Why Cascade Structure Needs Residual Diagnostics
A cascade model can trace the broad shape of a delayed response and still have the wrong internal structure. Too few transit stages, an omitted branch, a nonlinear transfer represented as linear, or drug action placed at the wrong stage can each produce predictions that look plausible at a glance. The mismatch may be clearest not in the response curve itself, but in the ordered pattern of its errors.
Residual diagnostics do not identify a unique biological mechanism. They help determine whether the fitted model leaves systematic features unexplained and suggest which structural alternatives are worth checking. This is especially important for cascades because changes in stage count, rates, branching, and input location can all alter delay and response shape.
Residuals should be examined in their original time order, and against time, predicted response, and relevant inputs such as exposure or a cascade-stage input. An observed-versus-predicted plot places \(y_j\) on one axis and \(\hat y_j\) on the other; agreement lies near the identity line \(y=\hat y\). A plot that discards time order can conceal a consistent early-to-late pattern. Conversely, an observed-versus-predicted curve that looks close can hide residuals that remain positive over one interval and negative over another.
Interpretation assumes that observations and predictions refer to the same response scale, time points, baseline convention, and input history. The error model must also be considered: residual magnitudes are only directly comparable when the variance is reasonably constant. If uncertainty grows with response size, standardized or appropriately weighted residuals are more informative than raw residuals. If measurements are correlated over time, apparent runs of residuals can arise from the error process as well as from misspecified cascade structure.
Why a Structural Error Can Leave a Pattern
A fitted model uses its parameters to absorb some discrepancies. Consequently, residuals are not simply the missing biological component minus random noise. A local approximation makes this distinction explicit. Suppose the true mean response contains a structural component \(g\) that is absent from the fitted model, and the observation error is \(\varepsilon\). Near the fitted parameter value, a small parameter change \(\delta\) changes the model prediction approximately by \(J\delta\), where \(J\) is the matrix of prediction sensitivities to the parameters.
For weighted least squares with fixed weight matrix \(W\), define \(A=W^{1/2}J\) and \(q=W^{1/2}(g+\varepsilon)\). The fitted local parameter change is \(\hat\delta=(A^\mathsf{T}A)^{-1}A^\mathsf{T}q\), assuming the columns of \(A\) are linearly independent. Substituting this into the weighted residual gives the following expression:
The bracketed matrix removes the part of the discrepancy that can be mimicked by changing fitted parameters. The remaining component may retain a time pattern, but its shape need not equal the omitted mechanism \(g\). This is why residual patterns are clues rather than one-to-one labels for particular biological structures. The approximation is local; strong nonlinearity, parameter boundaries, or poor identifiability can make the actual fitting behavior more complicated.
Read the Pattern Before Naming the Cause
A useful first pass is to look for persistent runs, smooth trends, repeated peaks, changes in spread, or patterns tied to input level. A single large residual may reflect an unusual observation. Several residuals with a coherent sequence or relation to a covariate are more suggestive of a systematic mismatch, although serially correlated errors can also produce runs.
| Residual feature | Structural possibility to investigate | Important alternative explanation |
|---|---|---|
| Early residuals mostly negative, followed by mostly positive residuals | Predicted response may rise too early or decay too quickly; stage count, rates, or input placement may be wrong | Time alignment error, misspecified input history, or correlated errors |
| Persistent positive residuals around a later shoulder or second peak | An omitted delayed branch or response component may be relevant | Incorrect main-branch rate, baseline drift, or an unmodeled time-varying input |
| Residuals become increasingly negative at high input or predicted response | A linear transfer may be inadequate if the data show saturation | Variance scaling, assay ceiling, or an exposure measurement problem |
| Timing mismatch begins after the input is applied | Drug action may be placed too near the output or too far upstream | Incorrect exposure timing or an unmodeled delay between measured and effective input |
These are not decision rules. Different structures can produce similar curves, and a flexible parameterization can mask structural errors over a limited observation window. Use biological plausibility, input variation, and the pattern across multiple time courses to evaluate a proposed explanation.
