Why Exposure-Effect Interpretations Need an Audit
An exposure-effect analysis can be calculated correctly and still be interpreted incorrectly. A dose may be presented as though it were a concentration, an exposure summary may refer to the wrong time window, or two effect measures may be compared as if their numerical scales were interchangeable. Even a genuine exposure-effect association does not, by itself, establish that changing exposure would cause the observed effect.
A useful audit checks what was measured, when it was measured, how the effect was represented, and what conclusion the design supports. The checks below focus on interpretation rather than on fitting a concentration-effect curve.
Before interpreting a result, state the exposure metric and its units, the effect measure and its units or scale, the time window for each, and whether the claim is descriptive, associational, or causal. These details make many errors visible before any model is considered.
Check That the Exposure Is Really the Exposure Named
Dose is an amount administered, often expressed as mg or mg/day. Concentration is an amount per volume, such as mg/L. An exposure summary such as area under the concentration-time curve (AUC) combines concentration and time and may have units such as mg·h/L. These are distinct quantities; dose cannot be substituted for concentration or AUC merely because both relate to drug administration.
Dose can be a suitable predictor when the question is explicitly about the effect of assigning or receiving different doses. But a dose-based result answers a dose-effect question, not automatically a concentration-effect question. Dose does not capture all the reasons people receiving the same dose may have different concentrations. If the intended question concerns concentration or another exposure metric, that metric must be defined and measured or appropriately estimated.
Worked Example: Same Dose, Different Measured Concentrations
In a hypothetical dataset, three participants each receive 100 mg/day. Their measured trough concentrations are 2, 4, and 8 mg/L, and their measured inhibition scores are 10, 21, and 31 percent. A report calls 100 mg/day the “concentration” for all three participants. Diagnose the problem.
The reported value is a dose, with units of mg/day. The measured values 2, 4, and 8 mg/L are trough concentrations. They cannot be represented by the same numerical value as the dose: their units and definitions differ.
The mean trough concentration in these three participants is
This descriptive mean and the range from 2 to 8 mg/L show that the same administered dose did not correspond to one measured trough concentration in this small hypothetical set. The mean inhibition score is \((10+21+31)/3=20.67\) percent, but these values alone do not establish that concentration caused the differences in inhibition. The correction is to label dose and concentration separately, then interpret the observed exposure-effect pattern without claiming more than the data support.
A diagnostic check is whether the exposure’s units match its name and whether its calculation matches its definition. If a report describes an AUC but provides only a dose, or describes a concentration but reports mg/day, the exposure has been misidentified. A dose-based comparison may still be useful for a dose question, but it cannot silently stand in for a concentration-based comparison.
Match the Exposure Window to the Effect Timing
An exposure metric is tied to a time period. For example, AUC from 0 to 8 hours summarizes concentration over that interval; AUC from 8 to 12 hours summarizes a different interval. An effect measured at a particular time, or over a specified period, should be interpreted in relation to an exposure window that is pharmacologically relevant to that effect. The relevant window depends on the question and the timing of the biological response; it should not be selected merely because data are available for it.
For a concentration-time profile with measured points, the trapezoidal rule approximates AUC by summing the areas of trapezoids between adjacent observations. For two observations separated by \(\Delta t\), that interval’s area is the average of the endpoint concentrations multiplied by \(\Delta t\).
Worked Example: An Effect Measured in the Wrong Window
Suppose, in a hypothetical profile, concentration is 10 mg/L at hour 0, 6 mg/L at hour 4, 3 mg/L at hour 8, and 1 mg/L at hour 12. An effect is assessed over hours 8 to 12. A summary instead relates it to AUC from hour 0 to hour 8. Calculate both AUCs and identify the mismatch.
