Goodness-Of-Fit Diagnostics Under Sparse Sampling

Why Sparse Sampling Changes the Diagnostic Question

In a population pharmacokinetic analysis, the goodness of fit is assessed from observed concentrations and model predictions across participants and sampling times. When each participant contributes many measurements, their residuals may show a recognizable within-person pattern. Under sparse sampling, a participant may contribute only a few concentrations, perhaps at different times from those of other participants. A pooled plot can still reveal a systematic population-level pattern, but it cannot turn a few measurements into a well-resolved individual concentration-time trajectory.

The key is to separate two questions: Do residuals show a systematic pattern across the data? and Do the available observations establish how well the model describes a particular participant’s trajectory? The first can be addressed by pooled plots if sampling coverage is adequate and the points are interpreted in context. The second is often weakly supported when each participant has only a few observations.

A residual is a difference between an observed value and a model prediction. For illustration, let \(C_{ij}\) be the observed concentration for participant \(i\) at observation \(j\), and let \(\widehat{C}_{ij}\) be the corresponding prediction. The raw residual is \(e_{ij}=C_{ij}-\widehat{C}_{ij}\), with the same concentration units as the data. A standardized or weighted residual scales that discrepancy using an error scale specified by the residual definition; a simple illustrative form is \(r_{ij}=e_{ij}/s_{ij}\), where \(s_{ij}\) is the relevant standard deviation in concentration units. Such a residual is dimensionless. Different residual types use different variance definitions, so a plotted value must be interpreted according to how it was constructed.

Under an adequate model and an appropriate residual definition, residuals should not exhibit a systematic trend against relevant predictors such as time after dose or prediction. This does not mean that every residual must be near zero, nor that a pattern-free plot proves the model correct. In population data, repeated measurements from the same participant share model components and are not automatically independent. Pooled plots are descriptive diagnostics, not a substitute for accounting for the data structure in formal inference.

What a Pooled Plot Can and Cannot Show

Pooling provides a broad view of whether discrepancies tend to vary with time, concentration, or another plotted quantity across the observed population. It is especially useful when individual schedules are sparse but collectively cover different parts of the dosing interval. For example, if many participants contribute early and late samples, the pooled plot can suggest a residual trend across those phases even though no single participant was sampled densely across the interval.

That strength depends on the data actually covering the x-axis. A smooth curve drawn through a pooled plot cannot supply information where observations are absent. Nor does a large total number of points guarantee broad representation: if a few participants contribute many observations, they may dominate the visible pattern. Where feasible, note the number of participants represented across time regions, not just the total number of observations.

Pooling can also conceal phase-specific bias. Positive residuals in one part of the interval and negative residuals in another may average to approximately zero overall. A single mean residual, histogram, or overall summary is therefore not a sufficient check for time-related misspecification. Inspect residuals against time after dose and consider whether the design provides observations in the phases being compared. Choosing time since dose versus clock time should follow the dosing and observation context; these axes do not answer identical questions.

Subject-level displays answer a narrower question: what discrepancies were observed for a particular participant at the sampled occasions? With two or three measurements, a plot may show where those points lie relative to predictions, but the gaps between them remain unobserved. The points alone generally cannot determine the participant’s peak concentration, terminal decline, or the shape of the intervening trajectory. This is a limitation of information, not proof that the model fits or fails for that participant.

Why an Overall Residual Can Hide a Time Pattern

The arithmetic behind this problem is simple. Suppose residuals are grouped into two phases, early and late. Their pooled mean is weighted by the number of observations in each phase. Opposing phase-specific means can cancel, while unequal numbers of observations can make one phase dominate the summary.

Here \(n_E\) and \(n_L\) are the numbers of residuals in the early and late phases, and \(\bar{r}_E\) and \(\bar{r}_L\) are their respective means. This identity is a weighted average, not an assumption that observations are independent. It explains why a near-zero pooled mean cannot rule out opposing phase patterns. The same issue arises when summarizing across participants: observation-weighted summaries give more weight to participants who contribute more records.

Subject Representation and Unequal Contribution

A pooled plot often treats each observation as a point, but the points do not necessarily represent the same number of participants across the x-axis. One subject with ten records contributes ten points; a subject with one record contributes one. Thus, an observation-level summary answers a different descriptive question from a summary that first averages within each participant and then gives each participant equal weight. Neither summary is universally correct: the appropriate view depends on whether the diagnostic question concerns observed records or typical participants. Showing participant counts or identifying unequal contributions helps avoid overinterpreting a dense cluster as broad population evidence.

Worked Interpretation Under Sparse Sampling

Assumptions, Diagnostics, and Failure Modes

Before interpreting a pooled or subject-level residual plot, check the conditions that make its visual message meaningful:

• Comparable residual definitions: points on the same plot should use a consistent residual definition and scale. Raw residuals in mg/L should not be read as directly comparable in magnitude across markedly different concentration levels when residual variability changes with concentration.
• Correct time reference: time should represent the relevant dosing interval, and dose timing should be known well enough to place observations meaningfully. A mislabeled or inconsistently defined time axis can create or hide apparent trends.
• Coverage: assess where observations occur and which participants contribute them. A blank time region is not evidence of good fit in that region.
• Dependence and representation: repeated residuals from one participant are related through shared individual features and should not be mistaken for independent replication. Dense contributions can dominate a pooled visual pattern.
• Appropriate prediction and variance: clarify whether residuals are based on population predictions, individual predictions, or another defined prediction, and how their scale was calculated. These choices affect what a pattern means.

Common failure modes include declaring good fit because the pooled residual mean is near zero, interpreting a smooth pooled trend as a trajectory followed by each participant, and inferring individual fit from only a few points. Another is treating the total number of concentrations as if it were the number of independent participants. The correction is to describe the evidence at the level it supports: pooled plots show patterns among observed records; sparse subject-level plots show discrepancies at sampled occasions.

A pooled plot is often the most informative first view of systematic population-level residual patterns. A subject-level plot adds context about concentration and residuals for a participant but becomes less conclusive as that participant’s sampling becomes sparse. Binned summaries and smooth curves can help reveal broad trends, but should not be mistaken for extra data; sparse bins and uneven sampling can make them unstable. Stratifying by phase or relevant group can uncover patterns hidden by aggregation, provided strata retain enough coverage and are not selected merely to make a plot appear cleaner. These displays complement rather than replace one another.

The practical limitation is fundamental: diagnostics only evaluate observed data under the sampling design and chosen model. Sparse sampling can leave an individual’s curve weakly identified, while a population-level pattern remains detectable if many participants collectively provide relevant coverage. Conversely, even a large pooled data set may be uninformative for a phase represented by few participants. State those limits when reporting conclusions rather than treating visual absence of a trend as proof of adequacy.

A Practical Reading Sequence