Why Residual Order Matters
A linked-pool model can reproduce a baseline value and still describe the wrong biological pathway. Its predictions may diverge from observations in an orderly way: for example, consistently too early during a decline, or too low in one pool while too high in another. Such patterns matter because a sequence of errors can reveal a structural mismatch that a single summary of overall fit would conceal.
Residuals are the differences between observed values and model predictions at the observation times. Their order is essential in turnover and precursor-pool models: a delayed response, an incorrectly directed conversion, and a misplaced production or loss process can all leave patterns that vary systematically over time. A residual pattern is a diagnostic clue, not a unique identification of mechanism. Sampling times, measured pools, the observation-error model, and other plausible pathway structures all affect what can be inferred.
A common starting assumption is that, conditional on the model and its parameters, additive observation errors have mean zero, common variance \(\sigma^2\), and are independent: \(Y_i=f(t_i;\theta)+\epsilon_i\), with \(\epsilon_i\sim N(0,\sigma^2)\). This assumption makes signed, standardized residuals interpretable against a common scale. It is not automatically appropriate for repeated measurements from the same individual, proportional variability, or assay error that changes with concentration. In those settings, residuals should be evaluated on a scale and with an error structure that match the observation model.
Even when the observation model is appropriate, residuals from a fitted model are not independent evidence that each assumption holds: parameter estimation can make residuals smaller or induce dependence. The useful question is whether their signs, magnitudes, or ordering show structure that the fitted model has not captured. The visual predictive checks for precursor-pool models provide a complementary view of model predictions; residuals make the signed discrepancies explicit.
How a Missing Delay Can Leave a Phase Pattern
An omitted precursor stage often changes when a response appears, not just its eventual magnitude. Consider a model that predicts a response curve \(f(t)\), while the actual pathway produces a similar curve delayed by \(\tau>0\). The residual expected from this timing mismatch is approximately \(f(t-\tau)-f(t)\). For a small delay, a first-order Taylor expansion gives a useful justification for the sign pattern.
During a falling response, \(f'(t)<0\), so the delayed observation tends to exceed the immediate model prediction: residuals tend to be positive. During a rising response, \(f'(t)>0\), so the delayed observation tends to be below the prediction: residuals tend to be negative. Near a flat maximum or minimum, the first-order difference is small, and the sign may be less informative. This is a local approximation, not a rule that every omitted stage must produce the same signs: model fitting can shift the curve, alter its amplitude, or change other parameters.
Worked Example: A Possible Omitted Precursor Stage
Suppose a hypothetical biomarker is measured in ng/mL after a perturbation. A fitted model without an explicit delay predicts the values below. The assumed residual standard deviation is \(\sigma=3\) ng/mL. Calculate residuals and standardized residuals, then assess whether their ordering is consistent with a phase lag.
| Time (days) | Observed (ng/mL) | Predicted (ng/mL) |
|---|---|---|
| 2 | 98 | 94 |
| 4 | 91 | 82 |
| 8 | 72 | 70 |
| 12 | 68 | 75 |
| 20 | 84 | 90 |
At day 2, \(r=98-94=4\) ng/mL and \(z=4/3=1.33\). At day 4, \(r=91-82=9\) ng/mL and \(z=9/3=3.00\). At day 8, \(r=72-70=2\) ng/mL and \(z=2/3=0.667\). At day 12, \(r=68-75=-7\) ng/mL and \(z=-7/3=-2.33\). At day 20, \(r=84-90=-6\) ng/mL and \(z=-6/3=-2.00\).
The predictions fall early and then rise; residuals are mainly positive during the fall and negative during the rise. That ordering is consistent with the small-delay argument: the predicted trajectory appears phase-advanced relative to the observations. The day-4 discrepancy is also large relative to the assumed error scale. These results motivate examining a delayed or staged pathway, but they do not establish that an omitted precursor stage is the cause. A different time course of drug action, an incorrectly specified observation-error model, or another structural feature could also produce systematic residuals.
