From a Single Dose to an Exposure-Response Question
After one analgesic dose, plasma concentration changes over time, and pain may change at the same time. An exposure-response analysis asks whether observations with different concentrations also show different responses. This is useful for describing a possible concentration-response shape and generating a dose hypothesis. By itself, however, it does not establish that the drug caused the observed pain change: time, natural symptom fluctuation, and other influences may also affect pain.
For the examples, define \(Y\) as pain reduction from baseline: baseline pain score minus pain score at the assessment. Thus, a larger \(Y\) means greater observed improvement. Pain reduction is expressed in points on the pain scale, and concentration \(C\) is measured in mg/L. Each concentration must be paired with the response measured at the same assessment time. This analysis treats those paired observations as independent, an assumption that may be poor when several observations come from the same participant.
A Linear Description
The linear model is \(\mathbb{E}(Y_i\mid C_i)=\beta_0+\beta_1C_i\), with observations \(Y_i=\beta_0+\beta_1C_i+\varepsilon_i\). Here \(\beta_0\) is the mean response at zero concentration under the model, and \(\beta_1\) is the change in mean pain reduction per mg/L. Ordinary least squares estimates the coefficients by minimizing the sum of squared residuals. For the usual coefficient standard errors and t-based intervals, assume a correctly specified conditional mean, independent errors with mean zero and constant variance, and normally distributed errors for exact small-sample inference. A linear mean also assumes that the response is adequately represented by a straight line over the studied concentration range.
Worked Example: Fitting a Linear Response
Suppose five hypothetical assessments after a single dose give the following paired concentrations and pain reductions: \(C=(1,2,3,4,5)\) mg/L and \(Y=(0.4,1.3,2.1,3.2,3.0)\) points. These are illustrative data, not results from a real study. The concentration-response scatter would need to be inspected for curvature and unusual observations; with only five pairs, those checks have little power.
The means are \(\bar C=3\) mg/L and \(\bar Y=2.00\) points. The slope is the cross-product sum divided by the concentration sum of squares:
The intercept is \(\hat\beta_0=\bar Y-\hat\beta_1\bar C=2.00-(0.710)(3)=-0.130\) points. The fitted mean is therefore \(\hat Y=-0.130+0.710C\). The residuals, in order, are \(-0.180,\ 0.010,\ 0.100,\ 0.490,\ -0.420\) points, giving \(SSE=0.459\) points squared. With \(n=5\) and two fitted coefficients, the residual degrees of freedom are \(df=3\), and \(s^2=SSE/df=0.153\) points squared.
The standard error of the slope is \(SE(\hat\beta_1)=s/\sqrt{\sum(C_i-\bar C)^2}=\sqrt{0.153}/\sqrt{10}=0.124\) points per mg/L. Using \(t_{0.975,3}=3.182\), a 95% t interval is \(0.710\pm(3.182)(0.124)=[0.316,\ 1.104]\) points per mg/L. Under the linear-model assumptions, this interval is compatible with a positive mean slope. It does not show that increasing concentration caused greater pain reduction, and its precision is limited by the small sample.
The intercept here is an extrapolation: the observed concentrations range from 1 to 5 mg/L, not zero. Its negative estimate does not imply that zero drug concentration produces negative pain relief. It is a feature of the fitted line and a warning against interpreting the model outside its data range.
A Saturable Emax Description
Analgesic response may rise rapidly at lower concentrations and then increase less as concentration grows. A simple saturable model for mean pain reduction is
Here \(E_{\max}\) is the limiting pain reduction as concentration becomes very large, and \(EC_{50}\) is the concentration associated with half of that modelled maximum. Both must be positive for the usual interpretation. This version assumes zero mean drug-related response at zero concentration; it has no separate baseline or placebo component. An additional baseline term may be appropriate in other designs, but this tutorial does not use it to separate placebo change from drug effect.
