BLRM Parameterization With Two-Parameter Logistic Models

Why Parameterization Matters

A Bayesian logistic regression model for dose-limiting toxicity (DLT) represents the probability of toxicity as a function of dose. Its parameterization determines what the model’s coefficients mean, how priors can be specified, and how a parameter value is translated into a predicted risk at a candidate dose. In a two-parameter BLRM, one parameter describes risk at a chosen reference dose and the other describes how log-odds change with log-dose.

The reference dose is a modeling choice, not a claim that it is the correct dose for treatment. Choosing it sensibly can make the intercept easier to interpret and the parameters easier to estimate. The model also makes substantive assumptions: the dose-response relationship is linear on the log-odds scale, and a positive dose slope implies increasing toxicity risk across positive doses.

Model, Data, and Assumptions

Let \(d>0\) denote dose, \(d_{\mathrm{ref}}>0\) a fixed reference dose, and \(Y_i=1\) if participant \(i\) experiences a DLT during the specified assessment window and \(Y_i=0\) otherwise. Define the centered log-dose covariate \(x(d)=\log(d/d_{\mathrm{ref}})\). Let \(\alpha\) be the intercept and \(\beta\) the dose slope. The model is

Y_i\mid d_i,\alpha,\beta \sim \operatorname{Bernoulli}(p(d_i)), \qquad \operatorname{logit}\{p(d)\}=\alpha+\beta\log(d/d_{\mathrm{ref}}), \qquad p(d)=\frac{\exp\{\alpha+\beta\log(d/d_{\mathrm{ref}})\}} {1+\exp\{\alpha+\beta\log(d/d_{\mathrm{ref}})\}}.

Definition: In this two-parameter logistic model, \(\alpha\) is the log-odds of DLT at the reference dose, because \(x(d_{\mathrm{ref}})=0\). The parameter \(\beta\) is the change in log-odds associated with a one-unit increase in log dose. Equivalently, multiplying dose by \(c\) multiplies the odds by \(c^\beta\), for any \(c>0\).

The model is Bayesian when a joint prior distribution is assigned to \((\alpha,\beta)\), and inference about these parameters is based on that prior and the likelihood. For example, one could specify a bivariate normal prior, allowing prior dependence between the intercept and slope. The prior, including its scale and any restrictions, is part of the model specification. If increasing dose is assumed not to decrease toxicity risk, the parameter space can be restricted to \(\beta>0\); a prior that permits negative values does not enforce that monotonicity.

For \(n_j\) participants treated at dose \(d_j\), let \(y_j\) be the number who experience DLT. Under conditionally independent outcomes with a common risk at each dose, the aggregated likelihood is

L(\alpha,\beta) = \prod_j {n_j\choose y_j} p(d_j)^{y_j}\{1-p(d_j)\}^{n_j-y_j}.

This likelihood assumes a consistently defined DLT outcome and assessment window, complete ascertainment for the analyzed participants, and independent outcomes conditional on dose and the model parameters. It also assumes that participants at the same dose share the same modeled DLT probability. The logistic form further assumes that the log-odds are linear in log dose over the modeled dose range. It does not account by itself for DLT timing or patient-specific covariates.

How the Parameters Translate Into Dose-Specific Risk

The inverse-logit function converts a linear predictor into a probability. At \(d_{\mathrm{ref}}\), the predictor is \(\alpha\), so \(p(d_{\mathrm{ref}})=\exp(\alpha)/\{1+\exp(\alpha)\}\). At any other positive dose, the dose ratio—not the numerical difference in dose—determines the covariate. If dose is multiplied by \(c\), then \(x\) increases by \(\log(c)\), and the log-odds change by \(\beta\log(c)\). Consequently, the odds are multiplied by \(\exp\{\beta\log(c)\}=c^\beta\).

The following derivation also gives the dose corresponding to a specified risk \(q\), where \(0<q<1\). Taking the logit of that risk and equating it to the model’s linear predictor gives

\operatorname{logit}(q)=\alpha+\beta\log(d/d_{\mathrm{ref}}), \qquad \log(d/d_{\mathrm{ref}})=\frac{\operatorname{logit}(q)-\alpha}{\beta}, \qquad d=d_{\mathrm{ref}}\exp\left\{\frac{\operatorname{logit}(q)-\alpha}{\beta}\right\}, \quad \beta\ne 0.

This algebra supplies a model-implied dose, not a recommendation to administer it. Whether a dose is appropriate for a trial depends on design rules and clinical considerations beyond the parameter-to-risk translation.

