Diagnosing Extrapolation Distance Across Species

Why Extrapolation Distance Matters

A human prediction can be numerically precise under a fitted model while depending on covariate combinations never observed in animals. The prediction may then rely on the model's functional form rather than on direct support from the data. Diagnosing this gap does not determine whether the model is biologically correct; it clarifies how far the human target lies from the animal evidence and where predictions warrant particular scrutiny.

The distance depends on how covariates are represented. A body-mass difference measured in kilograms is not equivalent to a difference on a logarithmic scale, and checking one covariate at a time does not reveal whether a human combination of covariates occurred jointly in the animal data. A useful assessment therefore combines interpretable range checks with geometry-based diagnostics such as regression leverage.

Definition: For a covariate $x$, let the observed animal range be $[a,b]$, where $a$ and $b$ are its minimum and maximum values in the data used to fit the model. The distance from a target value $x_*$ to that range is $$ d_{\mathrm{range}}(x_*)=\max(a-x_*,\,0,\,x_*-b). $$ It is zero inside the observed range and positive outside. If direction is useful, a signed measure can be defined explicitly as $x_*-a$ below the range, zero inside, and $x_*-b$ above it. The signed measure is negative below and positive above; it is not the same quantity as the nonnegative distance.

Range Checks on the Model Scale

A range check is meaningful only after specifying which covariate representation enters the model. If a model uses $\log_{10}(x)$, compare the human value and animal range on that transformed scale. A gap of 0.3 on a base-10 logarithmic scale corresponds to a factor of about two, not a difference of 0.3 units on the original scale. Always report the scale and, where useful, translate the gap back into a ratio or original units.

A univariate range check has a deliberately limited interpretation: it asks whether each target covariate lies between its observed minimum and maximum. It does not show that the model is well supported throughout that interval, nor that the target's joint covariate combination resembles the data. Sparse observations, gaps between clusters, and covariate interactions can all make an apparently in-range prediction fragile.

Worked Example: Distance Beyond the Animal Body-Mass Range

Hypothetical data. Suppose an animal scaling model uses $z=\log_{10}(\text{body mass in kg})$. The animal masses are 0.2, 2, and 20 kg, and the target is a 70-kg adult. Calculate the gap on the model scale and interpret it.

The smallest and largest animal values on the model scale are $$ a=\log_{10}(0.2)=-0.6990,\qquad b=\log_{10}(20)=1.3010. $$ The target value is $$ z_*=\log_{10}(70)=1.8451. $$ Because the target is above the animal range, the nonnegative distance is $$ d_{\mathrm{range}}=z_*-b=1.8451-1.3010=0.5441 $$ log-base-10 mass units. The signed measure under the stated convention is also $+0.5441$, indicating that the target is above the range. On the original scale, the target mass is $70-20=50$ kg above the largest animal mass and is $70/20=3.5$ times that mass.

These are different ways of describing the same boundary gap. For a log-linear model, the 0.5441 log-unit gap is often more relevant to the fitted relationship than the 50-kg difference. It flags extrapolation, but it does not by itself quantify prediction error or establish that the extrapolation is unacceptable.

Joint Support and Regression Leverage

With multiple covariates, being inside every marginal range is not enough to establish that a target is supported by the animal data. For example, two physiological measurements may each range from 0 to 2 in the animals, yet the animals may have high values of one only when the other is low. A target with both values high is marginally in range but outside the observed joint pattern.

One diagnostic for a linear regression is prediction leverage. Consider the model $Y=X\beta+\varepsilon$, where $X$ is the design matrix containing an intercept and the model covariates. Assume the mean is linear in these terms, $X$ has full column rank, and errors are independent with mean zero and common variance $\sigma^2$, conditional on the covariates. For a target with design vector $x_*$, the prediction leverage is $h_*=x_*^\mathsf{T}(X^\mathsf{T}X)^{-1}x_*$. It reflects how far the target's design vector lies, in the geometry induced by the observed design, from the covariate support used to estimate the regression.

Under these assumptions, the variance of the fitted mean at the target is $\sigma^2 h_*$. This is uncertainty in the estimated mean response, conditional on the model and design. Predicting a new individual response also includes that individual's residual variation, giving variance $\sigma^2(1+h_*)$. Leverage is therefore a geometric diagnostic, not a probability that the prediction is wrong, and not a measure of all sources of translational uncertainty.