Using the trapezoidal rule, the AUC from 0 to 8 hours is
For the interval in which the effect is assessed, the AUC from 8 to 12 hours is
The corresponding average concentrations are \(50/8=6.25\) mg/L over hours 0 to 8 and \(8/4=2.00\) mg/L over hours 8 to 12. The first summary describes earlier exposure, not exposure during the stated effect-assessment interval. If the intended question concerns exposure during the effect window, the second interval is the temporally aligned candidate. That does not prove it is the biologically correct exposure metric: the response could depend on earlier exposure or a different window. The analysis must state and justify the relevant timing rather than treating the two windows as interchangeable.
A practical diagnostic is to put the exposure interval and effect-assessment time or interval on the same timeline. Check for a gap, an unintended overlap, or an interval that ends after the effect was recorded. If the effect is delayed, same-time exposure may not be the right predictor; the biological rationale for the lag or window must be explicit. If timing information is absent, the interpretation should acknowledge that limitation.
Keep Effect Scales and Units Distinct
The effect measure determines what a numerical difference means. A change measured in mg/L is an absolute change in concentration; a percentage change is relative to a reference value. A score on a 0-to-10 symptom scale is neither of these. Numerical values from different scales cannot be compared directly simply because they are numbers. First identify the effect definition, its units or scale, and its reference point.
Worked Example: Absolute Change Versus Relative Change
In a hypothetical comparison, a biomarker is measured before and after treatment. Group A changes from 100 mg/L to 80 mg/L; Group B changes from 50 mg/L to 40 mg/L. Compare the absolute and relative changes, stating what each represents.
Define the absolute change as follow-up minus baseline. For Group A it is \(80-100=-20\) mg/L; for Group B it is \(40-50=-10\) mg/L. Both values represent decreases, but the absolute decrease is larger in Group A.
Define percentage change relative to baseline as
For Group A, the calculation is \(100\%\times(80-100)/100=-20\%\). For Group B, it is \(100\%\times(40-50)/50=-20\%\). Thus, the groups have different absolute decreases but the same relative decrease. The values \(-20\) mg/L and \(-20\%\) do not mean the same thing and should not be treated as directly comparable effect magnitudes.
This calculation describes the two groups’ observed changes; it does not establish a treatment effect. A causal comparison would require an appropriate design and analysis, including a suitable comparison condition.
Scale choice can change which contrast is emphasized, not just the way it is written. Before interpreting an effect, check whether it is an endpoint value, an absolute change, a change relative to baseline, or a score on a bounded or otherwise defined scale. If effects from different measures are compared, state the measurement definitions and justify any transformation or common scale. A shared numerical range does not make two distinct measures equivalent.
Association Is Not Automatically a Drug Effect
An observed association means that exposure and effect vary together in the data under study. A causal claim goes further: it says that changing exposure would change the outcome, under a defined intervention and for a specified population. Confounding occurs when another variable is associated with exposure and also predicts the effect, so the exposure groups differ in more than exposure. Disease severity, for example, may influence both the concentration achieved and the probability of improvement.
The following hypothetical counts show why a pooled association can be misleading. Suppose severity is categorized as severe or mild and the outcome is whether a participant improves.
| Severity | Higher exposure: improved / total | Lower exposure: improved / total |
|---|---|---|
| Severe | 6 / 30 | 2 / 10 |
| Mild | 8 / 10 | 40 / 50 |
| All participants | 14 / 40 | 42 / 60 |
Worked Example: A Pooled Association Driven by Group Composition
Using the table, calculate the improvement proportions for the higher- and lower-exposure groups, then examine them within each severity category. What can and cannot be concluded?
The pooled proportions are \(14/40=0.350\), or 35.0%, for higher exposure and \(42/60=0.700\), or 70.0%, for lower exposure. The crude difference, higher minus lower, is \(0.350-0.700=-0.350\), or \(-35.0\) percentage points.
Within the severe category, the proportions are \(6/30=0.200\) and \(2/10=0.200\): both are 20.0%. Within the mild category, they are \(8/10=0.800\) and \(40/50=0.800\): both are 80.0%. In this constructed example, the outcome proportions are equal across exposure groups within each severity category, while the pooled proportions differ.
