Probability of Target Attainment Across Candidate Doses

Why Calculate Target-Attainment Probability?

A dose recommendation is often summarized by a target dose: the dose predicted to produce a prespecified response. But parameter uncertainty means that the target dose is not known with certainty. At a fixed candidate dose, one useful question is therefore: across the plausible model parameters given the data, how often does the modeled response meet the target? Answering this question over a set of candidate doses produces a probability-of-target-attainment (PTA) curve.

Here, target attainment refers to whether the model-predicted response meets or exceeds a specified threshold. The calculation integrates over parameter uncertainty; it does not, by itself, include variability in an individual patient’s response around the model prediction. That distinction matters: a probability based on uncertain mean response parameters is not automatically the probability that a randomly selected patient will attain the target.

Definition: Let $m(d;\theta)$ be the model-predicted response at dose $d$, with parameter vector $\theta$, and let $T$ be a prespecified response target. Given data $\mathcal{D}$, the parameter-uncertainty probability of target attainment at dose $d$ is $$ \operatorname{PTA}(d)=P\{m(d;\theta)\geq T\mid\mathcal{D}\} =\int I\{m(d;\theta)\geq T\}\,p(\theta\mid\mathcal{D})\,d\theta, $$ where $I\{\cdot\}$ equals 1 when its condition is true and 0 otherwise, and $p(\theta\mid\mathcal{D})$ is the posterior distribution of the model parameters given the data.

This is a probability conditional on the fitted model, the data, and the parameter distribution used for uncertainty propagation. It is not a probability that the target or model is correct. In a Bayesian analysis, it is a posterior probability over parameters; in a simulation-based frequentist analysis, a corresponding fraction of bootstrap or other uncertainty draws can be reported, but its interpretation follows that procedure rather than posterior probability.

Estimating the Curve From Parameter Draws

Suppose $M$ draws $\theta^{(1)},\ldots,\theta^{(M)}$ represent the parameter uncertainty distribution. At each candidate dose, evaluate the model response for every draw, mark whether it meets the target, and average the indicators:

\widehat{\operatorname{PTA}}(d) =\frac{1}{M}\sum_{m=1}^{M} I\{m(d;\theta^{(m)})\geq T\}.

If draws have unequal simulation weights $w_m$, normalized so that $\sum_{m=1}^{M}w_m=1$, use $\sum_{m=1}^{M}w_m I\{m(d;\theta^{(m)})\geq T\}$. The equal-weight formula is the usual estimate from equally weighted posterior draws. For draws with meaningful dependence, such as successive draws from a Markov chain, the estimate remains an average, but its Monte Carlo precision depends on effective rather than raw sample size.

The essential assumptions are that the model and target are defined on compatible response scales, that the parameter draws adequately represent the intended uncertainty distribution, and that the model can be evaluated at every candidate dose. A model-based PTA also assumes the model is sufficiently credible for the dose range being examined. Boundary and extrapolation concerns are especially important when candidate doses extend beyond well-supported data.

Worked Example: Counting Target-Dose Draws at Candidate Doses

Suppose eight equally weighted parameter draws have been transformed into target-dose draws, in mg: 4, 5, 6, 7, 8, 9, 10, and 12. Assume that, for every draw, the modeled response is nondecreasing with dose, so a dose at or above that draw’s target dose meets the target. Estimate PTA at 5, 8, and 10 mg.

At 5 mg, two of the eight target doses are at most 5 mg (4 and 5), so $ \widehat{\operatorname{PTA}}(5)=2/8=0.250 $.

At 8 mg, five target doses are at most 8 mg (4, 5, 6, 7, and 8), giving $ \widehat{\operatorname{PTA}}(8)=5/8=0.625 $.

At 10 mg, seven target doses are at most 10 mg; only the draw of 12 mg is above 10 mg. Thus $ \widehat{\operatorname{PTA}}(10)=7/8=0.875 $.

These are empirical estimates from only eight draws, not exact probabilities. They show how a distribution of target doses can be converted into an attainment curve under the stated monotonicity assumption. The denominator is eight at every dose because all eight draws contribute.

Why the Target-Dose Shortcut Works

When each draw’s response is nondecreasing in dose, define its target dose as the smallest dose at which the modeled response meets or exceeds $T$. For that draw, attainment at $d$ is equivalent to its target dose being no greater than $d$. Averaging this equivalence gives the empirical target-dose shortcut.

I\{m(d;\theta^{(m)})\geq T\}=I\{D_T^{(m)}\leq d\}, \qquad \widehat{\operatorname{PTA}}(d)=\frac{1}{M}\sum_{m=1}^{M}I\{D_T^{(m)}\leq d\}.

This is an empirical cumulative distribution of the target-dose draws. The equivalence requires the defined target dose to exist and the response to remain at or above target after crossing it. If a response curve is nonmonotone, has multiple crossings, or never reaches the target in some draws, the shortcut needs careful definition and may be misleading. Directly evaluating $m(d;\theta^{(m)})$ at each candidate dose avoids assuming that a single target dose fully describes each draw’s response.

Justification: For a nondecreasing response curve, if $D_T^{(m)}\leq d$, then the response at $d$ is at least the response at $D_T^{(m)}$, which is at least $T$. Conversely, if the response at $d$ is at least $T$, the smallest dose that reaches $T$ cannot exceed $d$. Thus the two indicators are equal draw by draw, and their averages are equal.

Worked Example: Evaluating Model Draws Directly

For illustration, suppose six equally weighted hypothetical parameter draws produce linear mean-response curves $m(d)=a+bd$, with response units shown on the same scale as a target of 50 units. The draws $(a,b)$ are (20, 6), (30, 4), (10, 5), (25, 5), (15, 5), and (35, 3). Each slope is positive, so each curve increases with dose. Find the estimated PTA at 6 mg.

