Why Developmental Residual Patterns Need Diagnosis
A pediatric clearance model can describe overall concentrations reasonably well and still miss an important change in clearance during development. The practical question is not simply whether residuals vary with age. It is whether the pattern is consistent with a rapid change in maturation, whether the age variable used by the model represents development appropriately, or whether some other feature of the model or data is inadequate.
These explanations can produce similar-looking plots. A smooth maturation function that changes too slowly may leave a run of clearance residuals with the same sign during an early transition. An age metric that aligns poorly with developmental timing may create residuals that differ between infants with the same postnatal age but different gestational ages. Other model problems, such as an omitted covariate or an unsuitable structural model, can also produce age-associated residual patterns. A plot is therefore a diagnostic clue, not a diagnosis by itself.
The earlier material on maturation functions and on gestational age and postnatal age establishes the relevant modeling context: clearance may depend on body size and a developmental maturation function, and different age measures describe different aspects of development. This tutorial focuses on interpreting residual patterns to assess those choices.
What to Examine
The superscript “obs” in a clearance diagnostic may refer to a suitably estimated or summarized clearance value, not necessarily a directly measured individual quantity. In routine population pharmacokinetic analysis, the primary diagnostics are often based on observed concentrations and model predictions. For a concentration residual \(r_i^C=C_i^{\mathrm{obs}}-C_i^{\mathrm{pred}}\), positive residuals mean the model underpredicts concentration. Under a fixed dosing and size context, this can be consistent with the model predicting clearance that is too high, but the relationship is not one-to-one: dose history, distribution, sampling time, and random effects also affect concentration.
For age-pattern diagnosis, examine both individual and grouped displays where possible. Plot residuals against the age metric used in the model and plausible alternatives, such as postnatal age and postmenstrual age. Also stratify by relevant features that may be associated with age, including body size, renal function, illness severity, and concomitant treatment, when available and scientifically justified. A mean residual near zero across the whole data set can conceal opposing age-specific errors.
Here \(g\) denotes an age stratum, \(n_g\) its number of observations, and \(s_g\) the sample standard deviation of the residuals in that stratum. The usual standard-error expression is a descriptive approximation when observations are independent and have comparable variability. Repeated observations from the same child are correlated, so treating them as independent can make uncertainty look too small. In that setting, use an analysis that accounts for within-child dependence or summarize at the child level. Regardless of method, a few extreme observations or very small strata should not be treated as conclusive evidence of a developmental transition.
How to Distinguish the Main Explanations
| Diagnostic clue | Possible interpretation | What to check next |
|---|---|---|
| Residuals have a consistent sign through a narrow developmental interval and return toward zero later. | The maturation function may be too gradual, too early, or otherwise poorly shaped for the observed transition. | Check the age metric, the number and spread of observations across the interval, and whether a plausible alternative maturation curve improves diagnostics. |
| At the same postnatal age, residuals differ systematically by gestational age or postmenstrual age. | Postnatal age alone may not represent developmental timing adequately. | Compare predictions using the alternative age metric and assess whether the pattern persists after accounting for other covariates. |
| Residual patterns remain across age metrics, occur in multiple age ranges, or differ by another clinical group. | Consider broader model misspecification, omitted covariates, data issues, or a different structural pattern. | Check data quality, dosing and sampling records, covariate balance, and model fit within relevant groups. |
These clues are not exclusive. A model can use an inappropriate age metric and also have an overly restrictive maturation function. In addition, the apparent shape of maturation can be distorted when age is strongly associated with body size, illness, or treatment. The purpose of the checks is to narrow the explanations, not to declare a cause from one residual plot.
