Crossed Study And Site Effects

Why the Study–Site Relationship Matters

A multicenter population PK dataset may contain subjects from several studies and several clinical sites. The labels alone do not determine the random-effect structure. The key question is whether each site belongs to only one study or whether the same site contributes subjects to multiple studies. That relationship determines which groups share a site effect, and therefore which subjects the model treats as correlated through that effect.

This matters because a study effect and a site effect represent different sources of between-subject variation. For example, studies may differ in protocol or overall conduct, while sites may differ in local procedures. A model that treats a shared site as if it were study-specific can fail to represent the dependence in the data. Conversely, treating uniquely study-specific sites as shared across studies imposes a connection that the design does not contain.

Linear Predictors for Crossed Effects

Suppose subject \(i\) is enrolled in study \(j\) at site \(k\). For illustration, let the effects apply to clearance. Write the structural PK model's individual clearance as \(CL_{ijk}\), and let \(\theta_{CL}\) denote the population typical value. A crossed additive model on the log scale is

Here \(a_j\) is the study-level random effect, \(d_k\) is the site-level random effect, and \(\eta_{ijk}\) is the subject-level random effect. The divisor makes the logarithm's argument dimensionless. In this example, assume \(a_j\sim N(0,\tau^2_{\text{study}})\), \(d_k\sim N(0,\tau^2_{\text{site}})\), and \(\eta_{ijk}\sim N(0,\omega^2)\). Assume these effects are mutually independent across levels and independent across different studies, sites, and subjects. These are modeling assumptions, not consequences of a crossed design.

The PK structural model uses each subject's resulting clearance, along with the other relevant parameters, to predict concentrations. The crossed-effects expression specifies how clearance varies across study, site, and subject; it does not replace the structural concentration-time model. Other PK parameters could have their own study and site effects, but adding them requires scientific justification and enough information in the design.

For nested sites, use a site effect specific to its study, such as \(d_{k(j)}\), where \(k(j)\) identifies a site within study \(j\). The corresponding predictor includes \(a_j+d_{k(j)}+\eta_{ijk}\). The distinction is substantive: in the crossed model, a subject at site \(k\) in different studies shares the same \(d_k\); in the nested model, sites in different studies cannot share a site effect.

What Covariance Does the Crossed Model Imply?

The shared random effects induce covariance between subjects' log clearances. Let \(Z_{ijk}\) denote the log-scale quantity on the left side of the predictor. Consider distinct subjects, so their subject effects are independent. Under the assumptions above, the covariance between two such quantities is the sum of the variances of the random effects they share. To see this, write \(Z=\mu+a+d+\eta\) for each subject and expand covariance using additivity. Independent terms have covariance zero; a shared study effect contributes \(\operatorname{Var}(a_j)=\tau^2_{\text{study}}\), and a shared site effect contributes \(\operatorname{Var}(d_k)=\tau^2_{\text{site}}\).

The indicator \(\mathbf{1}(A)\) equals 1 when condition \(A\) is true and 0 otherwise. Thus two subjects in the same study and same site share both effects; subjects in the same study but different sites share only the study effect; subjects at the same site but in different studies share only the site effect. Subjects sharing neither have covariance zero under this model. For one subject, the log-clearance variance is \(\tau^2_{\text{study}}+\tau^2_{\text{site}}+\omega^2\). This calculation follows from the assumed independence; correlated study and site effects would require a different covariance expression.

Design Overlap, Identifiability, and Model Choice

Crossing describes the design, but it does not guarantee that both variance components can be estimated precisely. To separate study and site contributions, the data need informative replication and overlap. Multiple subjects within groups help distinguish group-level variation from subject-level variation. Site identities represented across studies provide information about a shared site contribution separate from study membership. If every site occurs in only one study, study and site membership are aligned: the design is nested, and the data do not contain cross-study reuse of sites that would support the crossed interpretation.

Even with a crossed design, sparse or highly unbalanced overlap can make variance components weakly informed. For example, if only one site is shared between two studies and nearly all subjects come from other sites, the shared-site information is limited. A model may then have difficulty separating study and site variation, particularly if the number of studies or sites is small. Multiple concentration samples from one subject do not create additional independent subjects for estimating these group-level components; those samples share that subject's random effects.

The main alternatives answer different questions. A nested random-effects model is appropriate when each site occurs in exactly one study. Treating study or site as fixed effects can be useful when inference concerns the particular included studies or sites, especially when their number is small; it does not estimate a population variance for a broader distribution of studies or sites. Omitting one grouping effect gives a simpler model, but assumes that the omitted source contributes no relevant systematic between-group variation. These approaches are not interchangeable just because they can all produce fitted values.

Assumptions, Diagnostics, and Common Pitfalls

A crossed random-intercept model usually assumes mean-zero effects with specified distributions, commonly normal distributions, and a stated covariance structure. The additive predictor assumes that study and site contributions combine without an unmodeled study-by-site interaction. It also assumes that the chosen grouping labels correctly represent shared groups. These assumptions should be considered alongside the PK model and the sampling and residual-error models; a plausible grouping structure alone does not guarantee that the full model is adequate.

Useful checks begin with the data structure. Tabulate subjects by study and site, verify whether repeated labels refer to the same actual site, and inspect how many groups and subjects support each component. After fitting, assess whether estimates are stable and whether the estimated variance components are plausible and meaningfully informed. Very large uncertainty, estimates at a boundary, or strong sensitivity to small changes in the grouping definition can indicate that the data do not separate the components well. Examine model fit across studies and sites as well: systematic differences not captured by the additive effects may indicate misspecification, although an apparent pattern can also arise from other parts of the model.

A common mistake is to infer a crossed structure merely because the dataset has both a study column and a site column. The incorrect reasoning is: “There are multiple studies and multiple sites, so the effects are crossed.” The correction is to examine actual membership: crossed effects require at least some site identity to recur across studies. Another mistake is to treat a site code reused for different clinics as a shared site. Verify identity before constructing the random effects.

A further limitation is that a random site effect models a distribution of site deviations, and a crossed additive model attributes shared variation according to its specified components. If the number of independent studies or sites is small, the variance estimates may be imprecise. If site effects differ substantially by study, an additive model may be insufficient; a study-by-site interaction is a possible extension only when the design and scientific question support it. The appropriate structure is determined by design and purpose, not by selecting the most complex formula available.