One way to estimate the two parameters is to take reciprocals: \(1/Y=1/E_{\max}+(EC_{50}/E_{\max})(1/C)\). A straight-line regression of \(1/Y\) on \(1/C\) then has intercept \(a=1/E_{\max}\) and slope \(b=EC_{50}/E_{\max}\), so \(E_{\max}=1/a\) and \(EC_{50}=b/a\). This transformation is only valid for positive concentrations and positive responses. It also changes the error structure: ordinary least squares on reciprocal responses assumes independent, constant-variance errors on the reciprocal scale, not on the original pain-reduction scale. It can give disproportionate influence to small responses. Direct nonlinear least squares on the original response scale is a common alternative.
Estimating the Emax Parameters and Their Uncertainty
Worked Example: A Reciprocal Emax Fit
Consider a second set of five hypothetical concentration-response pairs. The table gives the reciprocal values used in the fit, rounded to three decimals; \(z=1/C\) and \(w=1/Y\). Calculations below use the displayed \(z\) and \(w\) values.
| \(C\) (mg/L) | \(Y\) (pain-reduction points, approximately) | \(z=1/C\) (L/mg) | \(w=1/Y\) (1/point) |
|---|---|---|---|
| 4.00 | 3.390 | 0.250 | 0.295 |
| 2.00 | 3.226 | 0.500 | 0.310 |
| 1.00 | 1.923 | 1.000 | 0.520 |
| 0.667 | 1.538 | 1.500 | 0.650 |
| 0.500 | 1.220 | 2.000 | 0.820 |
The ordinary least-squares fit to the displayed reciprocal values is \(w=a+bz+\varepsilon\), with \(a=0.1917\) per point and \(b=0.3117\) mg per (L·point). Thus \(\hat E_{\max}=1/a=5.216\) points and \(\widehat{EC}_{50}=b/a=1.626\) mg/L. At \(z=(0.25,0.50,1.00,1.50,2.00)\), the fitted reciprocal values are approximately \((0.26963,0.34756,0.50341,0.65925,0.81510)\). The residuals are approximately \((0.02537,-0.03756,0.01659,-0.00925,0.00490)\), giving a sum of squared errors of \(0.002439\).
For this fit, \(\bar z=1.05\) and \(S_{zz}=\sum(z_i-\bar z)^2=2.05\). Thus \(s^2=0.002439/(5-2)=0.000813\), \(SE(b)=s/\sqrt{2.05}=0.0199\), and \(SE(a)=s\sqrt{1/5+\bar z^2/S_{zz}}=0.0245\), in reciprocal-response units. The covariance is \(\operatorname{Cov}(a,b)=-\bar z s^2/S_{zz}=-0.000416\). Using the delta method for \(E_{\max}=1/a\) gives \(SE(\hat E_{\max})=SE(a)/a^2=0.666\) points.
For \(EC_{50}=b/a\), the gradient with respect to \((a,b)\) is \((-b/a^2,1/a)\approx(-8.481,5.216)\). Applying that gradient to the coefficient covariance matrix gives \(SE(\widehat{EC}_{50})\approx0.301\) mg/L. With \(df=3\), approximate 95% delta-method intervals are \(5.216\pm(3.182)(0.666)=[3.10,\ 7.34]\) points for \(E_{\max}\), and \(1.626\pm(3.182)(0.301)=[0.667,\ 2.585]\) mg/L for \(EC_{50}\). These are approximate intervals based on the reciprocal-scale regression and a local normal approximation for the transformed parameters; the small sample and transformation make them fragile.
This worked fit is based on the displayed reciprocal values in the table. In particular, using \(w=0.820\) at \(z=2.00\) gives a final residual of approximately \(0.00490\), as included in the SSE above. Such details matter: a rounded table or an inconsistent residual can alter standard errors and intervals.
