope corresponds to a multiplicative change of \(\exp(0.0280)=1.028\), or about a 2.84% increase per occasion in the parameter under this log-linear description. It is a descriptive estimate from four summaries, not by itself evidence that the trend is statistically or clinically established.
Justifying the Trend Calculation
The slope calculation follows by minimizing the sum of squared differences between the observed summaries \(y_j\) and a fitted line \(a+\gamma u_j\). For a fixed \(\gamma\), the minimizing intercept is \(a=\bar{y}-\gamma\bar{u}\). Substituting this intercept gives centered residuals \((y_j-\bar{y})-\gamma(u_j-\bar{u})\). Differentiating their squared sum with respect to \(\gamma\) and setting the derivative to zero yields the slope formula used above:
This calculation assumes the summaries are meaningfully comparable on the same scale and that a straight line is an adequate description over the observed range. It does not establish that the individual occasion deviations are independent, nor does it replace fitting the PK model to the concentration data. The formula is useful for understanding the direction and magnitude of a pattern; inference should reflect the data structure and estimation uncertainty.
What Occasion-to-Occasion Variation Looks Like
After accounting for subject-level differences and any supported time trend, an occasion random effect should look like variation around zero rather than a persistent directional sequence. Its defining feature is not that every subject changes in the same way, but that the deviations are occasion-specific and, under the basic model, exchangeable across occasions. A run of positive deviations followed by negative deviations may indicate an unmodeled trend or other structure; random effects need not alternate signs, but their expected value at each occasion is zero.
Worked Example: Estimating Occasion Variation Around Subject Means
Consider hypothetical log-parameter deviations for three subjects, each observed on four occasions. Suppose there is no fitted trend for this calculation, and the deviations after removing each subject's mean are:
Subject A: \(0.08,-0.06,0.04,-0.06\). Subject B: \(-0.04,0.07,-0.05,0.02\). Subject C: \(-0.03,-0.02,0.06,-0.01\). Each subject's four values sum to zero.
Under independent, equal-variance, zero-mean occasion effects, a pooled estimate of the variance uses the sum of squared centered deviations divided by the residual degrees of freedom. The squared deviations sum to \(0.0152\) for A, \(0.0094\) for B, and \(0.0050\) for C, for a total of \(0.0296\). There are 12 values and three subject means have been estimated, so the degrees of freedom are \(12-3=9\). Therefore,
The estimated occasion standard deviation is about 0.057 log units under these assumptions. This is a small-sample illustration, not a substitute for estimating the variance in the nonlinear population model. The occasion means across subjects are \(0.00333,-0.00333,0.0167,-0.0167\), which do not form a steadily increasing or decreasing sequence. Their lack of a clear linear pattern is compatible with occasion variation, but three subjects cannot establish that a trend is absent.
Modeling Both a Trend and Occasion Effects
A fitted trend does not mean every subject follows the population pattern exactly. Likewise, adding occasion effects does not erase a trend. The two terms answer different questions: \(\gamma u_{ij}\) describes the systematic expected change, while \(\kappa_{ij}\) allows an individual occasion to depart from that expectation. The following hypothetical calculation illustrates their separation.
Worked Example: Interpreting Trend and Occasion Deviation Together
Suppose clearance has a reference typical value of \(5.00\ \mathrm{L/h}\), the subject's interindividual effect is \(\eta_i=-0.10\), and the population trend is \(\gamma=0.0200\) per month. On month 0, let the occasion deviation be \(\kappa_{i1}=0.050\); on month 3, let it be \(\kappa_{i2}=-0.030\). For simplicity, there are no additional covariates.
At month 0, \(u=0\), so
At month 3, \(u=3\), so
The population trend term alone multiplies the parameter by \(\exp(0.0200\times3)=1.062\), a 6.18% increase over three months at fixed subject and occasion effects. The realized value at month 3 is nevertheless lower because its occasion deviation is more negative. This is not a contradiction: a positive systematic trend does not require every observed pair of occasions to increase.
