Interpreting Individual Fits Across Shrinkage Levels

Why Individual Fits Need a Shrinkage Context

An individual prediction can track a subject’s observations closely for two very different reasons. The observations may contain substantial information about that subject’s parameters, or the estimated individual effect may be pulled toward the population because the subject’s data are sparse or noisy. The plotted fit alone does not distinguish these explanations.

This distinction matters when reviewing individual concentration-time profiles, interpreting empirical-Bayes estimates (EBEs), or comparing subjects. Two subjects can have nearly identical EBE-based predictions even though one has many informative observations and the other has little data support. Conversely, a subject’s EBE-based curve may look close to the population curve precisely because the subject’s estimate has been strongly shrunk toward the population.

The focus here is how to interpret those fits, not how to select a covariate model or diagnose a covariate relationship. The essential question is: How much of this individual fit is supported by this subject’s data, and how much reflects the population distribution?

Definition: A population prediction is the model prediction at the population-level random effects, conventionally $\eta=0$. An EBE-based individual prediction is the prediction obtained by inserting a subject’s estimated random effects, such as an empirical-Bayes mode, into the model. Data support describes how strongly that subject’s observations distinguish their random effects from other plausible values. Shrinkage is the resulting pull of an individual estimate toward the population when the data provide limited information.

A Simple Model for Seeing the Pull

Consider a one-compartment model after an intravenous bolus, with a hypothetical dose and volume fixed across subjects. Let clearance vary by a log-normal random effect:

CL_i=5\exp(\eta_i)\ \text{L/h}, \qquad \eta_i\sim N(0,0.09).

Here $CL_i$ is subject $i$’s clearance and the population typical clearance is $5$ L/h. For a dose of $100$ mg and volume $50$ L, the model prediction at time $t$ hours is

C_i(t)=\frac{100}{50}\exp\left(-\frac{CL_i}{50}t\right)\ \text{mg/L}.

Thus the population prediction uses $\eta_i=0$, giving $CL=5$ L/h and $C(10)=2\exp(-1)\approx0.7358$ mg/L. An EBE-based prediction instead uses the estimated $\hat{\eta}_i$. In real nonlinear models, the likelihood can be complicated, and conditional modes and conditional means need not agree. For transparent arithmetic, the examples below use a one-dimensional Gaussian approximation to the subject’s likelihood.

Gaussian approximation: Suppose the subject’s data yield a likelihood summarized by $z_i$, with $z_i\mid\eta_i\sim N(\eta_i,s_i^2)$, and suppose the random-effects distribution is $\eta_i\sim N(0,\omega^2)$. Then the conditional distribution is normal, and its mean and mode are $$ \hat{\eta}_i=\frac{\omega^2}{\omega^2+s_i^2}z_i. $$ The multiplier $\lambda_i=\frac{\omega^2}{\omega^2+s_i^2}$ is the fraction of the likelihood-based estimate retained in the EBE. In this simplified setting, smaller $s_i^2$ represents stronger data information and yields less shrinkage.

The formula makes the two sources of an EBE visible. The value $z_i$ reflects the subject’s likelihood; $\lambda_i$ reflects the balance between that likelihood’s precision and the population variance. In this approximation, the EBE is not a data-only estimate. It combines the likelihood with the random-effects distribution.

Worked Examples: Reading Individual Fits Across Support

Worked Example: A Sparse Profile Pulled Toward the Population

Suppose a hypothetical subject has a likelihood summary $z=0.40$ for the clearance random effect, with likelihood variance $s^2=0.36$. The population variance is $\omega^2=0.09$. Calculate the EBE and the predicted concentration at 10 h.

The shrinkage multiplier is $\lambda=0.09/(0.09+0.36)=0.09/0.45=0.200$. Therefore $\hat{\eta}=0.200(0.40)=0.0800$. The likelihood summary is $0.40$, but only one fifth of it appears in the EBE in this approximation.

