Why a Correct Baseline May Not Determine the Rates
An indirect response model can match a baseline observation exactly while leaving its underlying production and loss rates uncertain. At steady state, production balances loss. The observed baseline therefore constrains a ratio of rates, not necessarily either rate on its own. This matters because two models with the same baseline can predict different speeds of response and recovery after a perturbation.
Consider a response \(R(t)\) with constant production rate \(k_{\mathrm{in}}\) and first-order loss rate \(k_{\mathrm{out}}R(t)\). The units of \(k_{\mathrm{in}}\) are response units per time; \(k_{\mathrm{out}}\) has units of inverse time. Assume \(k_{\mathrm{in}}>0\), \(k_{\mathrm{out}}>0\), and that the baseline is at steady state.
This is a structural property of the baseline equation, not merely a consequence of noisy measurements. Even if \(R_0\) were measured without error, the equation \(k_{\mathrm{in}}=R_0k_{\mathrm{out}}\) admits infinitely many positive solutions. More precise baseline measurement narrows uncertainty about the ratio but does not, by itself, separate the rates.
How the Rescaling Preserves Baseline Predictions
At baseline, the differential equation is \(dR/dt=k_{\mathrm{in}}-k_{\mathrm{out}}R\). At equilibrium, its derivative is zero, yielding \(R_0=k_{\mathrm{in}}/k_{\mathrm{out}}\). Under the proportional rescaling,
The rates have not been left unchanged: both have been multiplied by \(c\), with time units held fixed. The new model turns over faster when \(c>1\) and slower when \(0<c<1\). Baseline observations alone cannot reveal this change because the steady-state balance remains exact.
Worked Example: Two Rate Pairs, One Baseline
Suppose a hypothetical response has a baseline of \(12\) response units. One candidate model uses \(k_{\mathrm{out}}=0.20\ \mathrm{h}^{-1}\). Find the corresponding production rate and a second rate pair that preserves the baseline after multiplying both rates by 3.
From the steady-state relation, \(k_{\mathrm{in}}=R_0k_{\mathrm{out}}\). Substituting the values gives
The original pair is therefore \((2.40\ \text{response units per hour},\,0.20\ \mathrm{h}^{-1})\). Multiplying both rates by 3 gives \(k_{\mathrm{in}}=7.20\ \text{response units per hour}\) and \(k_{\mathrm{out}}=0.60\ \mathrm{h}^{-1}\). Its predicted baseline is
Thus the baseline is identical, although the second model has a threefold higher turnover rate. A baseline observation cannot choose between these parameter pairs.
Why Response-Time Observations Can Break the Confounding
To see what additional information a time course provides, suppose the production rate is multiplied by a known, time-varying input function \(f(t)\). This is a deliberately specified model for the argument: \(f(t)=1\) means baseline production, and a value such as \(f(t)=0.5\) means production is half its baseline rate. Assume the response is continuous through a change in \(f(t)\), and begins at the baseline \(R(0)=R_0\).
Define the normalized response \(r(t)=R(t)/R_0\). Substituting \(k_{\mathrm{in}}=k_{\mathrm{out}}R_0\) and dividing the differential equation by \(R_0\) gives
The normalized trajectory depends on \(k_{\mathrm{out}}\), even though the baseline ratio does not. This is the central reason response-time data can distinguish baseline-equivalent rate pairs. The production rate can then be recovered as \(k_{\mathrm{in}}=R_0k_{\mathrm{out}}\), provided the model assumptions hold and \(R_0\) is known.
For a constant known input multiplier \(f(t)=a\) after time zero, with \(r(0)=1\), the solution is
When \(a\ne 1\), a response measured at an informative time can, in the ideal noise-free setting, determine \(k_{\mathrm{out}}\):
This inversion requires a positive logarithm argument and a response consistent with the assumed trajectory. It also requires that the input multiplier and time be known. With noisy data, a single measured response generally gives an uncertain estimate; multiple appropriately timed observations help assess the trajectory as well as estimate its rate.
