Why Oncology Dose Optimization Needs More Than a Dose-Response Plot
In oncology, increasing dose may increase drug exposure and the probability of tumor response, but it can also increase dose-limiting toxicity (DLT). Patients with more severe baseline disease may have different response and toxicity risks even at the same exposure. A dose comparison therefore needs to distinguish the exposure-response relationship from differences in the patients who received each dose.
This tutorial uses hypothetical trial data and a deliberately compact model to show the analysis. The endpoints are whether a patient has a confirmed tumor response by a prespecified assessment window and whether the patient experiences a DLT during a prespecified safety window. These are binary outcomes; the same principles apply only after an endpoint and its observation window have been clearly defined. This analysis does not replace a full longitudinal tumor-size model or a time-to-event analysis when those outcomes are the scientific target.
Specify the Models Before Comparing Doses
Suppose an illustrative trial evaluates three tested doses: 50, 100, and 150 mg. For a simple demonstration, their representative exposures are \(E=0.25\), \(0.50\), and \(0.75\), respectively. These values are hypothetical; in a real analysis, individual exposure measurements would be used. Assume severity is recorded at baseline and is not changed by treatment. A fitted logistic model for confirmed response is
The corresponding DLT model is
For either model, \(\operatorname{logit}(p)=\log\{p/(1-p)\}\), and converting a linear predictor \(\eta\) to a probability gives \(p=1/\{1+\exp(-\eta)\}\). The response model implies that, at a fixed severity, a one-unit increase in exposure multiplies the odds of response by \(\exp(1.20)\), about 3.32. It does not multiply the probability by 3.32. The DLT model's corresponding odds ratio is \(\exp(0.40)\), about 1.49. The severity coefficients represent adjusted differences in log odds at the same exposure.
These models assume that the log odds change linearly with exposure over the range studied, severity has the stated additive effect, and there is no exposure-by-severity interaction. They also assume the selected link and predictors adequately describe the conditional probabilities. In a trial analysis, the endpoints must be ascertained comparably across dose groups; model fit, missing outcomes, and the way exposure is measured must be examined. Exposure is not randomized even when dose is randomized, so interpreting its coefficient as a causal effect requires additional assumptions, including adequate control of exposure-outcome confounding.
Worked Example: Predicting Response and DLT at One Exposure
At \(E=0.50\), calculate the response and DLT probabilities separately for lower- and higher-severity patients. Then standardize each outcome to a population that is 50% lower severity and 50% higher severity.
Response, lower severity. The linear predictor is \(-1.50+1.20(0.50)-0.80(0)=-0.90\). Therefore,
Response, higher severity. The predictor is \(-1.50+1.20(0.50)-0.80(1)=-1.70\), so \(p_{\mathrm{response},S=1}=1/\{1+\exp(1.70)\}=0.1545\). With equal severity weights, the standardized response probability is \(0.50(0.2891)+0.50(0.1545)=0.2218\), or 22.18%.
DLT, lower severity. The predictor is \(-2.15+0.40(0.50)+0.20(0)=-1.95\). Thus \(p_{\mathrm{DLT},S=0}=1/\{1+\exp(1.95)\}=0.1246\).
DLT, higher severity. The predictor is \(-2.15+0.40(0.50)+0.20(1)=-1.75\), giving \(p_{\mathrm{DLT},S=1}=1/\{1+\exp(1.75)\}=0.1480\). The standardized DLT probability is \(0.50(0.1246)+0.50(0.1480)=0.1363\), or 13.63%. The response and DLT probabilities are separate outcome predictions; they do not by themselves specify how often both outcomes occur in the same patient.
Standardize to Make a Fair Dose Comparison
A dose group with more patients who have severe disease could have a lower response rate even if its exposure-response relationship were identical. Conversely, a dose group with fewer high-risk patients could appear safer for reasons unrelated to dose. Standardization addresses this compositional difference by predicting each dose's outcomes in the same target distribution of baseline severity.
