Why Reporting Requires More Than a Parameter Table
A fitted concentration-effect curve can be numerically precise in some respects and weakly supported in others. A small standard error for one parameter does not establish that the curve is appropriate across all concentrations, and a smooth fitted line does not show that concentrations outside the observed range have been validated. A useful report therefore connects the model, its uncertainty, the data’s concentration support, and diagnostic findings.
The aim is not to repeat every fitting detail. It is to let a reader determine what was estimated, how uncertain the estimate is, where the observations inform the curve, and which conclusions should remain tentative. The report should describe a modeled relationship between concentration and response; by itself, a fitted direct-effect curve does not establish that changing concentration causes a response change.
These expressions describe the fitted concentration-response component. They do not specify how any reference response is decomposed into baseline, placebo, or natural-history contributions. Report the response scale and any reference condition used by the analysis, but do not infer such a decomposition from the curve alone.
What a Concise Report Should Contain
A defensible report usually includes five elements: the model form; parameter estimates with uncertainty; the number and range of observed concentrations; the main residual findings; and a boundary on interpretation. State concentration and response units wherever applicable. If a parameter is defined on a transformed scale, identify that scale and report the interpretation on the original scale when useful.
Uncertainty intervals should match the estimation method and the parameter’s behavior. A Wald interval is an estimate plus or minus a critical value times its standard error. It is convenient when the estimate is sufficiently regular and approximately normal on the reported scale. Nonlinear parameters can be skewed, correlated, or weakly identified, so profile-likelihood or other suitable intervals may be more informative than a symmetric Wald interval. Name the interval method and confidence level rather than presenting bounds without context.
This formula uses a \(t\) critical value when the relevant residual degrees of freedom are used for the stated approximation. Its interpretation is conditional on the model and interval assumptions: over repeated comparable samples, the procedure is designed to cover the fixed parameter at the stated rate. It does not mean there is a 95% probability that a particular fixed parameter lies in the interval.
Worked Example: Reporting a Linear Concentration-Effect Estimate
Suppose a hypothetical analysis of 20 independent response observations uses \(m(C)=a+bC\), with concentration in mg/L and response in response units. The fitted slope is \(b=1.20\) response units per mg/L, with \(SE(b)=0.30\) response units per mg/L and \(df=18\). Assume the residual diagnostics do not show a clear pattern against fitted values or concentration, and the measured concentrations range from 0.5 to 4.0 mg/L.
For a two-sided 95% Wald interval, \(t_{0.975,18}=2.101\). The margin of error is \(2.101(0.30)=0.6303\) response units per mg/L, so the interval is \(1.20\pm0.6303\), or approximately 0.570 to 1.830 response units per mg/L.
A suitable report is: “Over the observed concentration range of 0.5–4.0 mg/L, the fitted linear model estimated an increase of 1.20 response units per mg/L (95% Wald CI 0.570–1.830). Residual plots showed no clear concentration-related pattern. This slope summarizes the fitted relationship within the observed range; it should not be assumed to continue linearly at higher concentrations.”
The interval calculation is justified only to the extent that the model and its standard-error approximation are adequate. The stated residual review is reassuring but does not establish linearity or guarantee the interval’s coverage in this particular analysis.
Concentration Support and Nonlinear Parameters
For an Emax curve, \(EC_{50}\) is the concentration at which the modeled change from \(E_0\) is half of \(E_{\max}\), provided the model and parameter estimates support that interpretation. A fitted \(EC_{50}\) can still be poorly determined if the observations do not adequately cover the rising portion and the approach toward saturation. Likewise, estimating a sigmoid steepness parameter \(\gamma\) requires information about the transition’s shape; a numerical estimate alone does not demonstrate that the data distinguish a steep transition from a shallow one.
Describe the observed concentration support directly. Give the range and, where relevant, note whether observations are sparse near the lower or upper end, whether concentrations cluster in a narrow band, and whether the data include a visible plateau. Do not call an estimated \(E_{\max}\) an observed maximum effect unless the data actually support that interpretation. When the upper observed concentrations remain on the rising part of the fitted curve, the estimated plateau can depend strongly on the chosen model and its assumptions.
