Why Individual Plots Need a Shrinkage Qualification
A subject-level prediction plot can make a fitted model look reassuring: the individual prediction follows the observations, or the residuals appear small and pattern-free. But an individual prediction commonly uses an empirical Bayes estimate (EBE) of that subject’s random effects. When the observations provide limited information about those effects, the EBE is pulled toward the population mean. The fitted individual curve may then reflect the population prediction more than a well-supported subject-specific estimate.
Shrinkage summaries help qualify what an individual plot can show. They do not diagnose a particular subject, prove a model is adequate, or replace inspection of the data. Their role is narrower: indicate whether the subject-level estimates or residuals may be too contracted for their apparent patterns to be read at face value.
For a random effect \(\eta\) with estimated population variance \(\omega^2\), a commonly used variance-based summary is \(1-\widehat{\mathrm{Var}}(\hat\eta)/\hat\omega^2\), where \(\hat\eta\) denotes the EBEs across subjects. The estimate is an aggregate for that random effect, not an estimate of the reliability of every subject’s EBE. A value near zero indicates little aggregate variance contraction by this measure; a larger value signals greater contraction. Epsilon shrinkage is also an aggregate summary, but it concerns residual information rather than the random-effect estimates used to form IPRED.
What Shrinkage Changes in a Plot
Consider a parameter represented as a typical population value plus a subject-specific random effect. The EBE estimates that effect using both the population distribution and the subject’s observations. If the observations are informative, the EBE may depart substantially from zero. If they are weakly informative, the population distribution has more influence and the EBE tends to lie closer to zero. Since IPRED uses that estimate, high eta shrinkage can make IPRED resemble PRED even when the subject’s true effect is not zero.
A simple normal model illustrates the pull. Suppose an observation-based summary \(z_i\) has conditional distribution \(z_i\mid\eta_i\sim N(\eta_i,s_i^2)\), and the random effect has prior distribution \(\eta_i\sim N(0,\omega^2)\). Here \(s_i^2\) represents uncertainty in the subject’s information. Completing the square in the product of the normal likelihood and prior gives the conditional mean:
The multiplier is between zero and one. When \(s_i^2\) is small relative to \(\omega^2\), the multiplier is near one and the estimate follows the subject’s information. When \(s_i^2\) is large, the multiplier is nearer zero and the estimate is pulled toward the prior mean. This simplified derivation is not a formula for every model’s EBE, but it shows why sparse or weakly informative subject data can yield highly contracted estimates.
The same logic guides plot interpretation. A strong difference between PRED and IPRED, accompanied by low eta shrinkage for the relevant random effect, is more compatible with a data-informed individual adjustment. It is not proof that the adjustment is correct. Conversely, similar PRED and IPRED curves with high eta shrinkage do not establish that the subject is typical; the EBE may simply lack information to move far from zero.
Worked Example: A Subject-Specific Adjustment With Supporting Information
Hypothetical data. For one subject, concentrations are 11, 9, and 5 mg/L at three sampling times. PRED values are 10, 7, and 4 mg/L; IPRED values are 10.8, 8.7, and 5.2 mg/L. The fitted model’s eta shrinkage for the random effect driving the individual adjustment is 12%.
At the first time, the PRED residual is \(11-10=1\) mg/L and the IPRED residual is \(11-10.8=0.2\) mg/L. At the second time, they are \(9-7=2\) mg/L and \(9-8.7=0.3\) mg/L. At the third time, they are \(5-4=1\) mg/L and \(5-5.2=-0.2\) mg/L. Thus the IPRED residuals are closer to zero at all three observations, and the individual prediction follows the observed values more closely.
The low aggregate eta shrinkage makes it less likely that the relevant EBEs are uniformly pulled close to zero across the population. It therefore supports, but does not guarantee, treating this subject’s individual adjustment as potentially informative. The judgment still depends on the number and timing of observations, the model’s adequacy, and whether the adjustment is plausible. Three close residuals alone do not validate the individual parameter estimate.
Use the Two Shrinkage Summaries for Different Questions
Eta shrinkage qualifies interpretation of the individual prediction because IPRED depends on EBEs. Epsilon shrinkage qualifies interpretation of residual diagnostics because residual-related quantities may themselves be contracted. These summaries are not interchangeable. A model can have substantial eta shrinkage but less epsilon shrinkage, or the reverse, and the plots should be read accordingly.
For residual plots, first identify exactly what is plotted. A plot of raw \(DV-PRED\) values is not the same as a plot of standardized residuals or residual estimates. Epsilon shrinkage should qualify the residual diagnostic it describes, using the definition reported for the analysis. In particular, a compact or pattern-free residual display should not automatically be taken as strong evidence of individual fit when residual information is highly contracted. The summary is a warning about interpretability, not a correction factor to apply to each plotted point.
Worked Example: High Eta Shrinkage and a Persistent Residual Pattern
Hypothetical data. For a subject, the observed concentrations are 13, 14, and 9 mg/L. The PRED values are 12, 8, and 4 mg/L, while IPRED values are 12.2, 9, and 5 mg/L. The eta shrinkage for the relevant random effect is 75%.
The PRED residuals are \(13-12=1\), \(14-8=6\), and \(9-4=5\) mg/L. Their root mean square is:
The IPRED residuals are \(13-12.2=0.8\), \(14-9=5\), and \(9-5=4\) mg/L. They remain positive and large at the second and third times. The EBE has adjusted the prediction somewhat, but the apparent individual prediction has not removed the residual pattern.
