Why a Carried Value Can Become a Wrong History
In an NLME data set, a covariate may be measured less often than the drug response. A common representation assigns each response the most recent earlier covariate value. This can be a sensible stepwise history when changes between measurements are plausibly small or when the covariate is known to remain fixed until updated. It can also silently assert that a value observed long ago still describes the subject now.
The risk is not merely a stale number in a table. A time-varying covariate can alter a model parameter, such as clearance, and that parameter can influence drug concentrations over an interval. A long gap, an abrupt clinical event, or a value accidentally inherited from another subject can therefore change a prediction substantially. A timeline plot makes the implied history visible; a prediction comparison quantifies its consequence.
The distinction matters: a measured value is an observation at a particular time, while the covariate history used by the model is an assumption over time. A plot should therefore show both the measurement times and the intervals over which each value is assigned. Mark dose and response times on the same individual timeline when they help explain which covariate history contributes to a prediction.
A Simple Model for Quantifying the Consequence
Consider a deliberately simple one-compartment example. After an intravenous bolus dose \(D\), assume volume \(V\) is fixed and there is no further input. Let clearance depend on a time-varying covariate \(x_i(t)\) through a log-linear relation. The concentration then follows the differential equation below. This setup isolates the consequence of the covariate history; it is not a claim that every NLME model has this structure.
Here \(CL_0\) is clearance at \(x=0\), \(\beta\) is the covariate effect per unit of \(x\), and \(C_i(t)\) is concentration. Assume \(CL_0>0\), \(V>0\), and a specified covariate path over the interval. Separating variables and integrating gives a prediction that depends on the accumulated clearance, not just on the covariate value at the endpoint.
Exponentiating and substituting the bolus initial concentration yields \(C_i(t)=(D/V)\exp\{-V^{-1}\int_0^t CL_i(u)\,du\}\). Thus, if LOCF assigns a stale covariate \(x_{\mathrm{stale}}(u)\) instead of the history \(x_{\mathrm{ref}}(u)\) used for comparison, the ratio of predictions is
This result makes the direction of the error interpretable. If \(\beta>0\) and the carried value is higher than the comparison value throughout an interval, the carried value implies higher clearance there and hence a lower predicted concentration at the end of that interval. The size of the difference depends on the duration as well as on the covariate contrast. A small discrepancy sustained for a long time can matter; a large discrepancy over a very short interval may matter less.
Worked Comparisons on Individual Timelines
Worked Example: A Long Gap Before the Next Measurement
Suppose a hypothetical subject receives a \(100\)-mg intravenous bolus at time \(0\). Let \(V=20\) L, \(CL_0=2\) L/h, and \(\beta=0.20\) per covariate unit. The covariate is \(x=0\) at time \(0\), but for this example assume it actually changes to \(x=-2\) at hour 6 and remains there. It is not measured again until hour 12. Compare carrying \(x=0\) through hour 12 with the stipulated history that changes at hour 6.
A timeline would show a measurement at hour 0, a long gap to hour 12, and the hypothesized change at hour 6 marked as an unobserved event. Under LOCF, \(CL=2e^{0}=2\) L/h for all 12 hours, so the accumulated clearance is \(2(12)=24\) L. Under the stipulated changing history, clearance is \(2\) L/h for 6 hours and \(2e^{-0.4}=1.3406\) L/h for the next 6 hours. Its accumulated clearance is \(2(6)+1.3406(6)=20.04\) L.
The corresponding predictions at hour 12 are
and
The ratio is \(\exp\{-(24-20.04)/20\}\approx0.8204\). Under these stated assumptions, the stale history predicts a concentration about 17.95% lower than the comparison prediction. The later measurement alone does not prove that the covariate changed at hour 6; that trajectory was stipulated to quantify what a missed change could do.
Worked Example: An Abrupt Change With a Known Time
Now suppose the same hypothetical model parameters apply, but a clinical event at hour 4 is known to coincide with a change from \(x=0\) to \(x=-2\). A concentration is predicted at hour 8 after the bolus. Compare carrying the baseline value through hour 8 with updating the covariate at hour 4. The assumed stepwise values and change time are explicit, so the comparison does not rely on guessing when the change occurred.
With the stale value, accumulated clearance is \(2(8)=16\) L, giving \(C_{\mathrm{LOCF}}(8)=5e^{-16/20}=5e^{-0.8}\approx2.247\) mg/L. With the change at hour 4, it is \(2(4)+2e^{-0.4}(4)=8+5.363=13.36\) L. Therefore
The ratio of stale to updated predictions is \(\exp\{-(16-13.36)/20\}\approx0.8763\). The stale value therefore gives a prediction about 12.37% lower than the updated-history prediction. The calculation illustrates why an abrupt change deserves a visible breakpoint on the timeline: applying the pre-event value after that breakpoint encodes the wrong exposure to clearance in the model.