Pool-Specific Residuals and Conversion Direction
A proposed conversion direction determines which pool supplies material to another and how their amounts should change together. If both pools are observed, their residuals can be examined jointly. A model that predicts too little of a proposed source pool and too much of a proposed product pool may be allocating material along the pathway incorrectly. The reverse pattern later in time can suggest that the model has the wrong transient transfer between pools.
This reasoning depends on what is measured. With only a terminal biomarker, different pool structures or conversion directions can produce similar output trajectories, so a residual sequence alone may not distinguish them. With measurements of multiple pools, reciprocal discrepancies are more informative, especially when they occur in a time-ordered way consistent with a transfer. Even then, an error in rates, initial conditions, or unmodelled input can imitate part of the pattern.
Worked Example: Reciprocal Residuals in Two Pools
Consider hypothetical measurements of a precursor pool \(P\) and product pool \(M\), both in nmol/L. A fitted pathway model predicts the values shown. Suppose the residual standard deviation for each pool is 5 nmol/L. Calculate the pool-specific residuals and describe what the pattern suggests.
| Time (hours) | Observed \(P\) | Predicted \(P\) | Observed \(M\) | Predicted \(M\) |
|---|---|---|---|---|
| 2 | 80 | 70 | 20 | 30 |
| 8 | 50 | 60 | 50 | 40 |
At 2 hours, \(r_P=80-70=+10\) nmol/L and \(z_P=10/5=+2.00\). For \(M\), \(r_M=20-30=-10\) nmol/L and \(z_M=-10/5=-2.00\). At 8 hours, \(r_P=50-60=-10\) nmol/L and \(z_P=-2.00\); \(r_M=50-40=+10\) nmol/L and \(z_M=+2.00\).
The early prediction has too little \(P\) and too much \(M\); later it has too much \(P\) and too little \(M\). The paired, reversing signs are consistent with a mismatch in the timing or direction of material transfer between pools. They do not prove that the conversion arrow is reversed: incorrect conversion rates, a missing input or loss route, or biased measurements could also cause reciprocal residuals. Further evaluation should check whether the implied pool totals and other measured time points are compatible with the proposed mass balance.
Production Versus Loss: Compare the Transient
A production pathway adds material to a pool; a loss pathway removes it. At steady state, the simple turnover equation \(dY/dt=R-kY\) has \(Y_{\mathrm{ss}}=R/k\). Consequently, a lower steady-state concentration could result from lower production \(R\), higher loss rate \(k\), or both. A steady-state fit alone may not separate these explanations. Their transient predictions can differ because \(k\) also determines the rate at which the pool approaches its new steady state.
Worked Example: A Loss Assignment That Predicts Too-Fast Decline
Suppose a hypothetical biomarker begins at 100 concentration units. Following a challenge, the true pathway reduces production from 10 to 5 concentration units/day while the loss rate remains \(k=0.10\) per day. A competing model leaves production at 10 units/day and increases the loss rate to \(0.20\) per day. Both models imply a new steady state of 50 concentration units. Calculate their predictions at 2 and 8 days, assuming each starts at 100.
For constant \(R\) and \(k\), solving \(dY/dt=R-kY\) gives \(Y(t)=R/k+[Y(0)-R/k]e^{-kt}\). Under the production-change pathway, the prediction at day 2 is \(50+(100-50)e^{-0.10(2)}=50+50e^{-0.2}=90.94\) units. At day 8 it is \(50+50e^{-0.8}=72.47\) units.
Under the loss-change pathway, the prediction at day 2 is \(10/0.20+[100-10/0.20]e^{-0.20(2)}=50+50e^{-0.4}=83.52\) units. At day 8 it is \(50+50e^{-1.6}=60.09\) units. If observations follow the production-change pathway, the residuals against the loss-change predictions are \(90.94-83.52=+7.42\) units at day 2 and \(72.47-60.09=+12.38\) units at day 8. The loss-change model predicts a decline that is too fast, so it underpredicts the observed biomarker during this transient.
This calculation shows why matching the new steady state is not sufficient to establish a pathway assignment. Measurements during the transition and, where possible, after the challenge ends can help distinguish the mechanisms. The residuals do not by themselves prove that production changed; other causes of a slower decline could be present.
