Two Roles for Baseline
In a concentration–effect analysis, baseline can serve either as a value that anchors the response trajectory or as a quantity the model estimates. These choices may produce similar-looking equations, but they answer different questions. Fixing baseline asks how the response evolves from the recorded starting value. Estimating baseline asks which baseline level, together with the other model parameters, best explains the observations included in the fit.
Let \(Y_i(t)\) denote the response for individual \(i\) at time \(t\), and let \(Y_{i0}\) denote that individual’s observed response at the reference time \(t=0\). Let \(C_i(t)\) denote concentration, and let \(g(t,C_i(t);\theta)\) describe the modeled response contribution after the reference time, with parameters \(\theta\). For a baseline-anchored model, the contribution is defined to be zero at the reference condition: \(g(0,C_i(0);\theta)=0\). In a model that estimates baseline, write the baseline level as \(B_i\) or \(B\), as appropriate to the model.
For a fixed-baseline fit, the post-baseline observations are commonly modeled conditionally on the recorded baseline; the baseline value itself is not a residual to be fitted. For a baseline-parameter fit, the baseline parameter is estimated jointly with the other parameters, and observations included in the likelihood contribute residuals. A fitting plan must therefore say which observations are modeled. Merely writing an intercept in an equation does not settle whether it is fixed or estimated.
What Changes in the Fitting Problem?
The distinction is especially clear with a linear trajectory. Suppose the modeled change is \(\beta t\), where \(\beta\) has units of response per unit time. With a fixed observed baseline \(Y_0\), the model is \(Y(t)=Y_0+\beta t+\epsilon(t)\), and the baseline is held at its recorded value. If the baseline is estimated, the model is \(Y(t)=B+\beta t+\epsilon(t)\), and both \(B\) and \(\beta\) are fitted.
Assume, for this derivation, independent residuals with mean zero and common variance \(\sigma^2\), and an adequately specified linear trajectory. For the fixed-baseline model, least squares minimizes the sum of squared post-baseline residuals. Differentiating that sum with respect to \(\beta\) gives the fitted slope:
Rearranging yields \(\hat{\beta}=\sum_j t_j[Y(t_j)-Y_0]/\sum_j t_j^2\), provided \(\sum_j t_j^2>0\). This estimate is conditional on the fixed baseline. If baseline is instead estimated, least squares fits \(B\) and \(\beta\) together to all included observations. With observations at times \(t_j\), the fitted slope is \(\hat{\beta}=\sum_j(t_j-\bar{t})(Y_j-\bar{Y})/\sum_j(t_j-\bar{t})^2\), and the fitted intercept is \(\hat{B}=\bar{Y}-\hat{\beta}\bar{t}\), provided the observation times are not all identical. Here \(\bar{t}\) and \(\bar{Y}\) are the sample means of the times and responses included in the fit.
These formulas show why the estimates need not agree. The fixed-baseline slope uses the recorded starting value as a fixed anchor. The jointly estimated slope is calculated around the means of the fitted observations, while the intercept can move to reduce their overall squared residuals. Which fit is appropriate depends on the scientific question and on how the baseline is represented in the model, not on which formula gives a preferred numerical result.
Worked Example: Fitting a Slope With the Baseline Fixed
Suppose, hypothetically, that response is measured in points and the recorded baseline is \(Y_0=100\) points. Post-baseline responses are 93 points at \(t=1\) day and 89 points at \(t=2\) days. Fit the linear slope with baseline fixed, using equal-weight least squares.
The fixed-baseline slope formula is \(\hat{\beta}=\sum_j t_j(Y_j-Y_0)/\sum_j t_j^2\). Substituting the two observations gives
The fitted trajectory is \(Y(t)=100-5.8t\) points. It predicts 94.2 points at day 1 and 88.4 points at day 2. The residuals, defined as observed minus predicted, are \(-1.2\) points and \(0.6\) points. The post-baseline sum of squared residuals is \((-1.2)^2+(0.6)^2=1.80\) points squared. The baseline is fixed by construction; it is not an additional fitted residual in this conditional fit.
Worked Example: Estimating the Baseline and Slope Together
Use the same hypothetical responses, now treating the baseline as an intercept to estimate from all three observations: 100 points at day 0, 93 points at day 1, and 89 points at day 2. Fit \(Y(t)=B+\beta t\) by ordinary least squares with equal weights.
The means are \(\bar{t}=(0+1+2)/3=1\) day and \(\bar{Y}=(100+93+89)/3=94\) points. The slope is
The intercept is \(\hat{B}=\bar{Y}-\hat{\beta}\bar{t}=94-(-5.5)(1)=99.5\) points. Thus the fitted trajectory is \(Y(t)=99.5-5.5t\) points, with predictions 99.5, 94.0, and 88.5 points at days 0, 1, and 2. The residuals are 0.5, \(-1.0\), and 0.5 points, so the sum of squared residuals is \(0.25+1+0.25=1.50\) points squared.
The fitted intercept differs from the recorded baseline, and the slope differs from the fixed-baseline estimate. This is not an arithmetic inconsistency: the models estimate different quantities and use different fitting rules. In particular, the conditional fit’s post-baseline sum of squares should not be compared directly with the all-observation sum of squares as if they came from the same likelihood.
Uncertainty and Interpretation
A fixed baseline is treated as given for the purpose of the conditional fit. Consequently, the fit does not provide a standard error for that fixed value as a model parameter. This does not establish that the recorded baseline is known without uncertainty in the real world. It means only that uncertainty in the baseline was not estimated or propagated by that particular fit. By contrast, a baseline-parameter model can estimate uncertainty in \(B\), as well as its covariance with \(\beta\) and other fitted parameters.