Worked Example: Reporting an Emax Fit With Limited Plateau Support
Consider a hypothetical Emax fit to response observations measured in response units, with concentration in mg/L. The model is \(m(C)=E_0+E_{\max}C/(EC_{50}+C)\). The estimates are \(E_0=10\) response units, \(E_{\max}=30\) response units, and \(EC_{50}=2.0\) mg/L. A profile-likelihood 95% interval for \(E_{\max}\) is 18–47 response units, and the profile interval for \(EC_{50}\) is 1.1–5.8 mg/L. Observed concentrations range from 0.2 to 3.0 mg/L, with few observations above 2.0 mg/L.
At \(C=EC_{50}=2.0\) mg/L, the model gives \(m(C)=10+30(2.0)/(2.0+2.0)=10+15=25\) response units. This calculation illustrates the half-maximum-change definition: 25 is halfway from \(E_0=10\) to the fitted asymptote \(E_0+E_{\max}=40\). It does not show that the observed responses reached 40.
A suitable report is: “The Emax model estimated \(E_{\max}=30\) response units (95% profile-likelihood interval 18–47) and \(EC_{50}=2.0\) mg/L (95% profile-likelihood interval 1.1–5.8). Observed concentrations were 0.2–3.0 mg/L, with few observations above 2.0 mg/L; therefore, the asymptotic maximum is model-based and the data provide limited direct support for a plateau.”
The asymmetric uncertainty method is stated because nonlinear parameter profiles need not be symmetric. The concentration range and sparse upper support make a restrained plateau interpretation appropriate even though the fitted curve has a finite asymptote.
Worked Example: Reporting a Sigmoid Emax Fit
In a separate hypothetical analysis, a sigmoid Emax model is fitted to response observations in response units over concentrations of 0.1–3.0 mg/L. The estimates are \(E_0=8\) response units, \(E_{\max}=24\) response units, \(EC_{50}=1.5\) mg/L, and \(\gamma=2.2\). Profile-based 95% intervals are 14–39 response units for \(E_{\max}\), 0.7–3.4 mg/L for \(EC_{50}\), and 1.1–5.6 for the dimensionless \(\gamma\). Residual plots show no strong systematic trend, but only a few observations are above 2.0 mg/L.
At \(C=EC_{50}=1.5\) mg/L, the concentration term in the sigmoid model is \(C^\gamma/(EC_{50}^\gamma+C^\gamma)=1/2\), because \(C^\gamma=EC_{50}^\gamma\). Thus the fitted response is \(8+24/2=20\) response units. This confirms the parameter interpretation at that concentration; it does not establish that the model’s curvature or asymptote is well determined.
A concise report is: “The sigmoid Emax fit estimated \(E_{\max}=24\) response units (95% profile-based interval 14–39), \(EC_{50}=1.5\) mg/L (0.7–3.4), and \(\gamma=2.2\) (1.1–5.6). Residual plots showed no strong systematic trend. Because observed concentrations were 0.1–3.0 mg/L and few exceeded 2.0 mg/L, estimates of the upper plateau and sigmoid steepness should be interpreted cautiously.”
The wide interval for \(\gamma\) signals uncertainty in steepness, while the limited high-concentration support cautions against treating the fitted asymptote as directly observed. A lack of obvious residual structure does not remove either limitation.
Justifying the Boundary on Interpretation
The distinction between interpolation and extrapolation follows from what concentrations inform the fit. Within the observed range, the fitted curve is constrained by data at concentrations in that range, although uncertainty can still be substantial. Outside the range, the curve’s form determines how it continues without direct observations there. The linear model continues with a constant slope; Emax forms approach an asymptote. Those are model properties, not empirical confirmation beyond the data.
This is a practical reporting distinction, not a claim that all in-range predictions are reliable or all out-of-range predictions are useless. A prediction near a sparsely sampled edge may be uncertain, and an out-of-range prediction may be useful for a stated modeling purpose. The report should make the evidential boundary visible and describe the prediction as dependent on the assumed model form.
Residual findings should also be summarized with restraint. State which plots or checks were examined and whether they showed a notable pattern, such as curvature, changing spread, or an outlier. Do not write “the model was validated” merely because residuals appear unstructured. Residual diagnostics can reveal incompatibility with a model; they cannot prove that the mean function, error assumptions, or extrapolated behavior are correct.
