Comparing Standard Errors From Staged and Joint Analyses

Why Standard Errors Can Differ

A staged analysis may estimate pharmacokinetic (PK) parameters first, then hold those estimates fixed while estimating pharmacodynamic (PD) parameters. A joint analysis estimates both sets together. Even when the point estimates are similar, their reported standard errors need not be: a PD standard error calculated with PK parameters fixed describes uncertainty conditional on those PK estimates. It does not automatically include the uncertainty in estimating PK parameters or their association with PD estimates.

The distinction matters when interpreting precision. A narrow conditional standard error is not necessarily evidence that the data determine the PD parameter precisely overall. Conversely, it is not correct to assume that every staged standard error is too small by a fixed amount: the size and direction of covariance contributions depend on the model and parameter of interest. The goal is to identify what each standard error measures and compare its assumptions with the analysis actually performed.

These standard errors are generally local, large-sample approximations. Their usual interpretation relies on a correctly specified likelihood or estimating model, an identifiable parameterization, a well-behaved optimum, and enough information for a quadratic approximation around that optimum to be useful. If the likelihood is flat, the optimum is on a boundary, or the model is misspecified, inverse curvature alone may give a misleading impression of precision.

How the Joint Information Matrix Carries Covariance

Let \(\theta\) denote PK parameters and \(\eta\) denote PD parameters. Near a regular joint maximum-likelihood estimate, partition the information matrix into blocks:

Here \(A\) describes information about \(\theta\), \(D\) describes information about \(\eta\), and \(B\) describes how the likelihood couples the two parameter sets. Under the local quadratic approximation, the joint covariance estimate is \(\mathcal{I}^{-1}\). Treating \(\theta\) as fixed instead uses \(D^{-1}\) for the conditional covariance of \(\eta\), because the PK coordinates are not allowed to vary.

When \(A\) and the relevant Schur complement are invertible, the PD block of the joint covariance is:

The subtracted term represents information about \(\eta\) that is not available once uncertainty in \(\theta\) is allowed for. Under a positive-definite information matrix, \(B^{T}A^{-1}B\) is positive semidefinite. Thus the joint PD covariance is at least as large as \(D^{-1}\) in the positive-semidefinite ordering, when both are calculated from this same local information matrix. This comparison does not establish a universal ordering between standard errors from arbitrary staged and joint procedures: they may use different likelihoods, data, assumptions, or estimators.

For a scalar PD parameter, the omitted contribution is especially easy to see. For a derived PD quantity \(g(\theta,\eta)\), a first-order delta-method approximation uses its gradient with the full covariance matrix:

Here \(V_{\theta}\), \(V_{\eta}\), and \(V_{\theta\eta}\) are the PK, PD, and cross-covariance blocks, and the derivatives are evaluated at the estimates. The cross term can be positive or negative. Dropping it is not generally conservative, and dropping the PK term is justified only if PK uncertainty is negligible for the quantity at hand or is deliberately outside the target of inference.

Propagating Uncertainty in Staged Results

A staged workflow does not have to stop at the conditional standard error. If the final estimate depends on an estimated PK quantity, its uncertainty can be propagated. One route is a delta method using the covariance of the complete staged estimator. Another is to formulate the PK and PD estimating equations together and calculate a sandwich covariance. In either case, a covariance matrix for the final estimators is needed; the conditional PD Hessian alone does not supply it.

For a two-stage estimator, the second-stage PD estimate responds to changes in the first-stage PK estimate. Its sampling variation therefore contains both first-stage variation and second-stage variation. A general way to represent this is to stack the estimating equations, calculate their sensitivity to all parameters, and combine that sensitivity with the covariance of the estimating contributions. This yields a sandwich covariance for the stacked estimator. It is especially relevant when the staged procedure is not simply equivalent to maximizing one joint likelihood.

Checking Model-Based Precision Against Resampling

A bootstrap or simulation can test whether a reported standard error is credible under a specified analysis procedure. For a bootstrap comparison, resample at the independent sampling-unit level (often subjects), retain the relevant PK and PD observations together, and refit the entire staged or joint procedure in every replicate. Repeating only the second stage while reusing one fixed PK estimate cannot reveal uncertainty introduced by the PK stage.

If \(B\) successful bootstrap replicates produce estimates \(\hat{\psi}^{*(1)},\ldots,\hat{\psi}^{*(B)}\), their empirical standard deviation estimates the sampling variability represented by that bootstrap scheme. Failed fits, boundary estimates, and extreme estimates should be investigated and reported; silently discarding them can make the bootstrap appear more stable than the estimator really is. A parametric bootstrap instead simulates data from the fitted model and refits the method. It assesses behavior under that model, not under every possible data-generating process.

Simulation is useful when the estimand, design, or estimator has features that make a bootstrap difficult to interpret. Simulate repeated data sets with known parameter values, apply the full estimation pipeline to each, and summarize bias, empirical standard deviation, and interval coverage. For nominal 95% intervals, coverage is the fraction of simulated intervals that contain the generating value. If the empirical standard deviation is much larger than the model-based standard error, or coverage is substantially below 95%, the model-based precision may be optimistic under the simulated conditions.

Diagnostics, Alternatives, and Limitations

Before comparing numbers, verify that the standard errors refer to the same estimand, parameter scale, subjects, observations, and uncertainty target. A conditional standard error and a joint standard error answer different questions even if both are labelled “SE.” If a parameter is log-transformed, for example, an SE on the log scale is not directly an SE on the original scale; transformation requires an appropriate calculation such as the delta method.

Useful warning signs include a nearly singular information matrix, large correlations among parameter estimates, very different standard errors from observed and expected information, or notable changes in estimates and standard errors across starting values. Profile likelihoods, bootstrap distributions, and simulation-based performance summaries provide alternatives or checks when a local quadratic approximation is doubtful. Each has limitations: profiles can be computationally demanding, bootstrap results depend on the resampling scheme and successful refitting, and simulations depend on the data-generating assumptions selected for evaluation.

Common mistake: “The staged PD standard error is smaller, so the staged analysis estimates the PD parameter more precisely.” This conclusion confuses a conditional calculation with uncertainty in the full procedure. First identify whether PK parameters were held fixed in the calculation; then compare like with like, using a joint covariance or a fully refitted bootstrap or simulation when overall uncertainty is the target. A smaller number alone does not establish greater precision.