The EBE-based clearance is $CL=5\exp(0.0800)=5.416$ L/h. Substituting this value into the concentration model gives $C(10)=2\exp[-(5\exp(0.08)/50)(10)]=2\exp(-1.083287)\approx0.6770$ mg/L. This is closer to the population prediction, $0.7358$ mg/L, than a prediction based on the unshrunk value $z=0.40$ would be. The closeness to the population curve is consistent with substantial shrinkage; it is not evidence that the subject’s data precisely identify $\eta$.

Worked Example: Similar EBEs, Unequal Data Support

Compare two hypothetical subjects. Subject A has $z_A=0.20$ and $s_A^2=0.36$; Subject B has $z_B=0.05$ and $s_B^2=0.01$. Both share the same population variance, $\omega^2=0.09$. Calculate their EBEs and compare their predicted concentrations at 10 h.

For Subject A, $\lambda_A=0.09/(0.09+0.36)=0.200$, so $\hat{\eta}_A=0.200(0.20)=0.0400$. For Subject B, $\lambda_B=0.09/(0.09+0.01)=0.900$, so $\hat{\eta}_B=0.900(0.05)=0.0450$. The EBEs are very similar, but their support differs: Subject A’s likelihood summary is much less precise, while Subject B’s EBE retains 90% of its likelihood summary in this approximation.

For Subject A, $CL_A=5\exp(0.04)=5.204$ L/h and $C_A(10)=2\exp[-\exp(0.04)]\approx0.7063$ mg/L. For Subject B, $CL_B=5\exp(0.045)=5.230$ L/h and $C_B(10)=2\exp[-(5\exp(0.045)/50)(10)]\approx0.7027$ mg/L. The predictions are close, but that similarity does not imply comparable reliability. Subject A’s estimate is more strongly influenced by the population distribution; Subject B’s is more data-informed under the stated approximation.

Worked Example: A Near-Population Fit Is Not Necessarily Strong Evidence

A third hypothetical subject has $z=0.05$ and $s^2=0.36$, with $\omega^2=0.09$. Compare the EBE-based concentration prediction at 10 h with the population prediction.

The multiplier is $\lambda=0.09/(0.09+0.36)=0.200$, so $\hat{\eta}=0.200(0.05)=0.0100$. The estimated clearance is $5\exp(0.0100)=5.050$ L/h. The individual prediction is $C(10)=2\exp[-(5\exp(0.01)/50)(10)]\approx0.7284$ mg/L, close to the population prediction $0.7358$ mg/L.

The small prediction difference should be interpreted alongside the weak data support. The subject’s likelihood summary is already near zero, and the EBE is further pulled toward zero. A curve close to the population prediction does not show that the subject has been measured precisely enough to confirm population-typical clearance. It may simply be what the conditional estimate looks like when the data do not strongly move it away from the population.

Why the Gaussian Result Holds

The shrinkage expression follows directly from combining the normal likelihood with the normal random-effects distribution. Suppressing the subject index, the conditional density is proportional to $\exp[-\eta^2/(2\omega^2)]\exp[-(z-\eta)^2/(2s^2)]$. Completing the square in $\eta$ gives the conditional variance and mean below.

\begin{aligned} \frac{\eta^2}{\omega^2}+\frac{(z-\eta)^2}{s^2} &=\left(\frac{1}{\omega^2}+\frac{1}{s^2}\right) \left(\eta-\frac{\omega^2}{\omega^2+s^2}z\right)^2 +\frac{z^2}{\omega^2+s^2},\\ \eta\mid z&\sim N\left(\frac{\omega^2}{\omega^2+s^2}z,\, \frac{\omega^2s^2}{\omega^2+s^2}\right). \end{aligned}

The coefficient of $\eta$ in the squared term gives the conditional mean, and a normal distribution has its mode at its mean. If $s^2$ is large relative to $\omega^2$, then $\lambda$ is small and the conditional estimate is close to zero. If $s^2$ is small, $\lambda$ approaches one and the conditional estimate approaches $z$. This result is exact for the stated Gaussian model, but serves as an approximation—not a universal formula—for nonlinear NLME likelihoods.

