Tutorials › Quantitative Systems Pharmacology (QSP) › Choosing Transit-Chain Order from Delay Dispersion

Home  /  Tutorials  /  Delay and Stochastic Extensions

Delay and Stochastic Extensions · Tutorial 96 of 799

Choosing Transit-Chain Order from Delay Dispersion

Use the target delay’s mean and variance to select a transit-chain order and rate, while checking whether the resulting Erlang approximation is adequate.

Advanced 12 min read

What You'll Learn

  • Derive the stage count implied by a target delay mean and variance.
  • Calculate the rate that preserves the target mean for any chosen integer stage count.
  • Check whether an exact match is possible and compare candidate integer orders.
  • Assess what a moment-matched transit chain does not capture about the delay distribution.
  • Identify common fitting, interpretation, and numerical pitfalls.

From Delay Dispersion to a Transit-Chain Order

A transit chain represents a distributed biological completion time using a sequence of stages. Its stage count controls how dispersed completion times are: at a fixed mean, more stages produce less variability. Choosing the order is therefore not just a matter of making the chain longer or shorter. It is a quantitative decision about how much dispersion the mechanism should represent.

Suppose a biological completion time has a specified target mean \(\mu_D\) and variance \(\sigma_D^2\), where \(\mu_D>0\) and \(\sigma_D^2>0\). The goal is to choose a stage count and rate so that the transit-chain completion time reproduces those moments exactly when possible, or as closely as a stated criterion allows when it is not.

Model and Assumptions

Definition: An equal-rate Erlang transit chain has \(k\) sequential stages, where \(k\) is a positive integer. Each stage has an independent exponential waiting time with the same rate \(\lambda>0\). The total completion time \(T\) is the sum of the \(k\) stage times. Thus \(T\) has an Erlang distribution with shape \(k\) and rate \(\lambda\).

This interpretation assumes that every completion follows the same ordered sequence, that the stage waiting times are independent and have the same rate, and that there is no branching, return to an earlier stage, or additional waiting time outside the chain. Under these assumptions, the chain’s completion-time distribution is Erlang. In a compartment model, the corresponding transit compartments pass material forward at rate \(\lambda\) per unit amount in each stage.

The stage count must be an integer because it counts stages. The rate is positive and has units of inverse time. Keep the time units consistent: if the target mean is in days, the rate must be in day\(^{-1}\), and the variance must be in days squared.

Deriving the Moment-Matching Rule

Let \(T_i\) be the waiting time in stage \(i\). For an exponential waiting time with rate \(\lambda\), \(E[T_i]=1/\lambda\) and \(\operatorname{Var}(T_i)=1/\lambda^2\). Since the total time is \(T=\sum_{i=1}^{k}T_i\), linearity of expectation gives its mean. Independence makes the variance of the sum equal to the sum of the stage variances:

$$ E[T]=\frac{k}{\lambda}, \qquad \operatorname{Var}(T)=\frac{k}{\lambda^2}. $$

To match a target mean \(\mu_D\), solve \(k/\lambda=\mu_D\) for the rate. Substituting that rate into the variance gives the variance implied by any chosen stage count:

$$ \lambda=\frac{k}{\mu_D}, \qquad \operatorname{Var}(T) =\frac{k}{(k/\mu_D)^2} =\frac{\mu_D^2}{k}. $$

Equating this implied variance to the target variance yields the stage count required for an exact match. It is useful to express the same condition using the squared coefficient of variation, the variance divided by the squared mean:

$$ k^*=\frac{\mu_D^2}{\sigma_D^2} =\frac{1}{\sigma_D^2/\mu_D^2}. $$

The target moments can be matched exactly by an equal-rate Erlang chain if and only if \(k^*\) is a positive integer. If it is not an integer, there is no stage count that reproduces both moments exactly. For any selected integer \(k\), choosing \(\lambda=k/\mu_D\) preserves the mean, while the resulting variance is \(\mu_D^2/k\).

Key result: The target squared coefficient of variation must equal \(1/k\) for an equal-rate \(k\)-stage chain. Therefore, an Erlang transit chain has squared coefficient of variation at most 1, with equality at \(k=1\). At a fixed mean, its variance is \(\mu_D^2/k\), so increasing the stage count narrows the completion-time distribution.

A Practical Selection Procedure

When \(k^*\) is not an integer, there is no single “best” integer independent of the modeling objective. A transparent default is to preserve the target mean and choose the admissible integer stage count with the smallest absolute variance error. Since \(\mu_D\) is fixed, this is equivalent to choosing the candidate with the smallest absolute error in squared coefficient of variation. Other objectives, such as matching a particular quantile or the full completion-time profile, may lead to a different choice.

