Why Different Parameter Values Can Fit Almost the Same Curve
A compartment model translates parameters such as clearance and volume into a concentration-time curve. But the data do not observe those parameters directly: they record concentrations at selected times, with measurement error. Consequently, a change in one parameter can sometimes be offset by changes in others, leaving predicted concentrations very similar over the observed time range. The fitted curve may then be well determined even though some parameter estimates are not.
This distinction matters in pharmacometrics. A model can describe the observed profiles adequately while its parameter values remain uncertain or strongly dependent on one another. The earlier material on structural and practical identifiability establishes the distinction: structural identifiability concerns whether ideal, exact observations can uniquely determine parameters; practical identifiability concerns whether the actual design and data determine them precisely. Here the focus is how parameter correlation and near-equivalent curves arise and how to interpret them.
From Microconstants to a Concentration Curve
Consider a linear two-compartment model after an intravenous bolus. Let \(V_c\) be central volume, \(CL\) systemic clearance, \(Q\) intercompartmental clearance, and \(V_p\) peripheral volume. Define \(k_{10}=CL/V_c\), \(k_{12}=Q/V_c\), and \(k_{21}=Q/V_p\). For dose \(D\), central concentration has the biexponential form \(C(t)=a e^{-\alpha t}+b e^{-\beta t}\), where \(\alpha>\beta>0\) are the fast and slow exponential rates, and \(a,b\) are their concentration amplitudes.
For this model the rates satisfy \(\alpha+\beta=k_{10}+k_{12}+k_{21}\) and \(\alpha\beta=k_{10}k_{21}\). The amplitudes also obey constraints imposed by the initial concentration and its initial decline. Immediately after a bolus, \(C(0)=D/V_c=a+b\). The initial slope is \(C'(0)=-k_{10}D/V_c=-\alpha a-\beta b\). Solving these two equations gives:
These relationships help explain why changes in microconstants can compensate for one another. A small change in a rate or amplitude may barely affect predictions at the sampled times, especially if the fast phase is already mostly over or the slow phase has not been observed for long. Similar curves over a finite range do not imply identical parameters, nor do they establish that two parameter sets would remain indistinguishable with ideal, densely sampled data.
Worked Example: Two Distinct Parameter Sets With Similar Predictions
Consider a hypothetical 25 mg IV bolus. The first parameter set has \(V_c=5.00\) L, \(CL=1.00\) L/h, \(Q=2.00\) L/h, and \(V_p=8.00\) L. Thus \(k_{10}=0.2000\), \(k_{12}=0.4000\), and \(k_{21}=0.2500\) h\(^{-1}\). The rate equation has sum \(0.8500\) h\(^{-1}\) and product \(0.05000\) h\(^{-2}\), so its roots are \(\alpha\approx0.7864\) h\(^{-1}\) and \(\beta\approx0.06358\) h\(^{-1}\).
Here \(D/V_c=5.00\) mg/L. The fast amplitude is \(a=5.00(0.2000-0.06358)/(0.7864-0.06358)\approx0.9437\) mg/L, and the slow amplitude is \(b=5.00-a\approx4.056\) mg/L. Therefore the predicted curve is approximately \(0.9437e^{-0.7864t}+4.056e^{-0.06358t}\) mg/L, with \(t\) in hours.
For comparison, take \(\alpha=0.9000\) h\(^{-1}\), \(\beta=0.06500\) h\(^{-1}\), and the same amplitudes \(a=0.9437\) mg/L and \(b=4.056\) mg/L. The implied \(k_{10}\), from the initial slope, is \((0.9437(0.9000)+4.056(0.06500))/5.00\approx0.2226\) h\(^{-1}\). Then \(k_{21}=\alpha\beta/k_{10}\approx0.2628\) h\(^{-1}\), and \(k_{12}=\alpha+\beta-k_{10}-k_{21}\approx0.4796\) h\(^{-1}\). These are positive, physically admissible microconstants. With \(V_c=5.00\) L, they correspond approximately to \(CL=1.113\) L/h, \(Q=2.398\) L/h, and \(V_p=9.13\) L.
At \(t=0.5\) h, the first set predicts \(0.9437e^{-0.7864(0.5)}+4.056e^{-0.06358(0.5)}\approx4.566\) mg/L. The second predicts \(0.9437e^{-0.9000(0.5)}+4.056e^{-0.06500(0.5)}\approx4.528\) mg/L, a difference of about \(0.038\) mg/L. At \(t=8\) h the predictions are about \(2.441\) and \(2.412\) mg/L, respectively. The sets differ appreciably in some parameters, yet these example predictions differ by only about 1% at 8 h. Whether that difference can be resolved depends on the actual assay error, design, and other observations; curve similarity at a few times is not proof of equivalence.
What a High Parameter Correlation Means
When the data constrain a combination of parameters more clearly than they constrain each parameter separately, the uncertainty in their estimates often has an elongated, tilted shape. Locally, this is summarized by a covariance matrix. For estimates \(\hat{\theta}_1\) and \(\hat{\theta}_2\), their correlation is their covariance divided by the product of their standard errors; it ranges from \(-1\) to \(1\). A value near \(1\) means that, within the estimated uncertainty, larger values of one estimate tend to accompany larger values of the other. A value near \(-1\) means they tend to move in opposite directions.