With 75% eta shrinkage, the small difference between PRED and IPRED should not be read as strong evidence that the subject truly has a near-zero random effect. The estimate may be pulled toward zero. At the same time, the sizeable residuals are visible in the data and merit investigation; high shrinkage does not make them disappear or justify dismissing them. This plot supports concern about fit at those times, but it does not by itself identify the cause.
Worked Example: A Quiet Residual Plot With High Epsilon Shrinkage
Hypothetical diagnostic output. A subject has observed concentrations of 8.1, 5.8, and 3.2 mg/L. The PRED values are 8, 6, and 3 mg/L, so the raw prediction residuals are \(0.1\), \(-0.2\), and \(0.2\) mg/L. Suppose the model’s reported epsilon shrinkage is 65%, while eta shrinkage for the random effect used in IPRED is 18%.
The raw residuals are all small in the units shown, and their signs do not form a sustained sequence. However, the epsilon shrinkage is high, so residual-related diagnostics from this fit warrant caution: apparently subdued residual behavior may not provide strong evidence about the subject-level residual pattern. The low eta shrinkage does not remove this concern, because it concerns the random-effect estimates rather than residual information.
A careful report would state both observations: the raw \(DV-PRED\) residuals are small for these three samples, and the reported epsilon shrinkage limits how confidently the residual diagnostic can be interpreted. It would not infer that the individual model is definitively adequate, nor treat the epsilon summary as proof that these particular residuals are biased. Epsilon shrinkage is a population-level qualifier, not a subject-specific explanation.
A Practical Reading Sequence
A disciplined reading sequence prevents a visually appealing individual plot from carrying more inferential weight than it can support. Check the model prediction and residual definitions, connect each relevant shrinkage summary to the quantity it qualifies, and then describe what the plotted data show independently of the model-based interpretation.
Record whether the panel shows PRED or IPRED, raw residuals or standardized residuals, and the units and sampling times. Do not compare unlike residual scales as if they were the same quantity.
Use eta shrinkage for the random effect influencing IPRED. Use epsilon shrinkage to qualify the residual-related diagnostic described by its calculation. A summary for one quantity does not automatically apply to another.
State whether observations track the prediction, whether residuals have a directional or time-related pattern, and how large deviations are in context. Do not call a plot informative merely because it looks smooth.
Explain how shrinkage changes confidence in an apparent individual adjustment or residual pattern. Preserve the distinction between a visible data feature and the strength of evidence for a subject-specific parameter.
Diagnostics, Comparisons, and Limitations
Several features should prompt extra caution: high eta shrinkage for the random effect used in IPRED; IPRED curves that barely differ from PRED despite a claim of subject-specific characterization; a small number of observations; or high epsilon shrinkage accompanying apparently unstructured residuals. These features do not prove that the model is wrong. They indicate that individual plots may have limited power to reveal subject-specific behavior.
PRED and IPRED answer related but different descriptive questions. PRED shows the population-level prediction for the covariates under the fitted model. IPRED shows the prediction after incorporating estimated subject-specific effects. Comparing both can reveal how much the EBE changes the fitted curve, but the comparison alone cannot establish whether that change is supported by sufficient information. A residual plot adds information about discrepancies, but it too depends on the residual definition and available samples.
Shrinkage values are summaries across subjects and can conceal heterogeneity in information. A reported value does not say that every EBE is equally unreliable, and no universal cutoff turns a subject-level plot into either interpretable or uninterpretable evidence. The relevant random effect, observation schedule, parameter correlations, model structure, and diagnostic construction all matter. Shrinkage should therefore qualify a plot’s interpretation rather than replace the plot, the data, or other model checks.
Common Mistake
Wrong reasoning: “IPRED is close to every observation, so the individual parameter estimates must be reliable.” IPRED is calculated using the same subject’s observations through EBEs; close fit can partly reflect that conditioning, and high shrinkage may mean the EBE is largely pulled toward the population value.
Correction: Inspect PRED and IPRED together, examine the residual pattern and sampling information, and report the relevant eta and epsilon shrinkage summaries. Describe what is observed, then qualify how strongly it supports a subject-specific interpretation. Neither low shrinkage nor a visually good fit proves that the model or an individual estimate is correct.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A model has high eta shrinkage for the random effect used in IPRED. What does that imply, and what does it not imply?
- For one hypothetical subject, \(DV=7\) mg/L, PRED is 5 mg/L, and IPRED is 6.5 mg/L. Calculate both residuals and state what their difference shows.
- A residual plot appears pattern-free, but epsilon shrinkage is high. Give an appropriate interpretation.
- Why should a low eta shrinkage summary not be treated as proof that a particular subject’s EBE is precise?
Solutions. (1) High eta shrinkage indicates aggregate contraction of EBEs toward zero for that random effect, so IPRED may be close to PRED even when individual information is weak. It does not prove that every subject’s estimate is poor or that the model is misspecified. (2) \(DV-PRED=7-5=2\) mg/L and \(DV-IPRED=7-6.5=0.5\) mg/L. The EBE-based prediction is closer to this observation, but one point cannot establish the reliability of the subject-specific estimate. (3) Report that the displayed residuals show no clear pattern, while noting that high epsilon shrinkage limits how strongly the residual diagnostic supports a conclusion about individual residual behavior. (4) Eta shrinkage is an aggregate summary; subjects can differ in sampling information and parameter identifiability, so a low overall value does not certify precision for every individual.