Why Combine Experimental Conditions?
A control experiment and a treatment experiment can each be informative about a QSP model, but not necessarily about the same parameter combinations. One condition may produce responses that change similarly when two parameters change, making their effects hard to distinguish. Another intervention may alter those responses in a different way. The central question is therefore not simply whether more observations are collected, but whether the combined conditions provide distinct, sufficiently precise information about the shared parameters.
The practical tool is to calculate each condition’s observation sensitivities and combine them with weights determined by measurement error. This gives a local assessment of how well the conditions jointly distinguish parameter changes. It complements the earlier material on Fisher information, replicates, sampling times, doses and perturbations, biomarkers, and sparse sampling: here the emphasis is on how information from different conditions fits together.
Assumptions and the Combined Information Matrix
Suppose condition \(c\) produces an observation vector \(y_c\) with mean \(f_c(\theta)\) and a known, parameter-independent error covariance matrix \(\Sigma_c\). Assume the observation errors are approximately multivariate normal and that the data from different conditions are independent. The covariance matrices may differ across conditions and may include within-condition correlation. In the simplest case, \(\Sigma_c\) is diagonal, with the observation-error variances on its diagonal.
Under these assumptions, the local Fisher information contributed by condition \(c\) is \(J_c^\mathsf{T}\Sigma_c^{-1}J_c\). For \(C\) independent conditions, the combined information is the sum of their contributions:
Equivalently, stack the observations, sensitivities, and covariance blocks:
The covariance weighting matters: a large predicted response change is not necessarily strong information if that observation is very noisy. If errors across conditions are correlated, the block-diagonal covariance assumption does not hold; the full covariance must be used instead. These equations also assume that the same \(\theta\) governs every condition. If a condition has its own unknown parameters, they must be represented in the parameter vector and accounted for rather than silently treated as shared.
Why Complementary Sensitivities Help
Each row of \(J_c\) describes the direction in parameter space to which one observation responds. If the rows from every condition point in nearly the same direction, the data may estimate one combination of parameters well while leaving another combination weakly determined. Adding a condition is especially useful when its sensitivities point in a direction that existing observations do not cover well. The rank and eigenvalues of the combined information matrix help reveal this: a zero eigenvalue indicates a locally uninformative parameter direction, while a small eigenvalue signals a weak direction.
The additive result follows from the likelihood under the independence assumption. The joint likelihood is the product of condition-specific likelihoods, so its log likelihood is their sum. The negative expected Hessian of a sum is the sum of the negative expected Hessians. For fixed-covariance Gaussian observations, the contribution from condition \(c\) is \(J_c^\mathsf{T}\Sigma_c^{-1}J_c\), giving the combined expression above.
There is also a useful precision implication. If the existing information matrix \(A\) is positive definite and a new independent condition contributes positive semidefinite information \(B\), then the new information is \(A+B\). Its approximate covariance is \((A+B)^{-1}\), which is no larger than \(A^{-1}\) in the positive-semidefinite ordering. To see why, write \(A+B=A^{1/2}(I+C)A^{1/2}\), where \(C=A^{-1/2}BA^{-1/2}\) is positive semidefinite. The eigenvalues of \(I-(I+C)^{-1}\) are nonnegative, so the added condition cannot increase the local covariance in this idealized setting. It may, however, add almost no information in a direction that it does not probe. This comparison presumes the model and error structure are correct and that the same parameterization is used.
Worked Example: A Treatment Condition Resolves a Parameter Trade-off
Consider two shared parameters, \(\theta_1\) and \(\theta_2\), and two hypothetical observations with independent errors of variance \(1\). At the parameter value used for the local calculation, the control condition has one observation with sensitivity row \([1,1]\). Its information matrix is
This matrix has rank 1: the control observation responds to the sum \(\theta_1+\theta_2\), but cannot distinguish changes in \(\theta_1\) from opposite changes in \(\theta_2\). Suppose a treatment condition adds a hypothetical observation with sensitivity row \([1,-1]\). Its information is
Adding the conditions gives
Under the local, known-variance approximation, each parameter has standard error \(\sqrt{0.5}\approx0.707\) in the parameter units used. The treatment condition adds a distinct sensitivity direction, making both parameters locally estimable in this two-observation example. This calculation is about expected local precision, not a guarantee that estimates from a particular noisy data set will be close to their true values.
Worked Example: Weight Information by Measurement Error
Suppose the shared parameter \(\theta\) is measured through a hypothetical control observation with sensitivity \(2\) and error variance \(4\). A treatment observation has sensitivity \(5\), but its error variance is \(100\). For a scalar parameter and one observation with sensitivity \(j\) and variance \(v\), information is \(j^2/v\).
The control contributes \(2^2/4=1\) unit of information. The treatment contributes \(5^2/100=0.25\). Together they contribute \(1+0.25=1.25\), giving approximate standard error \(1/\sqrt{1.25}\approx0.894\) parameter units. The control alone gives standard error \(1/\sqrt{1}=1\) parameter unit.
Although the treatment response is more sensitive to \(\theta\), its high observation variance makes its information contribution modest. Sensitivity magnitude alone is not an adequate basis for comparing conditions.
