Deriving Dose Decision Thresholds from Harm Tradeoffs

Why a Dose Decision Needs a Harm Threshold

A predicted toxicity risk does not, by itself, determine whether a dose should be reduced. The same risk can justify different actions when the consequences of the alternatives differ. For one patient and treatment, missing a toxicity risk may be much more harmful than a modest loss of treatment effect from reducing the dose. In another setting, unnecessary reduction may carry substantial clinical cost. A decision threshold makes that tradeoff explicit.

Here the decision is between maintaining a dose and reducing it, and the outcome of interest is a binary toxicity event. The central calculation compares the expected harm of false reassurance—maintaining the dose when toxicity would occur—with the expected harm of an unnecessary reduction—reducing the dose when toxicity would not occur. It is a decision rule for a specified clinical setting, not a universal toxicity cutoff.

Definition: Let \(T\) denote the event that toxicity would occur under the current dose, and let \(p=P(T\mid \text{available patient information})\) be its predicted probability. A false reassurance occurs when the dose is maintained despite \(T\); an unnecessary dose reduction occurs when the dose is reduced despite the absence of \(T\). Let \(h_F\) be the harm assigned to false reassurance and \(h_R\) the harm assigned to unnecessary reduction. Both harms must be expressed on the same scale for comparison.

Deriving the Threshold

Begin with a deliberately simple decision model. Assume the only harm of maintaining the dose is toxicity in state \(T\), with harm \(h_F\). Assume dose reduction prevents that toxicity for this calculation, and its harm \(h_R\) occurs when the patient would not otherwise have experienced toxicity. These are simplifying assumptions, not facts that hold for every treatment.

If the dose is maintained, the expected harm is the probability of toxicity multiplied by its assigned harm. If the dose is reduced, the expected harm is the probability of no toxicity multiplied by the harm of unnecessary reduction. The preferred action is the one with lower expected harm.

L(\text{maintain})=p h_F,\qquad L(\text{reduce})=(1-p)h_R

Reduce the dose when \(L(\text{reduce})<L(\text{maintain})\). Rearranging gives the threshold:

(1-p)h_R < p h_F \quad\Longleftrightarrow\quad h_R < p(h_F+h_R) \quad\Longleftrightarrow\quad p > p^*=\frac{h_R}{h_F+h_R}

At \(p=p^*\), the two actions have equal expected harm under this model. Above the threshold, reduction has lower expected harm; below it, maintaining the dose has lower expected harm. The threshold rises as the harm of unnecessary reduction rises, and falls as the harm of false reassurance rises. If the two harms are equal and positive, \(p^*=0.5\); equal harms do not imply a 5%, 10%, or other commonly used clinical risk cutoff.

Worked Example: A Lower Threshold When Missed Toxicity Is More Harmful

Suppose, hypothetically, a clinical team assigns 80 harm points to false reassurance and 20 harm points to unnecessary dose reduction. The points are a decision aid for this example, not a validated clinical scale. For simplicity, assume reducing the dose prevents the toxicity event and has no other harm when toxicity would occur.

The threshold is

p^*=\frac{h_R}{h_F+h_R} =\frac{20}{80+20} =\frac{20}{100}=0.20

A hypothetical patient has a predicted toxicity risk of \(p=0.28\). The risk exceeds the 0.20 threshold. Compare the expected harms directly:

L(\text{maintain})=0.28(80)=22.4\text{ harm points}

L(\text{reduce})=(1-0.28)(20)=0.72(20)=14.4\text{ harm points}

Under these assumptions, reducing the dose has 8.0 fewer expected harm points. This is not a statement that toxicity is certain, nor that the model establishes what will happen to this patient. It is the action favored by the stated risk estimate and harm assignments.

Reading the Threshold in Context

A threshold is meaningful only alongside its harm definitions. In this binary setup, \(h_F\) is the assigned harm of an event that could have been avoided by reducing the dose, while \(h_R\) is the assigned harm of reducing the dose when that event would not have occurred. The numbers should reflect the intended clinical decision and population. They should not be selected merely to produce a preferred action.

The threshold is not the predicted risk for a particular patient, and crossing it is not proof that toxicity will occur. It specifies how a risk estimate is compared with a tradeoff. A predicted risk of 0.28 means a 28% modeled probability of the defined event under the specified conditions; it does not mean a 28% probability that dose reduction is the correct action. The action depends on the harm model as well.

Worked Example: When Unnecessary Reduction Is More Costly

Consider a different hypothetical treatment context. The team assigns 30 harm points to false reassurance and 10 harm points to unnecessary reduction. Assume again that dose reduction prevents toxicity and that these are the only harms counted.

The action threshold is

p^*=\frac{10}{30+10}=\frac{10}{40}=0.25

For a hypothetical patient with \(p=0.18\), the risk is below the threshold. Maintaining the dose has expected harm

L(\text{maintain})=0.18(30)=5.4\text{ harm points}

Reducing the dose has expected harm

L(\text{reduce})=(1-0.18)(10)=0.82(10)=8.2\text{ harm points}

Maintaining the dose is favored in this model, with 2.8 fewer expected harm points. The conclusion does not mean that 18% is a negligible risk in all clinical settings. It means only that, given these particular harm assignments and assumptions, the expected harm of reduction is greater at this risk.

When Dose Reduction Does Not Eliminate Toxicity

The simple derivation treats reduction as if it prevents the event. In practice, toxicity may still occur after a reduction. Let \(d\) be the probability that reducing the dose prevents toxicity, conditional on toxicity occurring under the current dose. Thus \(d=1\) represents complete prevention, while \(d=0\) represents no prevention. Suppose the harm of toxicity is \(h_T\), and retain \(h_R\) as the harm of an unnecessary reduction when toxicity would not otherwise occur.