Derivation and Conditions for Information in the Time Course
The constant-input solution follows by subtracting the equilibrium response \(aR_0\). Let \(Z(t)=R(t)-aR_0\). Since \(k_{\mathrm{in}}=k_{\mathrm{out}}R_0\),
Separating variables and integrating from zero to \(t\) yields \(\ln[Z(t)/Z(0)]=-k_{\mathrm{out}}t\), or \(Z(t)=Z(0)e^{-k_{\mathrm{out}}t}\). Since \(R(0)=R_0\), \(Z(0)=(1-a)R_0\). Dividing \(R(t)=aR_0+Z(t)\) by \(R_0\) gives the stated normalized solution.
The solution also shows why not every response-time observation is equally useful. If \(a=1\), the system remains at baseline and \(r(t)=1\) for every value of \(k_{\mathrm{out}}\). If the observation is taken at \(t=0\), the response is still \(r(0)=1\), independent of the rate. For \(a\ne1\), a finite-time observation away from the initial instant can carry information, but observations very close to baseline may be hard to distinguish in the presence of measurement error.
For a general known input function, the integrating-factor solution makes the same dependence explicit:
A response measured while the input is changing can therefore carry information about the rate through both the exponential weighting and the integral. The interpretation relies on the assumed input function being known and the model's rate form being adequate. If the input history is unknown, a mismatch between the assumed and actual forcing can be mistaken for a different turnover rate.
Worked Example: Estimating Turnover From a Known Input Change
In a hypothetical experiment, the response begins at \(R_0=10\ \mathrm{mg/L}\). A known intervention changes the production multiplier to \(a=0.5\), and the response at \(t=2\ \mathrm{h}\) is \(R=7.247\ \mathrm{mg/L}\). Under the constant-input model, estimate \(k_{\mathrm{out}}\).
First normalize the observation: \(r(2)=7.247/10=0.7247\). Substituting \(a=0.5\), \(t=2\ \mathrm{h}\), and this response into the inversion gives
The implied production rate is \(k_{\mathrm{in}}=R_0k_{\mathrm{out}}=(10\ \mathrm{mg/L})(0.400\ \mathrm{h}^{-1})=4.00\ \mathrm{mg/L/h}\). This inference is conditional on the known input multiplier, baseline equilibrium, and first-order loss assumptions. It is not obtained from the baseline alone.
The contrast between baseline-equivalent models can be checked directly. With \(k_{\mathrm{out}}=0.400\ \mathrm{h}^{-1}\), the predicted normalized response is \(0.5+0.5e^{-0.8}=0.7247\). If both rates had instead been multiplied by 3, then \(k_{\mathrm{out}}=1.20\ \mathrm{h}^{-1}\), and the prediction would be \(0.5+0.5e^{-2.4}=0.5454\), or \(5.454\ \mathrm{mg/L}\). The two models share a baseline of \(10\ \mathrm{mg/L}\) but not this response-time prediction.
Washout Observations and Sampling-Time Limitations
A washout segment provides another direct illustration. If the response is \(r(t_0)\) at the start of washout and the input returns to baseline, \(f(t)=1\), then
When \(r(t_0)\ne1\), the distance from baseline decays exponentially. The rate is informed by how quickly that distance shrinks over elapsed time, not merely by the fact that the response eventually returns to baseline. A measurement taken after nearly complete recovery may be close to 1 under a range of plausible rates, so it can be weakly informative in noisy data. If washout begins exactly at baseline, it provides no rate information in this model.
Worked Example: Recovering the Rate From Two Washout Times
Suppose a hypothetical response has \(R_0=8\ \mathrm{mg/L}\), and its normalized value is \(r(0)=1.6\) at the start of washout. The response is \(r(1)=1.4445\) after \(1\ \mathrm{h}\) and \(r(4)=1.1807\) after \(4\ \mathrm{h}\). Assume the input has returned to baseline and first-order loss applies. Estimate \(k_{\mathrm{out}}\) using the ratio of deviations from baseline.
