Why Evaluate a Maturation Curve Against Observed Data?
A fitted maturation curve can look smooth and biologically plausible while being weakly supported by the observations at some ages. This matters in population pharmacokinetics because a model may use the curve to predict clearance or another parameter for patients whose ages are poorly represented in the data. A useful evaluation therefore asks two separate questions: does the curve agree with the observed age-stratified parameter information, and where are there enough observations to support that agreement?
The comparison must be made on a consistent scale. For example, if age-stratified summaries describe a size-standardized clearance relative to its adult value, the curve predictions must describe that same normalized quantity—not raw clearance values that still vary with body size. The age definition must also match the model: postnatal age and postmenstrual age are not interchangeable. The earlier material on age and weight trends in individual parameters provides the background for reading such summaries.
What a Curve Comparison Can and Cannot Show
A common maturation function is the sigmoid Emax form. For a normalized maturation fraction \(M(a)\), one version is
Here \(a\) is age in a specified unit, \(h>0\) is the Hill coefficient, \(TM_{50}>0\) is the age at which \(M(a)=0.5\), and \(\theta_0\) and \(\theta_{\mathrm{adult}}\) are the lower-age and adult parameter values. If the observed summaries are already normalized maturation fractions, compare them directly with \(M(a)\). If they are parameter values, compare them with \(\theta(a)\). The model assumes a particular smooth, monotone transition; agreement with data does not establish that this functional form is uniquely correct.
A plot is often the first diagnostic: show the age-stratified estimates, an uncertainty display such as standard errors or confidence intervals, the number of observations in each age stratum, and the fitted curve on the same axes. Numerical discrepancies can help identify ages where the fit deserves attention, but they do not automatically constitute formal hypothesis tests.
Worked Example: Checking Curve Predictions at Observed Ages
Suppose a hypothetical analysis summarizes a size-standardized clearance maturation fraction at several ages. The candidate curve is \(M(a)=a^2/(a^2+6^2)\), with age in months. The table gives the observed stratum estimate \(y_i\) and its standard error \(SE_i\). For a descriptive comparison, calculate \(r_i=(y_i-M(a_i))/SE_i\).
| Age (months) | Observed estimate \(y_i\) | \(SE_i\) | Curve prediction | Standardized discrepancy \(r_i\) |
|---|---|---|---|---|
| 1 | 0.040 | 0.030 | 0.0270 | 0.432 |
| 3 | 0.220 | 0.030 | 0.200 | 0.667 |
| 6 | 0.500 | 0.040 | 0.500 | 0.000 |
| 12 | 0.820 | 0.050 | 0.800 | 0.400 |
| 24 | 0.900 | 0.060 | 0.941 | -0.686 |
| 36 | 0.970 | 0.050 | 0.973 | -0.0595 |
At 1 month, \(M(1)=1/(1+36)=1/37=0.027027\). Thus \(r_1=(0.040-1/37)/0.030=0.4324\), or about 0.432 standard errors. At 3 months, \(M(3)=9/(9+36)=0.200\), so \(r_3=(0.220-0.200)/0.030=0.667\). At 24 months, \(M(24)=576/(576+36)=576/612=0.941176\), and \(r_{24}=(0.900-576/612)/0.060=-0.6863\). These calculations use the unrounded predictions before rounding the results.
The discrepancies in this hypothetical table are all within one reported stratum standard error of the curve prediction. That is a useful descriptive summary, not proof that the curve is correct. It also does not show that the data are equally informative at all ages: the table must be read alongside the sample counts and the way the stratum estimates were obtained.
Interpreting Discrepancies and Their Uncertainty
For an observed estimate \(Y_i\) and curve prediction \(f(a_i)\), the descriptive standardized discrepancy is
It expresses the difference in units of the standard error of the observed estimate. The sign indicates direction: positive values mean the observed estimate is above the prediction, and negative values mean it is below. This scaling is informative only when \(SE_i\) is meaningful for the quantity being compared. In particular, it does not account for uncertainty in the fitted curve parameters.
The reason is seen from the variance of a difference. If the observed estimate and fitted prediction are both uncertain, then
Dividing by \(SE_i\) alone is equivalent to treating the prediction as fixed and using only the observed estimate’s uncertainty. In a fitted model, the prediction is usually estimated from the same data, so its uncertainty and its covariance with the observed summary may matter. Moreover, empirical Bayes individual parameter estimates can be shrunk toward the population distribution, and estimates from the same fitted model need not behave like independent direct measurements. Unless the full uncertainty structure is handled, describe \(r_i\) as a descriptive standardized discrepancy, not as a z statistic with an automatic p-value or cutoff.
Sparse Ages and Unsupported Predictions
The curve should be judged against the distribution of information, not merely the minimum and maximum ages in the data. A broad age range can still contain a large gap, and a plotted point based on very few subjects can be imprecise. Counts by age stratum, the spread of individual ages, and the uncertainty of summaries help distinguish dense coverage from isolated observations. Predictions within a gap may be interpolation in the overall age range but remain weakly supported locally; predictions beyond the observed range are extrapolations.
Worked Example: Finding a Weakly Supported Age Range
Consider a hypothetical data set with observations at ages 2, 3, and 4 months, and then at 18 and 24 months. The corresponding counts are 2, 3, 2, 4, and 3 subjects. There are no observations from 5 through 17 months. Under \(M(a)=a^2/(a^2+36)\), the predictions at 2, 3, and 4 months are \(4/40=0.100\), \(9/45=0.200\), and \(16/52=0.308\). At 18 and 24 months they are \(324/360=0.900\) and \(576/612=0.941\).
