When Does Pooling Answer the Question?
A pooled concentration mean is easy to calculate: add the recorded concentrations and divide by the number of records. The harder question is what that number represents. If some subjects contribute more observations than others, the pooled mean gives those subjects more influence. If the scientific target is an average across subjects, that weighting may be unintended. Even when the pooled mean targets the right average, repeated measurements from one subject are dependent, so treating every record as independent can misstate uncertainty.
Pooling is not automatically wrong. It can be defensible when the concentration measurements refer to a comparable time and population, the weighting matches the target, and the analysis accounts for dependence when making inferences. The earlier material on subject identity, sampling weights, and individual trajectories motivates these checks; here the goal is to bring them together into a decision about whether a pooled summary is suitable.
The target estimand must be stated before deciding whether either summary is suitable. For example, an equally subject-weighted mean at a prespecified time describes the average concentration across the target population of subjects at that time. An observation-weighted mean describes the average among the recorded observations, giving each record equal weight. The latter is not necessarily an average across subjects.
A pooled value across different times also needs a clearly defined target. It may summarize the set of observed records, but it does not automatically estimate a population mean concentration at any one time. Concentrations at different times are not interchangeable simply because they share the same units.
Conditions to Check
A pooled summary is most defensible when the following conditions support its interpretation. These are checks on the target and the data collection, not a requirement that subjects be identical.
- Target match: The scientific question concerns the same quantity the pooled mean weights. If the target is the mean across subjects, each subject should contribute equally or the observations should be reweighted to achieve equal subject influence.
- Comparable timing: Concentrations being combined should refer to the same prespecified time or a scientifically justified time window. A mean across changing times must be described as such, rather than interpreted as a fixed-time mean.
- Suitable subject sampling: Included subjects should represent the target population for the intended descriptive or inferential claim. Subjects with more measurements should not receive extra weight merely because they were sampled more often.
- Non-informative observation patterns, or adjustment: The number and timing of observations should not systematically depend on concentration in a way that changes the target. If sampling, missed visits, or dropout alter which subjects contribute at a time, raw pooling can reflect that changing composition.
- Dependence recognized: Measurements from the same subject can be correlated. This does not by itself invalidate a pooled point estimate, but it means that uncertainty calculations must not treat all records as independent subjects.
- Heterogeneity considered: A population mean can be meaningful despite large differences among subjects. But it may be a poor stand-in for a typical subject, individual predictions, or decisions near a clinical threshold.
Worked Example: Unequal Numbers of Concentrations
Suppose a hypothetical PK dataset records concentrations in mg/L at the same nominal time for three subjects. Subject A contributes \(10\) and \(12\), subject B contributes \(20\), and subject C contributes \(30\), \(32\), and \(34\). Assume the measurements are comparable for this fixed-time summary.
The observation-weighted pooled mean is
The subject-specific means are \(11\), \(20\), and \(32\) mg/L. Giving each subject equal weight gives
The difference arises because subject C supplies half of all records, while subject B supplies only one. The value \(23.0\) mg/L describes the average recorded observation; \(21.0\) mg/L describes the equally weighted mean of these three subjects’ means. If the target is the average across subjects, raw pooling is not suitable here without a weighting adjustment. The example does not establish which summary is scientifically preferred; that depends on the stated target.
Balance, Dependence, and Uncertainty
When each subject contributes the same number of comparable observations, the two means can coincide. Suppose each of \(N\) subjects contributes \(m\) concentrations, with subject mean \(\bar C_i\). Then the pooled mean can be rewritten as
This is a short justification for why balanced sampling removes the difference in point-estimate weighting. It does not prove that the records are independent, nor does it mean that a balanced pooled curve captures each subject’s trajectory. If subjects contribute the same number of records but at different times, the equality may hold for the overall record average while the fixed-time interpretation still fails.
For uncertainty about an equally subject-weighted mean, the independent sampling units are the subjects, not the individual repeated records. One straightforward illustration is to calculate a mean for each subject at the target time and then calculate uncertainty across those subject means. The appropriate inferential method depends on the design and assumptions; the key condition is that within-subject correlation must be reflected rather than ignored.
Worked Example: Balanced Records Do Not Make Them Independent
In a hypothetical study, four subjects each have two comparable concentration measurements, in mg/L, at a prespecified nominal time on two protocol occasions. The values are: subject A, \(8\) and \(10\); B, \(12\) and \(14\); C, \(16\) and \(18\); D, \(20\) and \(22\). Assume the four subjects are independently sampled, but measurements within a subject may be correlated.
The pooled mean is
The subject means are \(9\), \(13\), \(17\), and \(21\) mg/L, whose equally weighted mean is also \((9+13+17+21)/4=15.0\) mg/L. Balance makes the point estimates agree.
