When Simultaneous Estimation Does Not Improve Precision

Why a Joint Fit May Add Little Precision

Simultaneous estimation can allow pharmacodynamic (PD) observations to inform pharmacokinetic (PK) parameters that matter for the exposure–response relationship. That information transfer can improve precision, as the earlier material on joint estimation explains. But a joint fit does not guarantee a material gain. If PD data barely inform PK parameters, if PK data already estimate them very precisely, or if the relevant PK and PD parameters are nearly independent, joint estimation may change standard errors very little.

The comparison must be made carefully. A joint fit should not be credited with a precision gain merely because a staged analysis fixed PK estimates and ignored their uncertainty. A staged analysis that properly carries PK uncertainty into the PD estimates is a fairer comparator. Under the same correctly specified likelihood, a joint estimator may use the data more efficiently than a particular staged estimator, but it cannot create information absent from the data. When the connection between PK and PD parameters is weak, even that efficiency difference may be negligible.

Information That Can Be Borrowed

Consider one PK parameter \(\theta\) and one PD parameter \(\psi\). Near the true parameter values, suppose the PK data contribute information \(A_0\) about \(\theta\), and the PD data contribute information \(C\) about \(\theta\), cross-information \(B\) between \(\theta\) and \(\psi\), and information \(D\) about \(\psi\). The joint Fisher information matrix is

Information is local: it describes how sharply the likelihood changes near the parameter values. The cross-information \(B\) captures how changes in \(\theta\) and \(\psi\) can produce similar changes in the PD likelihood. If \(B\) is small relative to the diagonal information, the parameters are nearly orthogonal in the local likelihood, and estimating them together will generally have little effect on either variance. A well-conditioned, positive-definite information matrix is also required for the inverse-information variance approximation to be meaningful.

For a staged procedure, suppose \(\theta\) is estimated from PK data alone, then \(\psi\) is estimated from the PD data while using that PK estimate. In the local asymptotic approximation, the staged variance for \(\psi\) is

The first term is the variance if \(\theta\) were known. The second is the extra variance from estimating \(\theta\) using PK data with information \(A_0\), then carrying that uncertainty into the PD estimate. This expression assumes the PK and PD data contributions are independent conditional on the model parameters, and that the local likelihood approximation is adequate. It is not the variance of a staged analysis that simply treats \(\hat{\theta}\) as known.

For the joint fit, the approximate variance of \(\hat{\psi}\) is the lower-right diagonal element of \(\mathcal{I}^{-1}\). Inverting the two-by-two matrix gives

The term \(B^2/A\) is the information about \(\psi\) lost to uncertainty in \(\theta\). If \(B\) is small, this loss is small. If PK information \(A_0\) is already large, PD information about \(\theta\) often changes the effective information only slightly. These are distinct reasons for a small gain.

Deriving the Comparison

The formulas also clarify why joint estimation can help a little without necessarily transforming precision. In a staged analysis, the uncertainty in \(\theta\) comes from PK information \(A_0\). The joint fit uses total information \(A=A_0+C\) about \(\theta\), including the contribution from PD data. When the PD likelihood connects \(\theta\) to \(\psi\), this additional information can reduce the variance contribution associated with estimating \(\theta\).

To see the joint result, the determinant of the information matrix is \(AD-B^2\), so its inverse is

Thus the variance for \(\psi\) is \(A/(AD-B^2)\), equivalently \(1/(D-B^2/A)\). This calculation requires \(A>0\) and \(AD-B^2>0\); otherwise the local information matrix is not positive definite and the variance approximation is not defined as a finite positive quantity. The result is a local, large-sample approximation, not a guarantee about finite-sample standard errors.

A useful limiting check is \(B=0\). Then the joint variance is \(1/D\): uncertainty in \(\theta\) does not affect the local precision of \(\psi\), and adding PK information to a joint fit cannot improve \(\psi\)'s precision through this coupling. Another limiting check is very large \(A_0\). The staged uncertainty term \(B^2/(D^2 A_0)\) approaches zero; if PK parameters are already known precisely, there is little uncertainty left for the PD data to resolve.

Diagnosing a Small Gain

A small standard-error difference is interpretable only if the fit and comparison are credible. Examine the estimated information or covariance structure, the model's local parameter correlations, and the stability of the estimates under reasonable changes in starting values. Information-matrix calculations can be distorted by poor scaling: a numerical cross-term \(B\) cannot be judged as “small” without considering its size relative to \(A\) and \(D\). Parameter rescaling changes the entries, although it does not change the underlying inferential question.

Check that both procedures use the same observations, likelihood assumptions, estimand, and treatment of PK uncertainty. A staged standard error calculated as \(1/\sqrt{D}\) assumes PK parameters are known; comparing it with a joint standard error answers a different question from comparing joint estimation with uncertainty-aware staging. Also check whether the joint fit converged to a stable solution and whether the information matrix is positive definite. An apparently tiny gain accompanied by unstable estimates, a nearly singular matrix, or substantial sensitivity to starting values is not reassuring evidence of independence.

Alternatives, Pitfalls, and Limits

The main alternatives are a staged fit that fixes PK estimates, a staged fit that propagates PK uncertainty, and a simultaneous fit. A fixed-PK staged fit can be convenient but typically understates uncertainty when PK estimates are imprecise. Uncertainty-aware staging avoids that particular omission, but may not use PD information to refine PK parameters. A joint fit can use the shared likelihood structure, yet it is not automatically more precise for every parameter: its advantage depends on parameter coupling and on which staged estimator is the comparator.

Common mistake: “The joint fit has almost the same standard error as the staged fit, so PK uncertainty does not matter.” This conclusion does not follow if the staged standard error treated PK estimates as fixed. The apparent agreement could instead reflect an invalid comparison. First calculate or otherwise account for uncertainty propagated from PK into PD. Only then can a small difference support the conclusion that joint estimation adds little precision for the target.

The information formulas are local approximations. They can be unreliable with small samples, strong nonlinearity, weak identification, boundary estimates, or model misspecification. A near-zero cross-information term at one parameter value does not establish independence throughout the parameter space. Nor does a small precision gain establish that the PK or PD model is scientifically adequate. These formulas diagnose where efficiency gains may be limited; they do not replace model checking or a design-specific comparison.