Turnover-Rate Estimation With Sparse Biomarker Sampling

Why Sparse Sampling Makes Turnover Estimation Difficult

A biomarker may be measured only at baseline, once during treatment, and once after treatment stops. Those observations can show a response and recovery, but they do not necessarily determine how quickly the biomarker turns over. Different combinations of turnover rate and treatment effect can produce similar values at the observed times, especially when the observations are noisy or the washout period is short.

This tutorial considers a simple first-order model to make the information in such a design explicit. The model is a diagnostic tool for asking whether the data constrain the rate; it is not a claim that every biomarker follows first-order turnover. In the notation of earlier material on analytic turnover responses to step changes, the turnover rate is $k$, with units of inverse time. The corresponding half-life is $\ln(2)/k$.

Model and assumptions: Let $R(t)$ be the biomarker concentration and $R_0$ its known baseline steady-state concentration. At treatment start, $R(0)=R_0$. During a constant treatment effect, production is a fixed fraction $q$ of baseline production, where $0\leq q\leq1$; loss remains first order with rate $k>0$. After treatment stops, production returns immediately to baseline and the same loss rate applies. Thus, $$ \frac{dR}{dt}=k(qR_0-R) \quad \text{during treatment}, \qquad \frac{dR}{dt}=k(R_0-R) \quad \text{during washout}. $$ These equations assume a stable baseline, a constant treatment effect over the treatment interval, no residual drug effect during washout, and no additional delay, feedback, or time-varying input.

If treatment lasts $T$ time units and the biomarker is observed at treatment end, the predicted treatment value is

R_T=R_0\left[q+(1-q)e^{-kT}\right].

At $s$ time units after withdrawal, the washout prediction is

R_W(s)=R_0+(R_T-R_0)e^{-ks}.

The first equation shows the central estimation problem: an observed treatment value can often be matched by several $k,q$ pairs. A washout observation adds information about $k$, because recovery depends on the decay factor $e^{-ks}$. But if washout is brief, or the biomarker is already close to baseline, that additional information may be weak.

Profile Likelihood for the Turnover Rate

Assume independent normally distributed measurement errors with known standard deviations. For observations $y_i$, model predictions $m_i(k,q)$, and standard deviations $\sigma_i$, the weighted sum of squared residuals is

S(k,q)=\sum_i\left(\frac{y_i-m_i(k,q)}{\sigma_i}\right)^2.

For each candidate $k$, minimize $S(k,q)$ over the allowed $q$ values. This gives the profile objective $S_p(k)=\min_{0\leq q\leq1}S(k,q)$. Profiling asks how well each rate can fit the data after allowing the nuisance parameter $q$ its best possible value. With Gaussian errors and fixed standard deviations, differences in this objective are differences in minus twice the log likelihood.

Profile-likelihood interval: A conventional approximate 95% interval for one profiled parameter contains the values of $k$ satisfying $S_p(k)-S_{\min}\leq3.84$, where $S_{\min}$ is the minimum over all allowed $k,q$. This chi-square cutoff is an approximation; boundary parameters, small samples, or uncertain error assumptions may require calibration by simulation.

The constraint $q\geq0$ matters. If a candidate $k$ would require negative production to match a low treatment value, that $q$ is inadmissible. The best fit at that rate then occurs at the boundary $q=0$, and the treatment observation contributes a residual rather than being fitted exactly. Ignoring that residual can make the lower end of a profile interval too high or otherwise distort the interval.

Worked Example: Profiling a Rate With One Treatment and One Washout Observation

Consider hypothetical normalized data with $R_0=1.00$, a 10-day treatment, and a washout observation 1.18 days after withdrawal. The treatment-end observation is $y_T=0.490$, with standard deviation $0.005$. The washout observation is $y_W=0.5896$, with standard deviation $0.030$. Assume the model and independent normal errors stated above.

For a candidate rate $k$, the treatment prediction is $m_T=q+(1-q)e^{-10k}$. When the unconstrained value of $q$ is admissible, setting $m_T=y_T$ gives

q(k)=\frac{y_T-e^{-10k}}{1-e^{-10k}}.

At $k=0.072\ \mathrm{day}^{-1}$, $e^{-10k}=e^{-0.72}=0.4868$, so

q(0.072)=\frac{0.490-0.4868}{1-0.4868}=0.0063.