To illustrate the uncertainty issue, the sample standard deviation of the four subject means is \(\sqrt{80/3}=5.164\) mg/L: their squared deviations from \(15\) sum to \(36+4+4+36=80\). The standard error across subjects is therefore \(5.164/\sqrt{4}=2.582\) mg/L. In contrast, treating all eight records as independent gives a record-level sample standard deviation of \(\sqrt{168/7}=4.899\) mg/L, and a standard error of \(4.899/\sqrt{8}=1.732\) mg/L. That calculation uses the wrong independence assumption for the stated design. The numerical difference is illustrative, not a universal direction or size of error.
Heterogeneity and the Meaning of an Average
A pooled mean can be an accurate estimate of a population mean and still describe few individual subjects well. Heterogeneity is not, by itself, a reason to discard a mean; it is a reason to state what the mean can and cannot represent. If a decision concerns the average population concentration, a mean may be directly relevant. If a decision concerns whether an individual subject exceeds a threshold, the mean alone may conceal important variation.
Worked Example: A Mean That No Subject Has
At one prespecified time, four hypothetical subjects have concentrations \(4\), \(6\), \(24\), and \(26\) mg/L. Assume the subjects are representative of the target population and each is measured once using the same protocol.
The pooled mean is
No subject has a concentration near \(15.0\) mg/L: the two lower values are \(9\) and \(11\) mg/L below it, and the higher values are \(9\) and \(11\) mg/L above it. The range is \(26-4=22\) mg/L. The mean remains the arithmetic average of the four subjects and can be appropriate for that target, but it is not a description of a typical individual in this small, clearly heterogeneous sample. A subject-level plot or distributional summary would reveal what the mean hides.
Diagnostics, Alternatives, and Limitations
Before reporting a pooled summary, check the data at the level at which they were collected. A subject-by-time table or plot can reveal unequal counts, missing visits, timing differences, and subjects who dominate the pooled calculation. Compare the observation-weighted and subject-weighted summaries when both can be computed; a material difference is evidence that weighting affects the answer. Inspect individual profiles or subject-level summaries for spread and distinct patterns. A single pooled mean cannot diagnose these issues on its own.
| Approach | What it summarizes | When it is useful | Main caution |
|---|---|---|---|
| Raw pooled mean | Average of all recorded concentrations, with each record equally weighted | The record-level average is the intended estimand and sampling supports it | Subjects with more records receive more weight; dependence still matters for uncertainty |
| Equally subject-weighted mean | Average of subject-level summaries, giving each subject equal weight | The target is an average across subjects and each summary refers to a comparable time or window | Requires defining subject summaries and handling missing or mistimed observations |
| Model-based analysis | A target population summary while representing repeated measures and relevant subject variation | Unequal schedules, dependence, or covariates require explicit adjustment | Depends on model assumptions and does not remove the need to define the target |
The model-based approach is an alternative when the design is not well represented by simple pooling, but it is not automatically superior: its assumptions must also be checked. Nor can a weighting adjustment solve every problem. If subjects are missing for reasons related to their concentrations, or if the observations do not cover the target time comparably, the intended population mean may require additional assumptions or data.
Check Your Understanding
Try each question, then compare your reasoning with the solutions.
- Two subjects have concentrations \(8\) and \(12\) mg/L; a third has \(20\), \(22\), and \(24\) mg/L at a comparable target time. Calculate the observation-weighted and equally subject-weighted means. Which targets does each describe?
- Four subjects each contribute three comparable observations. Does balance make the pooled mean equal to the average of subject means? Does it make all observations independent?
- At a fixed time, subject concentrations are \(3\), \(5\), \(19\), and \(21\) mg/L. Calculate the mean and range, and state one limitation of reporting only the mean.
- Why can a pooled mean across observations at several different times fail to estimate a fixed-time population mean?
- Why is it not generally valid to calculate a standard error by treating every repeated concentration record as an independent subject?
Solutions. (1) The pooled mean is \((8+12+20+22+24)/5=86/5=17.2\) mg/L. The subject means are \(10\), \(10\), and \(22\) mg/L, so the equally weighted mean is \((10+10+22)/3=14.0\) mg/L. The first is the record average; the second is the mean of these three subject summaries. (2) Yes: with three records per subject, each subject has equal weight in the pooled mean, so it equals the average of subject means. No: measurements within a subject can remain dependent. (3) The mean is \((3+5+19+21)/4=12.0\) mg/L and the range is \(21-3=18\) mg/L. The mean alone hides the substantial separation between lower and higher concentrations and may not describe a typical subject. (4) Concentration can change over time, so observations at different times do not share the same fixed-time target; the pooled value also depends on how many records occur at each time. (5) Repeated records from one subject can be correlated, so the number of independent subjects is smaller than the record count. An uncertainty calculation must reflect the subject-level dependence rather than assume every record is independent.