Give the equation or model family, response scale, concentration units, and the parameter estimates that are central to the interpretation.
Report confidence level and interval method; use a method that is defensible for the parameter’s shape and identifiability.
Give the observed concentration range and mention sparse regions or lack of plateau support when relevant.
Report notable residual findings, then qualify interpretations and predictions that rely on unsupported concentrations or model assumptions.
Choosing What to Emphasize
The model family changes what deserves emphasis in the report. For a linear fit, the slope and the range over which its constant-slope interpretation is used are central. For Emax, emphasize uncertainty in \(EC_{50}\) and \(E_{\max}\), and whether high concentrations support a plateau. For sigmoid Emax, also report uncertainty in \(\gamma\), because an apparently precise midpoint does not necessarily mean the transition shape is well identified.
| Model | Useful reporting emphasis | Common limit |
|---|---|---|
| Linear | Slope and its units; observed concentration range | A constant slope may not continue outside the observed range |
| Emax | \(E_{\max}\), \(EC_{50}\), interval method, and high-concentration support | The asymptote may be weakly supported without observations near a plateau |
| Sigmoid Emax | \(E_{\max}\), \(EC_{50}\), \(\gamma\), and uncertainty in curve shape | Parameters describing steepness and position may be uncertain when the transition is sparsely observed |
Reporting one preferred model does not require hiding plausible alternatives. If the analysis considered linear, Emax, or sigmoid Emax forms, state the basis for the reported choice and whether conclusions important to interpretation change under reasonable alternatives. Model comparison can inform that choice, but a favored fit is not proof that its structural assumptions are true.
Common Mistakes and Failure Modes
- Reporting estimates without units or intervals. A value such as \(EC_{50}=2\) is incomplete without concentration units and uncertainty.
- Calling the fitted asymptote an observed effect. \(E_0+E_{\max}\) is the model’s limiting response as concentration increases; if the data do not approach a plateau, say it is estimated by extrapolating the fitted curve.
- Over-reading a central estimate. A single \(EC_{50}\) or \(\gamma\) estimate does not convey weak identification. Report its interval and the concentration support that informs it.
- Treating unstructured residuals as proof. Diagnostics can fail to reveal a problem, especially with limited data; they do not establish causal interpretation or validate extrapolation.
- Reporting an interval without its method. A symmetric Wald interval and an asymmetric profile interval need not agree. Identify the method and avoid implying more precision than it provides.
- Confusing association with intervention effect. A direct-effect curve describes a fitted relationship under its analysis assumptions. Causal language requires a design and assumptions that support causal interpretation.
A practical check is whether a reader can tell which claims are supported by observed concentrations and which follow from the mathematical form of the model. If that distinction is absent, the report risks presenting model-based extrapolation as data-confirmed behavior.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A linear model estimates a slope of 0.8 response units per mg/L. Observations cover 1–5 mg/L. What should a report avoid claiming about concentrations of 10 mg/L?
- An Emax fit has \(E_{\max}=20\) response units and \(E_0=5\) response units. What response does the model give at \(C=EC_{50}\), and does that prove the data observed the asymptote?
- Why might a profile interval be preferable to a symmetric Wald interval for \(EC_{50}\)?
- Residual plots show no strong pattern, but only a few observations are near the highest measured concentration. Give one appropriate diagnostic statement and one limitation.
- A report gives \(EC_{50}=1.4\) without units or uncertainty. Name two missing details that would improve it.
Solutions. (1) It should not assert that the fitted slope continues to 10 mg/L as an empirically supported relationship; 10 mg/L is outside the observed range, so any prediction depends on the assumed linear form. (2) At \(C=EC_{50}\), the model response is \(E_0+E_{\max}/2=5+10=15\) response units. This is not evidence that observations reached the asymptote of 25 response units. (3) The likelihood profile for \(EC_{50}\) may be asymmetric or reveal weak identification, which a symmetric standard-error interval can conceal. (4) One could report that the examined residual plots showed no strong systematic pattern, while noting that upper-range support is sparse and the plateau is therefore uncertain. (5) State the concentration units and an uncertainty interval with its confidence level and method.