Choosing Transit-Stage Number and Rate

Why Stage Number and Rate Must Be Chosen Together

A transit chain represents a delayed signal as it passes through a sequence of first-order stages. Two parameters determine much of its timing: the number of stages and the rate at which signal leaves each stage. Increasing the number of stages generally makes the response more concentrated around its mean, while changing the rate can offset the resulting change in average delay. Consequently, two chains can have the same mean delay yet produce meaningfully different response profiles.

This is a practical model-selection problem, not just a matter of choosing a convenient delay parameter. A model that matches the average timing may still miss the observed onset, peak, or tail. Conversely, if observations are sparse or noisy, those shape differences may not be resolved well enough to determine a unique stage count. The earlier material on impulse responses and convolution provides the basic cascade framework; here the focus is how stage number and rate jointly control its timing and shape.

Definition: Consider a serial chain of $N$ identical, linear, first-order transit stages, each with rate constant $k_{\mathrm{tr}}>0$. Here $N$ is a positive integer and $k_{\mathrm{tr}}$ has units of inverse time. For a unit impulse input and unit total gain, the normalized output impulse response is $$ h_N(t)=\frac{k_{\mathrm{tr}}^N t^{N-1}e^{-k_{\mathrm{tr}}t}}{(N-1)!},\qquad t\geq 0. $$ Its area is one. A separate gain can multiply this response without changing its timing or shape.

The convention here counts every first-order stage in the chain: $N=1$ gives a single exponential response, and each added stage contributes another first-order delay. The parameter $k_{\mathrm{tr}}$ is not itself a delay. A single stage has mean residence time $1/k_{\mathrm{tr}}$, so the whole chain has a longer mean that depends on both $N$ and $k_{\mathrm{tr}}$.

How Stage Count and Rate Shape the Response

The impulse response has a gamma density form. Its three key timing summaries are the mean, variance, and mode:

\text{Mean}=\frac{N}{k_{\mathrm{tr}}},\qquad \text{Variance}=\frac{N}{k_{\mathrm{tr}}^2},\qquad \text{Mode}=\frac{N-1}{k_{\mathrm{tr}}}\quad (N>1).

The mean is often called the mean transit time in this model. It is not generally the time of the peak: for $N>1$, the peak occurs earlier. The standard deviation is $\sqrt{N}/k_{\mathrm{tr}}$, and the coefficient of variation, standard deviation divided by mean, is $1/\sqrt{N}$. Thus, at a fixed mean, increasing $N$ decreases relative spread and brings the peak closer to that mean.

For a constant unit step input, the normalized output is the cumulative integral of the impulse response. It rises from zero toward one:

y_{\mathrm{step}}(t)=1-e^{-k_{\mathrm{tr}}t}\sum_{j=0}^{N-1}\frac{(k_{\mathrm{tr}}t)^j}{j!}.

This is useful when a drug effect is driven by a sustained change in upstream signal. The impulse response describes how a brief input is distributed over time; the step response describes the accumulated output under a sustained input. Both reflect the same stage number and rate.

A Short Derivation of the Timing Results

A first-order stage has an exponential waiting-time distribution with rate $k_{\mathrm{tr}}$, mean $1/k_{\mathrm{tr}}$, and variance $1/k_{\mathrm{tr}}^2$. In a serial chain of $N$ identical stages, the total transit time is the sum of $N$ such waiting times. Under the model assumption that these stage delays act sequentially with the same rate, their means and variances add:

E[T]=\sum_{i=1}^{N}\frac{1}{k_{\mathrm{tr}}} =\frac{N}{k_{\mathrm{tr}}},\qquad \operatorname{Var}(T)=\sum_{i=1}^{N}\frac{1}{k_{\mathrm{tr}}^2} =\frac{N}{k_{\mathrm{tr}}^2}.

The sum has the gamma density $h_N(t)$ given above. To obtain its mode for $N>1$, maximize its log density. The terms that depend on $t$ are $(N-1)\log(t)-k_{\mathrm{tr}}t$, whose derivative is $(N-1)/t-k_{\mathrm{tr}}$. Setting this equal to zero gives $t=(N-1)/k_{\mathrm{tr}}$; the derivative changes from positive to negative there, so this is the maximum. For $N=1$, the exponential response instead has its maximum at $t=0$.

A useful parameterization fixes the mean transit time $T=N/k_{\mathrm{tr}}$. Then $k_{\mathrm{tr}}=N/T$, and the variance becomes $T^2/N$. This makes the joint control especially clear: at equal mean $T$, the stage count alone determines the model's relative spread, while the rate must change in proportion to $N$.

