Sequential Fits as Starting-Value Generators

Why Use Sequential Fits?

A population model can be difficult to initialize when many parameters must be estimated together. One practical source of starting values is to fit a simpler model separately to each subject, or separately within clinically meaningful strata, and then summarize the resulting estimates. This is a sequential fitting strategy: local fits are performed first, and their results are used to initialize a later population fit. The summaries are starting values, not automatically valid estimates of the population parameters.

For example, separate concentration-time fits may provide subject-specific clearance estimates. Their pattern can suggest a plausible typical clearance and the scale of between-subject variability. But subjects have different amounts of information, some estimates can be poorly identified, and the sampled subjects may not represent the intended target population. A careful workflow therefore records how each summary was constructed and checks whether it is robust to those features.

From Individual Estimates to Population Starting Values

Suppose a one-compartment model is fit separately to each subject, producing clearance estimates \(CL_i\), where \(i=1,\ldots,n\). A common population model is

Here \(TVCL\) is the typical clearance, \(\eta_i\) is the subject-specific deviation on the log scale, and \(\omega^2\) is the variance of those deviations. Under this model, \(\log(CL_i/ (1\ \mathrm{L/h}))\) has mean \(\log(TVCL/(1\ \mathrm{L/h}))\); therefore the median clearance is \(TVCL\). The arithmetic mean clearance is instead \(TVCL\exp(\omega^2/2)\). This distinction matters: the arithmetic mean and the typical value are not interchangeable when between-subject variability is appreciable.

If the individual estimates are sufficiently stable and their errors do not dominate the between-subject pattern, a simple starting value for \(TVCL\) is their geometric mean. Equivalently, average their log values and exponentiate. This is a practical initializer under the lognormal model, not a general guarantee of unbiasedness. A population fit accounts for the observation model and uncertainty in individual parameters; a summary of local estimates generally does not.

The log is formally applied to a dimensionless ratio, such as \(CL_i/(1\ \mathrm{L/h})\); the displayed shorthand is common when all values use the same unit. To seed variability, one can examine the spread of the log estimates. However, that spread combines genuine between-subject variation with estimation error. It should not be interpreted directly as \(\omega\) unless the estimation-error contribution is negligible or otherwise accounted for.

Shrinkage and the Meaning of Subject-Level Estimates

Some subject-level estimates are obtained independently, whereas others are conditional estimates informed by a population distribution. The latter can be pulled toward a population reference value when the subject contributes little information. This pull is called shrinkage. It is especially important when using conditional estimates as if they were direct measurements of each subject's true random effect.

A simple normal-normal model shows the mechanism. Let a subject's latent log-scale value be \(\theta+\eta_i\), where \(\theta\) is a fixed reference and \(\eta_i\sim N(0,\omega^2)\). Suppose the subject's local estimate \(z_i\) satisfies \(z_i\mid\eta_i\sim N(\theta+\eta_i,s_i^2)\). Under these assumptions, the conditional mean of \(\eta_i\) given \(z_i\) is

The multiplier is between zero and one. When the local estimate is precise, \(s_i^2\) is small and the multiplier approaches one. When it is imprecise, the multiplier approaches zero and the conditional estimate lies closer to the reference. Consequently, the distribution of conditional estimates may look less variable than the underlying population distribution. Treating that reduced spread as the true \(\omega\) can produce an overly small variability starting value.

Unequal Sampling and the Weighting Target

A summary can change depending on what receives equal weight. Giving each subject one vote targets a subject-level average in the sampled set. Giving each observation equal weight gives more influence to subjects with more records. Weighting strata by their sample proportions targets the sample's stratum composition; weighting them by prespecified target-population proportions targets a different composition. These targets coincide only under suitable conditions.

For a typical log-scale parameter, a stratified initializer can be formed by weighting stratum-specific log estimates by the intended population proportions. If stratum \(k\) has typical value \(TVCL_k\) and target proportion \(w_k\), with \(\sum_k w_k=1\), the mixture's weighted average log typical value is \(\sum_k w_k\log(TVCL_k)\). Exponentiating gives a geometric, not arithmetic, combination. This is a transparent initialization convention; it does not replace estimation of covariate effects or a population distribution.

Unstable Fits, Diagnostics, and Alternatives

Individual fits can be unstable when the subject has few observations, a narrow sampling window, noisy concentrations, or weak information about a parameter. Warning signs include boundary estimates, implausibly large standard errors, strong sensitivity to starting values, optimizer failure, and estimates that change sharply when a small number of records are omitted. A numerical result from a converged fit is not automatically informative: convergence only says that the algorithm met its stopping criteria.

Before summarizing local estimates, inspect their units, scale, uncertainty, boundary status, and fit diagnostics. Plot estimates against measures of information such as number and timing of observations, and examine whether unusual values are concentrated among poorly informed subjects. Refit suspicious subjects from multiple plausible initial values. If an estimate is highly unstable, do not silently clip it or give it equal weight; document the issue and consider using more reliable summaries or a different source of starting values.

Sequential summaries are useful for initialization because they can reveal plausible scales and expose problematic subjects before a more comprehensive fit. Their limitations should remain explicit: they may ignore unequal precision, estimation error, correlations among parameters, and the joint likelihood structure. A joint population fit uses all subject records together to estimate typical values and variability under a specified population model. It is generally preferable for final estimation when that model is appropriate and estimable, but it may be harder to initialize. Sequential estimates can help start that fit; they are not a substitute for it.

Stratified summaries are a useful alternative when covariates or design groups make pooling all subjects inappropriate. Their weights should reflect the intended population and scale of the parameter. Mechanistic calculations and stratified summaries can also provide independent checks on plausible starting magnitudes. Whichever approach is used, preserve the original local estimates and the rules used to summarize them so that the initialization is reproducible.

A Practical Workflow

Common Pitfall

Wrong reasoning: “The average of the subject estimates is the population parameter, so I can use it as the final answer.” This confuses an initializer with a population estimate and overlooks parameter scale, shrinkage, sampling imbalance, and uncertainty in local fits. Correction: state the intended population quantity, choose a summary consistent with its scale and target weights, screen the local fits, and treat the result as a starting value unless a justified analysis establishes otherwise.