The deviations are \(r(1)-1=0.4445\) and \(r(4)-1=0.1807\). Over the \(3\ \mathrm{h}\) interval, the washout solution gives
The estimated production rate is \((8\ \mathrm{mg/L})(0.300\ \mathrm{h}^{-1})=2.40\ \mathrm{mg/L/h}\). As a check, the model predicts \(r(1)=1+0.6e^{-0.3}=1.4445\) and \(r(4)=1+0.6e^{-1.2}=1.1807\), matching the stated values to rounding.
Diagnostics, Alternatives, and Common Pitfalls
The baseline-only model and a model informed by response-time data answer different questions. Baseline data estimate \(R_0\), or equivalently the ratio \(k_{\mathrm{in}}/k_{\mathrm{out}}\). A response trajectory under a known perturbation can inform the turnover scale \(k_{\mathrm{out}}\), after which \(k_{\mathrm{in}}\) follows from the baseline relation. If suitable response-time data are unavailable, an independently justified value for one rate can be used to calculate the other, but that is an added source of information, not something the baseline itself established.
In practice, inspect whether the observations capture a meaningful departure from baseline and its subsequent evolution. If estimates of \(k_{\mathrm{out}}\) change substantially when one time point is removed, or many distinct rate pairs fit nearly equally well, the rate scale may be weakly informed. Systematic departures from the predicted exponential approach or recovery can indicate that the assumed input history, first-order loss, baseline equilibrium, or measurement model is inadequate. A flexible numerical fit cannot repair a structurally uninformative design.
- Structural confounding: baseline-only predictions are exactly unchanged by proportional rescaling of both rates.
- Practical weak information: response-time data theoretically depend on the rate, but noise, timing, or a small response departure makes plausible rates difficult to distinguish.
- Model failure: observed trajectories conflict systematically with the assumed response equation; interpreting the fitted rate as a genuine turnover rate may then be misleading.
A common mistake is to argue that a precisely measured baseline identifies both \(k_{\mathrm{in}}\) and \(k_{\mathrm{out}}\). Precision does not remove the structural confounding: for every positive \(k_{\mathrm{out}}\), choosing \(k_{\mathrm{in}}=R_0k_{\mathrm{out}}\) gives the same baseline. The correction is to use observations that contain time-course information under a specified input, or to state explicitly which rate is fixed using independent information.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A model has \(R_0=6\ \mathrm{mg/L}\) and \(k_{\mathrm{out}}=0.10\ \mathrm{h}^{-1}\). What is \(k_{\mathrm{in}}\), and what happens to the baseline if both rates are multiplied by 4?
- Under a known constant input multiplier \(a=0.4\), a normalized response is measured at \(r(t)=0.7\) after \(t=2\ \mathrm{h}\). Calculate \(k_{\mathrm{out}}\).
- Why does observing a response at baseline when \(f(t)=1\) provide no information about \(k_{\mathrm{out}}\) in this model?
- During washout, the response starts at baseline. Can its later baseline values identify the loss rate under the stated model? Explain.
Solutions. (1) \(k_{\mathrm{in}}=R_0k_{\mathrm{out}}=(6\ \mathrm{mg/L})(0.10\ \mathrm{h}^{-1})=0.600\ \mathrm{mg/L/h}\). Multiplying both rates by 4 gives \(2.40\ \mathrm{mg/L/h}\) and \(0.400\ \mathrm{h}^{-1}\); their ratio remains \(6\ \mathrm{mg/L}\). (2) Use \(k_{\mathrm{out}}=-(1/t)\ln[(r-a)/(1-a)]\). Thus \(k_{\mathrm{out}}=-(1/2\ \mathrm{h})\ln[(0.7-0.4)/(1-0.4)]=-(1/2\ \mathrm{h})\ln(0.5)\approx0.347\ \mathrm{h}^{-1}\). (3) With \(r=1\) and \(f=1\), \(dr/dt=k_{\mathrm{out}}(1-1)=0\) for every positive \(k_{\mathrm{out}}\); the predicted response is constant. (4) No. Starting at baseline during washout gives \(r(t_0)=1\), so the washout solution remains \(r(t)=1\) for every \(k_{\mathrm{out}}\). There is no deviation whose decay could reveal the rate.