Concentration Assignment and PK Uncertainty in QTc Models

Why concentration assignment matters

A concentration–QTc model estimates how QTc changes with exposure. That slope depends not only on the QTc observations and the model form, but also on which concentration is assigned to each ECG. A measured concentration taken at the ECG time, an individual prediction from a subject-specific PK fit, and a population prediction based on typical PK parameters are different inputs. They can produce different slopes even when the QTc records are unchanged.

Let \(Y_{ij}\) denote the baseline-adjusted, time-matched QTc change for subject \(i\) at occasion \(j\), in milliseconds. Let \(C_{ij}\) denote the underlying concentration at the ECG time, in mg/L. For the comparisons here, suppose other relevant terms, such as the intercept and previously specified time or placebo effects, have been handled, and the intended exposure–response relationship is linear over the observed range:

Here \(\beta\) is the QTc change associated with a 1 mg/L increase in concentration, in ms per mg/L. This equation treats the concentration as the exposure relevant at the ECG time; it does not model a delayed response. The residual \(\varepsilon_{ij}\) contains variation not explained by the stated terms. Concentration assignment is a separate issue from the choice of repeated-measures model: a mixed-effects model does not, by itself, remove bias caused by a poor exposure input.

What each concentration input represents

A measured concentration is closest to a direct observation of exposure at a particular time, but it is not error-free. Assay error, sample handling, and imperfect alignment between blood draw and ECG time can all matter. If samples are sparse, measurements may also fail to represent concentration at unsampled ECGs without a model-based interpolation or prediction. A measured value should therefore be associated with its actual collection time, not merely with a nominal visit label.

An individual-predicted concentration can assign an exposure to every ECG time, including times without a blood sample. It can reflect a subject’s estimated clearance or other PK parameters. However, those parameters are estimated from limited PK information. Their uncertainty, and the way individual estimates are pulled toward the population distribution when information is sparse, can affect predicted concentrations. Treating the resulting predictions as known exact values does not carry that PK uncertainty into the QTc slope’s uncertainty.

A population-predicted concentration can be useful when individual PK data are unavailable, and it can provide predictions at ECG times using the population model and known covariates. But it does not generally capture each subject’s individual departures from typical PK. If it compresses the exposure range or fails to preserve the ordering of subjects’ concentrations, a slope estimated against it need not equal the slope against individual exposure. This is not a universal direction of bias: the result depends on how prediction errors relate to true concentration and to the QTc outcome.

When classical measurement error attenuates the slope

A useful benchmark is classical additive exposure error. Let the assigned input be \(C^*=C+u\), where \(C\) is the true concentration and \(u\) is a mean-zero assignment error. For a simple regression with an intercept, the population slope of \(Y\) on \(C^*\) is \(\operatorname{Cov}(C^*,Y)/\operatorname{Var}(C^*)\). To obtain the familiar attenuation result, several conditions are needed: \(\operatorname{Var}(C)>0\); \(\operatorname{Cov}(C,\varepsilon)=0\); \(\operatorname{Cov}(u,C)=0\); and \(\operatorname{Cov}(u,\varepsilon)=0\). These conditions hold, for example, if \(E[\varepsilon\mid C]=0\) and \(u\) is independent of \((C,\varepsilon)\). Independence between \(u\) and \((C,\varepsilon)\) alone is not enough; the condition on \(C\) and \(\varepsilon\) is also required.

Under these assumptions, substituting \(Y=\alpha+\beta C+\varepsilon\) gives the following derivation:

The ratio \(\lambda=\operatorname{Var}(C)/[\operatorname{Var}(C)+\operatorname{Var}(u)]\) is the reliability ratio under these assumptions. Because \(\operatorname{Var}(C)>0\), it satisfies \(0<\lambda\leq 1\); it equals 1 when \(\operatorname{Var}(u)=0\). Thus the expected fitted slope is closer to zero than \(\beta\) when there is nonzero classical error. This result is a benchmark, not a general formula for PK-predicted concentrations: estimation errors may be structured, and the required zero-covariance conditions may fail.

PK prediction error and exposure misalignment

An individual-predicted concentration is a generated input: it depends on estimated PK parameters, the PK model, covariates, and the timing information used to calculate the prediction. If a prediction error is larger at some concentrations than others, or is associated with subject-specific features that also relate to QTc, it is not classical additive error. It can attenuate, inflate, or otherwise distort the slope. For example, shrinkage of individual PK parameter estimates can make predictions less variable across subjects; model misspecification can create systematic differences by dose or time. Neither situation is summarized by a single reliability ratio without additional assumptions.

Timing matters as well. A blood sample taken before or after an ECG is not automatically the concentration at that ECG. During changing concentrations, an offset in collection time can assign the wrong exposure, and a record-linkage error can pair an ECG with a concentration from another time or subject. These errors can scramble the exposure–outcome pairing. The slope may then be pulled toward zero, increased, or even have the wrong sign; the direction depends on the pattern of mismatch.

Practical checks and analysis choices

Before interpreting a slope, establish how every concentration was obtained and matched to an ECG. A useful sequence is to check identifiers and timestamps, examine paired measured and predicted concentrations where both are available, and assess whether discrepancies vary with dose, time, or subject. Then compare slope estimates under defensible concentration assignments. Large changes are evidence that the conclusion depends on the exposure input, not proof that one assignment is correct.

• Check timing and linkage. Confirm subject identifiers, actual sample and ECG times, units, and the rule used to associate a concentration with an ECG. Investigate duplicate, missing, or implausibly distant matches.
• Check prediction behavior. Compare individual predictions with observed PK concentrations where available. Look for systematic discrepancies and an unexpectedly narrow range of predicted exposure. Consider whether sparse PK information makes individual predictions imprecise.
• Compare estimates transparently. When appropriate, report slopes from measured, individual-predicted, and population-predicted concentrations as sensitivity analyses, stating the input and units each time. Differences can reflect distinct exposure information and should not be treated as interchangeable estimates.
• Carry PK uncertainty forward. A two-stage analysis that inserts one set of predicted concentrations as fixed values can understate slope uncertainty. Joint PK–QTc estimation, or a procedure that propagates PK parameter uncertainty through repeated PK predictions and QTc fits, can better represent that uncertainty. The suitability of a method depends on the model and data.

These checks do not establish that the linear concentration–QTc model is otherwise adequate. They address whether the concentration assigned to each ECG is credible and how sensitive the estimated slope is to that assignment. A slope from a population prediction may answer a different practical question from one based on individual exposure; it should not be labelled an individual-exposure slope without justification.

Common mistake and limitations

No single concentration source is best in every dataset. Measured values may be sparse or imperfectly time-matched; individual predictions rely on subject-specific PK information that may be limited; population predictions may omit individual PK differences. Comparisons are most informative when their assumptions, timing rules, and uncertainty are made explicit. In particular, a plug-in slope estimate should not be interpreted as if its exposure values were observed without error.