Worked Example: Two Chains With the Same Mean Delay

Suppose two hypothetical unit-gain chains are required to have a mean transit time of 4 hours. Compare a two-stage chain with an eight-stage chain. The mean-matching condition is $k_{\mathrm{tr}}=N/T$.

For $N=2$, $k_{\mathrm{tr}}=2/4=0.500\ \mathrm{h}^{-1}$. Its mean is $2/0.500=4.00$ h, its variance is $2/(0.500)^2=8.00\ \mathrm{h}^2$, and its standard deviation is $\sqrt{8.00}=2.83$ h. Its mode is $(2-1)/0.500=2.00$ h.

For $N=8$, $k_{\mathrm{tr}}=8/4=2.00\ \mathrm{h}^{-1}$. Its mean is $8/2.00=4.00$ h, its variance is $8/(2.00)^2=2.00\ \mathrm{h}^2$, and its standard deviation is $\sqrt{2.00}=1.41$ h. Its mode is $(8-1)/2.00=3.50$ h.

Both models have the same mean, but the eight-stage response is more concentrated in time and peaks later. The two-stage response is more dispersed and peaks earlier. Therefore, reporting only a mean delay would conceal a substantial difference in predicted timing.

Worked Example: Comparing Impulse-Response Heights

Use the two hypothetical chains from the previous calculation to compare their normalized impulse-response heights at $t=4$ h. Since the area under each impulse response is one, the response height has units of $\mathrm{h}^{-1}$.

For the two-stage chain, substitute $N=2$, $k_{\mathrm{tr}}=0.500\ \mathrm{h}^{-1}$, and $t=4$ h:

h_2(4)=\frac{(0.500)^2(4)^1e^{-(0.500)(4)}}{1!} =1.00e^{-2} =0.1353\ \mathrm{h}^{-1}.

For the eight-stage chain:

h_8(4)=\frac{(2.00)^8(4)^7e^{-(2.00)(4)}}{7!} =\frac{4{,}194{,}304}{5{,}040}e^{-8} \approx 0.2793\ \mathrm{h}^{-1}.

At 4 hours, the eight-stage impulse response is about twice as high. This does not mean that its total response is twice as large: both curves have area one. It means that more of the eight-stage response is concentrated near that time. A model comparison should therefore examine the curve's shape, not interpret a local height as a change in total gain.

Using Shape Information to Choose Among Candidates

A measured mean delay alone determines only the ratio $N/k_{\mathrm{tr}}$. It cannot uniquely determine both parameters. Shape information, such as spread or peak timing, can help distinguish candidate stage counts, provided the data are informative and the assumptions of the chain are reasonable.

Worked Example: Matching Mean and Spread

Suppose a hypothetical, well-characterized pulse-response summary has an estimated mean of 6.00 h and standard deviation of 2.40 h. Treat these as approximate summaries of the normalized response, and compare integer-stage chains with mean fixed at 6.00 h.

For a gamma response, the moment relationship is $\mathrm{CV}=1/\sqrt{N}$. The observed coefficient of variation is $2.40/6.00=0.400$, suggesting $N\approx 1/(0.400)^2=6.25$. This is a diagnostic approximation, not a literal fractional stage count; the model requires an integer $N$.

For $N=6$, mean matching gives $k_{\mathrm{tr}}=6/6.00=1.00\ \mathrm{h}^{-1}$. The predicted standard deviation is $\sqrt{6}/1.00=2.45$ h, and the mode is $5/1.00=5.00$ h.

For $N=7$, mean matching gives $k_{\mathrm{tr}}=7/6.00=1.17\ \mathrm{h}^{-1}$. The predicted standard deviation is $\sqrt{7}/(7/6)=6/\sqrt{7}=2.27$ h, and the mode is $6/(7/6)=5.14$ h. The six-stage model's predicted spread is closer to 2.40 h in this comparison, but the difference is modest. These summaries favor six stages only weakly; they do not establish that the biological system literally contains six observable compartments.

In practice, stage number is a structural choice, while rate is a continuous timing parameter. A useful comparison keeps the mean delay matched across plausible values of $N$, then asks whether observed response onset, peak, spread, and tail favor one shape. If the data cannot distinguish the candidates, a simpler model may be more defensible than a highly specific stage count.

Assumptions, Diagnostics, and Alternatives

The formulas above apply to an identical-stage, linear, time-invariant transit chain with a common first-order rate. They presume a serial path, no branching or loss that changes the normalized response, and a meaningful impulse-response interpretation around the operating conditions being modeled. A gain may scale the output, but it does not alter the normalized timing formulas. If stage rates differ, the response is not this identical-rate gamma form; if transfer between stages is nonlinear or changes over time, the same expressions need not apply.

