Why Model Specification Must Be Explicit
A model-informed conclusion is only as interpretable as the model behind it. A report may state that clearance depends on body weight, for example, without specifying the mathematical relationship, reference weight, or units. Those omissions matter: a reader cannot tell what predictions the model makes for a particular patient, nor can another analyst implement the same model unambiguously.
A review of model specification asks whether the report clearly states the equations and assumptions that define the analysis. It is distinct from deciding whether the model fits well or whether its estimates are precise. The immediate question is more basic: what model was used, and what precisely does each part mean?
The structural model describes the expected system behavior, such as drug concentration over time. The residual-error model describes how an observation may differ from that expected value. These are separate components: a report that gives a concentration-time equation but does not explain the observation error has not fully specified how the data are represented.
What a Reviewer Needs to Find
A clear specification should let a qualified reader identify the model inputs, calculate a prediction in a stated situation, and understand the assumptions used to relate that prediction to measured data. The appropriate details depend on the model, but a review commonly checks for the following:
- Structural equations: the equations or algorithm defining the model predictions, including relevant initial conditions, dosing inputs, and time conventions.
- Parameter definitions: the meaning and units of each parameter, along with any population or individual-level interpretation.
- Covariate relationships: the mathematical form, covariate units, reference value, and interpretation of each included effect.
- Between-subject variability: how individual parameters relate to typical values, including the distribution and any stated dependence among random effects.
- Residual-error model: the distribution, scale, and dependence assumptions for deviations between observations and predictions.
- Scope and assumptions: the conditions under which the model is intended to represent the data, and any exclusions or simplifications that affect interpretation.
A list of parameter names is not a substitute for equations. Likewise, a statement that residuals were “proportional” is incomplete unless the report defines the proportionality and specifies the distribution and scale used. The aim is not to demand the same level of detail for every model; it is to make the implemented relationships unambiguous.
Worked Example: Reconstructing a Concentration Model
Hypothetical report fragment: “Concentrations were described by a one-compartment model with weight effects on clearance and volume.” The fragment does not provide equations or define the weight effects. Consider what additional specification would make its predictions interpretable.
Suppose the intended model is an intravenous bolus dose with first-order elimination, using body weight \(WT\) in kilograms. For an individual with no random deviation from the typical parameters, one complete specification is:
Here \(C(t)\) is concentration in mg/L at time \(t\) in hours after dosing; \(D\) is the intravenous bolus dose in mg; \(CL\) is clearance in L/h; and \(V\) is apparent volume of distribution in L. The typical parameters \(CL_{\mathrm{pop}}\) and \(V_{\mathrm{pop}}\) are the values at the reference weight of 70 kg. The exponents are fixed parts of this hypothetical model, not estimated here. The equation assumes a single dose, instantaneous mixing into one compartment, and first-order elimination.
Let \(D=100\) mg, \(CL_{\mathrm{pop}}=5.00\) L/h, \(V_{\mathrm{pop}}=50.0\) L, and \(WT=70\) kg. At \(t=4.00\) h, the covariate ratios equal 1, so \(CL=5.00\) L/h and \(V=50.0\) L. Therefore,
The units are consistent: \(D/V\) has units mg/L, while \(CL/V\) has units \(1/\mathrm{h}\), making the exponent dimensionless after multiplication by time. The original fragment is insufficient to reproduce this calculation: a different reference weight or exponent would give different predictions, even if the verbal phrase “weight effects” were unchanged.
Deriving and Checking the Covariate Relationship
In the example, the weight effect on clearance is a power relationship. Dividing clearance at a given weight by clearance at the reference weight gives a direct interpretation of the exponent:
The ratio \(WT/(70\ \mathrm{kg})\) is dimensionless, so raising it to the exponent does not change the clearance units. At \(WT=90\) kg, the ratio is \(90/70=1.286\), and the multiplier is \(1.286^{0.75}=1.207\). Thus, under this specified relationship, clearance is 1.207 times its typical value at 70 kg. This calculation does not establish that the relationship is scientifically appropriate; it shows what the stated equation implies.
For the same hypothetical parameters, a 90 kg individual has \(CL=5.00\times1.207=6.04\) L/h and \(V=50.0\times(90/70)=64.3\) L. At 4.00 h after the 100 mg dose, the model predicts:
This ratio-based check is useful in review because it exposes missing reference values and makes the covariate effect's direction and scale easier to inspect. A report should also say whether the relationship is applied to typical parameters, individual parameters, or another model quantity. Those choices define different models.
Parameter Variability and Residual Error
A population model may allow parameters to vary between individuals. For example, a log-normal random-effects model for clearance can be written as:
Here \(CL_i\) is the clearance for individual \(i\), \(CL_{\mathrm{typ},i}\) is that individual's typical clearance after applying any specified covariates, and \(\eta_{CL,i}\) is the individual's random deviation on the log scale. The variance parameter \(\omega_{CL}^{2}\) describes the distribution of that deviation. A report using such a formulation should identify the random effect, its distribution, and whether correlations with other random effects are included or assumed absent. It should not leave readers to infer those choices from software conventions.
