Why Separate Baseline and Placebo Errors?
A PD model can predict poorly even when its drug-effect component is adequate. The predicted response may start from the wrong baseline level, fail to represent change over time that would occur without active drug, or represent that placebo-period change with the wrong shape. If those errors are attributed immediately to the drug effect, estimates of its magnitude or time course can be distorted. Placebo-period observations are especially useful because they provide information about the response when the modeled active-drug contribution is absent.
The central diagnostic question is therefore comparative: does the prediction error appear at baseline, develop over time in placebo recipients, or occur specifically under active treatment? Residual plots can suggest which model component needs attention, but they do not by themselves establish a causal explanation. Treatment allocation, sampling times, covariates, and the way baseline enters the model all affect what a pattern can support.
Model Components and Residual Definitions
Let \(Y_{ij}\) denote the PD response for individual \(i\) at time \(t_{ij}\), and let \(\mu_{ij}\) be the model-predicted conditional mean for that observation. A useful schematic representation is
Here \(B_i\) is the modeled baseline level, \(P(t;\boldsymbol{\gamma})\) is the modeled change that can occur over time in the absence of active drug, and \(D_{ij}(\boldsymbol{\theta})\) is the modeled drug-related contribution. This is a diagnostic decomposition, not a claim that every PD model must be additive. In a placebo period with no active drug exposure, the drug contribution is set to zero under the model. A placebo response may be represented by \(P\), but “placebo-period” describes the treatment condition; it does not guarantee that the observed response is constant or that every change is caused by expectation.
Baseline-relative residuals ask whether the model predicts the change from baseline correctly, rather than whether it predicts the absolute response correctly. They are not interchangeable with raw residuals. In particular, subtracting an observed baseline carries its measurement error into every later contrast for that individual.
What the Main Patterns Suggest
| Plot or comparison | Pattern to look for | Diagnostic implication, not a verdict |
|---|---|---|
| Raw residuals at baseline and across placebo follow-up | Similar signed residuals at baseline and later times | A baseline-level discrepancy is plausible; the absolute prediction is shifted. |
| Baseline-relative residuals during placebo follow-up | Residuals move systematically with time | The modeled placebo-period trajectory or another time-related component may be inadequate. |
| Placebo and active-treatment residuals over comparable times | A similar time pattern occurs in both arms | A shared baseline or time-varying component is more plausible than a drug-specific explanation alone. |
| Active-treatment residuals after accounting for placebo patterns | Errors remain treatment-specific and relate consistently to exposure or predicted effect | A drug-related component may be misspecified, subject to design and covariate checks. |
A residual pattern should be evaluated against the model’s prediction and uncertainty, not merely against zero by eye. If residual variance changes with response, time, or another predictor, raw residual spread can be misleading; use the appropriate variance-standardized residuals. For a conditional residual, the prediction is conditional on the individual-level information included in the fit. A population-level prediction answers a different question and may show between-individual structure that is not the same as an individual’s baseline error.
Why Baseline-Relative Residuals Need Care
The variance of a baseline-relative residual follows directly from its definition. Write the raw residual at time \(t\) as \(e_{it}\) and at baseline as \(e_{i0}\). Then \(q_{it}=e_{it}-e_{i0}\), so
This identity follows by expanding the variance of a difference. If the two errors are independent, the covariance is zero and the variance is the sum of their variances. If they are correlated, omitting the covariance gives the wrong standardization. Further, two baseline-relative residuals from the same individual share the term \(-e_{i0}\), so they are generally correlated even when the follow-up measurement errors are independent.
There is a second practical issue. Some analyses condition on the observed baseline or use it to define the individual’s predicted baseline. The baseline residual may then be zero by construction, or may not be an ordinary model residual. In that setting, a flat baseline-relative plot cannot demonstrate that the model correctly predicts the absolute baseline level: that level was anchored to the observation. Inspect how baseline was handled before interpreting the plot.
A Structured Diagnostic Workflow
Confirm the response scale, time origin, baseline definition, treatment condition, and whether predictions are conditional or population-level. Check whether baseline observations were modeled or conditioned upon.
When baseline is an actual modeled observation, examine residual sign and magnitude at baseline and compare them with later residuals. A common offset is different from an error that develops over time.
Plot standardized residuals and baseline-relative residuals against time, retaining individual trajectories as well as summaries. Look for persistent direction, curvature, or changes shared across individuals.
At comparable times and covariate distributions, ask whether a residual pattern is also present in placebo recipients. Only patterns that remain specifically associated with active treatment support focusing next on drug-related prediction errors.
Use the correct residual variance, consider within-individual dependence, and assess whether conclusions change when influential times or individuals are displayed separately. Treat patterns as diagnostic evidence, not automatic model-selection rules.
Worked Examples
Worked Example: A Constant Baseline-Level Offset
Suppose a hypothetical placebo recipient has a baseline response of 102 units when the model predicts 100 units. At two later placebo visits, observed and predicted responses are 103 and 101 units, then 102 and 100 units. Assume the model-implied residual standard deviation is 2 units at all three occasions, and that these are modeled observations rather than a baseline-conditioned fit.
At baseline, the raw residual and standardized residual are
At the first follow-up, \(e_{i1}=103-101=2\) units and \(r_{i1}=2/2=1\). At the second follow-up, \(e_{i2}=102-100=2\) units and \(r_{i2}=2/2=1\). The residual is consistently positive rather than increasing with time.