Worked Diagnostic Examples
Worked Example: A Transition That May Be Too Gradual
Hypothetical data. Investigators compare body-size-adjusted clearance summaries with predictions from a model using postmenstrual age and a maturation function. The residual is observed minus predicted clearance, in L/h per 70 kg. For simplicity, each age-stratum mean has a standard error of 0.05 L/h per 70 kg.
| Postmenstrual age (weeks) | Observed summary | Predicted value | Residual | Standardized residual |
|---|---|---|---|---|
| 28 | 0.29 | 0.28 | 0.01 | 0.20 |
| 40 | 0.62 | 0.48 | 0.14 | 2.80 |
| 52 | 0.86 | 0.70 | 0.16 | 3.20 |
| 104 | 0.92 | 0.91 | 0.01 | 0.20 |
For the 40-week stratum, \(r^{CL}=0.62-0.48=0.14\) L/h per 70 kg, and the standardized residual is \(0.14/0.05=2.80\). For the 52-week stratum, \(r^{CL}=0.86-0.70=0.16\) L/h per 70 kg, giving \(0.16/0.05=3.20\). The model is close to the summaries at 28 and 104 weeks but predicts lower clearance than the summaries in the middle strata.
This pattern is consistent with a transition that is faster or differently timed than the fitted function represents. It does not establish that maturation is the cause: the age metric, group composition, and clearance-summary uncertainty also need review. A useful next check is whether an alternative, biologically plausible maturation shape removes the middle-age pattern without creating systematic residuals elsewhere.
Worked Example: Testing Whether Postnatal Age Is an Inadequate Metric
Hypothetical data. At two weeks postnatal age, two groups have different gestational ages at birth. Their postmenstrual ages are 30 weeks for the group born at 28 weeks and 42 weeks for the group born at 40 weeks. A model using postnatal age alone predicts size-adjusted clearance of 0.40 L/h per 70 kg for both groups. The group summaries are 0.32 and 0.57 L/h per 70 kg, respectively. Assume a standard error of 0.04 L/h per 70 kg for each summary.
With postnatal age alone, the clearance residuals are \(0.32-0.40=-0.08\) L/h per 70 kg for the earlier-gestation group and \(0.57-0.40=0.17\) L/h per 70 kg for the later-gestation group. Standardizing gives \(-0.08/0.04=-2.00\) and \(0.17/0.04=4.25\). Thus, at the same postnatal age, the model overpredicts clearance for one group and underpredicts it for the other.
Suppose a model using postmenstrual age instead predicts 0.33 and 0.56 L/h per 70 kg. The corresponding residuals are \(0.32-0.33=-0.01\) and \(0.57-0.56=0.01\) L/h per 70 kg. The residuals are much closer to zero in these hypothetical summaries. This supports investigating postmenstrual age as a more appropriate metric in this setting. It is not proof that the metric alone explains the data: the groups may differ in other ways, and the comparison should be evaluated across the full data set, not just two summaries.
Worked Example: A Pattern That Suggests Broader Misspecification
Hypothetical data. For children aged 8 to 12 months, a model predicts size-adjusted clearance of 0.80 L/h per 70 kg in each of two clinical groups. The observed group summaries are 0.72 for children receiving a particular concomitant treatment and 0.95 for those not receiving it. Each summary has a standard error of 0.05 L/h per 70 kg.
The residuals are \(0.72-0.80=-0.08\) L/h per 70 kg for the treated group and \(0.95-0.80=0.15\) L/h per 70 kg for the untreated group. The standardized residuals are \(-0.08/0.05=-1.60\) and \(0.15/0.05=3.00\). Because both groups are in the same age range, an age-metric change alone does not directly explain their contrast.
This pattern motivates checking whether the treatment is associated with clearance, whether it is a proxy for illness or another covariate, and whether the summaries are comparable. It does not show that the treatment causes a clearance change. If the contrast persists with adequate data and appropriate adjustment, the model may need to represent a relevant group difference. If it disappears after checking dose records, sampling times, or covariate imbalance, a structural change may not be warranted.
Why the Diagnostic Logic Works
A model that adequately represents the conditional mean and relevant variability should not leave a systematic residual trend with a covariate that the model has represented correctly. In idealized terms, for a correctly specified conditional mean, the expected residual given the modeled predictors is zero:
If residuals are repeatedly positive in one age range and negative in another, the model's predictions are systematically displaced relative to the observations in those ranges. The derivation is direct: within a stratum, the average residual is the average observed value minus the average predicted value. If the model systematically predicts too low, then \(E(r_i\mid g)>0\); if it predicts too high, then \(E(r_i\mid g)<0\). A trend in \(\bar r_g\) therefore flags a mismatch in predictions across strata.
