Stable Initialization for Two-Class NLME Fits

Why Initialization Matters in a Two-Class Fit

A two-class NLME model must estimate both the population parameters within each latent class and the probability that an individual belongs to each class. Because class membership is unobserved, the likelihood combines contributions from both classes rather than assigning each individual permanently to one group. The resulting objective can be difficult to optimize: a poor starting point may lead to slow progress, implausible parameter values, or a converged solution that depends strongly on the initial values.

Initialization is not a substitute for likelihood-based estimation. Its purpose is to give the optimizer plausible starting values in a region where both classes are represented and their clearance distributions can be distinguished. Stratifying a noisy clearance proxy can help create such starts, provided that the proxy is treated as a guide rather than as a definitive classification.

Model Components to Initialize

For a simple two-class model, let \(k\in\{1,2\}\) index the latent classes and let \(i\) index individuals. A useful parameterization for clearance is

Here \(CL_k\) is the typical clearance in class \(k\), and \(\omega_k^2\) is the variance of the individual random effect on log clearance. The exponential ensures positive individual clearance. This expression describes a possible class-specific random-effect model; if the intended model shares a variance across classes, initialize one shared \(\omega^2\) instead of two separate variances.

Write \(\pi_{ik}\) for the probability that individual \(i\) belongs to class \(k\), with \(\pi_{i1}+\pi_{i2}=1\). If membership has no covariates, then \(\pi_{i1}=\pi\) and \(\pi_{i2}=1-\pi\). If membership depends on covariates, the probability model specified for that analysis supplies these probabilities; stratified class counts can inform a starting prevalence but do not replace that model.

In this expression, \(y_i\) denotes individual \(i\)'s observations, \(f\) is their conditional data density under class \(k\), \(\theta\) collects the model parameters, and \(\phi\) is the normal density for the random effect. The individual likelihood sums over the possible classes and integrates over the random effect. The starting values discussed here are supplied for parameters in this likelihood; they are not estimates obtained by assigning each person to a class and then fitting two unrelated models.

A practical stratified initialization has three parts: use approximate group sizes to start prevalence, use the proxy values within each group to start \(CL_k\), and use within-group spread on the log scale to start \(\omega_k^2\). The assumptions are modest but important: clearance proxies should be positive and sufficiently informative to suggest two groups, and the planned model should use the same clearance scale and random-effect structure as the starting calculations. Proxy-based group assignment is not assumed to be error-free.

Constructing Stratified Starts

For a positive individual clearance proxy \(q_i\), the log scale is natural because the model expresses clearance multiplicatively. If a provisional stratum \(k\) contains \(n_k\) proxies, a simple starting value for typical clearance is their geometric mean:

The superscript \((0)\) indicates an initial value. A provisional variance start can be calculated as the sample variance of the log proxies within that stratum:

This variance is only a rough guide. It combines true individual variability with measurement error, uncertainty in the proxy, and any misspecification in the proxy calculation. It should not be interpreted as a final estimate of random-effect variance. When a stratum is small, or its proxy values are nearly identical, the empirical variance may be unstable or zero; use a scientifically defensible positive starting value or a shared variance start rather than asking the optimizer to begin at a boundary.

Why These Calculations Are Reasonable

The geometric mean follows directly from the log-clearance parameterization. If a stratum has log proxies \(z_i=\log(q_i)\), their arithmetic mean \(\bar z\) corresponds to a typical clearance \(\exp(\bar z)\). Exponentiating the mean log is exactly the geometric mean of the original positive proxies. Thus the initialization is on the same multiplicative scale as \(CL_k\exp(\eta_{ik})\).

Similarly, the sample variance of the \(z_i\) values measures their observed dispersion on the scale where the random effect is additive. It is not automatically equal to \(\omega_k^2\): the proxies are not the latent random effects, and their variation may include other sources. The calculation is justified as a practical starting heuristic, not as an estimator with guaranteed accuracy.

For constant prevalence, the logit transformation also has a direct justification. The logit parameter \(\gamma_0\) satisfies \(\pi=\exp(\gamma_0)/(1+\exp(\gamma_0))\). Solving for \(\gamma_0\) gives \(\gamma_0=\log\{\pi/(1-\pi)\}\), which is why a provisional fraction can be translated into an intercept start.

Assessing Whether Starts Reach Comparable Solutions

A run that reports convergence has passed the software's stopping rules; that alone does not establish that it found a satisfactory or reproducible solution. Compare runs made with the same model, data, estimation settings, and likelihood approximation. Examine at least the objective value (or log likelihood), convergence and gradient information, class prevalence, class-specific typical clearances, random-effect variances, and any boundary or covariance warnings.

Comparable solutions should have objective values that agree to a practically negligible tolerance and parameter estimates that are close enough for the scientific interpretation. The tolerance is context- and software-dependent; it should be chosen before reviewing results where possible and assessed relative to numerical precision and the size of scientifically important differences. A tiny objective difference does not make substantially different parameter estimates harmless: that pattern can signal a flat likelihood or weak information. Conversely, small numerical differences in parameters may be unimportant when the objective and predictions are stable.

Diagnostics, Alternatives, and Limitations

Stratified starts are especially useful when an observed proxy provides a plausible separation of the classes. They are not the only option. One alternative is to use several deliberately varied, scientifically plausible starts for prevalence, clearances, and variances without assigning individuals to proxy strata. Another is to use a simpler preliminary fit to obtain scale information, then construct mixture starts from that information. These approaches can be combined: use proxy strata for one run and perturbed or dispersed starts for additional runs. The essential comparison is among fits of the same final model, not among different preliminary models.

Warning signs include an estimated prevalence close to zero or one, a variance at or near its lower bound, a class clearance moving to an implausible range, large gradients, or estimates that change materially across starts. Also inspect whether provisional strata are driven by a few extreme proxies or have very small counts. If two classes are not well separated in the available data, initialization may be difficult regardless of the strategy; changing starts cannot create information absent from the observations.

Avoid interpreting the proxy strata as observed biological categories. Proxy error can place individuals in the wrong provisional group, and the mixture likelihood is intended to account for uncertain membership. Likewise, avoid setting a random-effect variance to zero simply because proxy values within a stratum happen to be similar. Such a start may put the optimizer at a boundary and can conceal meaningful uncertainty. Use a positive, defensible value and examine sensitivity to reasonable alternatives.