Common Errors in Hierarchical Parameter Equations

Why Hierarchical Equations Are Easy to Misread

A hierarchical parameter equation connects quantities defined at different levels. A population typical value describes the model’s reference population; an individual parameter describes one subject; and an observation is a measured outcome at a particular time. An equation can look plausible while quietly assigning a term to the wrong level. The resulting model may imply that a subject’s clearance changes from one blood sample to the next, or that an eta is measurement error, even though neither interpretation was intended.

The key is to check what each symbol describes before doing algebra. The notation below uses \(i\) for a subject and \(j\) for an observation from that subject. For a clearance parameter, \(\theta_{\mathrm{CL}}\) is the population typical value, \(\mathrm{CL}_i\) is subject \(i\)’s individual clearance, \(X_i\) is a subject-level covariate, and \(\eta_{\mathrm{CL},i}\) is that subject’s random effect on clearance. The observation-level prediction at time \(t_{ij}\) is \(f(t_{ij},\mathrm{CL}_i,\ldots)\); the measured outcome is \(Y_{ij}\).

A Reference Structure for Auditing

For a positive parameter such as clearance, a common structure combines a typical value, a covariate effect, and an exponential subject effect:

Here, \(X_{\mathrm{ref}}\) is the covariate reference value, \(\beta_{\mathrm{CL}}\) is the coefficient on the log scale, and \(\eta_{\mathrm{CL},i}\) is the subject-specific deviation on that scale. At the reference covariate value and with \(\eta_{\mathrm{CL},i}=0\), the equation gives \(\mathrm{CL}_i=\theta_{\mathrm{CL}}\). The observation model has a separate role:

The function \(f\) gives the model prediction using the subject’s parameter values. The residual term \(\epsilon_{ij}\) represents observation-level discrepancy under the chosen residual-error model. These equations are a checking framework, not a claim that every analysis uses exactly this covariate form or residual distribution.

A useful audit asks four questions: What level does each quantity describe? Does the covariate enter the parameter equation or the observation equation? Is the random effect on the parameter scale or log scale? Does the equation imply that a subject-level quantity changes across observations? The following examples apply these questions to common errors.

Covariates Belong Where Their Meaning Says

A covariate that explains differences in a parameter belongs in that parameter’s equation. If body weight is intended to explain clearance, it should affect \(\mathrm{CL}_i\), which in turn affects the prediction \(f\). Putting weight directly into the observation equation instead changes the meaning: it says weight affects the predicted outcome through an additional direct pathway, rather than only through clearance.

The index matters too. A subject-level covariate such as baseline weight is usually denoted \(X_i\), because it is shared across that subject’s observations. Writing \(X_{ij}\) suggests a value that can vary by observation. That may be appropriate for a time-varying covariate, but should not be used accidentally for a fixed subject characteristic. The equation should also make clear which covariate value is the reference, so that the typical value has an interpretable meaning.

Eta Is Not Measurement Error

A random effect and a residual error term can both be random, but they operate at different levels. In the parameter equation, \(\eta_{\mathrm{CL},i}\) represents subject-to-subject deviation in clearance. It is shared across that subject’s repeated observations when the model specifies a subject-level effect. In the observation equation, \(\epsilon_{ij}\) represents residual discrepancy for a particular measured outcome. It can differ across observations.

For example, if clearance is modeled as \(\mathrm{CL}_i=\theta_{\mathrm{CL}}\exp(\eta_{\mathrm{CL},i})\), then the individual’s predicted concentration at each time is computed using that same \(\mathrm{CL}_i\). Residual error is applied to the observation model, according to the chosen residual-error specification. Replacing \(\epsilon_{ij}\) with \(\eta_{\mathrm{CL},i}\) does not merely rename an error term: it changes the model to attribute observation-level departures to variation in clearance.

A Short Justification for the Multiplicative Form

The exponential form is useful when a parameter must remain positive and deviations are intended to act proportionally. Taking logs of the no-covariate equation gives \(\log(\mathrm{CL}_i)=\log(\theta_{\mathrm{CL}})+\eta_{\mathrm{CL},i}\). Exponentiating returns a positive clearance because an exponential is always positive. A positive eta multiplies the typical value by a factor greater than one; a negative eta multiplies it by a factor between zero and one.

Thus, for this value, clearance is about 10.5% above the typical value, not 0.10 L/h above it. This check catches a scale error: adding eta directly to a clearance measured in L/h would treat a dimensionless log-scale effect as an absolute difference in clearance. Additive parameter models can be appropriate in other settings, but the random effect’s scale and the parameter’s units must agree with the specified model.

Systematic Checks, Diagnostics, and Limits

When auditing an equation, do not rely on appearance alone. Write down the meaning, index, and units of each term. Then substitute a reference covariate and zero eta: the resulting individual parameter should equal the intended typical value. Finally, hold a subject fixed while considering multiple observations. A subject-level parameter should not change just because \(j\) changes.

These are structural checks, not proof that a model fits data well. Warning signs include parameter values changing with an observation index when no such change was intended, a covariate term with incompatible units, or eta appearing in the observation equation as if it were residual error. A computational implementation can also encode an equation differently from its written specification, so checking the actual code against the intended hierarchy is important.

The main alternatives are not interchangeable fixes. A multiplicative parameter model and an additive parameter model make different assumptions about scale and positivity. A covariate acting on a parameter and a covariate acting directly on a prediction describe different pathways. A subject random effect and residual error represent different sources and levels of variation. Choose among them based on the scientific model and the intended interpretation, not to make an equation look simpler. These checks also do not establish whether the model is identifiable, whether a covariate should be included, or whether its estimated effect is precise; those require separate evaluation.

Common Mistakes

• Using an observation index for a subject effect. This makes the effect appear to vary across samples. Use the subject index when the intended effect is shared within subject.
• Putting a parameter covariate directly onto the outcome. This changes the modeled pathway and may create a units mismatch. Put it in the parameter equation when the intended effect is on that parameter.
• Calling eta measurement error. Eta changes an individual parameter; residual error describes observation-level discrepancy. State which one is meant.
• Adding a log-scale eta to a natural-scale parameter. The scales and units do not match. Use the specified multiplicative form for a log-scale eta.
• Assuming that every random term is an eta. Randomness alone does not determine the level or meaning of a term. Its definition and index do.