Why the Intercept and Slope Matter
A concentration-QTc analysis relates drug concentration to change in corrected QT interval. The earlier material on baseline-adjusted and time-matched QTc change establishes the response used here: a change from baseline measured at a time matched to the concentration. The linear model summarizes how the expected change varies with concentration. Its two central quantities are the intercept, the modeled change at zero concentration, and the slope, the modeled change per concentration unit.
These quantities are clinically interpretable only when their reference conditions and units are explicit. The intercept is not automatically a drug effect at zero concentration, and the slope is not a prediction of an individual's response. Both are properties of the specified model and the data used to fit it.
If concentration is recorded in ng/mL, then \(\beta_1\) has units of ms per ng/mL, while \(\beta_0\) has units of ms. A slope of 1 ms per ng/mL means that the model's expected QTc change increases by 1 ms for a 1 ng/mL increase in concentration. It does not mean that every observation changes by exactly 1 ms for that increase.
Fitting the Linear Model
The ordinary least-squares (OLS) fit chooses an estimated intercept \(\hat{\beta}_0\) and slope \(\hat{\beta}_1\) to minimize the sum of squared residuals. A residual is the observed change minus the change predicted by the fitted line. For \(n\) observations, the fitted value and residual are \(\hat{Y}_i=\hat{\beta}_0+\hat{\beta}_1 C_i\) and \(e_i=Y_i-\hat{Y}_i\).
The OLS formulas use the sample means \(\bar{C}\) and \(\bar{Y}\), and the sums of squared deviations and cross-products:
The denominator must be positive: there must be variation in concentration to estimate a slope. These formulas yield the best-fitting straight line in the least-squares sense, not proof that a straight-line relationship is scientifically correct.
The assumptions do different jobs. A linear conditional mean makes the intercept and slope meaningful summaries of the modeled relationship. Mean-zero errors support unbiased estimation under the model. Independence and constant variance support the usual standard-error calculations. Normality is not needed to calculate the OLS line, but it matters for exact small-sample inference. For a basic illustrative fit, one can use one matched concentration-change pair per participant; repeated observations from the same participant need a model that accounts for their dependence.
Worked Example: Calculating the Intercept and Slope
Worked Example: Calculating the Intercept and Slope
Suppose a hypothetical dataset contains one matched observation from each of five participants. Concentration is measured in ng/mL and change from baseline QTc in ms.
| Concentration \(C_i\) (ng/mL) | Change \(Y_i\) (ms) |
|---|---|
| 0 | 1 |
| 2 | 3 |
| 4 | 4 |
| 6 | 8 |
| 8 | 9 |
The means are \(\bar{C}=4\) ng/mL and \(\bar{Y}=5\) ms. The concentration deviations are \(-4,-2,0,2,4\), and the response deviations are \(-4,-2,-1,3,4\). Thus,
Substituting these values gives
The fitted model is \(\hat{Y}=0.800+1.05C\), with \(C\) in ng/mL. For example, at 6 ng/mL the fitted mean change is \(0.800+(1.05)(6)=7.10\) ms. The observed change there is 8 ms, so its residual is \(8-7.10=0.900\) ms.
The fitted intercept is 0.800 ms. Because this dataset includes an observation at zero concentration, zero is within its observed concentration range; nonetheless, the intercept is the fitted conditional mean, not the observed value for every participant at zero. The slope describes the fitted average increase in change for each 1 ng/mL higher concentration over this range.
Uncertainty in the Slope
An estimated slope is subject to sampling variability. Under the assumptions stated above, the residual standard deviation is estimated by \(s_e=\sqrt{\sum e_i^2/(n-2)}\). The estimated standard error of the slope is \(SE(\hat{\beta}_1)=s_e/\sqrt{\sum(C_i-\bar{C})^2}\). The \(n-2\) degrees of freedom arise because two regression coefficients, the intercept and slope, were estimated.
This is a two-sided confidence interval for the population slope under the linear model and its error assumptions. In repeated samples under those conditions, intervals constructed by this procedure would cover the true slope at the stated confidence level at the nominal rate. A particular interval is not a probability statement that the fixed slope lies inside it.