Justification: In ordinary least squares, $\hat{\beta}=(X^\mathsf{T}X)^{-1}X^\mathsf{T}Y$. For a target mean estimate $x_*^\mathsf{T}\hat{\beta}$, its estimation error is $x_*^\mathsf{T}(\hat{\beta}-\beta)$. Since $\operatorname{Var}(\hat{\beta}\mid X)=\sigma^2(X^\mathsf{T}X)^{-1}$, it follows that $$ \operatorname{Var}(x_*^\mathsf{T}\hat{\beta}\mid X) =x_*^\mathsf{T}\operatorname{Var}(\hat{\beta}\mid X)x_* =\sigma^2 x_*^\mathsf{T}(X^\mathsf{T}X)^{-1}x_* =\sigma^2h_*. $$ Adding an independent future residual with variance $\sigma^2$ gives $\sigma^2(1+h_*)$ for a new response.

Worked Example: Leverage Beyond a One-Covariate Animal Range

Hypothetical model. Suppose a linear model for a physiological response uses one standardized covariate $x$ and an intercept. The four animal covariate values are $-1,0,1,2$. Assume the linearity and independent, constant-variance error assumptions are reasonable for this illustration. A target human value is $x_*=3$. Calculate its leverage and compare it with an in-range target at $x=1.5$.

For a regression with an intercept and one covariate, the leverage for a target $x_*$ is $$ h_*=\frac{1}{n}+\frac{(x_*-\bar{x})^2}{\sum_{i=1}^{n}(x_i-\bar{x})^2}. $$ Here $n=4$, $\bar{x}=(-1+0+1+2)/4=0.5$, and $$ \sum_{i=1}^{4}(x_i-\bar{x})^2 =(-1.5)^2+(-0.5)^2+(0.5)^2+(1.5)^2=5. $$ For $x_*=3$, which is above the observed range $[-1,2]$, $$ h_*=\frac{1}{4}+\frac{(3-0.5)^2}{5} =0.25+\frac{6.25}{5} =1.50. $$ For comparison, $x=1.5$ is inside the range, and its leverage is $$ h_*=\frac{1}{4}+\frac{(1.5-0.5)^2}{5} =0.25+0.20=0.450. $$ Thus the out-of-range target has substantially greater leverage in this design. If the fitted residual standard deviation were hypothetically $0.20$ response units, the standard error of the fitted mean at $x=3$ would be $0.20\sqrt{1.50}=0.245$ response units, conditional on the assumed model. This calculation does not include uncertainty in model structure or cross-species applicability. The leverage comparison flags dependence on the extrapolation; it does not certify the model or the prediction.

Marginal Ranges Can Miss Joint Extrapolation

Leverage can help expose joint extrapolation missed by separate range checks. The following example uses a plane with two covariates and an intercept. It also illustrates an important caution: a leverage calculation is possible even when the data are too sparse to estimate residual variance reliably.

Worked Example: In-Range Covariates Outside the Joint Data Support

Hypothetical design. Suppose animal observations have covariate pairs $(x_1,x_2)=(0,0),(0,0),(2,0),(0,2)$. A linear model includes an intercept and both covariates. The target human pair is $(1.5,1.5)$. Each target coordinate is between the observed minimum 0 and maximum 2. Determine whether the target is inside the joint support and calculate its design leverage.

The animal points satisfy $x_1+x_2\leq 2$, while the target has $1.5+1.5=3>2$. It is therefore outside the triangle containing the observed covariate combinations, despite being inside both marginal ranges.

For design rows $(1,x_1,x_2)$, the target vector is $x_*=(1,1.5,1.5)^\mathsf{T}$, and the design matrix gives $$ X^\mathsf{T}X= \begin{pmatrix} 4&2&2\\ 2&4&0\\ 2&0&4 \end{pmatrix}. $$ Solving $(X^\mathsf{T}X)v=x_*$ gives $v=(-0.25,0.50,0.50)^\mathsf{T}$: the second and third equations imply $v_2=v_3=0.375-0.5v_1$, and substitution into the first gives $v_1=-0.25$. Therefore $$ h_*=x_*^\mathsf{T}v =1(-0.25)+1.5(0.50)+1.5(0.50) =1.25. $$ The target lies outside the joint support and has leverage 1.25. There are only four observations for three regression coefficients, leaving one residual degree of freedom. The leverage is a geometric warning, not a basis for a stable uncertainty estimate from these sparse data.

How to Use the Diagnostics

Define the prediction target.
State the human covariates and physiological conditions that enter the model, including units, transformations, and any derived terms.

