From Concentrations to a Response Model
When both a parent drug and an active metabolite are present, a response measured at a given time may reflect exposure to both compounds. The analysis therefore has two linked parts: a pharmacokinetic (PK) model predicts parent and metabolite concentrations under a dosing regimen, and a pharmacodynamic (PD) model maps those concentrations to the response. A response model that ignores one active compound can give misleading predictions if that compound contributes materially.
This worked analysis uses a small, entirely hypothetical dataset. The concentrations are treated as PK outputs at six design points, each representing a measured response and its corresponding parent and metabolite concentrations. The concentration units are mg/L, and the continuous response is in response units; higher values are taken to be more desirable for the dosing comparison. The small sample is for illustrating calculations, not for supporting a real clinical decision.
The additive model assumes a linear mean response over the concentration range studied, additive contributions from the two concentrations, and no parent–metabolite interaction. For least squares to estimate the conditional mean relationship without systematic bias, the errors should have mean zero conditional on the concentrations. Usual model-based standard errors and tests additionally rely on independent errors with constant variance; normal errors are needed for exact small-sample normal-theory inference. The model also treats the concentrations as known without material measurement error. If these assumptions are implausible, a different model or an integrated model for concentration and response uncertainty may be needed.
| Design point | Parent \(C_P\) (mg/L) | Metabolite \(C_M\) (mg/L) | Response \(Y\) (response units) |
|---|---|---|---|
| 1 | 1 | 0 | 23 |
| 2 | 3 | 0 | 25 |
| 3 | 1 | 2 | 26 |
| 4 | 3 | 2 | 40 |
| 5 | 1 | 4 | 35 |
| 6 | 3 | 4 | 43 |
The PK link is important even though this table starts with concentration values. In a practical analysis, a PK model—potentially a linked parent–metabolite model—would be fitted to concentration observations or supplied as an established model. It would then generate \(C_P(t)\) and \(C_M(t)\) under each candidate dose schedule. Those time-varying concentrations, not dose alone, are passed into the PD model to predict response over time. Here the six concentration pairs serve as the observed PK outputs for estimating the PD relationship; later, hypothetical PK predictions are used to compare dosing options.
Fit and Verify the Additive Model
Least squares chooses coefficients to minimize the sum of squared residuals, where a residual is observed response minus fitted response. With three coefficients—intercept, parent slope, and metabolite slope—the design matrix must have rank three to estimate them separately. The six concentration pairs here provide variation in both compounds and have full rank. The normal equations for an intercept-containing least-squares fit require the residuals to sum to zero and to have zero sample cross-product with each predictor. These checks are useful for detecting arithmetic errors.
Worked Example: Estimating the Additive Parent–Metabolite Fit
For the table, least squares gives \(\hat{\beta}_0=16.5\) response units, \(\hat{\beta}_P=4.00\) response units per mg/L, and \(\hat{\beta}_M=3.75\) response units per mg/L. Thus the fitted mean response is:
For design point 1, the prediction is \(16.5+4(1)+3.75(0)=20.5\) response units. For point 2 it is \(16.5+4(3)+3.75(0)=28.5\). The remaining predictions are \(16.5+4(1)+3.75(2)=28.0\), \(16.5+4(3)+3.75(2)=36.0\), \(16.5+4(1)+3.75(4)=35.5\), and \(16.5+4(3)+3.75(4)=43.5\).
Subtracting each prediction from its observed response gives residuals \(2.5,-3.5,-2,4,-0.5,-0.5\) response units. Their sum is \(2.5-3.5-2+4-0.5-0.5=0\). Their cross-product with parent concentration is \(2.5(1)-3.5(3)-2(1)+4(3)-0.5(1)-0.5(3)=0\); with metabolite concentration it is \(2.5(0)-3.5(0)-2(2)+4(2)-0.5(4)-0.5(4)=0\). These checks are consistent with an intercept-containing least-squares fit.
The residual sum of squares is \(RSS=2.5^2+(-3.5)^2+(-2)^2+4^2+(-0.5)^2+(-0.5)^2=39.0\) squared response units. There are \(6-3=3\) residual degrees of freedom. The residual standard deviation estimate is \(\sqrt{39/3}=\sqrt{13}=3.606\) response units, under the constant-variance model.
A coefficient is a conditional model slope, not automatically a causal effect. For example, \(\hat{\beta}_M=3.75\) means that the fitted mean response rises by 3.75 response units for a 1 mg/L increase in metabolite concentration while parent concentration is held fixed, according to this linear additive model. Whether such a comparison is supported by the design and whether the model assumptions are credible must also be considered.
Compare Competing Response Models
The additive model can be compared with simpler alternatives that omit one compound. A parent-only model asks whether parent concentration alone describes response; a metabolite-only model asks the analogous question for the metabolite. These are useful diagnostic comparators, not proof that an omitted compound has no biological effect. The comparison below uses least-squares fits and RSS. Because the additive model has an extra coefficient, its smaller RSS is expected; a small RSS alone does not establish that it is the best predictive model.
Worked Example: Comparing Parent-Only and Metabolite-Only Fits
For the parent-only model, fitting \(Y=a+bC_P\) gives \(\bar C_P=2\) mg/L and \(\bar Y=32\) response units. The slope is \(b=\sum(C_P-\bar C_P)(Y-\bar Y)/\sum(C_P-\bar C_P)^2=24/6=4.00\) response units per mg/L, and the intercept is \(a=32-4(2)=24.0\) response units. The fitted values are 28, 36, 28, 36, 28, and 36. The residuals are \(-5,-11,-2,4,7,7\), so \(RSS=25+121+4+16+49+49=264\) squared response units.