Diagnosis, Alternatives, and Limitations
Residual diagnosis is most useful when it begins with a clearly stated observation model and asks whether errors are patterned by time, pool, or predicted level. A residual-time plot can reveal runs of positive or negative values, phase-related reversals, and isolated outliers. For linked pools, plot or tabulate residuals separately by pool and compare their signs at matched times. Also check whether variability grows with predicted concentration; if it does, raw residuals may appear larger at high values even when proportional error is appropriate.
A time-ordered pattern is not synonymous with a missing pathway. It may instead reflect an incorrect initial condition, baseline drift, inaccurate timing of the perturbation, a misspecified drug-effect time course, a poor error model, or influential observations. If subjects contribute repeated measurements, within-subject dependence also matters: a run of same-sign residuals may be expected under a correlated error structure. Check whether the pattern is shared across subjects or driven by one individual before attributing it to the structural model.
| Observed residual pattern | Possible structural clue | Important alternative |
|---|---|---|
| Positive residuals on a falling limb and negative residuals on a rising limb | Response may be delayed relative to the model | Incorrect timing of the driver or another phase-shifting feature |
| Opposite-sign residuals in linked pools that change over time | Pool transfer or conversion direction may be misrepresented | Incorrect rates, unmodelled input/loss, or measurement bias |
| Systematic errors during a transition despite a matched new steady state | Production and loss assignments may differ from the model | Incorrect turnover rate or time-varying input |
| Residual magnitude increasing with predicted concentration | Not necessarily a pathway problem | Proportional error may be more suitable than additive error |
The main alternative to changing pathway structure is to retain the structure and reassess its inputs, initial conditions, parameter estimates, and observation model. Another alternative is to compare a small set of biologically plausible pathway models using the same observations and error assumptions. A candidate model should be judged not only by whether it removes a visible residual pattern, but also by whether its parameters are estimable and its predictions are credible across the observed time range.
Residuals have important limits. Sparse sampling can miss the interval in which two pathways differ most. A terminal measurement may not identify the direction of conversion. Flexible models can reduce residuals by adding poorly supported parameters, while a visually appealing fit can still be structurally wrong. Residual signatures should therefore guide model checking and targeted data collection, not be treated as standalone mechanistic evidence.
Common Pitfall
Wrong reasoning: “The residuals are positive after dosing, so the model must be missing a precursor stage.” A positive residual only says that the prediction is below the observation at those times. It does not name the cause. A baseline offset, delayed drug action, incorrect loss rate, or error-scale problem could also produce positive residuals.
Correction: Interpret the signs in temporal and biological context. Ask whether the residuals reverse as the response changes direction, whether multiple pools show a coherent transfer pattern, and whether the observation model is adequate. Then compare plausible explanations using the measurements that distinguish them. State the conclusion as a hypothesis supported by a pattern, not as a unique diagnosis.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If \(y_i=42\) ng/mL, \(\hat y_i=46\) ng/mL, and \(\sigma=2\) ng/mL, calculate the raw and standardized residual and interpret its sign.
- A delayed observed response is compared with a model curve that is falling. Using \(r(t)\approx-\tau f'(t)\), what residual sign is expected, and why?
- What does an early positive precursor residual paired with an early negative product residual suggest? Give one alternative explanation.
- Two pathway models have the same predicted new steady state but different turnover rates. Why can transient observations distinguish them?
- Why is a time-ordered residual pattern insufficient to prove that a particular precursor stage is missing?
Solutions. (1) \(r=42-46=-4\) ng/mL and \(z=-4/2=-2.00\); the model overpredicts by 4 ng/mL, or two residual standard deviations. (2) A positive residual is expected because \(f'(t)<0\) and \(-\tau f'(t)>0\); the delayed observation has not fallen as far as the immediate model prediction. (3) The model may transfer too much material into the product too early, consistent with a timing or conversion mismatch. Incorrect rates, an unmodelled input, or measurement bias could also contribute. (4) The loss rate affects the speed of approach to steady state, so predictions during the transition can differ even when the limiting value is the same. (5) Other structural and observation-model errors can produce similar signs; residuals indicate a mismatch, not a unique cause.