Worked Example: Calculating Risk at a Dose Above the Reference

Suppose, hypothetically, that \(d_{\mathrm{ref}}=40\) mg, \(\alpha=-1.60\), and \(\beta=0.90\). Calculate the model-implied DLT risk at 80 mg.

The dose ratio is \(80/40=2\), so the centered log-dose is \(x(80)=\log(2)\approx0.6931\). The linear predictor is

\eta(80)=-1.60+0.90(0.6931)=-0.9762.

Therefore,

p(80)=\frac{1}{1+\exp\{0.9762\}}\approx0.2737.

The model assigns a DLT probability of approximately 0.2737, or 27.37%, at 80 mg. The positive slope means the modeled risk increases with dose; it does not mean that the risk increases by 0.90 probability points per unit of dose. The slope is on the log-odds scale.

Worked Example: Translating a Reference-Dose Risk Across Doses

Suppose, hypothetically, that the reference dose is 40 mg, the DLT risk there is 0.20, and the log-dose slope is \(\beta=0.80\). First express the reference risk in the intercept parameterization, then calculate risks at 20, 40, 80, and 160 mg.

At the reference dose, \(\alpha=\operatorname{logit}(0.20)=\log(0.20/0.80)=\log(0.25)\approx-1.3863\). The risk at any dose is obtained from

p(d)=\frac{1}{1+\exp\{-[-1.3863+0.80\log(d/40)]\}}.

For 20 mg, \(\log(20/40)=\log(0.5)\approx-0.6931\), giving a predictor of \(-1.3863+0.80(-0.6931)=-1.9408\) and risk \(p(20)\approx0.1256\). At 40 mg, the covariate is zero, so \(p(40)=0.2000\). At 80 mg, the predictor is \(-1.3863+0.80(0.6931)=-0.8318\), giving \(p(80)\approx0.3033\). At 160 mg, the predictor is \(-1.3863+0.80(1.3863)=-0.2773\), giving \(p(160)\approx0.4311\).

Thus the corresponding modeled risks are approximately 12.56%, 20.00%, 30.33%, and 43.11%. Each doubling of dose multiplies the odds by \(2^{0.80}\approx1.741\), not the probability by 1.741. Probabilities and odds are different quantities.

Bayesian Updating and What the Model Can Learn

The likelihood connects observed DLT counts to the dose-specific risks, while the prior describes plausible values of the intercept and slope before observing the trial outcomes. Bayes’ rule combines them: the posterior density is proportional to the likelihood multiplied by the prior density. A posterior draw \((\alpha^{(m)},\beta^{(m)})\) can be translated to a draw-specific risk at dose \(d\) by substituting both values into the inverse-logit formula. This is the direct link between parameter inference and dose-specific risk.

As a numerical illustration of the likelihood, suppose a hypothetical cohort at each of 20, 40, and 80 mg contains three participants, with observed DLT counts \(0,1,\) and \(2\), respectively. Using the risks from the preceding example, the likelihood at \(\alpha=-1.3863,\beta=0.80\) is

L = {3\choose0}(0.1256)^0(0.8744)^3 {3\choose1}(0.20)^1(0.80)^2 {3\choose2}(0.3033)^2(0.6967)^1 \approx0.0494.

This is the probability of those grouped outcomes under the specified parameter values, including the binomial coefficients. It is not the posterior probability that those parameter values are correct. The posterior compares likelihood contributions across parameter values while also incorporating the prior.

Estimating both parameters requires information about the dose-response relationship. If all participants receive a single dose, the likelihood depends on the combination \(\alpha+\beta\log(d/d_{\mathrm{ref}})\) at that dose. Many intercept-slope pairs can produce the same combination, so the data at that single dose cannot separately identify both parameters without prior information. Observations across distinct doses can help distinguish the intercept from the slope, although sparse data may still leave substantial posterior dependence between them.

Worked Example: Solving for a Model-Implied Target-Risk Dose

Using the hypothetical values \(d_{\mathrm{ref}}=40\) mg, \(\alpha=-1.60\), and \(\beta=0.90\), calculate the dose at which the model-implied DLT risk is 0.30.

The target log-odds are \(\operatorname{logit}(0.30)=\log(0.30/0.70)\approx-0.8473\). Substituting into the dose equation gives

\log(d/40)=\frac{-0.8473-(-1.60)}{0.90} =\frac{0.7527}{0.90} \approx0.8363.

