Why Species-Structured Misfit Needs Its Own Diagnosis
A model can describe the overall exposure-response relationship reasonably well and still miss systematically in one species. For example, predictions may be too low at every exposure in one species, or may miss the timing of response in another. If the data are pooled and summarized only by an overall goodness-of-fit measure, those patterns can cancel out. That is a serious concern in translation: a model whose errors differ by species may not provide a dependable basis for translating parameters or predictions.
Diagnosis begins with a specific comparison: for each observation, how does the observed response differ from the response predicted by the fitted cross-species model? The pattern of those differences can indicate whether to investigate scaling, a missing mechanism, or a parameter that does not transport as assumed. Diagnostics do not, by themselves, prove which explanation is correct.
Write \(y_{ij}\) for response observation \(j\) from species \(i\), and let \(\mu_{ij}(\theta)\) be the model-predicted conditional mean for that observation, given its species, dose or exposure, time, and other modeled conditions. The parameter vector \(\theta\) includes whatever shared and species-dependent parameters the proposed translation model allows. The observed design matters: comparisons should be made at the times, exposures, and conditions represented in the data, rather than at unrelated conditions.
Residuals: Direction, Shape, and Species
The raw residual is \(e_{ij}=y_{ij}-\mu_{ij}(\theta)\). A positive residual means the observed response exceeds the prediction; a negative residual means it is below. Raw residuals are useful when the response scale is meaningful and variability is roughly comparable across observations. If response variability changes with the mean, exposure, or species, raw residuals from different conditions are not directly comparable.
A standardized residual divides the discrepancy by an appropriate measure of observation variability. For a model with conditional observation variance \(V_{ij}\), a simple form is \(r_{ij}=e_{ij}/\sqrt{V_{ij}}\). This is a conditional standardized residual: it uses the model’s observation variance at that condition. For correlated repeated measurements, residuals should be evaluated using the model’s covariance structure; treating correlated observations as independent can make a run of residuals look more informative than it is. Uncertainty in estimated parameters also matters, and is more fully represented by posterior predictive checks.
Inspect residuals by species and against relevant predictors, especially time, exposure, and fitted response. A species-specific run of positive residuals suggests that predictions are systematically low in that species. A residual trend with exposure may indicate an incorrect slope, potency, or scaling relationship. A time-dependent pattern can point to a mismatch in onset, delay, or response turnover. Curvature suggests that the model’s functional form may be inadequate. These are clues for targeted follow-up, not unique diagnoses: different mechanisms can produce similar residual patterns.
Worked Example: Finding a Species-Specific Residual Pattern
Suppose a hypothetical model is fitted to four observations in each of two species. After standardization, the residuals are \( -0.4, 0.2, 0.1, -0.1\) for species A and \(1.1, 1.3, 0.9, 1.1\) for species B. Assume, for this initial diagnostic, that residuals within each species are independent and have variance 1.
For species A, the mean standardized residual is \( (-0.4+0.2+0.1-0.1)/4=-0.05\). For species B, it is \( (1.1+1.3+0.9+1.1)/4=1.10\). The expected standard deviation of either mean under the stated assumptions is \(1/\sqrt{4}=0.50\).
Species A’s mean is close to zero. Species B’s mean is 2.20 such standard deviations above zero, since \(1.10/0.50=2.20\). More importantly for diagnosis, all four residuals in B are positive, while A’s residuals are mixed and small. The pattern suggests that the model tends to underpredict responses in B. With only four observations, and without checking dependence or uncertainty in the fitted model, this is a diagnostic signal rather than a formal conclusion that a particular mechanism is wrong.
Species-Specific Parameter Estimates
Residual plots ask where the model misses. A complementary diagnostic asks whether estimates of a parameter allowed to vary by species are consistent with the restriction imposed by the translation model. Examples include a common potency parameter, a shared response maximum, or a clearance parameter after a proposed body-size scaling. The comparison must be made on the parameter’s scientifically relevant scale and with its uncertainty shown.
A large difference between species-specific estimates can be informative, but is not automatically evidence of nontransportability. Estimates may be imprecise, strongly correlated with other parameters, or weakly identified because a species has few observations or a narrow exposure range. Fitting each species separately can also discard useful shared information. A sound diagnosis compares estimates and uncertainty under the same parameterization, checks whether the data identify the parameter, and then examines whether the discrepancy corresponds to a visible residual pattern.
Worked Example: Checking a Proposed Clearance Scaling
Consider hypothetical species-specific estimates of clearance after applying a proposed body-size scaling. The resulting normalized clearances are 0.80, 1.05, and 1.62 L/h/kg0.75 for species A, B, and C, with standard errors 0.10, 0.12, and 0.20 L/h/kg0.75, respectively. The translation model assumes the normalized clearance is common across species. For a simple exploratory comparison, assume the estimates are approximately normal and independent.
Compare species B and C. Their estimated difference is \(1.62-1.05=0.57\) L/h/kg0.75. Under independence, the standard error of that difference is \(\sqrt{0.12^2+0.20^2}=\sqrt{0.0544}=0.233\) L/h/kg0.75. The difference is therefore \(0.57/0.233=2.45\) standard errors from zero.
This contrast raises concern that the proposed scaling does not yield a common normalized clearance, particularly for species C. It does not establish that body size is the wrong scaling variable: other explanations include a species-specific physiological difference, estimation bias, or a misspecified PK structure. The contrast also depends on the assumptions of approximate normality and independence; shared data or fitted constraints can induce covariance. Residuals and model-based uncertainty should be examined before treating the estimate as evidence for a particular explanation.
