Combining Conflicting Parameter Estimates with Uncertainty

Why Conflicting Estimates Need More Than Averaging

PBPK parameters are often supported by several measurements that do not agree exactly. One study may report a tissue partition coefficient of 8, another 10, and a third 15. The spread could reflect sampling error, assay calibration, differences in the populations or conditions studied, or genuine variation between study settings. Simply averaging the values treats these explanations as interchangeable and can produce a precise-looking number that the evidence does not justify.

The first question is not how to average, but whether the estimates refer to the same parameter on the same scale under sufficiently comparable conditions. The earlier material on measurement error, biological variability, and knowledge uncertainty helps distinguish these sources conceptually. Here, the practical task is to combine compatible evidence—or explain why it should remain separate—and to describe what uncertainty remains.

Check Comparability Before Choosing Weights

A numerical estimate is interpretable only in relation to its measurand: what was measured, in what material, under what conditions, using what method, and on what scale. Before pooling, check that units and transformations are aligned; that the parameter definitions match; and that assay procedures, sample handling, and relevant population or experimental conditions are sufficiently comparable. For example, an estimate of unbound tissue concentration cannot be pooled directly with one for total concentration just because both are reported in the same concentration units.

An assay difference is not automatically a reason to discard a study, but it must be considered. If methods have a known calibration relationship, estimates may be transformed to a common scale, with uncertainty in that transformation accounted for where possible. If a method has a suspected systematic bias that cannot be quantified, inverse-variance weights do not correct it: they only reflect stated precision. If the studies target materially different conditions, separate condition-specific estimates or a transparent range may be more defensible than a single pooled value.

The standard formulas below assume estimates are unbiased for the quantity they claim to measure, their reported standard errors are appropriate, and studies are independent unless dependence is explicitly represented. Estimates derived from overlapping samples or shared calibration data may be correlated. Treating them as independent can make the combined standard error too small.

Fixed-Effect Combination: Precision Within a Common Target

When a common underlying value is scientifically reasonable, estimates with smaller standard errors receive more weight. The weight for study \(i\) is the inverse of its within-study variance, \(w_i=1/v_i=1/SE_i^2\). The combined estimate is the weighted mean, and its standard error follows from the variance of a weighted sum of independent estimates.

The interval \(\hat{\theta}_{FE}\pm1.96\,SE(\hat{\theta}_{FE})\) is a conventional approximate 95% confidence interval when the normal approximation is adequate. Its coverage interpretation relies on the common-target model and the standard-error assumptions; it is not a probability statement that the fixed parameter lies in the calculated interval.

When Study Results Differ Beyond Their Stated Precision

The fixed-effect model does not allow genuine study-to-study differences in underlying values. A simple diagnostic compares the dispersion of study estimates around the fixed-effect summary with what their within-study variances predict. Define \(Q=\sum_i w_i(y_i-\hat{\theta}_{FE})^2\). Under the common-target model and approximate normality, \(Q\) is compared with a chi-squared distribution with \(k-1\) degrees of freedom, where \(k\) is the number of studies. A large value relative to that reference suggests that the observed differences are greater than the reported within-study errors alone would typically explain.

This diagnostic is not a test of whether pooling is scientifically appropriate. It can have low power when few studies are available, and a large value does not identify the cause of disagreement. Examine assay procedures, parameter definitions, populations, and conditions as well as the statistic. Do not interpret a non-large \(Q\) as proof that all studies are equivalent.

Random Effects: Representing Between-Study Heterogeneity

When studies plausibly differ in their underlying values, a random-effects model represents study \(i\)'s underlying value as varying around an overall mean \(\mu\), with between-study variance \(\tau^2\). In the usual independent-study formulation, the variance of an estimate around that mean is \(v_i+\tau^2\). The random-effects weight is therefore \(w_i^*=1/(v_i+\tau^2)\). The combined estimate describes the mean of the study-level distribution under this model, not necessarily a universal value that applies to every population, assay, or condition.

The between-study variance \(\tau^2\) must be estimated or otherwise specified; it is not an extra precision adjustment chosen just to make the studies agree. Different estimation methods can yield different values, especially with few studies. A conventional approximate 95% interval uses \(\hat{\mu}_{RE}\pm1.96\,SE(\hat{\mu}_{RE})\), but this interval is conditional on the model and the treatment of \(\tau^2\). With few studies, uncertainty in \(\tau^2\) can make this interval overly optimistic.

Justification for the Weighting Formulas

For a fixed-effect model, suppose \(y_i=\theta+\varepsilon_i\), where the independent errors have mean zero and variances \(v_i\). Consider a linear estimator \(\hat{\theta}=\sum_i a_i y_i\) with \(\sum_i a_i=1\), so it is unbiased for \(\theta\). Its variance is \(\sum_i a_i^2v_i\). Minimizing that variance subject to the sum-to-one constraint gives weights proportional to \(1/v_i\); normalizing them gives \(a_i=w_i/\sum_jw_j\).

Consequently,

For random effects, the same argument applies after replacing each variance \(v_i\) by \(v_i+\tau^2\). The calculation explains why inverse-variance weighting is efficient under its assumptions; it does not establish that those assumptions hold for a particular collection of assays.

Practical Decisions, Diagnostics, and Limitations

Useful warning signs include a large dispersion statistic, estimates that shift substantially when one study is omitted, markedly different assay procedures, unusually narrow standard errors, or a pooled value that conflicts with biological plausibility checks. Sensitivity analysis can reveal dependence on a study or modeling choice; it cannot resolve an unidentified systematic bias. A random-effects interval for the overall mean also does not describe the full spread of study-specific values. A prediction interval would address a different target and depends on the heterogeneity model and its uncertainty.

When estimates are correlated, the independent-study formulas above are not appropriate without adjustment. In general, if the vector of estimates has covariance matrix \(V\), the generalized least-squares estimate under a common mean is \((\mathbf{1}^{T}V^{-1}\mathbf{y})/(\mathbf{1}^{T}V^{-1}\mathbf{1})\), with variance \(1/(\mathbf{1}^{T}V^{-1}\mathbf{1})\). If the covariance is unknown, reporting a simple independent-study result as definitive is not justified.

Common Pitfall: Precision Is Not the Same as Credibility

A common mistake is to say, “The estimate with the smallest standard error is the best one, so it should dominate the combined value.” The correction is that inverse-variance weighting is justified only when estimates target the same quantity, are unbiased on a common scale, and have valid uncertainty estimates. A precise assay with a systematic calibration error can receive a large weight and pull the pooled estimate away from the intended parameter. Conversely, a less precise but well-matched study may contain important information about the target.

Another mistake is to report a narrow pooled interval while ignoring substantial disagreement or assay differences. The interval quantifies uncertainty under a specified model; it does not absorb all uncertainty about whether that model is appropriate. When evidence cannot be reconciled, retaining separate estimates, explaining the discrepancy, and giving a suitably cautious range may communicate more than a single summary.