Choosing The Number Of Compartments

Why the Number of Compartments Is a Modeling Choice

A compartment is a modeling construct: it represents a kinetically distinct part of the body in a simplified description of drug distribution and elimination. Choosing one, two, or three compartments is therefore not a claim that the body literally contains that number of uniform spaces. It is a choice about how many distinguishable concentration-time phases the available data and scientific context can support.

After an intravenous bolus in a linear disposition model, the concentration in the sampled central compartment can often be expressed as a sum of exponential terms. One term produces a single declining phase; two terms can represent a faster distribution phase followed by a slower phase; three terms can represent three distinguishable phases. These profiles motivate the choice, but the number of visible phases depends on when and how often concentrations were measured. A phase that occurs before the first sample, or is too small relative to measurement noise, may not be identifiable from the data.

In this tutorial, the comparison is about disposition after intravenous input. The important practical question is not simply which model has the most flexible curve. It is whether the study design contains observations that can reveal the additional phase, and whether that phase is credible for the drug and context.

What the Candidate Structures Can Represent

For a one-compartment IV-bolus model with first-order elimination, the concentration has the form \(C(t)=A e^{-kt}\), where \(A\) is a concentration and \(k\) is a positive rate constant with units of inverse time. On a plot of log concentration against time, this gives a straight line. In the model, drug is treated as instantaneously mixed throughout one apparent space; that can be a useful approximation even when real distribution is more complex.

A two-compartment model can produce a rapid decline associated with distribution from the central space and a slower decline associated with the later disposition phase. A three-compartment model can represent two faster phases before the terminal decline. For the example profiles below, write the concentration as a sum of exponentials, \(C(t)=\sum_i A_i e^{-\lambda_i t}\), where each \(\lambda_i\) is a positive rate constant. The corresponding phases are not equally visible at every time: a term with a large rate constant fades quickly, while a term with a small rate constant persists longer.

The exponential coefficients and rates in these expressions are macroparameters describing the concentration profile. They are not, by themselves, a direct measurement of anatomical spaces. The earlier material on Microconstants and Macroconstants distinguishes the exponential description from the underlying exchange and elimination rate constants. That distinction matters when a fitted phase seems physiologically implausible: a profile can fit well without each exponential term mapping one-to-one to an organ or tissue.

Sampling Coverage Determines What Can Be Distinguished

The different phases place different demands on sampling. Early samples are needed to observe rapid distribution; intermediate samples help show a transition between faster and slower declines; sufficiently late samples help characterize the terminal phase. A design with only late samples might estimate a terminal slope reasonably well while providing almost no evidence about how many faster phases occurred earlier. Conversely, a design that ends too soon may fail to establish the terminal behavior.

A useful way to see the issue is to calculate a slope between two concentrations. If the profile were a single exponential over that interval, the apparent rate would be \(\lambda_{\text{app}}=\ln[C(t_1)/C(t_2)]/(t_2-t_1)\). For a sum of exponentials, this is an interval-specific summary; it equals the terminal rate only when faster contributions are sufficiently small at both times.

Compare Models Against the Observations, Not Just Their Flexibility

A higher-compartment model can describe more curvature, but the extra flexibility is useful only when the data contain information about it. Before comparing fits, check that candidate models use the same observations and a compatible residual-error model. Then inspect whether the simpler model leaves systematic structure in the residuals, especially over time, and whether the more complex model estimates a distinct phase with reasonable precision. A numerical improvement in fit is not enough if the extra rate or amplitude is poorly determined, highly sensitive to a few samples, or biologically implausible.

The main alternative to adding a compartment is to retain a simpler structure and acknowledge its limits. A one-compartment model may be adequate when a rapid distribution phase is unobserved or unimportant for the intended predictions. A two-compartment model may capture a clearly supported distribution phase without forcing the data to identify a third. A three-compartment model is most persuasive when the design covers all relevant phases and the additional phase is stable and plausible. The introductory material on Components of a Population PK Structural Model describes the role of structural assumptions in a population model; here the focus is the number of disposition spaces.

Mechanistic Plausibility, Diagnostics, and Limits

Mechanistic plausibility is a check on interpretation, not a substitute for evidence in the concentration data. Ask whether the proposed spaces and exchange processes make sense for the drug, route, and population, and whether the implied rates and volumes are credible. A model can be a useful empirical approximation without a literal anatomical interpretation; it should not be presented as a demonstrated physiological mechanism unless the evidence supports that claim.

Several warning signs suggest that the chosen structure may be going wrong:

• Time-ordered residual patterns: repeated underprediction early and overprediction later, or the reverse, can indicate that the model misses a phase or transition.
• Unidentified extra phases: an added compartment may produce unstable parameter estimates, large uncertainty, or strong dependence on a small number of observations.
• Inadequate time coverage: no early samples means rapid distribution may be invisible; no sufficiently late samples means the terminal phase may not be established.
• Overinterpretation of a straight segment: a log-linear late decline supports a terminal exponential approximation, not a specific total compartment count.
• Poor prediction outside the observed times: a complex model can fit observed points yet extrapolate implausibly if its rates and amplitudes are weakly identified.

A common mistake is to choose three compartments because a three-exponential curve can fit a profile more closely than a one-exponential curve. The correction is to ask whether the study sampled the time ranges that distinguish the extra phase and whether estimates are sufficiently stable and plausible. If the first sample occurs after the rapid phase has passed, increasing model complexity may only make the unobserved phase difficult to estimate rather than reveal it.

No single compartment count is best for every purpose. The appropriate structure depends on the scientific question, sampling design, observed concentration-time behavior, and intended predictions. A simpler model can be preferable when added phases are unsupported; a richer model can be necessary when data clearly show distinct phases that matter for the analysis. Whichever structure is selected, describe what the data support and where the model's applicability is limited.