For a target population with severity weights \(w_s\), where \(w_s\) is the proportion with severity \(s\) and \(\sum_s w_s=1\), the standardized probability at exposure \(E_d\) is
This follows by partitioning the target population into severity groups: the overall probability is the group-specific probability multiplied by the group's population share, summed over groups. The same weights must be used for each dose if the aim is to compare doses in one common population. Equal weights are only an illustrative choice; a trial population or clinically relevant target population may have a different severity mix.
Worked Example: Comparing the Three Tested Doses
Use the models above, the representative exposures \(0.25\), \(0.50\), and \(0.75\), and equal severity weights. For each dose, calculate the two severity-specific probabilities and average them. For example, at 50 mg, the response predictors are \(-1.20\) for \(S=0\) and \(-2.00\) for \(S=1\), giving probabilities 0.2315 and 0.1192. Their standardized average is 0.1753.
| Tested dose | Representative exposure | Standardized response probability | Standardized DLT probability | Response minus DLT |
|---|---|---|---|---|
| 50 mg | 0.25 | 0.1753 (17.53%) | 0.1250 (12.50%) | 0.0504 |
| 100 mg | 0.50 | 0.2218 (22.18%) | 0.1363 (13.63%) | 0.0855 |
| 150 mg | 0.75 | 0.2761 (27.61%) | 0.1485 (14.85%) | 0.1276 |
The 150 mg prediction has the greatest response probability and the greatest DLT probability. From 100 to 150 mg, the standardized response probability increases by \(0.2761-0.2218=0.0543\), or 5.43 percentage points; the DLT probability increases by \(0.1485-0.1363=0.0122\), or 1.22 percentage points. If, solely for illustration, the decision rule were to choose the dose with the largest response probability subject to a standardized DLT probability below 15%, 150 mg would meet the point-estimate criterion, at 14.85%. It is close to the boundary, however, and an uncertainty interval could change the decision.
The final column is a simple illustrative score, \(U=P(\text{response})-P(\text{DLT})\). It ranks these predictions in the same order, but it assumes equal numerical value for a response and a DLT event and is not a universal clinical utility. A decision should use clinically justified consequences, thresholds, and uncertainty rather than treating this score as self-evident.
Use Individual Exposures When Exposure Varies
A representative or mean exposure is not generally enough to predict a population probability. Because the logistic transformation is nonlinear, the average of predicted probabilities need not equal the probability calculated at the average exposure. This matters when exposure varies substantially within a dose group or is associated with baseline severity.
Worked Example: Why Averaging Individual Predictions Matters
Consider four hypothetical patients receiving 100 mg, with exposure and baseline severity pairs \((0.20,0)\), \((0.40,0)\), \((0.60,1)\), and \((0.80,1)\). Their mean exposure is 0.50 and half have each severity. Calculate the mean of their individual model predictions and compare it with prediction at mean exposure.
For response, the four linear predictors are \(-1.26\), \(-1.02\), \(-1.58\), and \(-1.34\). The corresponding probabilities are approximately \(0.2210\), \(0.2650\), \(0.1708\), and \(0.2075\). Their mean is \((0.2210+0.2650+0.1708+0.2075)/4=0.2161\), or 21.61%. Predicting at mean exposure with equal severity weights instead gives 22.18%, as calculated earlier.
For DLT, the four predictors are \(-2.07\), \(-1.99\), \(-1.71\), and \(-1.63\); the probabilities are approximately \(0.1121\), \(0.1203\), \(0.1532\), and \(0.1638\). Their mean is about 0.1373, or 13.73%, rather than the 13.63% obtained by predicting at the mean exposure with equal severity weights. The difference here is modest, but the principle is general: calculate predictions for the actual covariate-exposure records and then average over the target population. The four-person calculation is illustrative, not a reliable estimate of a trial-wide rate.
Diagnostics, Alternatives, and Limits
Model-based dose comparisons are only as credible as their data and assumptions. Examine exposure distributions by dose and severity before interpreting an adjusted curve. If the exposure ranges barely overlap, the model may be extrapolating when it estimates what a higher dose would do to patients whose exposures were observed only at lower doses. Check the number of responses and DLTs in each exposure and severity region; sparse events can yield unstable estimates. Compare observed and predicted outcome rates across clinically meaningful exposure groups, inspect calibration, and assess whether exposure-response patterns appear curved or differ by severity. Check missing response assessments and safety follow-up, since unequal observation can bias simple binary rates.