Useful diagnostics ask whether the chosen structure reproduces the whole time course, rather than just its mean. A predicted peak that is consistently too early or late, an onset that is too abrupt, or a tail that is systematically too long can indicate an unsuitable stage count or a failure of the equal-rate chain assumption. When fitting a model, inspect residuals over time and whether plausible changes in $N$ can be offset by changes in $k_{\mathrm{tr}}$ with little change in overall fit. Sparse sampling around the rise and peak, or substantial measurement noise, can make those shape distinctions difficult to resolve.

The main alternatives have different implications:

Approach	What it represents	Key limitation or distinction
One first-order lag ($N=1$)	A single exponential delay with mean $1/k_{\mathrm{tr}}$	Has its maximum at time zero for an impulse and may be too dispersed or too abrupt for the observed profile.
Identical-stage chain ($N>1$)	A gamma-shaped distribution of transit times	Convenient and interpretable, but imposes equal rates and a specific relationship among mean, spread, and peak.
Unequal-rate transit chain	A serial response with a distinct rate at each stage	Can represent more varied shapes, but requires more parameters and may be harder to distinguish from other structures with limited data.
Fixed or empirical delay	A specified shift in time or a flexible response shape	A fixed shift does not itself represent a spread of transit times; a flexible shape may have less mechanistic interpretation.

These alternatives are not interchangeable merely because they can be tuned to the same average delay. The appropriate choice depends on the available time-course information, the intended interpretation, and whether the additional shape flexibility is supported by data. Stage number should not be presented as a directly observed biological count unless independent evidence supports that interpretation.

Common Pitfall: Treating Mean Delay as the Entire Response

A common mistake is to calculate $T=N/k_{\mathrm{tr}}$, match $T$ across candidate models, and conclude that those models make the same prediction. That reasoning ignores variance and mode: at fixed $T$, the variance is $T^2/N$, and the mode for $N>1$ is $T(1-1/N)$. The correction is to compare the full response shape and to acknowledge when the data support only a mean delay, not a unique stage count and rate.

Key takeaway: In an identical first-order transit chain, $N$ and $k_{\mathrm{tr}}$ jointly determine timing. Matching the mean $N/k_{\mathrm{tr}}$ does not match the response shape; spread and peak timing carry additional information, but may be weakly identified by limited data.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A four-stage chain has $k_{\mathrm{tr}}=0.80\ \mathrm{h}^{-1}$. Calculate its mean transit time and variance.
Two chains have the same mean $T=5$ h, with $N=1$ and $N=5$. What are their respective rates and standard deviations?
For an impulse response with $N=3$ and $k_{\mathrm{tr}}=0.50\ \mathrm{h}^{-1}$, calculate the mode. Explain why the mode is not the mean.
Why can a mean delay estimate alone not determine both $N$ and $k_{\mathrm{tr}}$? Name one type of additional information that could help.

Solutions. (1) The mean is $N/k_{\mathrm{tr}}=4/0.80=5.00$ h. The variance is $N/k_{\mathrm{tr}}^2=4/(0.80)^2=6.25\ \mathrm{h}^2$. (2) For $N=1$, $k_{\mathrm{tr}}=1/5=0.200\ \mathrm{h}^{-1}$ and the standard deviation is $1/0.200=5.00$ h. For $N=5$, $k_{\mathrm{tr}}=5/5=1.00\ \mathrm{h}^{-1}$ and the standard deviation is $\sqrt{5}/1.00=2.24$ h. The means match, but the spreads do not. (3) The mode is $(N-1)/k_{\mathrm{tr}}=2/0.50=4.00$ h. The mean is $N/k_{\mathrm{tr}}=6.00$ h; the gamma response is right-skewed, so its peak occurs before its mean. (4) A mean gives only the ratio $N/k_{\mathrm{tr}}$, allowing multiple stage-and-rate pairs. Information about response spread, peak timing, or the time-course shape can help distinguish candidates, although noisy or sparse data may still leave them practically confounded.

Approach	What it represents	Key limitation or distinction
One first-order lag (\(N=1\))	A single exponential delay with mean \(1/k_{\mathrm{tr}}\)	Has its maximum at time zero for an impulse and may be too dispersed or too abrupt for the observed profile.
Identical-stage chain (\(N>1\))	A gamma-shaped distribution of transit times	Convenient and interpretable, but imposes equal rates and a specific relationship among mean, spread, and peak.
Unequal-rate transit chain	A serial response with a distinct rate at each stage	Can represent more varied shapes, but requires more parameters and may be harder to distinguish from other structures with limited data.
Fixed or empirical delay	A specified shift in time or a flexible response shape	A fixed shift does not itself represent a spread of transit times; a flexible shape may have less mechanistic interpretation.