Worked Example: Quantifying Slope Uncertainty
Use the five hypothetical observations above. The fitted values are 0.800, 2.90, 5.00, 7.10, and 9.20 ms. The residuals are 0.200, 0.100, \(-1.00\), 0.900, and \(-0.200\) ms, so
With \(n=5\), \(df=n-2=3\), and \(s_e=\sqrt{1.90/3}=0.796\) ms. Therefore,
For a 95% interval, \(t_{0.975,3}=3.182\). The resulting interval is
Under the model and its assumptions, this interval is compatible with slopes from approximately 0.649 to 1.45 ms per ng/mL. The interval's width also signals substantial uncertainty in this very small hypothetical dataset. The calculation does not establish that the model is linear, that the data are independent, or that the estimated relationship is causal.
Interpreting Zero and Clinically Relevant Concentrations
At \(C=0\), the fitted conditional mean is \(\hat{\beta}_0\). Its interpretation depends on what zero concentration represents in the study and on whether zero is supported by the data. It may be a useful reference prediction if observations near zero, such as an appropriate zero-exposure condition, are represented. If all measured concentrations are far above zero, the intercept is an extrapolation and may be imprecise or scientifically uninformative.
A zero drug concentration does not by itself imply zero change from baseline QTc. The response is a change measured under the study's conditions; variation unrelated to drug concentration may remain. Therefore, forcing \(\hat{\beta}_0=0\) is a substantive constraint, not a harmless way to simplify the equation. It should be used only when the scientific definition and design justify a zero expected response at zero concentration.
For a clinically relevant concentration \(C^*\), the model prediction is \(\hat{Y}(C^*)=\hat{\beta}_0+\hat{\beta}_1C^*\). Check that the concentration unit is consistent and that \(C^*\) lies within or near the range supporting the fit. The slope gives a change per unit; multiplying it by concentration gives a change in ms, and adding the intercept gives the predicted mean change.
Worked Example: Prediction and Unit Conversion
Using the hypothetical fitted equation \(\hat{Y}=0.800+1.05C\), predict change at a clinically relevant concentration of 5 ng/mL. This concentration is within the observed range of 0 to 8 ng/mL.
The prediction is a fitted mean change, not an individual-level guarantee. If concentration is instead expressed in mg/L, then 1 ng/mL equals 0.001 mg/L, so the numerical concentration at this point is 0.005 mg/L. The slope must change to preserve the same prediction: \(1.05/0.001=1{,}050\) ms per mg/L. Thus \(0.800+(1{,}050)(0.005)=6.05\) ms. The intercept remains in ms because changing the concentration unit does not change the response unit.
Common Pitfalls and Model Checks
A fitted line can conceal problems in the data or assumptions. Before interpreting the coefficients, examine the concentration range and the pattern of observations around the line. A plot of change against concentration can reveal whether a straight-line summary is plausible; residuals plotted against concentration can show curvature or changing spread. A small number of high-concentration observations can have substantial influence on the fitted slope, particularly when most participants have low concentrations.
Also ask whether the observations really are independent. If several ECGs from the same participant are treated as unrelated observations, the nominal standard errors can be misleading because those responses may be correlated. The simple model presented here does not handle that correlation. Nor does a straight line address possible time patterns, delayed responses, or uncertainty in concentration measurement; those features require additional modeling choices.
The main alternatives answer different questions. A constant-mean model sets the concentration slope to zero and represents no linear concentration association. A line constrained through zero fixes the intercept and encodes a stronger assumption about the response at zero concentration. More flexible exposure-response models can represent nonlinearity, but their use requires adequate data and additional assumptions. These alternatives should be chosen for scientific reasons, not merely because they produce a preferred coefficient or prediction.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A fitted model is \(\hat{Y}=1.2+0.80C\), where \(Y\) is in ms and \(C\) is in ng/mL. Interpret the intercept and slope.
- Using that model, what is the fitted mean change at 4 ng/mL?
- Why does a zero concentration not automatically justify forcing the intercept to zero?
- What is one concern with using the simple independent-observation model for several ECGs from each participant?
Solutions. (1) The fitted mean change at zero concentration is 1.2 ms; for each 1 ng/mL higher concentration, fitted mean change increases by 0.80 ms. (2) \(1.2+(0.80)(4)=4.4\) ms. (3) Change from baseline can be nonzero at zero concentration, and the data may not support the value at zero; forcing the intercept encodes an additional scientific constraint. (4) ECG observations from one participant may be correlated, violating independence and making ordinary standard errors unreliable.