Why Sparse Maturation Data Require Care
A maturation curve summarizes how an age-dependent biological activity changes across development. In a pediatric PBPK model, it may describe an enzyme activity or a clearance component after accounting for the separate effect of body size. Sparse age-stratified data make this task difficult: a curve can pass close to the few observed age groups while giving very different predictions at ages that were not observed.
The introductory material on choosing a mathematical form for enzyme maturation distinguishes candidate curve shapes; the earlier material on separating body-size scaling from ontogeny distinguishes size effects from age effects. Here the task is to fit candidate functions to observations, quantify uncertainty, and decide how much the data support. The analysis targets an age-dependent mean curve. It does not, by itself, estimate variation among children within an age group.
Age must use one consistent definition and unit throughout the fit—for example, postnatal age in years rather than a mixture of postnatal and postmenstrual age. The response scale must also be explicit. A fraction of adult activity is dimensionless; a clearance has units such as L/h. A model fitted to one scale should not be interpreted as if it were fitted to another.
Fitting and Its Assumptions
A common fitting method is weighted nonlinear least squares. Suppose the stratum-level observations have independent errors with mean zero and known standard errors \(s_i\). Then weights \(w_i=1/s_i^2\) give greater influence to more precise observations, and the fitted parameters minimize the weighted residual sum of squares:
This procedure assumes the chosen response scale is appropriate, the age and response measurements are meaningful, and observations from different strata have independent errors with the stated variances. If standard errors are not known, an unweighted or otherwise justified weighted fit may be used, but uncertainty then depends on additional assumptions about the residual variance. Strata with very different sample sizes or measurement precision should not automatically receive equal weight.
Candidate functions should reflect plausible maturation behavior and be kept as simple as the data allow. Two examples, with a fixed adult plateau of one, are:
The exponential model has one parameter and approaches one as age increases. The Hill model has two parameters: \(T_{50}\) is the age at which the predicted fraction is one-half, and \(h\) controls the steepness. Fixing the plateau at one is a substantive assumption, not a harmless mathematical convenience. If the observations are not expressed relative to a defensible adult reference, the fixed plateau may be inappropriate. Estimating a free plateau adds a parameter and requires data that can inform it.
A Practical Fitting Sequence
Specify whether the response is enzyme activity, clearance, or a normalized fraction, and use a consistent age definition and unit.
State which parameters are fixed, which are estimated, and what the asymptote means biologically.
Minimize the residual objective using known stratum-level standard errors when available; check that the optimizer has converged to a plausible solution.
Examine residuals, parameter uncertainty, the number and spread of ages, and predictions at the ages relevant to the PBPK application.
For nonlinear least squares, uncertainty is often approximated locally. Let \(J\) be the Jacobian matrix whose \(i,j\) entry is the derivative of \(f(a_i;\theta)\) with respect to parameter \(\theta_j\), evaluated at the fitted parameters. If observation variances are known, a first-order covariance approximation is \((J^\mathsf{T}WJ)^{-1}\), where \(W\) is the diagonal matrix of weights. If a common residual variance is estimated instead, this matrix is multiplied by that variance estimate. This approximation requires a locally identifiable fit and may be unreliable when data are sparse, parameters are near boundaries, or the curve is highly nonlinear over the uncertainty range.
Worked Fits and Uncertainty
Worked Example: Comparing an Exponential and a Hill Fit
Suppose a hypothetical study reports mean enzyme activity as a fraction of an adult reference at four age strata. The data are \((a_i,y_i)=(1,0.20),(3,0.55),(6,0.82),(12,0.95)\), with age in years. For this illustration, assume equal precision and fit both models by unweighted least squares, with the plateau fixed at one.
For the exponential candidate, numerical minimization of \(Q(k)=\sum_i[y_i-\{1-\exp(-ka_i)\}]^2\) gives approximately \(k=0.266\ \mathrm{year}^{-1}\). Its predictions are approximately \(0.234, 0.550, 0.797,\) and \(0.959\). The residuals (observed minus predicted) are \(-0.034, 0.000, 0.023,\) and \(-0.009\). Thus:
For the Hill candidate with \(h=1\), numerical minimization over \(T_{50}\) gives approximately \(T_{50}=2.20\) years. The fitted values are \(1/(1+2.20)=0.3125\), \(3/(3+2.20)=0.5769\), \(6/(6+2.20)=0.7317\), and \(12/(12+2.20)=0.8451\). The residuals are approximately \(-0.1125,-0.0269,0.0883,\) and \(0.1049\), giving \(Q_{\mathrm{Hill},h=1}\approx0.03219\) using the unrounded fitted values. On this unweighted sum-of-squares criterion, the fitted exponential tracks these hypothetical observations more closely.
That comparison does not establish that the exponential is biologically correct. It compares these two specified candidates for this data set and weighting scheme. Both models have residual degrees of freedom \(n-p=4-1=3\) here, because each has one estimated parameter in this comparison. The fit should still be checked against biological knowledge and the age range where predictions will be used.
Worked Example: Exact Fit Does Not Mean Zero Measurement Error
Consider three hypothetical age-stratum means generated by a Hill curve with \(T_{50}=2\) years and \(h=1.5\), at ages \(0.5,2,\) and \(8\) years. The underlying values are approximately \(0.1111,0.5000,\) and \(0.8889\). Suppose each reported mean has a known standard error of \(0.040\), and the values displayed here are rounded representations of the unrounded curve values.
A two-parameter Hill fit can recover \(T_{50}=2\) and \(h=1.5\) with zero residual sum of squares for these unrounded observations. There are \(n=3\) observations and \(p=2\) fitted parameters, so the residual degrees of freedom are \(3-2=1\), not zero. The residual-based estimate of a common error variance, \(Q/(n-p)\), is zero because the fit is exact. That zero is not evidence that the measurements have no error: the standard errors supplied for the stratum means remain \(0.040\) and are the assumed observation uncertainties.