This is a statement about joint uncertainty under the fitted model and data. It is not a causal claim, does not mean the parameters must vary together biologically, and does not show that either parameter is individually well estimated. A high correlation can coexist with a precise prediction of concentration over the observed times: the data may constrain the curve while leaving several parameter combinations compatible with it.
Worked Example: Reading Correlation From a Covariance Matrix
Suppose a hypothetical fit reports standard errors of \(0.090\) for \(\log Q\) and \(0.120\) for \(\log V_p\), and their estimated covariance is \(0.01037\). Because these are log parameters, their standard errors are dimensionless. The correlation is \(\rho=0.01037/(0.090\times0.120)=0.960\) (to three significant figures).
This strong positive correlation says that parameter combinations with both log estimates higher, or both lower, are relatively compatible with the estimated uncertainty. It does not say that the individual values of \(Q\) and \(V_p\) are known precisely, and it does not establish a biological correlation between subjects’ clearances and volumes. Those are different questions requiring different interpretations and, for the latter, an appropriate random-effects model.
Checking Whether the Data Can Separate the Curves
Parameter estimates are informed by the curve only where observations are available. If the fast decline occurs before the first sample, differences in the fast rate or amplitude may have little observable effect. If late observations are absent, the data may provide little information about the slow rate. The earlier material on sampling coverage, early samples, and terminal samples addresses those time regions; the practical question here is whether the measured differences between plausible fitted curves are large relative to the data’s uncertainty.
Worked Example: Comparing a Curve Difference With Measurement Noise
Use the two predictions at 8 h in the first example: \(2.441\) and \(2.412\) mg/L. Their absolute difference is \(2.441-2.412=0.029\) mg/L. Suppose, hypothetically, the residual standard deviation around this concentration is \(0.150\) mg/L. The discrepancy is \(0.029/0.150\approx0.193\) residual standard deviations.
At this time point, the difference is small relative to the assumed residual variability, so that observation alone would not strongly separate the two predictions. This calculation is a scale comparison, not a formal test or proof that the parameter sets are equivalent. Other times, repeated observations, or a different error model could change the evidence. In particular, a design that samples during the fast decline may distinguish sets that look close at later times.
Useful checks include plotting predictions from plausible parameter combinations over the actual sampling times, examining the estimated covariance or correlation matrix, and checking whether uncertainty intervals are wide or parameter estimates change substantially under small reasonable changes in the data or model. A fit that converges is not necessarily informative about every parameter. Warning signs include very high parameter-estimate correlations, unstable estimates across starting values, large standard errors, or a long, shallow likelihood region in which parameter values change while the fit changes little.
Ways to Respond, and Their Limitations
There is no single remedy for correlated estimates. The appropriate response depends on the scientific objective and available information. A useful first step is to establish which quantities the data support, rather than treating every reported estimate as equally reliable.
| Approach | What it can clarify | Important limitation |
|---|---|---|
| Improve sampling coverage | Observations in informative time regions can reveal curve differences. | More samples do not guarantee separation if the predictions remain close relative to error. |
| Report joint uncertainty and predicted profiles | Shows parameter dependence and whether relevant concentrations are stable. | A precise fitted profile does not make every parameter precise. |
| Fix a parameter using defensible external information | Can reduce the number of poorly distinguished quantities estimated from these data. | An incorrect fixed value can distort the remaining estimates and their interpretation. |
| Use a simpler model when scientifically justified | Can avoid estimating details unsupported by the design. | Simplification changes the scientific model and must be justified, not chosen solely to obtain convenient estimates. |
The introductory material on practical identifiability motivates a further caution: adding a constraint may make estimates appear more stable without adding information to the data. A fixed value or informative prior can help estimation, but the resulting inference depends partly on that supplied information. Report the constraint and distinguish data-supported conclusions from conclusions that rely on assumptions or external knowledge.
The practical importance of this distinction depends on the goal. If the task is to predict concentrations within the observed range, similar profiles may support similar predictions despite uncertain parameters. If the goal is to interpret clearance, distribution, or between-subject variability, correlated and weakly determined estimates demand caution. Predictions outside the observed range may also diverge, so similarity in the sampled window should not automatically be extrapolated.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- What does a high correlation between two parameter estimates describe, and what does it not establish?
- In a two-compartment IV bolus model, which quantities determine the initial concentration: the exponential rates or the sum of their amplitudes?
- Two curves differ by \(0.06\) mg/L at a sampled time, and the residual standard deviation there is \(0.12\) mg/L. How large is the difference in residual standard deviations?
- Why can a model give stable concentration predictions but uncertain clearance and volume estimates?
- Name one way to improve separation of parameter combinations and one limitation of that approach.
Solutions: (1) It describes joint uncertainty in the estimates; it does not prove biological association among individuals or causation. (2) The sum of the amplitudes, since \(C(0)=a+b=D/V_c\). (3) \(0.06/0.12=0.50\) residual standard deviations. (4) Different parameter combinations can produce similar predictions at the observed times, even when individual parameters vary substantially. (5) For example, add samples in a time region where plausible predictions differ; this helps only if the differences are sufficiently large relative to measurement error and the added observations are feasible.