From Model Output to a Coherent Report
A population pharmacokinetic report is not a list of estimates or a collection of figure captions. Its methods should explain what was analyzed and how the model was constructed; its results should state what the analysis produced; and its interpretation should explain what those results support, and what they do not. Limitations make the boundaries of those claims explicit.
This workshop uses a single, entirely hypothetical analysis. Imagine a study summary, a final-model output, and evaluation figures supplied to a report writer. The task is to connect them accurately, not to invent missing details or to repeat every available output. The examples develop separate parts of one report so the links between sections remain visible.
The hypothetical study enrolled 72 adults who received an intravenous drug. There were 432 quantifiable concentration observations, with samples collected at varying times after dosing. The supplied final model was a one-compartment model with linear elimination. Clearance (CL) and volume of distribution (V) had separate log-normal between-subject random effects; their covariance was not estimated. Residual error was proportional. The supplied summary reports first-order conditional estimation with interaction, also called FOCE-I.
The supplied parameter output gives typical CL of 5.2 L/h (relative standard error, RSE, 12%) and typical V of 42 L (RSE 15%). Between-subject variability was reported as coefficients of variation (CVs): 31% for CL and 28% for V. The proportional residual-error standard deviation was 0.18. A nonparametric bootstrap had 500 successful runs out of 500; its 95% percentile intervals were 4.1–6.7 L/h for CL and 31–58 L for V.
The supplied evaluation figures show conditional weighted residuals generally centered near zero against time and predictions, with a mild negative trend at late times. The prediction-corrected visual predictive check (pcVPC) shows the observed central trend broadly within the simulated prediction intervals, but only 76% of observations in the latest-time bin fall within its 90% prediction interval. Across the displayed bins, 87% fall within the corresponding 90% intervals. These are descriptive summaries of the supplied figures, not proof that the model is correct.
Establishing What the Report Can Say
Before drafting, separate supplied facts from interpretation. The study summary supports statements about the population, treatment, and observed data. The model output supports statements about the model structure, estimation method, and estimates. The figures support statements about the displayed diagnostic patterns. A report should not fill gaps by assuming, for example, that a covariate was tested or that a separate validation dataset was used when neither appears in the supplied material.
The model assumptions also matter to the wording. In this hypothetical analysis, the structural model assumes one compartment and linear elimination; the random effects represent between-subject variation in CL and V on a log scale; and the residual error is proportional to predicted concentration. These are assumptions of the fitted model, not facts established about every individual. Conclusions should therefore be phrased as findings under this model, and evaluation figures should be used to discuss whether the model describes the supplied data adequately.
Worked Example: Drafting a Methods Paragraph
The study summary gives 72 adults and 432 quantifiable observations. First calculate the average number of observations per participant:
This is an average, not a claim that every participant contributed exactly six samples. The summary says sampling times varied, so the paragraph should not imply a common schedule or uniform number of samples. The model and estimation details are also supplied, so they can be stated directly.
Draft: “Data from 72 adults who received the study drug intravenously were analyzed using a one-compartment model with linear elimination. The dataset contained 432 quantifiable concentration observations (a mean of 6.00 observations per participant); sampling times varied across participants. Between-subject random effects were included separately on CL and V, without estimating their covariance, and residual error was modeled as proportional. Parameters were estimated using FOCE-I.”
This paragraph reports only supplied facts and the transparent arithmetic from them. If the study summary does not specify the handling of below-quantification observations, dose range, or software version, those details should not be guessed. They can be added only if verified from the analysis record.
A concise methods section is selective, but it must retain details needed to understand and reproduce the reported analysis. A useful check is whether the reader can tell what data were modeled, the main structural and variability assumptions, and how the parameters were estimated. Avoid turning the methods into a second results section: parameter estimates and diagnostic findings belong later.