Choosing Among Alternative Structures
Different structures encode different assumptions about the pattern across occasions. The choice should follow the scientific question and the sequence visible in the data, not simply which model can fit the observed estimates most closely.
| Structure | What it represents | Main caution |
|---|---|---|
| Independent occasion random effects | Zero-mean deviations with a common variance and no systematic ordering across occasions | A persistent time pattern can be absorbed into the estimated occasion variance |
| Fixed linear time trend | A common directional change per unit of time or occasion index | A linear trend may be inappropriate if change is nonlinear or limited to particular periods |
| Fixed occasion-specific effects | A separate average shift for each labeled occasion, without requiring a straight-line pattern | Many occasion effects use more parameters and may not generalize beyond the observed occasions |
| Random subject-specific slope | Subjects differ in their rate or direction of change over time | Requires sufficient repeated information per subject to distinguish slopes from occasion noise |
A common fixed trend is a reasonable first representation when the scientific question concerns average change and a linear pattern is plausible. Occasion-specific fixed effects are useful when particular occasions have distinct meanings or a linear pattern is not credible, but they do not represent random fluctuation in the same way as an occasion variance. A random slope can be considered when subjects plausibly have different trends and the design supplies enough repeated observations to estimate that variation. These structures are not interchangeable.
Diagnostics, Assumptions, and Failure Modes
Assess the sequence after accounting for subject effects and the proposed trend. Plotting occasion-level parameter estimates against elapsed time, with subject identities shown, can reveal a shared direction, subject-specific trajectories, or irregular scatter. Such plots are diagnostic rather than definitive: empirical parameter estimates can be noisy, especially when each occasion has sparse concentration sampling. A model-based assessment should account for uncertainty in those estimates rather than treating them as error-free observations.
- Check for ordered residual patterns. After fitting a trend, occasion deviations should not retain a common increasing or decreasing sequence. A remaining pattern suggests that the trend form, time scale, or another systematic feature may be inadequate.
- Check whether occasions are comparable. Changes in dose, formulation, sampling schedule, assay, concomitant treatment, or study procedures can align with occasion number and mimic a time trend. The model and interpretation must distinguish these design changes where possible.
- Check the time scale. A slope per occasion is not a slope per month if intervals differ. Use elapsed time when the process is plausibly time-driven and occasion number only when it has a defensible interpretation.
- Check coverage and balance. A trend is difficult to distinguish from subject or occasion effects if few subjects contribute late occasions, or if occasion number is confounded with another change in the study.
- Check the random-effect assumptions. The basic occasion model assumes a common variance and no temporal dependence after modeled effects. A visibly changing spread or serial pattern is a warning that the simple structure may not describe the data.
The central limitation is identifiability: a pattern in observed concentrations can have several explanations, and data cannot separate them if the design provides no information to do so. For example, if all subjects receive a changed formulation only at the last occasion, an apparent time trend may be indistinguishable from a formulation effect without additional design information. A fitted slope is therefore not automatically a biological explanation or a causal effect.
Common Mistake
Wrong reasoning: “The parameter estimates differ across occasions, so I should add an interoccasion random effect.” This conclusion ignores whether the differences are ordered in time. If estimates rise steadily across occasions, a zero-mean random effect alone may wrongly describe systematic change as extra random variability.
Correction: First examine the sequence against a meaningful time variable. Represent a supported systematic pattern with an appropriate trend term, then use occasion effects for remaining occasion-specific deviations if the data support them. If the sequence is irregular, that is compatible with occasion variability, but it does not prove that the random-effect assumptions hold.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In a log-parameter model, what does a coefficient \(\gamma=0.01\) per month mean over two months, holding other model terms fixed?
- Why can a positive population trend coexist with a lower estimated parameter on a later occasion for one subject?
- What pattern would make an independent, zero-mean occasion effect alone questionable?
- When might fixed occasion-specific effects be preferable to a common linear trend?
Solutions. (1) The two-month multiplicative change is \(\exp(0.01\times2)=\exp(0.02)=1.020\), about a 2.02% increase, holding other terms fixed. (2) The subject's occasion-specific random deviations can offset the positive trend; a trend describes the expected pattern, not a required increase at every observed occasion. (3) A consistent ordered increase or decrease across occasions after accounting for other modeled terms suggests remaining systematic structure that a zero-mean independent effect alone does not represent. (4) They may be preferable when particular occasions have distinct scientific or design meanings, or when the average pattern is not plausibly linear and there are enough data to estimate the separate shifts.