Why Diagnose a Concentration-QTc Model?
A concentration-QTc model summarizes how QTc changes with drug concentration while accounting for features such as baseline, time, covariates, and repeated measurements within a participant. A plausible estimated slope is not enough to establish that the model describes the data adequately. A misspecified concentration function, an unmodeled time pattern, one unusual participant, or residual dependence on heart rate can all affect interpretation. Diagnostics help identify these possibilities; they do not by themselves establish a cause or select a replacement model.
The diagnostics below assume the data have been assembled correctly: ECG observations are paired with the intended concentration measurements and time points, QTc is calculated consistently, and the response variable is defined as intended. A plot cannot repair a mismatch in those inputs. The model and residual definitions also matter, especially because repeated ECGs from one participant are not independent observations.
Conditional residuals ask whether the model describes an individual's observations after accounting for the estimated participant effect. Marginal residuals retain between-participant differences and are useful for checking the population-level description, but may look systematically different when genuine participant heterogeneity is present. In mixed-effects models, residual variance and leverage can depend on the model and the observation. A raw residual plot is therefore a diagnostic view, not a test with a universal reference distribution. Where standardized or simulated residuals are used, their calculation should reflect the fitted model's variance and within-participant structure.
Residuals Against Concentration and Time
Plot residuals against concentration to look for a pattern that the fitted concentration-effect function has not captured. A satisfactory mean structure should not leave a consistent curve, trend, or step pattern across the observed concentration range. Also inspect the spread: residual variability that grows or shrinks with concentration may indicate a variance structure problem, even if the average residual is near zero. Because many observations may come from the same participant, display participant identity or use summaries that do not conceal clustering.
Plot residuals against time as well. Time can reveal a remaining circadian, placebo, adaptation, or delayed-response pattern, among other possibilities. A concentration plot alone may not distinguish these explanations when concentration and time move together. Residuals should also be considered against other variables included in the model, such as relevant covariates, and against fitted values to inspect changing spread. The aim is to locate structure, then assess whether it is reproducible and scientifically plausible.
Worked Example: Curvature in Residuals Versus Concentration
Suppose a hypothetical model predicts baseline-adjusted QTc change, in milliseconds, as \(\hat{Y}=1.0+0.8C\), where concentration \(C\) is measured in mg/L. The following eight hypothetical observations are used only to illustrate the diagnostic.
| \(C\) (mg/L) | Observed \(Y\) (ms) | Fitted \(\hat{Y}\) (ms) | Residual \(e=Y-\hat{Y}\) (ms) |
|---|---|---|---|
| 0 | 0.8 | 1.0 | -0.2 |
| 0 | 1.2 | 1.0 | 0.2 |
| 2 | 2.4 | 2.6 | -0.2 |
| 2 | 2.8 | 2.6 | 0.2 |
| 5 | 6.5 | 5.0 | 1.5 |
| 5 | 6.9 | 5.0 | 1.9 |
| 10 | 7.0 | 9.0 | -2.0 |
| 10 | 7.4 | 9.0 | -1.6 |
For example, at \(C=5\) mg/L, the fitted change is \(1.0+0.8(5)=5.0\) ms, so the first residual there is \(6.5-5.0=1.5\) ms. The mean residuals at concentrations 0, 2, 5, and 10 mg/L are 0, 0, 1.7, and -1.8 ms, respectively. The positive residuals at 5 mg/L followed by negative residuals at 10 mg/L suggest curvature or another concentration-related mismatch, rather than an isolated observation. One possibility is a saturating response, but this plot alone cannot establish that mechanism. A candidate model change should be evaluated with the design, uncertainty, and scientific rationale in mind.
Subject-Level Fits and Influence
A population summary can hide an individual pattern. Plot each participant's observed QTc response and fitted trajectory against concentration or time, using conditional fits to see how well the model describes that participant and marginal fits to see how the population structure performs. Look for participants whose observations consistently lie on one side of their fitted trajectory, whose response shape differs from the population pattern, or whose fit depends heavily on a few observations. Such patterns can indicate model inadequacy, unusual data, or real heterogeneity; they are not interchangeable explanations.
Influence asks how much an estimate changes when a participant is removed. For repeated ECG data, the participant—not a single ECG—is usually the appropriate deletion unit for a leave-one-out sensitivity check. Refit the model after omitting each participant, then compare key estimates, such as the concentration slope, with the full-data estimates. This is a direct sensitivity analysis. A large change signals dependence on that participant, but does not prove that the participant's data are erroneous. A change can also arise from a genuinely influential part of the exposure range or from weak information in the remaining data.
Worked Example: Leave-One-Participant-Out Influence
In a hypothetical analysis, the full-data concentration slope is \(0.62\) ms per mg/L, with standard error \(0.15\) ms per mg/L. Omitting participant 14 and refitting gives a slope of \(0.48\) ms per mg/L. The change is \(0.48-0.62=-0.14\) ms per mg/L. Expressed relative to the full-data standard error, the absolute change is:
This is a descriptive sensitivity measure, not a universal cutoff or a formal significance test. The slope decreases by \(0.14\) ms per mg/L when that participant is removed, so the estimate is meaningfully sensitive to this deletion and merits investigation. Check the participant's concentration range, ECG values, timing, and data provenance; then assess whether the model fits that participant and whether other participants support the same exposure range. Do not remove the participant solely to obtain a preferred estimate.