Reporting Estimates and Evaluation
Results should name the parameter, estimate, unit, and uncertainty measure. The RSE describes the estimated standard error relative to the estimate; it is not an interval containing a fixed proportion of the parameter. For an estimate \(\hat{\theta}\), the usual definition is:
The bootstrap intervals provide another description of uncertainty. A 95% percentile interval is formed from the 2.5th and 97.5th percentiles of the successful bootstrap estimates. Under repeated sampling and the assumptions of the bootstrap procedure, this construction is intended to capture the parameter about 95% of the time. It is not a statement that there is a 95% probability that the fixed parameter lies in this particular interval.
Worked Example: Turning Parameter Output into Results
For CL, the reported estimate is 5.2 L/h and the RSE is 12%. Rearranging the RSE definition gives an approximate standard error:
The output also supplies a bootstrap interval of 4.1–6.7 L/h. The report should preserve both forms of uncertainty and identify the bootstrap result as an interval, rather than presenting the RSE-derived standard error as though it were the same quantity.
Draft: “Typical CL was estimated as 5.2 L/h (RSE 12%; bootstrap 95% percentile interval, 4.1–6.7 L/h), and typical V as 42 L (RSE 15%; bootstrap 95% percentile interval, 31–58 L). Estimated between-subject CVs were 31% for CL and 28% for V; the proportional residual-error standard deviation was 0.18. All 500 bootstrap runs were successful.”
The final sentence is justified by the supplied count: 500 successful runs out of 500 attempted. It describes computational success for these runs; it does not by itself establish model adequacy or eliminate uncertainty in the estimates.
Evaluation findings should likewise identify what is shown, including relevant departures from the overall pattern. “The diagnostics were good” hides the evidence and can conflict with visible features. Conversely, a single imperfect diagnostic should not be inflated into a conclusion that the model is unusable. State the pattern, its location, and its potential relevance.
Worked Example: Writing a Balanced Evaluation Result
The pcVPC’s target interval is 90%, while the observed fraction inside the interval in the latest-time bin is 76%. The difference is:
The latest-time bin therefore has 14 percentage points less observed coverage than the nominal interval level. This descriptive comparison does not supply a formal hypothesis test, and it does not show by itself why coverage is lower. The overall displayed-bin summary, 87%, is three percentage points below 90%:
Draft: “Conditional weighted residuals were generally centered near zero across time and predictions, although a mild negative trend was visible at late times. In the pcVPC, 87% of observations across the displayed bins fell within their corresponding 90% prediction intervals; coverage was lower in the latest-time bin (76%). Thus, the overall predictive pattern was broadly compatible with the observed central trend, but late-time coverage warrants attention.”
The wording distinguishes the broad pattern from the localized concern and avoids claiming that the figure proves predictive performance for new populations or settings. Because the supplied figure summaries are descriptive, they should not be reported as inferential test results.
Connecting Interpretation and Limitations
Interpretation is not a more emphatic version of the results paragraph. It explains what the evidence suggests while preserving its conditions. Here, estimates quantify typical CL and V under the stated model, and the diagnostics show broadly reasonable but not uniform agreement with the observations. They do not establish that the one-compartment model is the only plausible structure, that the parameters apply to a different population, or that the model predicts every time range equally well.
A short justification helps keep claims calibrated: a conclusion cannot be stronger than the evidence that supports it. The chain is: the analysis produces estimates conditional on a model; uncertainty summaries describe precision under specified procedures; evaluation figures describe fit or predictive behavior in the supplied data; interpretation combines those pieces while retaining their limitations. If a link is missing, the report should not jump to a broader claim.
Worked Example: Linking a Finding to a Limitation
The late-time pcVPC coverage is 76% against a nominal 90%. The difference is 14 percentage points, as calculated above. The mild negative late-time residual trend points in a compatible direction, but neither result alone identifies a cause. A report can therefore flag a possible late-time mismatch without asserting that elimination is misspecified or that the model fails for all purposes.
Draft interpretation: “The model captured the overall concentration trend in the evaluated data, but the late-time diagnostics suggest that predictions in this region may be less reliable than predictions over the rest of the displayed range.”
Draft limitation: “Evaluation was based on the supplied diagnostic figures for this study population. The lower coverage in the latest-time pcVPC bin and the mild late-time residual trend limit confidence in late-time predictions. The available summary does not establish performance in an external population or identify the source of the late-time pattern.”
