Building a Joint Calibration Objective

Why a Joint Objective Needs More Than One Residual Formula

A QSP model may be calibrated against data from several assays and studies. One source may report a continuous concentration with approximately additive measurement error; another may report a positive biomarker with multiplicative error; a third may record responders out of a fixed number of subjects. These observations do not share a common raw scale or necessarily follow the same probability model. Adding their raw residuals, or squaring every raw residual, therefore has no general statistical justification.

A joint calibration objective combines evidence by first expressing each source’s observations under an appropriate observation model, then combining the resulting likelihood contributions. The model prediction for source \(s\), observation \(i\), is written \(f_{si}(\theta)\), where \(\theta\) denotes the biological model parameters being calibrated. The observation model specifies how the recorded data vary around that prediction. This tutorial assumes the data definitions, units, replicate structure, and source weights have already been considered; the task here is to formulate the combined objective consistently.

From Source Models to a Joint Objective

Let \(y_s\) denote all observations from source \(s\), and let \(L_s(\theta)\) be their likelihood under the QSP predictions and that source’s observation model. If the sources are conditionally independent given the model parameters, then

The independence condition concerns the observations after accounting for the modeled quantities and any specified dependence. It does not mean observations within a source must be independent. Within-source correlation can be represented in \(L_s\), for example by a covariance matrix. If sources share dependent observations or errors, their likelihoods cannot simply be multiplied as though independent; a joint distribution must represent that dependence.

For a source with a multivariate normal observation model, let \(r_s=y_s-f_s(\theta)\) be the residual vector and let \(\Sigma_s\) be its covariance matrix. Then

If \(\Sigma_s\) is fixed, the last two terms do not change with \(\theta\), so minimizing the source contribution is equivalent to minimizing \(r_s^{\mathsf T}\Sigma_s^{-1}r_s\). If covariance parameters are estimated, the log-determinant term generally depends on those parameters and cannot be dropped. The quadratic term is dimensionless: residuals are measured relative to their variances and correlations, rather than being compared in raw units.

Other observation models lead to different contributions. For positive measurements with multiplicative error, a lognormal model is one standard option: \(\log Y_{si}\sim N(\log m_{si}(\theta),\tau_s^2)\), where \(m_{si}(\theta)\) is the median on the original scale. Its residual is \(\log y_{si}-\log m_{si}(\theta)=\log(y_{si}/m_{si}(\theta))\), not \(y_{si}-m_{si}(\theta)\). For a binomial outcome with \(k_s\) responses among \(N_s\) individuals and model probability \(p_s(\theta)\), the likelihood is proportional to \(p_s(\theta)^{k_s}[1-p_s(\theta)]^{N_s-k_s}\). Each model preserves the observation scale and error assumptions relevant to its source.

Worked Calculations

Why the Likelihood Sum Works

The key justification is the logarithm’s conversion of products into sums. If the sources are conditionally independent, their joint probability density or mass function is \(\prod_s L_s(\theta)\). Taking minus twice its logarithm gives \(-2\sum_s\log L_s(\theta)=\sum_s[-2\log L_s(\theta)]\). Each term retains its source’s distribution and measurement scale; the sum is possible because each term is a log-likelihood contribution, not because the original measurements have the same units.

For independent normal observations with known standard deviations \(\sigma_{si}\), the parameter-dependent part reduces to \(\sum_i[(y_{si}-f_{si}(\theta))/\sigma_{si}]^2\). This follows by substituting the normal density into \(-2\log L\); its remaining terms are fixed when the standard deviations are known. For a multivariate normal source, the same argument yields the covariance-weighted quadratic form. These forms are special cases of a likelihood-based objective, not a rule to square every kind of residual.

If a source weight \(a_s\) has been justified in advance, a weighted objective can be written \(J(\theta)=\sum_s a_s[-2\log L_s(\theta)]\), with \(a_s\geq0\). Such weighting changes the influence of a source; it is not the same thing as using its measurement variance, which already appears inside its likelihood. A weighted likelihood is generally a chosen calibration criterion rather than the ordinary likelihood for independent observations. Its interpretation depends on the rationale and scale of the weights. Do not add weights merely because the component objectives appear numerically different.

Diagnostics, Alternatives, and Limitations

A numerically calculable objective can still be poorly specified. Inspect source-specific contributions and residuals on the scales implied by their observation models. For Gaussian data, standardized residuals should be examined relative to the assumed variance and correlation. For lognormal data, inspect log-scale residuals and verify all observations and predictions used in log ratios are positive. For binomial data, retain both the response count and the denominator; a proportion alone does not describe its sampling precision. Systematic residual patterns may indicate a model or observation-model problem rather than a need to rescale the objective.

Several warning signs merit investigation: one source dominates because its variance was entered in the wrong units; a log-scale model was applied to zero or negative values; a correlated series was treated as independent; a source was counted twice; or an estimated variance was optimized while its log-determinant term was omitted. Also check whether the source weights are being applied consistently and whether a reported total can be traced back to its source contributions. A single total objective hides which source is driving a calibration.

The main alternative is to minimize a raw sum of squared residuals. That approach is appropriate only under restrictive conditions, such as comparable measurement scales and a common, constant Gaussian error variance, or after a defensible standardization that corresponds to the assumed error model. An alternative is a weighted least-squares objective; it can be a useful approximation when residuals are approximately Gaussian and weights reflect their variances, but it does not automatically represent count likelihoods, lognormal errors, or parameter-dependent covariance. The full likelihood is more general, although it too depends on correctly specifying the observation distributions.

The objective does not guarantee that the QSP model is identifiable, that the assumed observation models are adequate, or that a good numerical minimum is scientifically credible. It also does not resolve conflicts between sources by itself. A low total may conceal poor fit to one source, especially when source weights differ. Keep the component contributions and source-specific diagnostics available when assessing a fitted model.

Common Pitfall

Wrong reasoning: “All measurements are numeric, so I can subtract every observation from its prediction, square the differences, and add them. If one source looks too influential, I can divide its residuals by an arbitrary constant.”

Correction: A residual’s contribution depends on the data-generating error model. Standardize additive normal errors by their standard deviations; use log residuals for a specified lognormal model; use the count likelihood for binomial data; and encode correlations through the covariance matrix. Any additional source weight needs a stated rationale and should not replace the variance or likelihood already required by the source model.