From a Response Target to a Candidate Dose
Worked Example: Translating a Target Into a Dose
Use the illustrative Emax estimates above to find the concentration associated with a target pain reduction of 4.00 points. For \(Y^*=E_{\max}C/(EC_{50}+C)\), solving for concentration gives \(C=EC_{50}Y^*/(E_{\max}-Y^*)\), provided \(0<Y^*<E_{\max}\). Substitution gives \(C=(1.626\ \text{mg/L})(4.00\ \text{points})/(5.216-4.00\ \text{points})=5.35\) mg/L.
Suppose, purely for illustration, that a simple PK calculation predicts a concentration of \(0.200\) mg/L per mg of dose at the chosen assessment time. Then \(C=0.200D\), with \(D\) in mg, and the candidate dose is \(D=5.35/0.200=26.8\) mg. This is a model-based candidate, not a dosing recommendation. The assumed concentration-per-dose relationship must be appropriate for the intended time and population, and the dose must be assessed against safety, formulation, and clinical evidence.
The target is close to the estimated maximum. The estimated \(E_{\max}\) interval includes values below 4.00 points, in which case this target is not attainable under the model. That uncertainty makes the 26.8 mg calculation especially fragile. A full uncertainty analysis would propagate joint uncertainty in \(E_{\max}\), \(EC_{50}\), and the PK relationship; substituting point estimates alone does not do so.
Diagnostics, Alternatives, and Limits
Before interpreting either fit, plot pain reduction against concentration and inspect residuals against fitted values and concentration. Curvature in a linear-model residual plot suggests that a straight-line mean may be inadequate. A pattern in residual spread suggests nonconstant variance. Check for influential observations, especially at the high-concentration end, because those observations can strongly determine an apparent plateau. For the reciprocal fit, also inspect residuals against \(1/C\) and \(1/Y\); a visually good reciprocal-scale fit does not guarantee a good fit on the pain scale. Small or noisy \(Y\) values can dominate after inversion.
The linear model is easy to interpret within the observed concentration range but predicts unbounded response as concentration increases. The Emax model represents a plateau, but its parameters may be weakly identified if observations do not cover both the rising portion and the flattening portion of the curve. A direct nonlinear fit avoids the reciprocal transformation but requires numerical optimization and suitable residual assumptions. Model choice should be informed by the design, plots, residuals, and parameter uncertainty, not simply by which curve appears more appealing.
A common mistake is to treat a positive slope or an Emax curve as proof that raising the dose will produce the predicted benefit. The correction is to distinguish association from causal effect, check that the model fits over the relevant range, and account for uncertainty in both the PK and response relationship. Extrapolating beyond observed concentrations—especially toward an estimated plateau—can produce misleading dose targets.
Specify the pain scale, response direction, assessment time, and concentration matched to that time.
Estimate linear and, when justified, saturable responses; examine residuals and the range of observed concentrations.
Report parameter uncertainty and avoid presenting a point-estimate dose as a dependable or validated regimen.
Check Your Understanding
Answer the questions, then compare your reasoning with the concise solutions.
- If pain reduction is defined as baseline score minus post-dose score, what does a larger response mean?
- For an Emax model with \(E_{\max}=6\) points and \(EC_{50}=2\) mg/L, what response is predicted at \(C=2\) mg/L?
- Why can reciprocal regression give a different fit from direct nonlinear least squares on the original response scale?
- If a desired response equals or exceeds the estimated \(E_{\max}\), what happens to the target-concentration calculation?
- Why is a positive concentration-response association after one dose not, by itself, evidence that increasing dose causes greater analgesia?
Solutions. (1) Greater observed improvement in pain. (2) \(6(2)/(2+2)=3.00\) pain-reduction points, half the modelled maximum. (3) It minimizes squared errors in \(1/Y\), not in \(Y\), and assumes a different error structure. (4) The target is unattainable if it exceeds \(E_{\max}\); at equality the required concentration tends to infinity in the model. (5) Concentration may co-vary with time and other causes of pain change, and the analysis lacks a design feature that isolates the drug effect.