This is a diagnostic implication, not a proof of which model component is wrong. Residuals can also be affected by measurement error, sparse sampling, shrinkage in individual estimates, model-dependent uncertainty, and selection into age strata. The inference is strongest when the pattern is reproducible, supported by enough observations, visible in relevant concentration diagnostics as well as clearance summaries, and consistent with plausible pharmacology.
A Practical Diagnostic Sequence
Check age definitions, units, dose histories, sampling times, body-size inputs, and whether the clearance summaries or residuals are comparable across strata. Confirm that a small number of observations are not driving the apparent trend.
Inspect residuals against the age variable used in the maturation function and against biologically plausible alternatives. Look for direction, location, and persistence of the pattern rather than relying on a single summary statistic.
Compare children of similar age who differ in gestational age, body size, renal function, illness, or treatment when data support those comparisons. If age groups lack overlap in these characteristics, acknowledge that the effects cannot be cleanly separated.
Assess whether a reasonable alternative age metric or maturation shape resolves the trend, and whether it worsens other diagnostics. If the pattern remains or tracks another factor, investigate broader structural, covariate, or data problems.
Where feasible, assess predictions in held-out subjects or a separate relevant sample. A more flexible function can improve the fitted residual plot while still predicting poorly outside the observed ages.
Comparisons, Limitations, and Common Pitfalls
A residual plot is quick and useful for locating a trend, but it cannot by itself distinguish a wrong age metric from an incorrect maturation curve. Age-stratified summaries make the size and direction of discrepancies concrete, but binning discards information and can hide within-band patterns. Refitting with alternative age metrics or maturation shapes tests whether a specified change improves the model, but improvement in fit does not establish a biological mechanism. Clinical covariate stratification may expose a missing pattern, yet sparse overlap between groups limits what can be separated.
Diagnostics are also limited by the data design. If nearly all younger children were born at an earlier gestational age and nearly all older children at term, effects of age at birth and postnatal development may be difficult to disentangle. Few samples per child can leave individual clearance estimates uncertain. Repeated measurements from the same children require attention to dependence. A flexible age function may appear to capture a rapid transition simply because it is fitting noise in a poorly sampled interval.
- Mistake: “Residuals rise with age, so maturation must be too rapid.” A residual trend only shows that predictions differ systematically from observations across the plotted data. Check the sign and scale, alternative age metrics, covariates, and data quality before attributing it to maturation.
- Mistake: “The whole-data mean residual is near zero, so the model is adequate.” Positive residuals in one age band can cancel negative residuals in another. Inspect conditional and stratified patterns.
- Mistake: “A better-fitting alternative proves the biological explanation.” Better fit supports the alternative as a statistical description of these data; it does not alone establish causality or guarantee reliable predictions in unobserved groups.
- Mistake: “Postmenstrual age and postnatal age are interchangeable.” They encode different information. Compare their implications in the population and developmental question at hand rather than substituting one without checking.
Check Your Understanding
Answer each question, then compare with the concise solutions below.
- A clearance residual is defined as observed minus predicted. In one age band, the mean residual is \(-0.12\) L/h and its standard error is \(0.04\) L/h. What is the standardized residual, and what does its sign indicate?
- At the same postnatal age, residuals are mostly negative for children born earlier and positive for children born at term. Name one age metric to investigate and one reason not to treat this pattern as proof.
- Residuals are positive in one middle-age band and negative in the next, with a whole-data mean near zero. Why is the whole-data mean insufficient?
- A model with a more flexible maturation function removes an age trend in the fitting data but performs poorly in a separate sample. What limitation does this illustrate?
Solutions. (1) The standardized residual is \(-0.12/0.04=-3.00\). Its negative sign means the observed clearance summary is below the prediction by 0.12 L/h; the standardized value expresses that discrepancy as three standard errors under the stated approximation. (2) Investigate postmenstrual age, which incorporates gestational age as well as postnatal age. The groups may also differ in other covariates or data quality, so the pattern alone does not identify the cause. (3) Opposing age-specific residuals can cancel when averaged, concealing systematic errors within the bands. (4) This illustrates overfitting or poor generalization: improved fit in the data used to adjust the model does not ensure reliable predictions in new data.