Why Panel Composition Matters
Two biological mechanisms can predict different values for a biomarker and still be difficult to distinguish in practice. The predicted difference may be small compared with measurement error, or a second biomarker may largely repeat information already present in the first. Conversely, markers with distinct response times can provide complementary evidence when measured at appropriate times. Panel design therefore asks not just which markers change, but how much the joint observations separate the competing predictions.
The comparison below assumes that two candidate mechanisms, \(M_0\) and \(M_1\), have already been specified for the same experimental context: the same perturbation, dose, population, and observation schedule. For a proposed panel, each mechanism predicts a vector of measurements. This tutorial compares those predictive distributions; it does not treat a large predicted difference as proof that either mechanism is biologically correct.
The covariance matrix must represent uncertainty on the scale of the measurements being compared. Depending on the prediction being assessed, it can include measurement error and other modeled predictive variation. If reported measurement errors are independent, their variances contribute to the diagonal. Correlated errors or shared biological variation can also create off-diagonal covariance. The covariance should match the intended data-generating and measurement process; treating it as diagonal simply for convenience can exaggerate panel separation.
Redundancy and the Value of an Additional Marker
The inverse covariance in \(D^2\) adjusts for redundancy. If two markers tend to rise and fall together, their joint observation contains less independent evidence than two equally precise, uncorrelated markers. A useful way to see this is to calculate the increment from adding one marker \(Y_j\) to an existing panel \(Y_S\).
This decomposition explains why a marker’s marginal mean difference is not enough to judge its value. If the added marker closely tracks the existing panel, much of its mechanistic difference may already be predictable from that panel. The residual difference \(r_{j\mid S}\), rather than the raw difference \(\delta_j\), measures what remains to distinguish the mechanisms after accounting for the existing markers.
Derivation and Likelihood Interpretation
The decomposition follows from the conditional distribution of a multivariate normal vector. Under either mechanism, the conditional variance of \(Y_j\) given \(Y_S\) is \(v_{j\mid S}\). At a fixed observed \(y_S\), the conditional mean under mechanism \(k\) is $$ \mu_{k,j}+\Sigma_{jS}\Sigma_{SS}^{-1}(y_S-\mu_{k,S}),\qquad k\in\{0,1\}. $$ Subtracting the conditional means gives \(r_{j\mid S}\), which does not depend on \(y_S\). To justify the separation formula, factor the joint normal density into the marginal density of \(Y_S\) and the conditional density of \(Y_j\mid Y_S\). Equivalently, the squared Mahalanobis distance is the sum of the squared standardized marginal and conditional mean differences. The marginal term is \(D^2_S\), while the conditional term is \(r_{j\mid S}^2/v_{j\mid S}\), yielding the stated result.
This is a decomposition of squared Mahalanobis separation, not a claim that the conditional log-likelihood ratio equals \(r_{j\mid S}^2/v_{j\mid S}\). In fact, let \(m_0\) and \(m_1\) be the two conditional means at the observed \(y_S\), and let \(y_j\) be the observed added measurement. The conditional log-likelihood ratio is $$ \log\frac{f_1(y_j\mid y_S)}{f_0(y_j\mid y_S)} =\frac{r_{j\mid S}}{v_{j\mid S}}\left(y_j-\frac{m_0+m_1}{2}\right). $$ It varies with the observed value \(y_j\). Its expectation under mechanism 1 is \(r_{j\mid S}^2/(2v_{j\mid S})\); that expected log-likelihood ratio is a conditional Kullback–Leibler divergence. The squared-separation increment is twice that expected value. Keeping these quantities distinct prevents a constant separation score from being mistaken for an observation-specific likelihood ratio.
Worked Example: A Correlated Marker Adds Only Modest Separation
Suppose two hypothetical mechanisms predict two biomarkers with mean difference \(\delta=(2,3)^\mathsf{T}\), in their respective concentration units. The common covariance matrix is $$ \Sigma=\begin{pmatrix}4&3\\3&9\end{pmatrix}. $$ Its diagonal entries are variances, and its positive off-diagonal entry represents correlated variation. The determinant is \(4(9)-3(3)=27\), so $$ \Sigma^{-1}=\frac{1}{27}\begin{pmatrix}9&-3\\-3&4\end{pmatrix}. $$
The two-marker separation is $$ D^2=\delta^\mathsf{T}\Sigma^{-1}\delta =\frac{1}{27}\left(9(2^2)-3(2)(3)-3(3)(2)+4(3^2)\right) =\frac{36}{27}=1.333. $$
For the first marker alone, the separation is \(D^2_S=2^2/4=1\). For the second marker, conditional on the first, the variance and residual mean difference are $$ v_{2\mid 1}=9-\frac{3^2}{4}=6.75,\qquad r_{2\mid 1}=3-\frac{3}{4}(2)=1.5. $$ Thus the increment from adding marker 2 is \(1.5^2/6.75=0.3333\), and the total is \(1+0.3333=1.333\). The second marker has a raw mean difference of 3, but much of its variation and mechanistic difference overlaps with the first marker. It adds separation, but substantially less than its marginal difference alone might suggest.
Measurement Error and Comparing Candidate Panels
Measurement error reduces separation because it increases the variance relative to the between-mechanism difference. It also matters in combination with correlation: a noisy marker may still help if its residual difference is distinct, while a precise marker may add little if its information is redundant. When panels have different numbers of markers, the same \(D^2\) criterion can be calculated for each panel, provided each covariance matrix reflects the measurement properties of that panel and the same target context.
Worked Example: A Noisy Marker Can Be Less Useful Than a Smaller Difference Suggests
Consider two hypothetical candidate markers measured independently. Marker A has a predicted mean difference of 2 concentration units and a measurement standard deviation of 1 unit under either mechanism. Marker B has a mean difference of 1 concentration unit and a measurement standard deviation of 0.5 units. Under the stated independent, equal-variance assumptions, the panel covariance is diagonal and the total squared separation is the sum of the individual contributions.