A linear exposure term on the log-odds scale is a modeling assumption, not a biological law. A plot suggesting curvature may motivate a prespecified nonlinear term or a flexible curve, but flexibility must be supported by enough data and should not be used to chase noise. An exposure-by-severity interaction may be relevant if the exposure association differs by severity; it should be assessed with adequate data and interpreted through predicted probabilities, not only a coefficient. Investigators should also examine influential observations and whether exposure was measured consistently.
Several alternative comparisons answer different questions. A randomized comparison by assigned dose estimates the effect of assignment to those dose regimens, subject to trial conduct and endpoint assumptions; it does not require treating measured exposure as randomized. A categorical-dose model avoids assuming a smooth exposure trend but may be less precise and cannot directly characterize individual exposure variation. A continuous exposure model can describe exposure-response more efficiently, but can be confounded because clearance, disease status, dose interruptions, and other patient characteristics may affect both exposure and outcome. Standardization adjusts for measured baseline factors included in the model; it does not remove bias from unmeasured confounding or repair lack of exposure overlap.
Separate response and DLT models provide two useful margins of a benefit-risk comparison, but do not describe their joint occurrence or the timing of events. If a decision depends on competing event times, treatment discontinuation, or simultaneous efficacy and toxicity outcomes, an analysis suited to those features is needed. Parameter and sampling uncertainty also matter: intervals for standardized probabilities and dose contrasts can be obtained by resampling patients or by drawing model parameters from their estimated joint distribution and recalculating the standardization. A point estimate just below a safety threshold is not assurance that the true risk is below it.
A Practical Analysis Sequence
Prespecify response and DLT windows, the exposure metric, baseline severity coding, and the patient population whose probabilities will guide the decision.
Summarize dose, exposure, severity, endpoint ascertainment, and missingness. Identify sparse regions before fitting an exposure-response model.
Model response and DLT separately when appropriate; check calibration, functional form, interactions, and sensitivity to influential records.
Predict each endpoint at each tested dose for the same target patients or covariate distribution, average predictions, and calculate uncertainty for the resulting contrasts.
Compare efficacy, safety, and uncertainty using clinically justified thresholds or utilities, and state the limitations of the evidence.
Common Mistake
A frequent error is to compare crude response or DLT percentages across dose groups and attribute every difference to dose. If the higher-dose group contains more patients with severe disease, a crude response comparison mixes dose-related differences with baseline case mix. The correction is to define a common target population, predict each group's outcome under the model at each tested dose, and average using the same severity weights. This adjustment improves comparability only for measured, adequately modeled factors; it does not turn an observational exposure association into a randomized causal effect.
Check Your Understanding
Use the models and calculations in this tutorial. The solutions follow the questions.
- At \(E=0.50\), what are the modeled response and DLT probabilities for a higher-severity patient?
- If the target population is 30% lower severity and 70% higher severity, what are the standardized response and DLT probabilities at \(E=0.50\)?
- Why can prediction at mean exposure differ from the mean of patient-level predictions?
- What does an exposure odds ratio describe, and why does it not by itself establish a causal effect of increasing dose?
Solutions. (1) At \(E=0.50,S=1\), the response predictor is \(-1.70\), so response probability is \(1/\{1+\exp(1.70)\}=0.1545\). The DLT predictor is \(-1.75\), giving \(1/\{1+\exp(1.75)\}=0.1480\). (2) The response probability is \(0.30(0.2891)+0.70(0.1545)=0.1949\), or about 19.49%. The DLT probability is \(0.30(0.1246)+0.70(0.1480)=0.1410\), or about 14.10%. (3) The logistic transformation is nonlinear, so averaging probabilities over the observed exposure distribution generally differs from transforming the average exposure; exposure may also be associated with severity. (4) It describes the multiplicative change in conditional odds associated with a one-unit exposure increase under the specified model. Measured exposure is not randomized, so confounding and model assumptions can prevent that association from representing the causal effect of assigning a higher dose.