Exponentiating and multiplying by the reference dose gives \(d=40\exp(0.8363)\approx40(2.3078)\approx92.31\) mg. Substitution verifies the result: \(x=\log(92.31/40)\approx0.8363\), the predictor is \(-1.60+0.90(0.8363)\approx-0.8473\), and the inverse logit is approximately 0.30.

If \(\beta=0\), risk is constant over positive doses. When that constant risk differs from \(q\), no dose solves \(p(d)=q\); when it equals \(q\), every positive dose solves the equation. Thus there is no unique target-risk dose when the slope is zero.

Diagnostics, Alternatives, and Limitations

Parameterization errors can produce plausible-looking but incorrect risk predictions. Before using the model, check that the reference dose is positive, all modeled doses are positive, the logarithm uses the intended dose ratio, and the same dose units are used for \(d\) and \(d_{\mathrm{ref}}\). Confirm that \(\alpha\) is the log-odds at the reference dose rather than at dose zero or at an uncentered numerical dose. Dose zero cannot be inserted into a log-dose model.

The slope also provides a direct structural check. If monotone increasing toxicity is required, \(\beta\) must be positive. If the fitted or sampled slope is negative, the model implies decreasing toxicity over the dose range; that is a substantive feature to investigate, not a harmless change in notation. Predictions at doses far beyond those observed can be especially sensitive to the linear-on-log-dose assumption. A two-parameter model may be too restrictive if the dose-response curve bends or plateaus.

The two-parameter BLRM differs from a CRM model based on a fixed skeleton and a single calibration parameter: here the intercept and log-dose slope both vary, and their joint prior matters. It also differs from a logistic model with an uncentered dose covariate, where the intercept refers to the log-odds at dose zero. Centering at a positive reference dose changes the intercept’s interpretation and often improves its practical meaning; it does not change the underlying logistic family when the parameter values are transformed consistently.

A two-parameter model is not automatically well determined merely because it has an interpretable form. With few participants or few observed DLTs, the posterior can depend strongly on the prior, and the intercept and slope can be correlated. Prior predictive calibration and examination of the observed dose range help reveal implausible implications. The model also assumes a common dose-response relationship across participants; it does not represent heterogeneity, covariate effects, delayed outcomes, or uncertainty about the assessment window unless those features are separately modeled.

Mistake: Treating \(\beta\) as the increase in DLT probability for each milligram. Correction: \(\beta\) is a slope in log-odds per unit log-dose; compute probability changes through the inverse-logit.
Mistake: Using \(\log(d)\) in one calculation and \(\log(d/d_{\mathrm{ref}})\) in another while keeping the same intercept. Correction: Keep the covariate definition consistent; changing the origin requires changing the intercept interpretation and value.
Mistake: Claiming a unique dose reaches a target risk when \(\beta=0\). Correction: Risk is constant, so either no dose reaches a different target or all positive doses reach the matching target.
Mistake: Reading an increasing log-odds as a constant increase in risk. Correction: The logistic transformation is nonlinear, so a fixed change in log-odds does not give a fixed change in probability.

Key takeaway: Specify the reference dose and centered log-dose covariate explicitly. Then interpret \(\alpha\) at that reference dose, interpret \(\beta\) on the log-odds scale, and obtain every dose-specific risk by applying the inverse-logit to \(\alpha+\beta\log(d/d_{\mathrm{ref}})\).

Check Your Understanding

Answer each question, then check the solutions below.

For \(d_{\mathrm{ref}}=25\) mg, \(\alpha=-1.20\), and \(\beta=0.70\), calculate the DLT risk at 50 mg.
If \(\alpha=\operatorname{logit}(0.10)\) at the reference dose and \(\beta=1\), by what factor do the odds change when dose is doubled?
Explain why data from participants all treated at one dose may not identify \(\alpha\) and \(\beta\) separately.
For \(\beta=0\), describe the solutions to \(p(d)=q\) when the constant risk is less than \(q\), and when it equals \(q\).

Solutions. 1. The covariate is \(\log(50/25)=\log 2\approx0.6931\), so the predictor is \(-1.20+0.70(0.6931)=-0.7148\). Thus \(p(50)=1/\{1+\exp(0.7148)\}\approx0.3285\), or 32.85%. 2. Doubling dose changes log-odds by \(\beta\log 2=\log 2\); the odds are multiplied by \(\exp(\log 2)=2\). 3. At one dose, the likelihood depends on the single linear combination \(\alpha+\beta\log(d/d_{\mathrm{ref}})\), so distinct pairs can give the same risk at that dose. 4. If the constant risk is less than \(q\), no positive dose solves the equation. If it equals \(q\), every positive dose solves it, so there is no unique solution.