The distinction matters for uncertainty calculations. Using \(\theta=(\log T_{50},h)\), the Hill-model derivatives are \(\partial f/\partial\log T_{50}=-h f(1-f)\) and \(\partial f/\partial h=f(1-f)\log(a/T_{50})\). At the three ages, the Jacobian rows are approximately \((-0.148,-0.137),(-0.375,0),(-0.148,0.137)\). With weights \(1/0.040^2=625\), the local covariance approximation is diagonal, with variances about \(0.00867\) for \(\log T_{50}\) and \(0.0427\) for \(h\). Thus the approximate standard errors are \(0.0931\) and \(0.207\), respectively. This uses the supplied measurement variances; it does not multiply by the zero residual-based variance estimate.
For illustration, the fitted activity fraction at age one year is \(f(1)=0.261\). The delta-method gradient there is approximately \((-0.289,-0.134)\); using the parameter covariance gives a prediction standard error of about \(0.0386\). A rough normal 95% interval for the mean curve value is \(0.2612\pm1.96(0.0386)\), or approximately \(0.1855\) to \(0.3369\). This is a local approximation conditional on the model and known standard errors, not a guarantee that the chosen curve is correct.
Age Coverage, Diagnostics, and Extrapolation
Worked Example: Why Similar Fits May Extrapolate Differently
Suppose hypothetical observations are concentrated near age three years, with little or no information at younger ages or near adulthood. A Hill curve with \(h=1\) and \(T_{50}=2\) predicts \(f(3)=3/(3+2)=0.600\). A second Hill curve with \(h=3\) and \(T_{50}=3/(1.5)^{1/3}\approx2.621\) also predicts \(f(3)=0.600\), since \((3/T_{50})^3=1.5\).
The curves agree at the observed age of three years but differ away from it. At age one, the first predicts \(1/(1+2)=0.333\), while the second predicts \(1/[1+(2.621)^3]\approx0.053\). At age eight, they predict \(8/(8+2)=0.800\) and \(8^3/(8^3+2.621^3)\approx0.966\), respectively. A cluster of observations near age three may constrain the response there while leaving the shape and the predictions at the extremes weakly determined. A visually smooth curve does not resolve that lack of information.
Inspect residuals against age and fitted values. A systematic run of positive residuals at young ages and negative residuals at older ages, or the reverse, can indicate a mismatched shape or an inappropriate plateau. A single large residual may reflect data quality, a stratum with a different composition, or model inadequacy; it should be investigated rather than automatically deleted. Also inspect fitted parameters and their uncertainty: very large or boundary estimates, strong parameter correlation, or failure to converge can signal weak identification.
The number of age strata alone is not enough. A curve with two free parameters cannot be credibly characterized by two age points: it may interpolate them exactly while leaving no information to assess residual error. More points do not solve the problem if they all lie in a narrow age range. Age coverage should span the portion of maturation that drives the intended PBPK predictions. Predictions beyond the observed range are extrapolations, even when the function has a biologically plausible asymptote.
Comparing Approaches and Knowing Their Limits
| Approach | Useful feature | Main caution |
|---|---|---|
| Weighted nonlinear least squares on the original response scale | Fits the stated mean function directly and can use stratum-specific standard errors. | Requires a defensible error model; sparse data can leave parameters weakly identified. |
| Transformation, such as fitting a logit of a fraction | May turn some curve forms into a simpler regression problem. | Changes the error scale and weighting; transformed least squares is not generally the same fit as least squares on the original scale. |
| More flexible curve with extra parameters | Can represent shapes that a simple function misses when data are rich enough. | Can overfit sparse strata, inflate uncertainty, and produce unstable extrapolation. |
Information criteria can supplement, but not replace, scientific judgment. AIC or AICc comparisons require likelihoods calculated under a stated error distribution and comparable response data. With very few age strata, an information criterion may be unstable or unable to distinguish candidates meaningfully. Do not select a function solely because its residual sum of squares is smaller when it also has more parameters; assess the added flexibility, uncertainty, residual pattern, and intended prediction range.
A common mistake is to say, “The curve passes through every observed age group, so it is well validated and its predictions at unobserved ages are reliable.” The correction is that interpolation fit is only one part of evaluation. With sparse data, a flexible curve may fit exactly while its parameter uncertainty or extrapolated predictions remain large. Report the observed age range, the model assumptions, uncertainty, and any age range outside which predictions are unsupported.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A Hill curve has \(T_{50}=4\) years and \(h=1\). What fraction does it predict at age four years?
- A fit uses \(n=5\) age strata and estimates \(p=2\) parameters. What are the residual degrees of freedom?
- Three observations are fitted exactly by a two-parameter curve. Does this establish zero measurement error? Explain.
- Why can observations clustered around one age provide weak support for predictions at much younger ages?
- Why is least squares after a logit transformation not generally equivalent to least squares on the original response scale?
Solutions. (1) At \(a=T_{50}\), \(f=a^h/(a^h+T_{50}^h)=1/2\), so the prediction is \(0.500\). (2) \(df=n-p=5-2=3\). (3) No. The residual degrees of freedom are \(3-2=1\), and the residual-based variance estimate is zero only because the observed residual sum is zero; measurement uncertainty must come from the assumed or supplied observation-error model. (4) The data constrain the response near the sampled ages but may not identify the curve’s early-age shape; candidate curves can agree locally and diverge outside that region. (5) A transformation changes the scale on which errors are measured and squared, so it changes the objective function and implicitly changes error and weighting assumptions.