Worked Example: A Possible Stage-Count Mismatch
Consider hypothetical response measurements after a single input. The fitted predictions from a three-stage candidate and a four-stage candidate are shown below. The values are response units; all observations use the same sampling times.
| Time (h) | Observed | Three-stage prediction | Three-stage residual | Four-stage prediction | Four-stage residual |
|---|---|---|---|---|---|
| 0 | 0 | 0 | 0 | 0 | 0 |
| 1 | 2 | 4 | -2 | 3 | -1 |
| 2 | 7 | 8 | -1 | 7 | 0 |
| 3 | 13 | 11 | 2 | 12 | 1 |
| 4 | 12 | 10 | 2 | 12 | 0 |
| 6 | 6 | 5 | 1 | 7 | -1 |
| 8 | 2 | 1 | 1 | 2 | 0 |
For the three-stage candidate, the residual at 1 h is \(2-4=-2\), while at 3 h it is \(13-11=2\). The early negative residuals followed by positive residuals indicate that this fitted curve rises too soon or has the wrong shape around the peak. Its residual sum of squares is \(0^2+(-2)^2+(-1)^2+2^2+2^2+1^2+1^2=15\) response-units squared. For the four-stage candidate, the residuals are smaller and the sum of squares is \(0^2+(-1)^2+0^2+1^2+0^2+(-1)^2+0^2=3\) response-units squared.
The observed-versus-predicted points for the four-stage candidate are closer to the identity line for these data, and the ordered residuals show less systematic discrepancy. This supports checking a four-stage structure, but does not establish that four biological stages exist: the candidates may differ in more than stage count, and the data are hypothetical and sparse. A fair comparison also requires consistent fitting assumptions and attention to the extra flexibility of the more complex model.
Worked Example: A Late Residual Shoulder and a Possible Branch
Suppose a single-branch model is fitted to a hypothetical response time course, with residuals defined as observed minus predicted.
| Time (h) | Observed response | Single-branch prediction | Residual |
|---|---|---|---|
| 0 | 0 | 0 | 0 |
| 1 | 2 | 2 | 0 |
| 2 | 4 | 5 | -1 |
| 3 | 7 | 7 | 0 |
| 4 | 9 | 7 | 2 |
| 5 | 9 | 6 | 3 |
| 6 | 7 | 5 | 2 |
The later residuals are positive at 4, 5, and 6 h, so the model underpredicts a broad late shoulder. The sum of squares is \(0^2+0^2+(-1)^2+0^2+2^2+3^2+2^2=18\) response-units squared. A delayed parallel branch could plausibly contribute a late component, so a branched candidate is worth evaluating. However, the same pattern could result from a slower main-branch rate, a drifting baseline, or an incorrect input history. The residual plot motivates a structural comparison; it does not prove branching.
Amplitude and Input-Placement Diagnostics
Worked Example: Residuals Suggesting Nonlinear Transfer
At a matched time after exposure, suppose a hypothetical experiment observes responses at four input levels. A linear transfer predicts response proportional to input. The residuals below are observed minus predicted.
| Input level (arbitrary units) | Observed response | Linear prediction | Residual |
|---|---|---|---|
| 1 | 2 | 2 | 0 |
| 2 | 4 | 4 | 0 |
| 4 | 7 | 8 | -1 |
| 8 | 9 | 16 | -7 |
The high-input observations fall increasingly below the linear predictions. The residual sum of squares is \(0^2+0^2+(-1)^2+(-7)^2=50\) response-units squared. A saturating transfer is one plausible alternative to investigate. This comparison is only diagnostic: if response variance also grows with input, standardized residuals and the error model matter, and an assay ceiling could create a similar pattern. The input-response relationship should be examined across replicated or otherwise informative conditions before assigning the pattern to transfer nonlinearity.
Worked Example: Testing a Suspected Input-Placement Error
After a hypothetical input step at time zero, observed responses are \(0,1,4,8,7,4\) at times \(0,1,2,3,4,5\) h. A candidate with drug action placed distally predicts \(0,3,7,8,5,2\). Its residuals are \(0,-2,-3,0,2,2\), and its sum of squares is \(0+4+9+0+4+4=21\) response-units squared. It overpredicts early and underpredicts late.