For the metabolite-only model, \(\bar C_M=2\) mg/L. The slope is \(60/16=3.75\) response units per mg/L, and the intercept is \(32-3.75(2)=24.5\) response units. The fitted values are 24.5, 24.5, 32, 32, 39.5, and 39.5. The residuals are \(-1.5,0.5,-6,8,-4.5,3.5\), giving \(RSS=2.25+0.25+36+64+20.25+12.25=135\) squared response units.
The additive fit has \(RSS=39\), smaller than either one-compound fit in this dataset. This is evidence that including both measured concentrations describes these six responses better in-sample. It does not by itself establish precise coefficients, external predictive performance, or a causal contribution from each compound. With only six observations, estimates and diagnostics are particularly sensitive to individual points.
In a fuller analysis, alternative models should be chosen for a scientific reason and compared using appropriate diagnostics and prediction goals. A nonlinear concentration-response relationship, a delayed effect, a baseline trend, or an interaction could make the additive straight-line model inadequate. Earlier material on effect-site delay and parent–metabolite contributions provides useful mechanistic distinctions; the present fit does not resolve those mechanisms merely by using contemporaneous concentrations.
Residual Checks and Failure Modes
Residual diagnostics ask whether the discrepancies left after fitting look compatible with the assumed model. Plot residuals against fitted response, parent concentration, metabolite concentration, and time or design order where available. A curved pattern can suggest nonlinearity; a fan-shaped spread can suggest changing variance; a sequence or time trend can indicate dependence, drift, or a missing delayed-effect component. A few large residuals can dominate RSS and should be checked against data records and the study context, not automatically deleted.
For this dataset the additive residuals range from \(-3.5\) to \(4\) response units. There are positive residuals at points 1 and 4, negative residuals at points 2, 3, 5, and 6, but six residuals are far too few to establish that the pattern is random or that variance is constant. In particular, design points are not a long time series, so a time-dependent diagnostic cannot be made from this table alone. The fitted residual standard deviation is a scale summary, not a guarantee that future prediction errors will be that small.
Use Linked Predictions for a Dosing Comparison
To compare dosing options, first obtain the concentration profiles each option is expected to produce from the PK model. Then substitute those concentrations into each candidate PD model. The following calculation uses hypothetical PK-predicted concentrations at a decision-relevant time; it is an illustration of model-based comparison, not a treatment recommendation. The two options are assumed to yield \((C_P,C_M)=(3,4)\) mg/L for the standard schedule and \((2,4)\) mg/L for a parent-lowering schedule.
Worked Example: Comparing Predicted Responses Under Two Schedules
Under the additive fit, the standard schedule predicts \(16.5+4(3)+3.75(4)=43.5\) response units. The parent-lowering schedule predicts \(16.5+4(2)+3.75(4)=39.5\) response units. The model therefore predicts a difference of \(43.5-39.5=4.00\) response units in favor of the standard schedule at this concentration pair.
The alternatives show why model choice matters. The parent-only fit predicts \(24+4(3)=36\) response units under the standard schedule and \(24+4(2)=32\) under the parent-lowering schedule. The metabolite-only fit predicts \(24.5+3.75(4)=39.5\) response units for either schedule because both have the same metabolite concentration. Thus the predicted response levels differ across models, and the metabolite-only model predicts no difference between these options.
The observed dataset includes parent concentrations of 1 and 3 mg/L and metabolite concentrations of 0, 2, and 4 mg/L. Both proposed concentration pairs are within those separate observed ranges. However, model-based prediction still depends on the adequacy of the additive relationship, the accuracy of the PK projections, and whether the response at that time is the relevant decision endpoint. A decision would also require an acceptable response target, uncertainty assessment, safety information, and validation beyond this six-point illustration.
The principal advantage of linking PK and PD is that a dosing change can alter parent and metabolite concentrations differently. Comparing dose amounts alone misses this distinction. The principal limitation is that a precise-looking response prediction can inherit uncertainty from both models: PK parameter uncertainty affects the concentration inputs, while PD parameter uncertainty and structural misspecification affect the concentration-to-response mapping. Treating predicted concentrations as exact may understate total uncertainty.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In the additive model, what does the parent coefficient describe, and what are its units?
- Why must the residuals sum to zero for an intercept-containing least-squares fit?
- Calculate the additive-model prediction at \(C_P=2\) mg/L and \(C_M=3\) mg/L.
- Why does the smaller additive-model RSS not, by itself, prove that it is the best model for a dosing decision?
- If two dosing options have equal predicted metabolite concentrations but different parent concentrations, which of the fitted models predicts equal responses for the two options?
Solutions. (1) It is the fitted change in mean response per 1 mg/L increase in parent concentration while metabolite concentration is held fixed; its units are response units per mg/L. (2) Least squares with an intercept makes the residual vector orthogonal to the intercept column, which is a column of ones, so the residual sum is zero. (3) The prediction is \(16.5+4(2)+3.75(3)=35.75\) response units. (4) The additive model has an extra parameter and is evaluated on only six observations; fit, assumptions, uncertainty, and prospective predictive performance also matter. (5) The metabolite-only model predicts equal responses because its prediction depends on metabolite concentration alone.