Interpreting and Checking Fits

A useful interpretation compares the population prediction, the EBE-based prediction, and the observed data together. The population prediction shows the model’s typical trajectory. The EBE-based prediction shows the trajectory at the subject’s estimated random effects. Observations show how well either trajectory describes the recorded data, but visual agreement alone does not quantify the uncertainty or data support behind the EBE.

When reviewing individual fits, consider the number, timing, and precision of observations in relation to the parameters being estimated. Many observations do not automatically provide strong information about every parameter: for example, observations concentrated in a time region insensitive to clearance may offer limited clearance information. Conversely, a smaller number of strategically informative observations can be more useful than a larger number of redundant ones. The relevant question is whether the observations constrain the parameter direction that drives the apparent individual difference.

In practice, compare the EBE-based predictions with the population predictions and inspect the subject’s observations and residual pattern. A large difference between the two predictions may reflect a data-supported estimate, but can also signal model tension or an unstable conditional estimate. A small difference may reflect genuinely typical behavior or strong shrinkage. Use shrinkage and information diagnostics already introduced in the course alongside the profile; do not infer reliability solely from the visual distance between curves.

Establish the reference.
Identify the population prediction and the subject-specific EBE-based prediction, and state which random effects differ between them.

Assess the subject’s information.
Consider observation timing, amount, precision, and sensitivity to the parameter of interest. A close curve is not a substitute for this assessment.

Interpret the contrast cautiously.
Describe whether the EBE moves the prediction away from the population and whether the data appear capable of supporting that movement. State uncertainty or shrinkage limitations where relevant.

Alternatives, Limitations, and a Common Pitfall

The population prediction is a stable reference but does not express the subject-specific shift captured by the EBE. An EBE-based prediction is useful for describing a fitted subject-level trajectory, but it is a conditional point estimate and does not display its uncertainty. A likelihood-based, data-only estimate can help reveal how the subject’s observations pull before the random-effects distribution is applied, but with sparse data it may be highly variable or poorly determined. These views answer different questions; none should be treated as a complete replacement for the others.

A further limitation is that a prediction obtained by inserting an EBE is generally not the same as averaging predictions over the conditional distribution of the random effects. Because the model is nonlinear, in general $f(\hat{\eta})\ne E[f(\eta)\mid\text{data}]$. The distinction can matter when uncertainty is substantial or the prediction function is strongly nonlinear. The examples use plug-in predictions to make the shrinkage effect transparent; they do not imply that an EBE curve contains all relevant predictive uncertainty.

Common mistake: “These two subjects have almost the same individual prediction, so their individual parameters are equally well estimated.” This conclusion does not follow. Similar predictions can arise from different likelihood information, different likelihood summaries, and different degrees of shrinkage. The correction is to evaluate each subject’s data support separately and interpret the EBE-based fit as a population-informed conditional estimate, not as a direct measurement of that subject’s true parameter.

Key takeaway: An individual fit is an estimate, not a direct view of an individual’s true trajectory. Interpret its distance from the population prediction together with the subject’s data information and the amount of shrinkage; visual closeness alone cannot establish reliability.

Check Your Understanding

Try each question, then compare with the brief solutions below.

Under the Gaussian approximation, let $\omega^2=0.16$, $s^2=0.04$, and $z=0.25$. Calculate $\lambda$ and the EBE.
Two subjects have nearly identical EBEs, but one has much larger likelihood variance. What can and cannot be concluded about their individual predictions?
Explain why a prediction close to the population prediction may reflect weak data support rather than strong evidence of typicality.
In a nonlinear model, why might a prediction at the EBE differ from the conditional mean prediction?

Solutions. (1) $\lambda=0.16/(0.16+0.04)=0.800$, so $\hat{\eta}=0.800(0.25)=0.200$. (2) Their plug-in predictions may be similar, but the subject with larger likelihood variance is more strongly influenced by the population distribution under this approximation; equal reliability cannot be inferred. (3) Sparse or noisy data can leave the conditional estimate near $\eta=0$, making the prediction resemble the population curve without precisely establishing the subject’s parameter. (4) Nonlinearity means averaging and plugging in do not generally commute: $f(E[\eta\mid\text{data}])$ need not equal $E[f(\eta)\mid\text{data}]$.