Evaluate each response at 6 mg: $20+6(6)=56$, $30+4(6)=54$, $10+5(6)=40$, $25+5(6)=55$, $15+5(6)=45$, and $35+3(6)=53$ response units. Four of these six values meet or exceed 50: 56, 54, 55, and 53. Therefore $ \widehat{\operatorname{PTA}}(6)=4/6=0.667 $.

As a check, the target dose for each draw is $(50-a)/b$: 5, 5, 8, 5, 7, and 5 mg. Four of those six doses are no greater than 6 mg, giving the same estimate. The check works because all six response curves are increasing and each target dose exists.

Monte Carlo Precision and Plotting

A finite set of draws introduces simulation error in the estimated curve. At one fixed dose, if the draws are independent and identically distributed, the attainment indicators behave as Bernoulli variables with success probability $\operatorname{PTA}(d)$. The estimated standard error is therefore approximately

\widehat{SE}_{MC}\{\widehat{\operatorname{PTA}}(d)\} =\sqrt{\frac{\widehat{\operatorname{PTA}}(d)\{1-\widehat{\operatorname{PTA}}(d)\}}{M}}.

This quantifies Monte Carlo error in estimating the probability, not uncertainty about the biological response or the target itself. With correlated simulation draws, this formula using raw $M$ can understate simulation error; use an appropriate effective sample size or a chain-based Monte Carlo error estimate.

Worked Example: Assessing Monte Carlo Error at One Dose

At a candidate dose, suppose 280 of 400 independent, equally weighted parameter draws meet the target. The estimated PTA is $280/400=0.700$. Its estimated Monte Carlo standard error is $ \sqrt{0.700(1-0.700)/400}=\sqrt{0.000525}=0.0229 $.

A rough two-standard-error range for simulation precision is $0.700\pm 2(0.0229)$, or approximately 0.654 to 0.746. This is a diagnostic approximation for the numerical estimate at this dose, not a 95% interval for a patient response or a statement that the true parameter-conditional PTA has a 95% probability of lying in that range.

If the analysis needs substantially finer Monte Carlo precision, increase the number of effective draws and recalculate. Simply reporting more decimal places does not reduce simulation error.

To plot the curve, evaluate PTA on a prespecified grid of candidate doses and plot dose on the horizontal axis and estimated PTA, from 0 to 1, on the vertical axis. Label dose units, response target and scale, and the source of parameter draws. With target-dose draws under monotonicity, the empirical curve is a step function; with direct response evaluation, show the grid estimates without implying that uncomputed intermediate values are exact. A smooth connecting line can aid visual reading, but it does not add information or justify interpolation beyond the model.

Specify the estimand.
State the target $T$, the response scale, the candidate doses, and whether attainment means meeting or strictly exceeding the threshold.

Obtain parameter draws.
Use draws that represent the intended fitted-model uncertainty and retain their weights or dependence structure.

Evaluate attainment.
At every candidate dose, evaluate the response for each draw and average the indicators, or use target-dose draws when the monotonicity conditions hold.

Check and display.
Inspect Monte Carlo precision, dose-range support, and curve shape before plotting and interpreting the estimates.

Diagnostics, Alternatives, and Limitations

Several checks help identify a PTA calculation that is not trustworthy. First, verify that the target and modeled response use the same scale and that all doses use the same response definition. Second, compare direct response indicators with target-dose indicators where monotonicity is assumed; disagreement suggests a coding error, inconsistent target-dose definition, or a failure of the assumption. Third, inspect the number of attaining draws at each dose. Estimates near 0 or 1 can still have relevant uncertainty, particularly with few effective draws. Finally, examine whether the model is being evaluated within a dose range supported by data and whether the resulting curve is scientifically plausible.

The main computational alternative is to calculate each draw’s target dose and then estimate the fraction of target doses below each candidate dose. This can be efficient when every draw has one well-defined target crossing and the response is monotone. Direct evaluation at the candidate doses is more general and makes the exact event being counted explicit. Neither approach removes model dependence: both summarize uncertainty under the chosen model and parameter distribution.

A common mistake is to call the PTA the probability that a patient will respond. The calculation above integrates over uncertain parameters but uses a modeled response, not an individual’s residual outcome. A patient-level predictive probability would also need to account for the model’s response variability and possibly other sources of heterogeneity. Another mistake is to read a high PTA as proof that a dose is optimal. PTA addresses one target and one uncertainty calculation; it does not, on its own, compare safety, treatment burden, or other decision criteria.

Key takeaway: At each candidate dose, PTA is the fraction of uncertainty draws whose modeled responses meet the target. Its interpretation depends on the response definition, parameter draws, model adequacy, and—when using target-dose draws—monotonicity.

Check Your Understanding

Answer each question, then compare with the solutions below.

At one dose, 45 of 60 equally weighted draws meet the target. What is the estimated PTA?
Under what condition can the target-dose distribution be used directly to estimate PTA across doses?
For an estimate of 0.50 from 100 independent draws, calculate the approximate Monte Carlo standard error.
Does a PTA of 0.80 mean that 80% of future patients will attain the target? Explain.

Solutions. (1) $\widehat{\operatorname{PTA}}=45/60=0.750$. (2) Each draw must have a well-defined target dose and a response that is nondecreasing across the dose range, so attainment at $d$ is equivalent to $D_T\leq d$. (3) $\widehat{SE}_{MC}=\sqrt{0.50(1-0.50)/100}=\sqrt{0.0025}=0.0500$. (4) No. It means 80% of the represented parameter draws yield a modeled response meeting the target at that dose. A future-patient probability would also require appropriate outcome variability and a predictive calculation.