These statements are narrower than “the model accurately predicts concentrations” and more informative than “the model has limitations.” They identify the evidence, the scope of concern, and an unresolved question. A limitation should not be added as boilerplate: it should follow from a feature of the analysis or the evidence.
The main alternatives when assembling a report are not competing statistical models but competing ways of presenting the same analysis. A narrative alone may omit numerical details; a table alone may leave the reader to infer the conclusion; and figure captions alone may not connect evaluation to the report’s interpretation. Use each form for its purpose: prose states the claim, tables organize estimates, and figures show patterns that prose can summarize. The report should not duplicate every number in both text and table, but the central findings and qualifications should be understandable without requiring the reader to reconstruct them from scattered outputs.
A useful assembly sequence is to draft methods from the study record, results from model output and evaluation figures, and interpretation only after checking those results. Then write limitations against the exact claims made. Finally, compare the sections for consistency: the stated model must match the model whose estimates are reported, units must agree, and interpretation must reflect any diagnostic concerns.
Separate study facts, model specifications, numerical output, and figure-based findings. Mark anything not supplied rather than filling it in by assumption.
Describe the analysis and its assumptions, then report estimates, uncertainty, and evaluation patterns with units and appropriate qualifiers.
Interpret what the estimates and diagnostics support, state the scope of the evidence, and make limitations specific to the analysis.
Check that every interpretation traces to a reported result and that no result contradicts the methods, figures, or stated limitations.
Common Pitfalls and Final Checks
A common mistake is to convert a diagnostic into a verdict: “87% were inside the 90% interval, so the model passed validation.” The correction is to report the observed coverage and its variation by bin, then describe what that pattern suggests. The supplied summary does not define a pass threshold, and one overall percentage can conceal localized disagreement.
- Do not confuse precision with validity. A bootstrap interval or low RSE addresses uncertainty in estimates; neither proves that the model structure is adequate.
- Do not turn descriptive coverage into a probability about the model. The 76% figure describes the displayed latest-time bin; it is not the probability that the model is correct.
- Do not omit units or uncertainty labels. “CL was 5.2” is incomplete; write “typical CL was 5.2 L/h” and identify the uncertainty summary.
- Do not imply causation from a pattern. A late-time residual trend and lower pcVPC coverage indicate a feature to investigate, not a proven cause.
- Do not report absent information as fact. If external validation, covariate testing, or data-handling details are not in the supplied materials, do not claim them.
A report is ready when the reader can trace its central claims from study description to model, output, evaluation, and appropriately bounded interpretation. Coherence is not achieved by making every section sound positive; it is achieved by making each section answer a distinct question while remaining consistent with the same evidence.
Check Your Understanding
Answer each question, then compare your response with the solutions below.
- Why should a methods paragraph say that 432 observations represent a mean of six per participant rather than stating that every participant contributed six?
- For a typical CL estimate of 5.2 L/h with RSE 12%, calculate the approximate standard error.
- The latest-time pcVPC bin has 76% coverage within a 90% interval. What is the coverage difference in percentage points, and what can the report conclude?
- Why do 500 successful bootstrap runs not establish that the model is adequate?
- Write one limitation supported by the supplied evaluation summary without claiming a cause for the late-time pattern.
Solutions. (1) \(432/72=6\) is an average; the study summary says sampling times varied and does not establish identical counts for all participants. (2) \(\mathrm{SE}=0.12\times5.2=0.624\) L/h, approximately 0.624 L/h. (3) \(90\%-76\%=14\) percentage points. The report can say coverage was lower in the latest-time bin and warrants attention, but the descriptive figure does not identify a cause or prove general model failure. (4) Bootstrap success indicates that all 500 specified runs completed; it does not assess whether the structural assumptions or predictive behavior are adequate. (5) For example: “The lower late-time coverage and mild late-time residual trend limit confidence in late-time predictions; the supplied evaluation does not establish the cause or performance in an external population.”