A candidate that places the input upstream predicts \(0,1,4,7,7,4\). Its residuals are \(0,0,0,1,0,0\), giving a sum of squares of \(1\) response-unit squared. This temporal pattern is consistent with the distal placement producing a response that is too early and too brief. The upstream candidate fits these hypothetical values better, but the comparison does not isolate input placement unless the candidates are otherwise comparably specified and fitted. An error in input timing or in transit rates could also shift the response.
A Practical Diagnostic Sequence
Confirm that observed and predicted values use the same time origin, input record, baseline, and response units. A mismatch here can imitate structural error.
Overlay observed and predicted response curves, then plot residuals in time order. Include an observed-versus-predicted plot and, where relevant, residuals against input or predicted response.
Look for changing residual spread or serial structure. Use standardized or whitened residuals when the specified error model requires them, and do not interpret a run as structural evidence without considering correlated errors.
Describe the feature first, such as “early overprediction followed by late underprediction.” Then identify plausible alternatives, such as stage count, rates, or input placement, rather than naming one mechanism as established.
Fit the alternative structures under comparable assumptions. Check whether the pattern improves across time courses or input conditions, and whether added complexity is justified rather than merely reducing residual sum of squares.
Alternatives, Failure Modes, and Limitations
Residual plots provide local, visually interpretable evidence. A numerical fit summary such as residual sum of squares compresses the discrepancies into one number, so it cannot show whether errors occur in a coherent time pattern. Conversely, a visually striking pattern can be unstable when there are few observations. Formal tests of residual dependence or model comparison criteria can supplement plots, but they also depend on assumptions and do not identify the biological cause of a mismatch. Predictive checks under new input profiles can be more discriminating when competing structures make different predictions there.
- Do not interpret residual signs backward. With \(e_j=y_j-\hat y_j\), a positive value means underprediction, not overprediction.
- Do not infer structure from one point. A single large residual may be an observation error or unusual measurement; inspect the sequence and the corresponding assay and input records.
- Do not equate smaller residual sum of squares with a better structure. More flexible models generally have more opportunity to fit the observed data. Consider added complexity, parameter identifiability, and performance across conditions.
- Do not treat a smooth residual run as automatic proof of misspecification. Serially correlated errors, sparse sampling, baseline error, and uncertainty in input timing may produce similar patterns.
- Do not overinterpret a clean plot. Several structures can make nearly indistinguishable predictions over the observed time window, especially if input conditions do not vary enough to separate them.
The most defensible conclusion is usually conditional: a residual pattern is inconsistent with the current model and error assumptions, and it makes specified alternatives worth testing. Residual diagnostics matter because they prevent a plausible-looking response curve from being mistaken for an adequate structural explanation, while preserving the distinction between statistical evidence and biological proof.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If \(y_j=6.2\) response units and \(\hat y_j=7.0\) response units, calculate the residual and state its direction.
- A model has negative residuals early and positive residuals late. Name two structural possibilities and one nonstructural alternative.
- Why can a late sequence of positive residuals be consistent with an omitted branch but fail to prove one?
- What pattern in residuals against input level could motivate checking a saturating transfer?
- Why might a model with more stages have a lower residual sum of squares without being the preferred structural model?
Solutions. (1) \(e_j=6.2-7.0=-0.8\) response units, so the model overpredicts by 0.8 response units. (2) Possible structural explanations include too few stages or input placed too distally; an input-time alignment error is a nonstructural alternative. (3) A branch could add a delayed component, but a slower main-branch rate, baseline drift, or an incorrect input history could also produce late underprediction. (4) Residuals that become increasingly negative as input rises indicate that a linear model increasingly overpredicts at high input, which can motivate checking saturation. (5) Extra stages increase flexibility and can fit observed noise; the comparison must account for complexity, identifiability, assumptions, and behavior across conditions.