1
Compute the implied order.
Calculate \(k^*=\mu_D^2/\sigma_D^2\), using consistent time units.
2
Check feasibility.
If \(k^*\) is a positive integer, use that order for an exact two-moment match. If not, no integer order gives an exact match.
3
Compare admissible candidates.
For a noninteger \(k^*\geq 1\), compare the neighboring integer orders, such as \(\lfloor k^*\rfloor\) and \(\lceil k^*\rceil\), by calculating their implied variances \(\mu_D^2/k\).
4
Set the rate and report the mismatch.
For the chosen order, set \(\lambda=k/\mu_D\). State the implied variance and how it differs from the target, rather than describing the fit as exact.

When \(k^*<1\), the target squared coefficient of variation exceeds 1. No positive integer stage count can match it: even the one-stage exponential chain has squared coefficient of variation 1, and adding stages reduces dispersion further. If variance is the chosen matching criterion, \(k=1\) is the least dispersed chain available and is closer in variance than any larger order.

Worked Example: An Exact Four-Stage Match

Suppose a hypothetical biological completion time has target mean \(\mu_D=8\) days and target variance \(\sigma_D^2=16\) days\(^2\). Find the equal-rate transit-chain order and rate that match both moments.

First calculate the implied order:

$$ k^*=\frac{(8\ \text{days})^2}{16\ \text{days}^2} =\frac{64}{16}=4. $$

Because 4 is a positive integer, exact matching is possible. The rate that matches the mean is:

$$ \lambda=\frac{k}{\mu_D} =\frac{4}{8\ \text{days}} =0.5\ \text{day}^{-1}. $$

Checking both moments, the chain mean is \(4/(0.5\ \text{day}^{-1})=8\) days and its variance is \(4/(0.5\ \text{day}^{-1})^2=16\) days\(^2\). Both targets are reproduced.

Worked Example: Choosing Between Three and Four Stages

Suppose a hypothetical completion-time distribution has target mean \(\mu_D=10\) days and target variance \(\sigma_D^2=30\) days\(^2\). Use the mean-preserving rate and choose between the neighboring integer orders by minimizing absolute variance error.

The implied order is \(k^*=(10\ \text{days})^2/(30\ \text{days}^2)=100/30=3.333\). It is not an integer, so exact matching is impossible. For \(k=3\), the rate and implied variance are:

$$ \lambda_3=\frac{3}{10\ \text{days}}=0.3000\ \text{day}^{-1}, \qquad \operatorname{Var}(T\mid k=3)=\frac{(10\ \text{days})^2}{3} =33.33\ \text{days}^2. $$

The variance is \(33.33-30=3.33\) days\(^2\) above the target, an 11.11% relative error. For \(k=4\):

$$ \lambda_4=\frac{4}{10\ \text{days}}=0.4000\ \text{day}^{-1}, \qquad \operatorname{Var}(T\mid k=4)=\frac{(10\ \text{days})^2}{4} =25.00\ \text{days}^2. $$

This variance is \(25-30=-5\) days\(^2\), or 16.67% below the target. Under the stated absolute-variance-error criterion, choose \(k=3\), because \(3.33<5\). Report the resulting model as having a mean of 10 days and variance 33.33 days\(^2\), not as a match to both target moments.

Worked Example: A Target More Dispersed Than Any Erlang Chain

Suppose a hypothetical completion-time distribution has target mean \(\mu_D=6\) hours and target variance \(\sigma_D^2=54\) hours\(^2\). Determine whether a mean-preserving equal-rate Erlang chain can match both moments.

The implied order is \(k^*=(6\ \text{hours})^2/(54\ \text{hours}^2)=36/54=0.6667\). This is below the minimum admissible order of 1. The target squared coefficient of variation is \(54/36=1.5\), exceeding the maximum value of 1 available for an Erlang chain.

If the objective is to keep the mean and get as close as possible in variance, use the minimum order \(k=1\). Its rate is \(\lambda=1/(6\ \text{hours})=0.1667\ \text{hour}^{-1}\), and its variance is \(1/(0.1667\ \text{hour}^{-1})^2=36\) hours\(^2\). It underestimates the target variance by 18 hours\(^2\). Increasing the order would lower the implied variance further, so it cannot improve the match under this criterion.

What Moment Matching Does—and Does Not—Guarantee

Matching the mean and variance constrains two features of the completion-time distribution, but it does not ensure that the entire distribution is a good representation of the biological process. An Erlang chain has a specific shape determined by its integer order. For example, its skewness is \(2/\sqrt{k}\); therefore, selecting an order also fixes a shape feature that was not independently specified by the target mean and variance. Tail probabilities, quantiles, and the fraction completing by a particular time can still differ from those of the target distribution.