Worked Example: A Value Crossing a Subject Boundary
A hypothetical data-processing error carries the final covariate value from subject A into subject B. At B’s baseline, B’s own covariate is \(x=-1\), but the mistakenly inherited value is \(x=3\). Let \(CL_0=2\) L/h, \(\beta=0.20\), \(V=20\) L, and consider a \(100\)-mg bolus prediction 6 hours after dosing. Assume each subject’s covariate is constant over those 6 hours.
For B’s own covariate, \(CL=2e^{-0.2}=1.637\) L/h, so
For the inherited value, \(CL=2e^{0.6}=3.644\) L/h and
The erroneous-to-correct ratio is \(\exp\{-(3.644-1.637)(6)/20\}\approx0.5477\). The inherited value produces a prediction about 45.2% lower than the prediction using B’s own covariate. A plot with separate subject panels would make the error apparent: B’s timeline must begin with B’s measurements, not with the last row belonging to A.
How to Audit a Carried-Forward History
The aim of an audit is to find unsupported assignments before interpreting model predictions. First verify which subject each measurement belongs to and its time. Then display the measurement points and the assigned step history for each subject, preferably with dose and response times on the same time axis. Finally, compare predictions under the implemented history with predictions under a clearly specified alternative history. The alternative is a diagnostic comparison, not automatically the truth.
Confirm that every carried value comes from the same subject and from a measurement at or before the prediction time. Check sorting and that a new subject starts a new history.
Show actual measurement times as points and LOCF intervals as steps. Mark long gaps and known clinical changes; do not draw an unobserved change as if it were measured.
Keep dose, model parameters, and prediction time fixed while changing only the covariate history. Report the absolute difference or ratio with units and state what each history assumes.
For unexpected jumps, implausible covariate values, or prediction discontinuities, inspect subject identifiers, timestamps, missing-value handling, and the rule that creates the step history.
The main warning signs are a value applied across a gap much longer than the usual measurement spacing, a step that ignores a known clinical change, and a subject’s first assigned value matching the prior subject’s last value. Other clues include covariates changing at data-row boundaries rather than clinically meaningful times, or predictions shifting sharply when the covariate is updated despite no plausible intervening change. These clues identify records and assumptions to review; they do not by themselves establish the correct unobserved history.
Alternatives, Limitations, and Common Mistakes
LOCF is not the only representation of a covariate between measurements. A step function is appropriate when the covariate is expected to remain stable until an update or when that is the intended modeling assumption. Linear interpolation can be considered for a continuously changing quantity when intermediate values are plausibly represented by a straight line, but it uses future measurements to define values between observations and may be inappropriate for abrupt changes. A model for the covariate process or an explicit scenario comparison may be more suitable when the unobserved trajectory is important and cannot be inferred from the sparse measurements alone.
| Approach | What it assumes | Important caution |
|---|---|---|
| Carry last value forward | Covariate stays at the most recently observed level until updated. | A long interval does not make that assumption more credible; clinical changes may invalidate it. |
| Interpolate between measurements | Values change smoothly between observed endpoints, often linearly. | May smooth over abrupt events and can use a later measurement to assign earlier values. |
| Represent a known change explicitly | A documented event or measurement defines a new level at a specified time. | Event timing and the interpretation of the measurement must be correct. |
| Compare plausible histories | The intervening path is uncertain and more than one scenario deserves evaluation. | Predictions describe the scenarios; the comparison does not identify which history occurred. |
A common mistake is to say, “The last recorded value is the best available value, so it must be used until the next measurement.” That reasoning confuses availability with validity. The correction is to describe LOCF as an explicit assumption, inspect the gap and any known events, and quantify whether a plausible alternative changes predictions materially. Another mistake is to let the data-processing order define the history. Covariate carry-forward must be grouped by subject and ordered by time; a value from one subject must never be inherited by another.
Even a careful timeline cannot recover a covariate trajectory that was never measured. Prediction comparisons are conditional on their specified histories and model. They do not account automatically for uncertainty in the covariate process, measurement error, or other model misspecification. In particular, a small difference between two selected histories does not prove that either history is correct. The purpose of the audit is to expose consequential assumptions and prevent obvious assignment errors, not to manufacture certainty about missing information.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- What does LOCF assert about a covariate between two measurement times?
- In the clearance model, if \(\beta>0\) and a stale value is higher throughout an interval than the comparison value, what happens to the predicted concentration at the interval’s end?
- Why can a timeline plot help detect a subject-boundary error that may be hard to notice in a sorted data table?
- What does a prediction comparison under two covariate histories establish, and what does it not establish?
- Name one situation in which interpolation might be less appropriate than an explicitly updated step history.
Solutions. (1) LOCF asserts that the covariate remains equal to its latest observed value until another value is assigned. (2) Higher \(x\) implies higher clearance when \(\beta>0\); accumulated clearance is larger, so the predicted concentration is lower. (3) A subject-specific timeline exposes values that continue from another subject’s endpoint instead of beginning from the new subject’s own measurements. (4) It quantifies how predictions differ under the stated histories, but does not determine which unobserved history actually occurred. (5) Interpolation can be inappropriate when a known abrupt clinical change occurs between measurements, because a smooth transition would conceal the documented step.