Worked Example: A Noisy Condition Adds Little in a Weak Direction
Consider two parameters and an existing set of observations with information matrix \(A=\begin{bmatrix}4&0\\0&0.25\end{bmatrix}\). The approximate standard errors are \(1/\sqrt{4}=0.5\) and \(1/\sqrt{0.25}=2\), respectively. A proposed condition contributes only \(B=\begin{bmatrix}0.01&0\\0&0\end{bmatrix}\). The combined matrix and standard errors are
This hypothetical condition barely changes precision for \(\theta_1\) and does not inform \(\theta_2\) at all. It is not enough for an intervention to be different in name or protocol: its weighted sensitivities must contribute information in parameter directions that matter.
How to Assess a Combined Design
A practical assessment begins by evaluating each condition’s model predictions and sensitivities at a scientifically plausible parameter value, often a fitted or nominal value. Use the anticipated observation-error covariance for each condition, then calculate and add the weighted information contributions. If parameters differ greatly in scale, raw matrix entries and eigenvalues can be misleading. Use scientifically meaningful parameter scales or dimensionless coordinates, such as log-parameters for positive parameters, and report the coordinate system used. Changing coordinates changes numerical values of the matrix; it should not change the underlying question about which parameter combinations the observations can distinguish.
List the parameters assumed to govern all conditions. Make condition-specific effects explicit if they are unknown rather than attributing their influence to shared parameters.
For each condition, calculate \(J_c\) and use its covariance \(\Sigma_c\) to form \(J_c^\mathsf{T}\Sigma_c^{-1}J_c\).
Check rank and weak eigen-directions in appropriate parameter scales. Ask which shared parameter combinations remain poorly determined after stacking.
Confirm that the anticipated errors, independence structure, observation model, and shared-parameter assumptions are defensible. Compare the local assessment with fitted-data uncertainty when available.
A near-zero eigenvalue, a very large approximate standard error, or a strong parameter correlation can indicate that conditions still leave a weak direction. These are warnings rather than definitive global conclusions: sensitivities are local, and they can change as parameters move. A numerical derivative that changes substantially with step size can also make a sensitivity-based assessment unreliable. Check derivatives and model predictions before interpreting a matrix.
Alternatives, Limitations, and a Common Pitfall
Combining conditions is related to, but distinct from, several other ways to strengthen an experiment. More replicates in one condition can reduce noise in that condition’s directions, while a new condition can supply different directions. Changing sampling times, doses, or measured biomarkers may also create complementary sensitivities. The matrix calculation gives a common local language for comparing these contributions, but it does not decide the scientific value of an intervention or account automatically for its cost.
| Approach | What it changes | Potential strength | Main caution |
|---|---|---|---|
| Add replicates | More observations under an existing condition | Reduces uncertainty along directions already informed | May leave the same parameter trade-off unresolved |
| Add a distinct condition | Model sensitivities through a different intervention | Can add a new direction and separate parameter effects | Benefit depends on weighted sensitivities and correct shared-parameter assumptions |
| Use a local information calculation | Summarizes sensitivity and error assumptions near a parameter value | Fast comparison of local precision | May not represent nonlinear, bounded, or multimodal uncertainty |
| Examine likelihood-based uncertainty | Uses the fitted likelihood over parameter values | Can reveal non-quadratic or asymmetric uncertainty | Requires fitting and evaluating the likelihood, and may be computationally demanding |
A common mistake is to argue, “The treatment produces a larger response, so it must improve estimability substantially.” The correction is to evaluate the derivative of the predicted observations with respect to the parameters and weight it by measurement error. A large response can have little information if it barely changes with the parameter, is measured imprecisely, or changes in the same parameter direction as existing observations.
A further limitation is that the local covariance approximation relies on a well-behaved, approximately quadratic likelihood near the parameter value and a nonsingular information matrix. It can be misleading near parameter bounds or in strongly nonlinear models. Model misspecification, condition-specific shifts, or unmodeled cross-condition correlation can also make the apparent gain optimistic. When these issues are plausible, the combined matrix is a screening and interpretation tool, not a substitute for checking uncertainty under the fitted model.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A new independent condition contributes a positive semidefinite information matrix. Under what assumptions does adding it avoid increasing the local covariance of shared parameters?
- Two conditions have identical sensitivity rows but different measurement-error variances. What determines which condition contributes more information?
- A combined information matrix has rank one for two shared parameters. What does this imply about local estimability?
- Why might a condition-specific unknown offset need to be included in the parameter vector?
Solutions. (1) The parameterization and shared-parameter model must be consistent; the information matrices must describe the same parameters; the new data must be independent of the existing data, with correctly specified error covariance; and the existing information matrix must be positive definite for the stated inverse-covariance comparison. (2) The sensitivity is weighted by the inverse error variance; for one observation, information is the squared sensitivity divided by its variance. (3) At least one parameter direction has zero local information, so the two parameters are not both locally distinguishable from the observations represented by that matrix. (4) If the offset is unknown and differs by condition, changes in it could otherwise be mistaken for changes in shared parameters, overstating how informative the condition is about those shared parameters.