Heart-Rate Diagnostics
QT correction is intended to account for the relationship between QT interval and heart rate, but a correction may not remove that relationship equally well across all observations or conditions. Examine model residuals against heart rate, and, where useful, against change in heart rate from baseline. A systematic trend may signal residual dependence associated with the correction, a heart-rate-related data issue, or another omitted feature. It does not by itself prove that the QT correction is the cause. Check the pattern by participant and time as well, since a pooled trend can reflect differences between participants or study periods.
Worked Example: Residual Trend With Heart Rate
Consider five hypothetical residuals from a concentration-QTc model, summarized against heart-rate deviations from baseline. The deviations are \(-10,-5,0,5,10\) beats/min, and the corresponding residuals are \(2,1,0,-1,-2\) ms. Their means are both zero. The least-squares slope of residual on heart-rate deviation is:
In this deliberately simple illustration, residuals decline by 0.200 ms for each 1 beat/min increase in heart-rate deviation across these points. That is a diagnostic pattern, not a causal effect estimate or a formal test: five points do not account for repeated observations or other covariates. A real analysis should examine the individual-level pattern and the model's dependence structure, and should check whether the signal persists after accounting for relevant design features.
Why Residual Patterns Signal Misspecification
The basic justification follows directly from the residual definition. Suppose the true mean response at an observation is \(m_{\text{true}}\), while the fitted model uses \(m_{\text{fit}}\), and write the observation as \(Y=m_{\text{true}}+\epsilon\), where the error has conditional mean zero. Then:
If the model mean is correct, the expected residual at that condition is zero, subject to the fitted model and its random-effects structure. If the mean model is wrong, the difference \(m_{\text{true}}-m_{\text{fit}}\) can vary with concentration, time, or heart rate, leaving a systematic residual pattern. Random variation can also produce apparent patterns, which is why diagnostics are interpreted alongside the design, repeated structure, and uncertainty rather than as automatic proof.
Inspect residuals against concentration, time, fitted values, and heart rate, and review participant-level profiles.
Determine whether the signal is isolated to one participant or observation, or recurs across participants and conditions.
For a potentially influential participant, compare full-data and participant-deleted fits without treating deletion as a correction.
Consider the measurement, timing, functional form, covariates, and variance or correlation structure; refit only with a defensible rationale.
Common Pitfalls, Alternatives, and Limitations
A residual pattern is evidence to investigate, not a diagnosis by itself. The same pattern can have multiple explanations, and a model can look adequate in a plot while still being uncertain in an exposure range with sparse data. Residuals estimated after fitting are also constrained by the fit; in particular, conditional residuals can appear reassuring when participant-specific effects absorb systematic individual differences. Review both conditional and population-level behavior.
- Do not label every large residual an outlier. Check the measurement and its context, and distinguish an unusual observation from a systematic trend or a legitimately extreme response.
- Do not treat every influential participant as removable. Influence describes sensitivity of the estimate; exclusion requires a defensible data-quality or analysis reason.
- Do not use an arbitrary residual threshold as a verdict. Interpret scale, model structure, repeated observations, and the full pattern.
- Do not interpret a heart-rate trend as proof of a defective correction. Check time, participant, and other relevant structure before assigning an explanation.
Residual plots and subject-level fits are visual diagnostics. Leave-one-participant-out refitting is a sensitivity analysis. These complement, rather than replace, comparison of scientifically plausible model forms, uncertainty assessment, and predictive checks. A formal comparison of nested models can address a specified model change under its assumptions, but cannot make an implausible model scientifically appropriate. No diagnostic can fully resolve a lack of information, such as little overlap in concentration across time or too few observations to distinguish competing explanations.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- A residual plot shows positive residuals at low and high concentrations and negative residuals in the middle. What type of concern does this pattern raise?
- Why is omitting one ECG at a time usually less informative than omitting one participant at a time when assessing influence in repeated-ECG data?
- A QTc residual plot shows a trend with heart rate. Does this establish that the QT correction is responsible? What should be checked?
- The full-data slope is \(0.50\) ms per mg/L with standard error \(0.10\) ms per mg/L. After omitting one participant, it is \(0.44\) ms per mg/L. Calculate the absolute change in units of the full-data standard error and state what it means.
Solutions. (1) The pattern suggests a curved mean relationship or another concentration-related mismatch; it does not identify the correct alternative by itself. (2) ECGs within a participant are repeated and related, so deleting one participant assesses how the estimate depends on that participant's cluster of observations. (3) No. Check the trend by participant and time, and consider other relevant design and model features before attributing it to QT correction. (4) The absolute change is \(|0.44-0.50|=0.06\) ms per mg/L, and \(0.06/0.10=0.600\) full-data standard errors. This describes sensitivity to that deletion; it is not a universal threshold or a reason by itself to exclude the participant.