Posterior Predictive Checks
A posterior predictive check asks whether data generated from the fitted model resemble the observed data in features relevant to the suspected misfit. For each draw of parameters from the posterior distribution, generate replicated observations at the observed design. Calculate the same discrepancy for the observed data and each replicated data set. Useful species-focused discrepancies include the largest absolute species mean residual, a species-specific residual slope against exposure, or a difference in mean response between two species at comparable exposure.
A posterior predictive tail area can be estimated as the fraction of replicated discrepancy values at least as extreme as the observed discrepancy. It is a model-checking summary, not the probability that the model is true or false. Because the data inform the posterior used to generate replications, posterior predictive tail areas are not generally calibrated like classical p-values under a null hypothesis. A check can still reveal that an important pattern is rarely reproduced by the fitted model.
Worked Example: Checking the Largest Species Residual Mean
Suppose a hypothetical fitted model has three species. The observed species mean standardized residuals are 0.2, \(-0.4\), and 2.6. Define the discrepancy \(T\) as the largest absolute species mean standardized residual: \(T=\max_i|\bar r_i|\). Thus the observed value is \(T_{\mathrm{obs}}=\max(0.2,0.4,2.6)=2.6\).
Generate 1,000 replicated data sets from posterior draws at the observed design and compute \(T\) for each. Suppose 18 replicated values are at least 2.6. The estimated posterior predictive tail area is \(18/1{,}000=0.018\). This indicates that a discrepancy at least this large is uncommon in these replications, so the model warrants investigation for species-level misfit.
The result does not mean there is a 1.8% probability that the model is correct, nor does it identify the cause. Inspecting the residuals shows that the largest contribution comes from one species; plotting that species’ residuals against exposure and time can help distinguish a constant offset from a trend or delay. The check is only useful if the replicated data preserve the relevant design and observation structure, including repeated-measure correlation when present.
Interpreting Patterns and Choosing Follow-Up
Several diagnostic views should agree before a strong mechanistic interpretation is made. A species-specific offset in residuals, together with a species-specific parameter estimate displaced from the transported value, supports investigating a nontransportable parameter or a missing species effect. Residual trends with exposure, combined with systematic differences in normalized PK parameters, make incorrect scaling a candidate explanation. A time pattern that persists after exposure alignment may point toward omitted delay or response dynamics. None of these signatures is unique, and parameter estimates can shift when the model’s structure changes.
The main alternatives answer different questions. Residual plots are direct and local: they show where predictions miss, but can be noisy and depend on residual construction. Species-specific fits or estimates expose differences hidden by shared constraints, but may be unstable when data are sparse and do not automatically account for partial pooling. Posterior predictive checks can assess complex, species-level features while propagating parameter uncertainty, but their results depend on the chosen discrepancy and do not provide a single objective pass/fail rule. Cross-species exposure-response meta-analysis can summarize between-species variation, but a summary of variation does not replace examining the structure of the underlying model errors.
| Diagnostic approach | Strength | Important limitation |
|---|---|---|
| Residual plots by species and predictor | Shows the direction and shape of local discrepancies | Can be distorted by dependence, heteroscedasticity, or fitted-parameter effects |
| Species-specific parameter estimates | Directly assesses assumptions of shared or scaled parameters | May be imprecise, correlated, or weakly identified |
| Posterior predictive checks | Compares complex features while reflecting posterior uncertainty | Depends on the chosen discrepancy; tail areas are not classical p-values |
| Aggregate fit summaries | Compact description of overall agreement | Can conceal opposing errors across species or conditions |
Failure Modes and Practical Limits
A common mistake is to see a good overall fit statistic and conclude that translation assumptions are adequate. For instance, positive residuals in one species and negative residuals in another can cancel in a pooled mean. The correction is to examine residuals and relevant discrepancies by species, not only in aggregate.
Other warning signs include residuals with unexplained autocorrelation, a species represented only over a narrow exposure range, estimates changing sharply with small modeling choices, or parameter intervals so broad that apparent differences are not informative. Posterior predictive checks can also miss a problem if the discrepancy was chosen without regard to the suspected failure. Plot multiple scientifically motivated summaries, and report uncertainty rather than interpreting a point estimate alone.
Diagnostics identify where a model is inadequate relative to its assumptions; they do not establish that a revised model will predict better in new species or conditions. A proposed correction should have a biological or pharmacological rationale, preserve appropriate uncertainty, and be checked against the data available for that purpose. Avoid adding species-specific parameters merely to eliminate residual patterns when the data cannot identify them.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A model has mostly positive standardized residuals for one species and residuals centered near zero for the others. What is the direct interpretation, and what cannot yet be concluded?
- For a species with nine independent standardized residuals, what is the standard deviation of their mean under a correctly specified model?
- A posterior predictive check produces an estimated tail area of 0.04 for a species-level discrepancy. State an appropriate interpretation and one inappropriate interpretation.
- Why might a species-specific parameter estimate differ from the transported value even when the parameter is transportable?
- Residuals are positive at low exposure and negative at high exposure in one species. What broad class of model feature should be investigated, and why is the cause not uniquely determined?
Solutions. (1) The model tends to underpredict that species’ responses at the observed conditions. This does not establish which mechanism or parameter causes the discrepancy. (2) The standard deviation is \(1/\sqrt{9}=0.333\). The calculation assumes independent residuals with variance 1. (3) The observed discrepancy is relatively uncommon among replications from the fitted model and merits follow-up. It is incorrect to say there is a 4% probability that the model is true. (4) Sampling variability, weak identification, parameter correlation, or differences in data coverage can produce an apparent difference. (5) Investigate the exposure-response form or scaling because the discrepancy changes with exposure. Several features, including potency, slope, exposure measurement, or omitted mechanisms, could produce such a pattern.