This value satisfies $0\leq q\leq1$, so the treatment observation can be fitted. The predicted washout value is

m_W=1+(0.490-1)e^{-(0.072)(1.18)} =1-0.510e^{-0.08496} \approx0.5315.

The washout residual is approximately $0.5896-0.5315=0.0581$. If the treatment is fitted exactly, the increase in weighted squared residuals is

\left(\frac{0.0581}{0.030}\right)^2\approx3.75.

Thus $k=0.072\ \mathrm{day}^{-1}$ is inside the conventional profile cutoff of 3.84; it is not the lower endpoint. Profiling over the constrained $q$ values, including the treatment residual when $q$ reaches zero, gives an approximate interval of $0.071$ to $0.315\ \mathrm{day}^{-1}$ for these hypothetical data. The rate at $0.072$ is compatible with the data, but the broad interval shows that the sparse observations do not pin down a precise rate. The corresponding half-lives across this interval are approximately 9.76 to 2.20 days, using $t_{1/2}=\ln(2)/k$.

The boundary can be checked directly. Setting $q=0$ makes the treatment prediction $e^{-10k}$; the rate at which this equals 0.490 is $k=-\ln(0.490)/10\approx0.07134\ \mathrm{day}^{-1}$. Just below that rate, the treatment prediction cannot be reduced further by choosing an allowed $q$, so its residual must be included in the profile objective. This is why the lower endpoint can lie slightly below 0.072.

Why Washout Can Help—and When It Does Not

If the baseline is known and two washout measurements are available, the model gives a simple rate calculation. Let $D(s)=R_0-R_W(s)$ be the remaining distance from baseline during recovery. Then $D(s)=D(0)e^{-ks}$. For two times $s_1<s_2$, provided both distances are nonzero and have the same sign,

k=-\frac{\ln\left[D(s_2)/D(s_1)\right]}{s_2-s_1}.

This calculation is useful as a transparent check, but it is sensitive to measurement error when either distance is close to zero. It also relies on the same washout assumptions as the model: in particular, the baseline must be known and recovery must follow the specified first-order process.

Worked Example: A Two-Point Washout Check

Suppose a hypothetical biomarker has known baseline $R_0=1.00$, is $0.600$ when treatment stops, and is measured at $0.683$ after 1 day and $0.841$ after 4 days. These values are consistent, to rounding, with a washout rate of $\ln(2)/3\ \mathrm{day}^{-1}$: the distance from baseline is halved over the 3-day interval.

Using the two observations directly, $D(1)=1.00-0.683=0.317$ and $D(4)=1.00-0.841=0.159$. Therefore,

\hat{k}=-\frac{\ln(0.159/0.317)}{4-1} \approx0.230\ \mathrm{day}^{-1}.

The implied half-life is $\ln(2)/0.230\approx3.01$ days. The calculation is reasonable under the stated assumptions, but two measurements alone provide no robust assessment of whether the exponential form fits the recovery profile. Additional sampling can reveal curvature, delayed recovery, or measurement values inconsistent with this model.

A washout observation adds little rate information if the biomarker is near baseline: the predicted change with $k$ is then small. It may also add little if taken very soon after withdrawal, because $e^{-ks}$ remains close to 1 for a wide range of plausible rates. A useful design therefore samples when recovery is expected to be appreciable, while retaining enough observations to assess whether the assumed recovery shape is plausible.

Simulation as an Identifiability Check

A fitted value and a narrow local standard error can be misleading when the likelihood is asymmetric, constrained by $q$, or nearly flat over a wide range of $k$. Simulation provides a direct way to ask whether the proposed sampling schedule would recover rates reliably under the assumed model. Choose plausible $k,q$, calculate the expected observations at the actual sampling times, add measurement errors with the assumed distribution, and refit the model to each simulated data set. Examine the spread of fitted rates, the frequency of boundary estimates, and the profile intervals. This assesses the design conditional on the model and error assumptions; it cannot establish that those assumptions are true.

Worked Example: Comparing Rate Sensitivity at a Sparse Washout Time

Return to the hypothetical profile example and suppose the treatment-end value is fitted at 0.490. At a washout time of 1.18 days, the predicted values under $k=0.072$ and $k=0.315\ \mathrm{day}^{-1}$ are approximately 0.5315 and 0.6483. Their difference is about 0.117, roughly 3.9 times the washout standard deviation of 0.030. Yet the observation lies between these predictions, and the fitted treatment effect can adjust with $k$. The profile interval is therefore a more complete measure of rate information than comparing predictions alone.