The fitted curve is anchored by few observations at either end, and there is a long unobserved interval between 4 and 18 months. At 10 months the curve predicts \(100/(100+36)=0.735\), but that value is not directly checked by any age-stratified observation. It is a model-based prediction across a data gap. A graph that only shows the full age range and a smooth line could conceal this lack of local support; displaying counts or the observed age distribution makes the limitation visible.
The appropriate conclusion is not that the prediction at 10 months is wrong. Rather, the available observations provide little direct evidence about the curve’s transition through that interval. A curve can pass close to all five observed summaries and still have its intermediate shape largely determined by the chosen model form.
Shape Checks and Alternative Descriptions
For the conventional sigmoid function above, positive \(h\) and \(TM_{50}\) imply a nondecreasing maturation fraction between 0 and 1. It approaches 0 as age approaches 0 and approaches 1 as age becomes large; it reaches 0.5 at \(a=TM_{50}\). These constraints provide simple plausibility checks. A fitted parameter curve that turns downward when the model is supposed to represent monotone maturation, leaves the biologically or mathematically permitted range, or places its transition where the observations show no support merits investigation. Whether a pattern is biologically plausible depends on the parameter and context, not on smoothness alone.
Worked Example: Comparing Two Candidate Transition Shapes
Suppose hypothetical normalized estimates are 0.040 at 1 month and 0.900 at 24 months. Compare two candidate curves with \(TM_{50}=6\) months: one has \(h=1\), and the other has \(h=2\). At 1 month, the \(h=1\) curve predicts \(1/(1+6)=0.1429\); the \(h=2\) curve predicts \(1/(1+36)=0.02703\). At 24 months, the predictions are \(24/(24+6)=0.800\) and \(576/(576+36)=0.9412\), respectively.
If the hypothetical standard errors are 0.030 at 1 month and 0.060 at 24 months, the \(h=1\) descriptive discrepancies are \((0.040-0.1429)/0.030=-3.429\) and \((0.900-0.800)/0.060=1.667\). For \(h=2\), they are \((0.040-1/37)/0.030=0.4324\) and \((0.900-576/612)/0.060=-0.6863\). In these two locations, the \(h=2\) curve lies closer to the summaries in standard-error units.
This small comparison does not establish that \(h=2\) is the correct model: only two ages have been considered, and the curve’s uncertainty and the other data have not been included. It illustrates how observed ages can challenge a candidate shape and why a visually smooth curve should not be selected by appearance alone.
Alternatives answer somewhat different questions. Age-stratified summaries show the data pattern with relatively few structural assumptions, but can be noisy or unstable in small strata. A smooth curve borrows strength across ages and gives predictions at ages without observations, but those predictions depend on the assumed functional form. A more flexible smoother can reveal departures from a rigid sigmoid, but may be unstable where data are sparse and may behave poorly at the boundaries. These approaches are complementary checks; none removes the need to show where observations occur.
Common Pitfalls and Practical Diagnostics
- Rounding before calculating discrepancies. Subtracting a rounded prediction and then dividing can change the reported standardized value. Keep full precision through the calculation and round only the final result.
- Treating every age point as equally informative. A point based on a few subjects is not as strong a check as a precise summary from many subjects. Show counts and uncertainty.
- Calling agreement validation. Agreement at observed ages is evidence of compatibility with those summaries, not proof that the curve will predict well in a different population or in an unobserved age range.
- Over-interpreting a standardized discrepancy. A value such as 1.5 is not automatically a formal test result. The fitted prediction, sampling process, shrinkage, and dependence among estimates affect its interpretation.
- Ignoring the parameter scale or age definition. Comparing raw clearance with a normalized curve, or aligning summaries by a different age measure than the model uses, can create apparent disagreement or conceal real disagreement.
A disciplined evaluation therefore checks the scale and age definition first, inspects the observed age distribution and stratum sizes, compares predictions at observed ages without premature rounding, and then asks whether the curve’s behavior between and beyond those ages is supported or primarily imposed by its form. This is a diagnostic assessment of support, not a substitute for considering uncertainty in predictions.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A normalized observed estimate is 0.35 at an age where the fitted curve predicts 0.30; the reported standard error is 0.10. Calculate and interpret the descriptive standardized discrepancy.
- There are many observations from 1 to 4 months and from 20 to 24 months, but none between 5 and 19 months. What can be said about a curve prediction at 12 months?
- Why is dividing an observed-minus-predicted difference by the observed estimate’s standard error not necessarily a formal z test?
- For a conventional sigmoid maturation fraction, what do positive \(h\) and \(TM_{50}\) imply about its range, direction, and half-maturation age?
Solutions. (1) \(r=(0.35-0.30)/0.10=0.50\). The observed estimate is 0.50 of its reported standard error above the prediction; this is descriptive, not automatically a test. (2) The 12-month value lies in a data gap. It may be a model prediction, but there is little direct local evidence for the transition shape there. (3) The fitted prediction is uncertain and may be correlated with the observed estimate; using only the estimate’s standard error omits those contributions. (4) The fraction increases from 0 toward 1 as age increases and equals 0.5 at \(a=TM_{50}\), under the stated conventional model.