The reason is visible in the group composition: \(30/40=75.0\%\) of the higher-exposure group is severe, compared with \(10/60=16.67\%\) of the lower-exposure group. The crude rates average the severity-specific rates using different weights. In general, for severity categories \(j\), the pooled rate is \(\sum_j w_j r_j\), where \(r_j\) is the category-specific rate and \(w_j\) is that category’s proportion of the exposure group. Here the rates \(r_j\) match across exposure groups, but the weights \(w_j\) do not, so the pooled rates differ. The crude association therefore does not show that higher exposure reduced improvement, nor does this example prove that exposure has no effect in other populations or settings. It illustrates how confounding by severity can produce an association that is absent within the displayed categories.
A diagnostic review asks whether plausible common causes of exposure and effect were measured, whether exposure groups were comparable, and whether the design supports a causal interpretation. Random assignment can support causal interpretation under its design assumptions; observational data generally require explicit consideration of confounding and other biases. Adjustment for measured variables may help, but it cannot guarantee removal of unmeasured confounding or correct measurement and selection problems.
A Practical Diagnostic Sequence
The checks can be applied in sequence before accepting a written interpretation. Each step identifies a different failure mode; passing one check does not guarantee that the others are satisfied.
Decide whether the quantity is dose, concentration, AUC, average concentration, or another defined metric. Check that the units and calculation agree with that label.
Record the exposure window and the effect’s measurement time or interval. Justify any lag or window linking them.
Distinguish endpoint values, absolute changes, relative changes, and scores. Include units and reference values where relevant.
Describe an association as an association unless the design and assumptions justify a causal effect claim. Consider confounding and other sources of bias.
Different problems require different remedies. A dose-versus-concentration error calls for a correctly defined exposure metric or a clearly stated dose question. A timing mismatch calls for a justified, aligned window, not simply a different label. A scale mismatch calls for clear units and a suitable contrast, not an unmotivated comparison of raw numbers. A causal overstatement calls for a design-based causal argument and assessment of its assumptions; rewording an observational association cannot make it causal.
These checks also have limits. A well-labeled exposure can still be measured inaccurately; a temporally aligned metric can still be biologically inappropriate; and adjustment for recorded confounders may leave residual confounding. Diagnostics help identify unsupported interpretations, but they do not by themselves establish the correct exposure metric, prove a mechanism, or validate a causal claim.
Common Pitfalls
- Calling dose “concentration.” The correction is to use the measured or estimated concentration and its units, or to describe the analysis as dose-based.
- Using whichever exposure window is available. The correction is to define the effect timing and justify the exposure window in relation to it.
- Comparing raw numbers from different effect scales. The correction is to state what each number measures and use a defensible common contrast if comparison is needed.
- Interpreting a crude association as a drug effect. The correction is to assess design, confounding, and other assumptions before making a causal statement.
- Treating a corrected label as a corrected analysis. Renaming a dose as concentration or an association as an effect does not change the underlying measurement or evidence.
Check Your Understanding
Answer each question by naming the interpretation problem and stating a suitable correction.
- A report calls 75 mg/day a trough concentration in mg/L. What is wrong?
- An effect is assessed from hour 4 to hour 8, but the reported AUC is from hour 0 to hour 4. What should be checked before interpreting the result?
- Two groups have biomarker changes of \(-12\) mg/L and \(-12\%\). Are these directly comparable?
- Higher exposure is associated with less improvement in an observational dataset. Does this establish that higher exposure reduced improvement?
Solutions. (1) The value is dose, not concentration; report dose as mg/day or use a measured concentration in mg/L. (2) The windows differ; determine and justify the exposure window relevant to the effect, and report the interval accurately. (3) No. One is an absolute change in mg/L and the other is a relative change; define the measures and choose a justified comparison. (4) No. The result is an association; assess confounding and other design assumptions before making a causal claim.