The equal-rate chain also imposes structural assumptions. It cannot represent a target squared coefficient of variation above 1, and an integer order may prevent exact moment matching even when that ratio is below 1. If matching the full distribution matters, assess the chain against the relevant completion-time data or predictions, not only against its first two moments. Depending on the scientific purpose, alternatives include a different distributed-delay family or a model that represents the delay in another way. The choice should be guided by the biological question and by which parts of the delay distribution drive the model’s predictions.

When the target moments are estimated from data rather than specified, they also carry uncertainty. A value of \(k^*\) near an integer does not by itself establish that the biological mechanism has that exact number of stages. It says that, for the supplied moments, that integer may be a convenient approximation. Sensitivity analysis across plausible orders can show whether conclusions depend on the approximation.

Diagnostics and Common Pitfalls

After selecting an order, check the model against the quantities it is meant to represent. Verify that the calculated chain mean equals \(\mu_D\), calculate its implied variance explicitly, and compare both with the targets in the same units. If completion-time observations or a target distribution are available, compare relevant quantiles or completion fractions as well. A large discrepancy in those features is evidence that moment matching alone is insufficient.

  • Rounding \(k^*\) and claiming an exact match. Rounding produces an integer order, but the resulting variance is \(\mu_D^2/k\). Check that value before reporting a match.
  • Using the target mean as the stage rate’s reciprocal. The rate is \(k/\mu_D\), not \(1/\mu_D\), except when \(k=1\). Using the wrong rate changes the chain’s mean.
  • Ignoring units. A rate expressed per hour cannot be combined with a mean in days without conversion. The computed variance must have units of time squared.
  • Assuming a noninteger order is a literal number of stages. \(k^*\) is an implied continuous value used to assess moment compatibility; a physical equal-rate chain requires integer \(k\).
  • Treating two matched moments as proof of a correct mechanism. Different delay distributions can share a mean and variance. Check distributional features that matter for the application.
Key takeaway: For a target mean \(\mu_D\) and variance \(\sigma_D^2\), the implied Erlang order is \(k^*=\mu_D^2/\sigma_D^2\). Exact matching requires a positive integer order; otherwise, choose and justify an integer approximation, preserve the mean with \(\lambda=k/\mu_D\), and report the resulting variance mismatch.

Check Your Understanding

Answer each question, then compare your work with the solutions below.

  1. A hypothetical target delay has mean 9 days and variance 27 days\(^2\). Can an equal-rate transit chain match both moments exactly? If so, give its order and rate.
  2. A hypothetical target delay has mean 4 minutes and variance 4 minutes\(^2\). Find the exact order and rate, and verify the variance.
  3. A hypothetical target delay has mean 5 hours and variance 50 hours\(^2\). Explain why exact matching is impossible and state the minimum variance attainable while preserving the mean.
  4. A hypothetical target delay has mean 12 hours and variance 8 hours\(^2\). Find the exact order and rate.
  5. A hypothetical target delay has mean 7 days and variance 10 days\(^2\). Compare \(k=4\) and \(k=5\) by absolute variance error, using a mean-preserving rate in each case. Which would you select under that criterion?

Solutions. (1) \(k^*=9^2/27=3\), so exact matching is possible; \(\lambda=3/(9\ \text{days})=0.3333\ \text{day}^{-1}\). Its variance is \(3/(0.3333\ \text{day}^{-1})^2=27\) days\(^2\). (2) \(k^*=4^2/4=4\), and \(\lambda=4/(4\ \text{minutes})=1\ \text{minute}^{-1}\); the variance is \(4/(1\ \text{minute}^{-1})^2=4\) minutes\(^2\). (3) \(k^*=25/50=0.5<1\), so no admissible order matches both moments. The minimum variance at mean 5 hours occurs at \(k=1\): \(5^2/1=25\) hours\(^2\). (4) \(k^*=12^2/8=18\), and \(\lambda=18/(12\ \text{hours})=1.5\ \text{hour}^{-1}\). (5) For \(k=4\), \(\lambda=4/7=0.5714\ \text{day}^{-1}\) and the variance is \(49/4=12.25\) days\(^2\), an error of \(2.25\) days\(^2\). For \(k=5\), \(\lambda=5/7=0.7143\ \text{day}^{-1}\) and the variance is \(49/5=9.8\) days\(^2\), an error of \(0.2\) days\(^2\). Choose \(k=5\) under the stated criterion.

Previous

Next

Back to Tutorials
Clinical Biostats