Check marginal ranges.
For each modeled covariate, compare the target with the observed animal minimum and maximum on the model scale. Report nonnegative distance or explicitly state a signed convention.

Check joint geometry.
Ask whether the target combination lies among the observed combinations. Consider plots or a model-based leverage calculation when several covariates act together.

Interpret in model context.
Review the functional form, data density near the target, influential observations, and whether the error structure is suitable. Treat a flag as a reason for scrutiny, not as an automatic rejection rule.

These checks complement, rather than replace, uncertainty analysis. A wide prediction interval can reflect uncertain parameters or residual variation even for an in-range target; conversely, a narrow model-based interval can coexist with serious extrapolation if the model's assumptions do not represent the human setting. Global sensitivity analysis, covered in the earlier material on sensitivity analysis, addresses how uncertain inputs affect predictions; range and leverage diagnostics address whether the target covariates are supported by the data geometry. They answer different questions.

Leverage thresholds should not be treated as universal pass-or-fail cutoffs for human prediction. The scale of leverage depends on the design, number of parameters, covariate coding, and target. A useful comparison is with leverage at observed design points and with typical leverage in the fitted data. In ordinary least squares, the average diagonal leverage for the observed design is $p/n$, where $p$ is the number of fitted coefficients and $n$ is the number of observations. A target leverage much larger than typical observed leverage merits attention, but context and model adequacy remain central.

Diagnostics, Alternatives, and Limitations

Range checks are transparent and easy to communicate, but can miss sparse regions and joint extrapolation. Convex-hull checks address whether a target is inside the smallest convex region containing the observed covariate vectors, but can be difficult to interpret in many dimensions and may be overly permissive if the hull spans large unsupported gaps. Leverage summarizes distance relative to the regression design and fitted terms; it can detect unusual combinations but is conditional on the specified model matrix. None of these diagnostics establishes biological similarity across species.

When data come from several species, repeated measurements or species-level dependencies can violate the independent-error assumption of the simple regression derivation. A leverage value computed from a pooled ordinary least-squares design may still describe its geometry, but the stated variance interpretation need not hold. The model's covariance and species structure must be respected for inferential uncertainty. Likewise, nonlinear models require diagnostics appropriate to their own design and estimation procedure; applying the linear formula mechanically is not justified.

Other failure modes include checking raw covariates when the model uses transformed covariates, omitting interaction or polynomial terms from the design vector, and allowing a large sample from one species to conceal poor coverage of other species. A target inside an observed range can still be weakly supported if most observations cluster at the boundaries or if animal species occupy different parts of the range. Always identify which observations define the support and make the support summary species-aware where relevant.

Common Pitfall

Wrong reasoning: “The human value lies within every animal covariate range, so the prediction is interpolation and is well supported.” This ignores the joint arrangement of covariates and the density of data around the target.

Correction: Treat marginal ranges as a first screen. Check joint covariate geometry and model leverage, then examine whether the fitted relationship and error assumptions are credible for that target. Also distinguish the nonnegative distance from a range from a signed measure: a distance cannot be negative, while a signed convention must specify which direction is positive.

Key takeaway: Extrapolation distance is a property of the target relative to the data representation and support. Report the covariate scale, distinguish marginal from joint support, and use leverage as a model-conditional geometric warning rather than a universal measure of predictive reliability.

Check Your Understanding

Answer each question, then compare with the solutions below.

Animal values for a modeled covariate range from 4 to 9. A human target is 11. What is its distance from the interval, and what is its signed distance if values above the interval are positive?
Why can a target be inside every marginal covariate range yet outside joint animal support?
In a linear regression with an intercept and one covariate, what does a high target leverage indicate, and what does it not establish?
A model uses the natural logarithm of body mass, but the analyst checks the target against the raw-mass range only. What is the problem?

Solutions. (1) The distance is $\max(4-11,0,11-9)=2$ covariate units; under the stated convention the signed distance is $11-9=+2$. (2) Marginal ranges ignore how covariates occur together. If animal observations occupy a restricted region of the multidimensional space, a target may be within each coordinate's minimum and maximum but outside that region. (3) High leverage indicates that the target design vector is geometrically unusual relative to the fitted design and that the fitted mean can be sensitive to the model fit there. It does not establish that the prediction is wrong, quantify all uncertainty, or assess biological validity. (4) The comparison uses a different scale from the fitted model. Transform the target and animal values using the natural logarithm before assessing the range; interpret any resulting gap on that transformed scale.