For example, under the independent, equal-variance linear model above, the conditional slope variance is \(\sigma^2/\sum_j t_j^2\). For a jointly estimated intercept and slope, the slope variance is \(\sigma^2/\sum_j(t_j-\bar{t})^2\), and the intercept variance is \(\sigma^2[1/n+\bar{t}^{\,2}/\sum_j(t_j-\bar{t})^2]\), where \(n\) is the number of observations included in that fit. These expressions rely on the stated linear model and residual assumptions. They also show that uncertainty depends on the design and fitting specification; one should not assume that either baseline treatment always gives smaller uncertainty.
Predictions inherit this distinction. A prediction from a fixed-baseline model is conditional on the specified \(Y_0\). A prediction from a baseline-parameter model reflects uncertainty in the estimated baseline and its dependence on the other parameters. Thus, a reported trajectory or concentration-effect prediction should state whether its baseline was fixed or estimated. The baseline parameter \(B\) is interpreted as the model’s reference response under its defined reference condition; it should not be described as an independently observed biological constant unless the model and data support that interpretation.
Worked Example: Comparing Baseline Uncertainty Under a Linear Model
Consider a hypothetical study with responses at days 0, 1, and 2, fitted either with a free intercept and slope or, conditionally, with the day-0 value fixed. Suppose the residual standard deviation is known to be \(\sigma=2\) response points. Compare the model-based standard error of the estimated baseline in the free-intercept fit with the uncertainty assigned to the fixed baseline by the conditional fit.
For the free-intercept fit, \(n=3\), \(\bar{t}=1\) day, and \(\sum_j(t_j-\bar{t})^2=2\) days squared. Therefore,
The standard error is \(\sqrt{10/3}=1.826\) points. In the fixed-baseline fit, the model-based standard error for a baseline parameter is not calculated because no such parameter is estimated. It would be incorrect to conclude that the baseline has zero real-world uncertainty; the conditional fit simply does not quantify it. The comparison is about what uncertainty the two fitting specifications represent, not about whether one value is inherently more certain.
Choosing and Checking a Baseline Treatment
A fixed initial condition is useful when the scientific target is change from a specified recorded starting response and conditioning on that value is appropriate. It can keep the trajectory anchored to the individual’s starting observation and avoid spending parameters on a baseline level that is not the focus of the fit. An estimated baseline is useful when the reference response itself is part of the model’s target, when observations are intended to inform that level, or when the recorded starting value should not be imposed as the trajectory’s exact reference.
Neither choice repairs a misspecified response function. Before interpreting a fit, check that the reference condition is defined clearly, that the modeled contribution is zero there when required, and that the observations included in the objective function match the stated fitting strategy. Examine residuals against time and concentration, and check whether the baseline choice produces systematic residual patterns or implausible parameter estimates. A slope or concentration-effect parameter can shift when the baseline is changed, so its interpretation must be tied to the baseline specification.
| Feature | Baseline fixed as initial condition | Baseline estimated as a parameter |
|---|---|---|
| Reference response | Set to the recorded baseline value | Estimated from observations included in the fit |
| Baseline contribution to fitting | Often conditioned on; not a fitted residual | Included through residuals for observations in the likelihood |
| Baseline uncertainty | Not estimated as parameter uncertainty | Estimated, with covariance with other parameters |
| Interpretation | Trajectory conditional on the chosen starting value | Trajectory with an estimated reference level |
A common mistake is to call the fixed baseline “known exactly” and then interpret conditional prediction intervals as if they included uncertainty in that baseline. The correction is to say that the baseline was held fixed for the analysis, and that uncertainty from estimating it is absent from that model’s parameter uncertainty. Another mistake is to compare fit statistics calculated from different sets of observations or different likelihood definitions. Such comparisons are not justified simply because both fits use the same response data somewhere in their calculations.
The central limitation is that these formulations address how baseline enters the model, not whether the recorded baseline is a perfect representation of the response at the reference condition. That separate issue requires modeling the observation process. In practice, document whether baseline is fixed or estimated, identify which observations enter the fit, and state whether predictions are conditional on the recorded baseline.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In \(Y(t)=Y_0+\beta t\), with \(Y_0\) held at the observed response, is \(Y_0\) an estimated parameter? What does the fitted trajectory condition on?
- For observations at times 1 and 2 days with responses 94 and 88 points, respectively, and fixed baseline \(Y_0=100\) points, calculate the least-squares slope.
- For observations at days 0, 1, and 2, suppose a free-intercept fit gives \(\hat{B}=98\) points. What does \(\hat{B}\) represent, and how does it differ from fixing the recorded day-0 response?
- Why is it incorrect to report that a fixed-baseline fit estimates the baseline with zero standard error?
- Name one diagnostic check that is useful under either baseline specification.
Solutions. (1) No. \(Y_0\) is fixed, and the trajectory is conditional on the recorded baseline value. (2) \(\hat{\beta}=[1(94-100)+2(88-100)]/(1^2+2^2)=(-6-24)/5=-6.0\) points per day. (3) \(\hat{B}\) is the fitted reference-level parameter under the model; fixing baseline instead imposes the recorded day-0 response as the reference and does not estimate it. (4) The fixed-baseline fit does not estimate a baseline parameter, so it supplies no parameter standard error for that value; this does not mean the recorded baseline has no real-world uncertainty. (5) Check residuals against time or concentration for systematic patterns, while confirming that the reference condition and included observations match the specified model.