A simulation check would generate treatment and washout values around their model predictions with standard deviations 0.005 and 0.030, refit $k,q$ for each replicate, and record the resulting profiles. If many replicates yield broad intervals or estimates at $q=0$, the design is weak for estimating $k$, even if one data set happens to produce a single best-fitting rate. If the simulated fits are consistently concentrated near the generating rate and their intervals have appropriate coverage, the schedule is more informative under the assumed model.

For reproducibility, use a fixed random-number seed, retain the same sampling times and error model in every replicate, and report how often fitting fails or reaches a parameter boundary. A simulation that changes the assumed error or model structure at the same time as the sampling schedule does not isolate which design feature caused the result.

Diagnostics, Alternatives, and Limitations

Inspect the profile across the full plausible rate range rather than relying only on an optimizer’s reported estimate. A broad, shallow profile signals weak information; a long tail indicates that some rates remain compatible with the observations. A sharp bend or truncation can indicate a parameter constraint. Also check residuals by phase and time. Systematic residuals during washout may indicate that the assumed single first-order recovery is inadequate, rather than simply showing an imprecise $k$.

Approach	What it provides	Main caution with sparse data
Washout-only log-ratio calculation	A direct rate estimate from two distances to a known baseline	Unstable when either distance is small; does not use treatment data or diagnose the full model
Nonlinear least-squares point fit	A best-fitting parameter combination	A point estimate alone does not show whether other rates fit nearly as well
Local standard error from curvature	A local approximation to uncertainty near the optimum	May be unreliable for a flat, asymmetric, or boundary-limited profile
Profile likelihood and simulation	Rate-specific compatibility and design-based recovery behavior	Results depend on the model, error assumptions, and appropriate likelihood cutoff

A common mistake is to report a precise turnover rate because an optimizer converged to several decimal places. Convergence means the algorithm found a numerical optimum; it does not mean the data distinguish that value from alternatives. The correction is to examine the profile and, when practical, simulate the proposed design. Another mistake is to interpret a washout decrease from a value equal to baseline as evidence for a positive rate. Under this model, if withdrawal starts exactly at $R_0$, then $R_W(s)=R_0$ for every $s$; a subsequent fall cannot estimate $k$ and signals inconsistency with the assumptions or data variability.

Finally, the profile interval is conditional on the model structure, known baseline, fixed treatment multiplier, and specified measurement errors. A broad interval cannot be repaired by reporting more digits, and a narrow interval does not protect against a misspecified model. Sparse sampling is most defensible when the model is scientifically plausible, the sampling times create measurable change, and the uncertainty assessment reflects parameter constraints.

Key takeaway: Sparse baseline, treatment, and washout data can produce a numerical turnover-rate estimate without strongly identifying that rate. Profile the rate while respecting parameter bounds, inspect the entire profile, and use simulation to test whether the actual sampling design can recover rates under the assumed model.

Check Your Understanding

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

If the estimated turnover rate is $0.20\ \mathrm{day}^{-1}$, what is the implied half-life?
For the treatment model with $R_0=1$, $T=5$ days, $k=0.10\ \mathrm{day}^{-1}$, and $q=0.20$, calculate the treatment-end prediction.
At a candidate rate, the best constrained fit has $q=0$. Which observations must be included in the profile objective, and why?
Why can a washout observation taken very soon after withdrawal be weak evidence about $k$?
A biomarker starts washout at its baseline value of 1.00 and is observed at 0.80 later. Can this estimate $k$ under the stated washout model? Explain.

Solutions. (1) $t_{1/2}=\ln(2)/0.20=3.47$ days. (2) $R_T=0.20+0.80e^{-0.50}=0.20+0.80(0.6065)=0.6852$, in the same concentration units as baseline. (3) Include both treatment and washout residuals, each divided by its standard deviation and squared. At $q=0$, the treatment value may not be fitted exactly, so omitting its residual would overstate the fit. (4) For small $s$, $e^{-ks}$ is close to 1 across a range of rates, so predicted recovery changes little with $k$. (5) No. With $R_W(0)=R_0$, the model predicts $R_W(s)=R_0+(R_0-R_0)e^{-ks}=1.00$ for all $s$. A fall to 0.80 is inconsistent with this washout model; the baseline and assumptions should be checked rather than treating that fall as a rate estimate.