The observed baseline-relative change at visit 1 is \(103-102=1\) unit, and the predicted change is \(101-100=1\) unit. Thus \(q_{i1}=1-1=0\) units. At visit 2, the observed change is \(102-102=0\) units and the predicted change is \(100-100=0\) units, giving \(q_{i2}=0-0=0\) units. In this small hypothetical example, the change predictions agree while the absolute predictions are 2 units too low. That pattern is consistent with a baseline-level discrepancy, not evidence of a changing placebo response. One individual cannot establish a population-wide defect.
Worked Example: A Placebo Trajectory That Drifts More Than Predicted
In a hypothetical placebo period, an individual has baseline observed and predicted responses of 50 units. At week 1, the observed response is 54 units and the model predicts 52 units. At week 2, the observed response is 58 units and the model predicts 54 units. Assume a residual standard deviation of 2 units at each occasion and a baseline residual of zero.
The raw follow-up residuals are \(54-52=2\) units and \(58-54=4\) units. Their standardized values are \(2/2=1\) and \(4/2=2\). The baseline-relative residuals give the same values here because the baseline residual is zero:
The model predicts an increase of 2 units per week, while these observations increase by 4 units per week. The residual grows by 2 units from week 1 to week 2, a descriptive residual trend of \(2\text{ units}/1\text{ week}=2\) units per week. This is consistent with an underpredicted placebo-period time trend. It does not show that the placebo condition caused the increase, and a single trajectory cannot distinguish a systematic trend from individual variation or chance.
Worked Example: Comparing Placebo and Active-Treatment Prediction Errors
Consider hypothetical group means at one matched follow-up time. Both groups have a baseline mean of 40 units. The placebo group changes by an observed mean of \(+3\) units, while the model predicts \(+2\) units. The active-treatment group changes by an observed mean of \(-5\) units, while the model predicts \(-4\) units. Suppose each group has 4 individuals, the residual standard deviation for individual changes is 2 units, and the groups are independent.
The placebo mean change residual is \(3-2=+1\) unit. The active-treatment mean change residual is \(-5-(-4)=-1\) unit. The treatment-minus-placebo discrepancy in predicted versus observed change is therefore \(-1-(+1)=-2\) units. Equivalently, the observed treatment-minus-placebo contrast is \(-5-(+3)=-8\) units, while the model predicts \(-4-(+2)=-6\) units; the contrast residual is \(-8-(-6)=-2\) units.
For a rough scale comparison under the stated independent, equal-variance assumptions, the standard error of the difference in mean residuals is
The discrepancy is \(-2/1.414=-1.414\) standard errors from zero. This calculation is descriptive here, not a formal hypothesis test or a claim that the drug effect is wrong. The placebo group itself has a positive residual, so shared model error remains plausible. To argue for a drug-specific error, one would look for a repeatable treatment-specific pattern across appropriate times or exposure levels, after checking that the groups and predictions are comparable.
Comparisons, Diagnostics, and Limits
Raw residual plots, baseline-relative plots, and placebo-versus-treatment contrasts answer related but distinct questions. A raw residual plot retains information about absolute calibration, including a baseline offset. A baseline-relative plot emphasizes prediction of change and removes a constant residual offset algebraically, but it also incorporates baseline measurement error. A treatment-placebo contrast can reduce shared time-related discrepancies when groups are comparable, but it may hide a defect common to both groups and can be confounded by differences in sampling, covariates, or baseline distributions. Observed-predicted plots provide an overall calibration view but may conceal when in the course of response the error arises.
Useful diagnostics include plotting individual trajectories rather than only pooled averages, marking baseline and placebo intervals, and showing residuals on a scale that accounts for the modeled variance. Check whether the same individuals contribute repeatedly to a pattern, whether sampling times are comparable across groups, and whether the apparent trend is driven by a few observations. A time-ordered pattern may also reflect serial dependence; residuals should not be treated as independent merely because they are plotted at separate visits.
Several limits are important. Placebo data may be sparse, missing, or collected at times that do not match the active-treatment schedule. A treatment-specific residual pattern may arise from an unbalanced covariate distribution or an incorrect observation model rather than the drug component itself. Conversely, a placebo-period pattern does not rule out simultaneous drug-related misspecification. Residual plots identify where predictions fail; they do not uniquely identify which scientific mechanism is responsible.
The practical value of this separation is that it prevents one model component from compensating for another. A drug-effect parameter should not be used to absorb a baseline offset or a time pattern that is already present without active drug. The most defensible interpretation comes from agreement among raw residuals, baseline-relative residuals, and placebo-treatment comparisons, with the modeling assumptions made explicit.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A modeled baseline observation has residual \(+3\) units, and placebo follow-up residuals remain near \(+3\) units. What pattern is suggested? What additional plot is informative?
- At baseline, observed and predicted values are both 20 units. At follow-up, observed change is \(+6\) units and predicted change is \(+4\) units. Calculate the baseline-relative residual.
- Why can baseline-relative residuals from the same individual be correlated even if follow-up measurement errors are independent?
- A residual trend appears in both placebo and active-treatment groups at comparable times. Why is it premature to assign the trend to the drug effect?
- If a model conditions on the observed baseline so the baseline residual is zero by construction, can that residual establish correct absolute baseline calibration? Explain.
Solutions. (1) A roughly constant positive offset is consistent with a baseline-level discrepancy; inspect baseline-relative residuals to see whether prediction of change is better calibrated. (2) \(q=(Y_t-Y_0)-(\mu_t-\mu_0)=6-4=+2\) units. (3) Each contrast contains the same subtracted baseline error, \(-e_{i0}\), which induces shared variation across follow-up contrasts. (4) The shared pattern may reflect a baseline or time-varying component common to both conditions; compare placebo structure before attributing it to active drug. (5) No. Conditioning makes the baseline fit untestable from that residual; assess absolute calibration using data or predictions that do not impose the observed baseline as the anchor.