The residual-error model is a separate layer. One common proportional-error specification is:
where \(Y_{ij}\) is observation \(j\) for individual \(i\), \(f_{ij}\) is the model prediction for that observation, and \(\epsilon_{ij}\) is a mean-zero residual term. Conditional on the predictions, this implies an observation variance of \(f_{ij}^{2}\sigma^{2}\), so the residual standard deviation grows in proportion to the prediction. By contrast, an additive-error specification \(Y_{ij}=f_{ij}+\epsilon_{ij}\), with \(\epsilon_{ij}\sim N(0,\sigma^2)\), has constant residual standard deviation \(\sigma\) in the observation units. A combined model can allow both components, but its exact equation must be stated.
Worked Example: Checking a Proportional-Error Statement
Hypothetical report statement: “A proportional residual error was used, with \(\sigma=0.20\).” Suppose the report also states that observations are conditionally normally distributed around the prediction, independently given the model predictions. For a prediction \(f=2.00\) mg/L, the model is \(Y\mid f\sim N(f,(f\sigma)^2)\). The conditional standard deviation is:
If the observed concentration is \(Y=2.30\) mg/L, its standardized difference from the prediction is \((2.30-2.00)/0.400=0.750\). This calculation follows from the stated error model; it does not by itself establish that the model describes the data adequately.
The unit check is important: \(\sigma=0.20\) is dimensionless in this formulation, whereas the residual standard deviation is in mg/L. If the report gave only the word “proportional” and the number 0.20, but did not state the distribution or define the role of \(\sigma\), a reader could not know whether the number is a standard deviation, a variance, or a differently parameterized quantity.
Assumptions, Alternatives, and Failure Modes
Every model specification makes assumptions, whether or not the report names them. A reviewer should look for the assumptions that determine how predictions and observations are generated. For the examples above, these include the one-compartment structure, bolus input, first-order elimination, stated weight scaling, distribution of between-subject effects, and conditional residual distribution. If residual independence is assumed, the report should say so; if a model includes serial correlation or another dependence structure, it should define that structure instead.
Alternative formulations can be compared by asking what each assumes, not by treating one as universally preferable. An additive residual model gives a constant standard deviation across predictions. A proportional model gives a standard deviation that scales with the prediction. A combined model includes both a prediction-independent and a prediction-dependent contribution. Similarly, a covariate may be represented by a linear, power, or categorical relationship. The report must state the chosen form and its reference or coding; the verbal label alone does not identify it.
Common specification failures include an equation without units, a covariate effect without a reference value, an unexplained parameterization of residual error, or a random effect whose distribution is omitted. Other warning signs are inconsistent notation across sections, unexplained symbols, and a mismatch between the stated model and the calculations or diagrams used to describe it. These are not all evidence that the model is wrong. They are evidence that the reported specification is not yet clear enough to evaluate or reproduce.
A focused review can proceed in a fixed order: identify the structural equations and their inputs; check parameter meanings and units; trace each covariate effect from its equation to a prediction; identify between-subject and residual variation separately; and list the assumptions needed to interpret the resulting predictions. This complements the earlier material on data provenance: knowing which data were analyzed is necessary, but it does not by itself tell a reader how the model transformed those data into evidence.
Find the structural, covariate, random-effects, and residual-error equations, including definitions of every symbol.
Verify that parameters, predictors, and predictions have interpretable units and that equations are dimensionally consistent.
Substitute a simple, stated dose, time, or covariate value to see whether the equations yield a traceable prediction.
Identify any unspecified distribution, reference value, dependence assumption, or model condition that prevents an independent reader from reproducing the interpretation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A report states, “Clearance was weight-adjusted using an allometric model.” What information is still needed to reproduce the covariate relationship?
- For a proportional-error model with prediction \(f=3.00\) mg/L and dimensionless \(\sigma=0.15\), calculate the implied residual standard deviation.
- A model report gives a concentration prediction equation but does not state the residual distribution. What can a reader conclude, and what remains unspecified?
- Why should a reviewer check the units of \(CL/V\) in a concentration-time equation?
Solutions. (1) The equation and exponent, the covariate's units and reference value, and an explanation of which clearance quantity is adjusted are needed; random effects must also be described if they modify that relationship. (2) Under the stated proportional-error formulation, the standard deviation is \(f\sigma=(3.00\ \mathrm{mg/L})(0.15)=0.450\ \mathrm{mg/L}\). (3) The reader can inspect the stated structural prediction, but cannot determine how observations are distributed around it, including the residual scale or any residual dependence assumptions. (4) \(CL/V\) must have units of inverse time so that multiplying it by time gives a dimensionless exponent; if the units do not agree, either the equation or the reported parameter units require clarification.