For marker A, \(D_A^2=2^2/1^2=4\). For marker B, \(D_B^2=1^2/0.5^2=4\). Despite their different raw differences and precision, each provides the same standardized separation. Together, independence gives \(D^2=4+4=8\).
Now suppose marker B instead has a measurement standard deviation of 2 units. Its contribution becomes \(1^2/2^2=0.25\), so the two-marker panel has \(D^2=4.25\). Marker B still adds separation, but its contribution is small relative to marker A. This comparison depends on the stipulated variances being appropriate: if the errors were correlated, the off-diagonal covariance would have to be included rather than adding the separate scores.
For equal prior probabilities and equal covariance under two multivariate normal mechanisms, a related interpretation is available: the optimal classifier that assigns an observation to the closer predictive mean in the covariance-adjusted metric has error probability \(\Phi(-D/2)\), where \(\Phi\) is the standard normal cumulative distribution function. This is a model-based classification result, not a guarantee of real-world performance. Unequal prior probabilities, unequal covariance matrices, misspecified predictions, or a different decision loss can change the appropriate comparison. In particular, \(D^2\) should not be reported as though it were a probability of choosing the correct mechanism.
Biomarker Response Times and Sampling Schedules
A panel is defined by both which biomarkers are measured and when they are measured. A biomarker that responds quickly may separate mechanisms at an early time and converge later; a delayed marker may show the reverse. Comparing panel options at an arbitrary common time can therefore favor the wrong marker or overlook a complementary combination. Each candidate measurement should be matched to the schedule at which its predictions are being compared.
Worked Example: Pairing Each Biomarker With Its Discriminating Time
Suppose two hypothetical biomarkers, A and B, are measured with independent errors of standard deviation 1 unit at either 1 hour or 8 hours. The following are the predicted differences between two mechanisms at those times:
| Measurement | Predicted mean difference | Standard deviation |
|---|---|---|
| A at 1 hour | 3 units | 1 unit |
| A at 8 hours | 0.5 units | 1 unit |
| B at 1 hour | 0.5 units | 1 unit |
| B at 8 hours | 2 units | 1 unit |
If both biomarkers are measured at 1 hour, independence gives \(D^2=3^2+0.5^2=9.25\). If both are measured at 8 hours, \(D^2=0.5^2+2^2=4.25\). But measuring A at 1 hour and B at 8 hours gives \(D^2=3^2+2^2=13\). Measuring A at 8 hours and B at 1 hour gives \(D^2=0.5^2+0.5^2=0.5\). Thus the same two biomarkers can yield very different discrimination depending on timing. These calculations assume independence across the selected measurements; if repeated measurements or shared processes induce covariance, the full covariance matrix must be used.
Diagnostics, Alternatives, and Limitations
Before comparing panels, check that the mechanistic predictions refer to the same intervention, population, units, and sampling schedule. Confirm that covariance estimates are plausible, symmetric, and positive definite. Strong correlations, very small conditional variances, or a covariance matrix close to singularity can make \(D^2\) unstable: small changes in the assumed covariance may cause large changes in its inverse. Check sensitivity to plausible measurement-error and correlation values, and inspect residual patterns or predictive checks when observations are available. Poor predictive fit means the panel score may be precise about the wrong predictive distributions.
The equal-covariance normal framework is useful because it gives a transparent comparison and an exact redundancy decomposition. It is not universally appropriate. If mechanisms predict different covariances, a single Mahalanobis distance between means omits differences in spread and shape. If predictions are strongly non-normal or multimodal, a mean-and-covariance summary can conceal important distributional structure. In such cases, compare the full predictive distributions using a criterion suited to the question, or simulate observations and evaluate classification or decision performance under stated assumptions. Those comparisons require specifying the distributions and decision rule; they cannot be recovered from mean differences alone.
Other simple approaches answer narrower questions. Ranking markers by absolute mean difference ignores units and precision. Ranking by a single-marker standardized difference accounts for marginal variance but ignores redundancy in a joint panel. A covariance-adjusted \(D^2\) accounts for both under its assumptions, but remains a summary of predictive separation, not a complete experimental-design objective. Practical constraints also matter: assay availability, sample volume, burden, cost, and timing feasibility can make a theoretically strong panel unusable. The earlier material on choosing sampling times and selecting perturbations addresses complementary design choices; here, the focus is on the panel and its joint predictive separation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A single marker has a predicted mechanism difference of 1.2 units and a common standard deviation of 0.6 units. What is its \(D^2\)?
- An added marker has conditional mean difference \(r_{j\mid S}=2\) units and conditional variance \(v_{j\mid S}=4\) square units. What is its increment to \(D^2\)?
- Why can a marker with a larger raw mean difference add less panel separation than a marker with a smaller difference?
- Under what assumption can separate single-marker \(D^2\) values be added to obtain a panel \(D^2\)?
- Why does the conditional log-likelihood ratio differ from the conditional increment in squared Mahalanobis separation?
Solutions. (1) \(D^2=1.2^2/0.6^2=1.44/0.36=4\). (2) The increment is \(2^2/4=1\). (3) The larger-difference marker may have greater measurement or predictive variance, or its difference may be largely redundant with the existing panel; the conditional residual difference and variance determine its increment. (4) The measurements must have independent errors and no other covariance among them, with the stipulated common variances. (5) The log-likelihood ratio depends on the observed measurement and is linear in its deviation from the midpoint of the conditional means; the separation increment is the constant \(r_{j\mid S}^2/v_{j\mid S}\). Under the added mechanism its expected conditional log-likelihood ratio is half that increment.