Maintaining the dose has expected toxicity harm \(p h_T\). Reducing the dose leaves residual expected toxicity harm \(p(1-d)h_T\), and also incurs expected unnecessary-reduction harm \((1-p)h_R\). Comparing these two expressions gives the adjusted threshold:

L(\text{maintain})=p h_T,\qquad L(\text{reduce})=p(1-d)h_T+(1-p)h_R

L(\text{reduce})<L(\text{maintain}) \quad\Longleftrightarrow\quad p d h_T>(1-p)h_R \quad\Longleftrightarrow\quad p>\frac{h_R}{d h_T+h_R}

The decision depends on the expected toxicity harm that reduction can avert, \(d h_T\), rather than on toxicity harm alone. When \(d=1\) and \(h_T=h_F\), this reduces to the earlier formula. When \(d=0\), the threshold is 1 for positive \(h_R\): reduction cannot prevent toxicity, so this model never favors it on the basis of toxicity prevention.

Worked Example: Accounting for Partial Prevention

Suppose, hypothetically, toxicity under the current dose is assigned 50 harm points. A dose reduction is estimated to prevent 60% of such events, so \(d=0.60\). The harm of an unnecessary reduction is assigned 12 points. Assume these values are applicable to the patient and decision being considered, and that no other harms are included.

The adjusted threshold is

p^*=\frac{12}{(0.60)(50)+12} =\frac{12}{30+12} =\frac{12}{42} \approx 0.2857

For a hypothetical patient with predicted risk \(p=0.32\), the risk is above the threshold. The expected harm of maintaining the dose is

L(\text{maintain})=0.32(50)=16.0\text{ harm points}

The expected harm after reduction includes both residual toxicity and unnecessary reduction:

L(\text{reduce})=0.32(1-0.60)(50)+(1-0.32)(12)

=0.32(0.40)(50)+0.68(12)=6.4+8.16=14.56\text{ harm points}

Reduction is favored by 1.44 expected harm points under the stated assumptions. Omitting the residual-toxicity term would overstate the benefit of reduction; treating \(d\) as 100% would give an unjustifiably low threshold.

Assumptions, Diagnostics, and Alternatives

Before applying a threshold, check that the probability and harm quantities refer to the same decision. The risk \(p\) must be the probability of the defined toxicity event under the current dose and relevant patient information—not a risk under a different regimen or time window. The harm values must describe the consequences relevant to this decision, and the reduction-prevention probability \(d\), when used, must concern the same event and action.

The calculation also assumes that the stated losses adequately represent the decision and that the actions being compared are available. The basic model treats the two states as mutually exclusive and exhaustive: toxicity would occur under the current dose, or it would not. It does not automatically account for several toxicities, graded severity, treatment response, patient preferences, or additional actions. If these matter, collapsing them into one binary event and two harm values may hide clinically important differences.

Several checks can reveal a fragile decision. Confirm how the risk was estimated and for which population, dose, and follow-up interval. Examine whether the predicted risks are credible for the patients to whom the threshold will be applied; systematic overprediction or underprediction can shift actions in the wrong direction. Check that harm assignments are defensible to relevant clinical decision makers and that changing a plausible harm value does not reverse the conclusion unexpectedly. When the conclusion changes under small, reasonable changes in inputs, report that the decision is sensitive rather than presenting the threshold as precise.

A single fixed risk cutoff is simpler to communicate, but without an explicit harm rationale it may conceal an implicit tradeoff. A threshold derived from equal harms is a special case, not a neutral default. A rule that reduces every dose or maintains every dose avoids individualized comparison but ignores patient-specific risk. The present calculation makes the tradeoff transparent; it does not remove uncertainty in the risk estimate, establish that the harm scale is complete, or replace clinical judgment.

Common pitfall: A reader may say, “The toxicity risk is 20%, so the dose should be reduced,” as though 20% were an intrinsic action boundary. The correction is to compare the risk with a threshold derived from the relevant harms. In the simple model, \(p^*=h_R/(h_F+h_R)\); the same 20% risk can favor reduction under one harm balance and maintaining the dose under another.

Check Your Understanding

Use the simple model unless a question says otherwise. Solutions are included after the questions.

If false reassurance has harm 45 points and unnecessary reduction has harm 15 points, calculate the threshold.
With those harms, a patient has predicted risk \(p=0.20\). Which action has lower expected harm, and by how much?
If the two harms are equal and positive, what is the threshold? Explain what that result does and does not imply.
With toxicity harm 40 points, reduction-prevention probability \(d=0.50\), and unnecessary-reduction harm 10 points, calculate the adjusted threshold.
Name two assumptions or inputs that should be checked before applying a calculated threshold to patients.

Solutions. (1) \(p^*=15/(45+15)=15/60=0.25\). (2) \(L(\text{maintain})=0.20(45)=9.0\) points; \(L(\text{reduce})=0.80(15)=12.0\) points. Maintaining the dose has 3.0 fewer expected harm points. (3) If \(h_F=h_R>0\), then \(p^*=h_R/(2h_R)=0.50\). This follows from equal assigned harms in this model; it is not a universal clinical cutoff. (4) \(p^*=10/[(0.50)(40)+10]=10/30\approx0.3333\). (5) Appropriate checks include whether the predicted risk matches the event, dose, and time window; whether the risk estimate is credible for the intended patients; whether the harm assignments reflect the decision; and whether reduction’s prevention probability is supported and relevant.