Why Model Progression Exponentially?
A linear trajectory represents a constant absolute change per unit time. But some disease measures may change at a rate that depends on their current level: a larger burden can correspond to a larger absolute increase, while a smaller remaining amount can correspond to a smaller absolute decline. An exponential trajectory represents this pattern with a constant proportional rate. The distinction matters because a constant proportional rate does not mean a constant number of response units gained or lost each week.
Here, \(Y(t)\) denotes the modeled untreated response at time \(t\), and \(Y_0\) denotes its value at the chosen time origin \(t=0\). The response could be a measure of disease burden, so increasing values indicate worsening, or a measure of function, so decreasing values might indicate worsening. The direction of clinical worsening must therefore be determined from the endpoint, not from the sign of the rate alone.
From Proportional Rate to Trajectory
The exponential model follows from the assumption that the instantaneous change in response is proportional to the current response. For a constant \(k\), this assumption is written as a differential equation. It is the model’s defining assumption, not a conclusion that can be inferred merely because two measurements are available.
To solve it when \(Y(t)>0\), divide both sides by \(Y(t)\) and integrate from time zero to time \(t\). The initial condition sets the integration constant.
This solution requires a constant proportional rate over the modeled time interval. It also requires a meaningful positive response scale for the logarithmic derivation and ratio interpretation. If the measured response can be zero or negative, or if ratios are not meaningful on its scale, this formulation may not be appropriate without a scientifically justified transformation. The model’s curve is smooth and never crosses zero when \(Y_0>0\).
For two positive modeled responses \(Y(t_1)\) and \(Y(t_2)\), the same equation gives \(k=\ln\!\big(Y(t_2)/Y(t_1)\big)/(t_2-t_1)\), provided the exponential assumption applies over that interval. When \(k>0\), the doubling time is the elapsed time for the response to multiply by two. When \(k<0\), the halving time is the elapsed time for it to multiply by one-half. These times follow by solving for the elapsed time in the corresponding response ratio.
Worked Example: Increasing Disease Burden
Hypothetical data. Suppose an untreated disease-burden measure is 80 units at time zero and 120 units 10 weeks later. Assume a constant exponential rate over these 10 weeks. Find the rate constant, doubling time, and predicted burden at 20 weeks.
The ratio after 10 weeks is \(120/80=1.5\). Thus, the rate estimate under the model is \(k=\ln(120/80)/10=\ln(1.5)/10=0.0405465\ \text{week}^{-1}\), or approximately \(0.04055\ \text{week}^{-1}\).
The doubling time is \(t_{\mathrm{double}}=\ln(2)/0.0405465=17.09\) weeks. This means that the modeled burden doubles in about 17.09 weeks if the same rate continues; it does not mean that 17.09 weeks is a guaranteed time to a clinical event.
At 20 weeks, the trajectory predicts \(Y(20)=80e^{(0.0405465\ \text{week}^{-1})(20\ \text{weeks})}\approx 80e^{0.81093}\approx 180\) units. The units cancel in the exponent, leaving the prediction in response units.
The instantaneous absolute rate at time zero is \(kY_0=(0.0405465)(80)=3.244\) units per week. At 10 weeks it is \(kY(10)=(0.0405465)(120)=4.866\) units per week. The proportional rate stays constant, while the absolute rate rises with the modeled burden.
Worked Example: Declining Response and Halving Time
Hypothetical data. An untreated measure of remaining functional capacity is 100 units initially and 60 units after 6 weeks. Assume exponential decline over this interval. Calculate \(k\), the halving time, and the prediction at 12 weeks.
The rate is \(k=\ln(60/100)/6=\ln(0.60)/6=-0.0851376\ \text{week}^{-1}\), approximately \(-0.08514\ \text{week}^{-1}\). The negative sign indicates decline in this response measure.
The halving time is \(t_{\mathrm{half}}=\ln(2)/|k|=0.693147/0.0851376=8.142\) weeks. This is the time for the modeled response to fall to half its current value, conditional on the constant-rate assumption.
For 12 weeks, it is useful to retain the exact ratio \(0.60\) rather than rely on a rounded rate: \(Y(12)=100(0.60)^{12/6}=100(0.60)^2=100(0.36)=36.00\) units. Equivalently, using the unrounded rate gives \(100e^{(\ln(0.60)/6)(12)}=36.00\) units. The prediction is consistent with the two-interval ratio: the response falls by a factor of 0.60 in each 6-week interval.
How Exponential and Linear Predictions Differ
The linear disease-progression model has a constant absolute slope: \(Y(t)=Y_0+\beta t\), where \(\beta\) is measured in response units per time. By contrast, the exponential model has constant fractional change and an absolute slope that varies with response level. Both can describe change from the same baseline, but their predictions between observed times need not agree.
Worked Example: Comparing Trajectories Between Two Measurements
Hypothetical data. A response is 50 units at time zero and 90 units at week 10. Compare the linear and exponential predictions at week 5, using each model to match both endpoint values.
For the linear model, \(\beta=(90-50)/10=4.000\ \text{units per week}\). Therefore, \(Y(5)=50+(4.000)(5)=70.0\) units.
For the exponential model, \(k=\ln(90/50)/10=\ln(1.8)/10=0.058778\ \text{week}^{-1}\). Its week-5 prediction is \(Y(5)=50e^{(0.058778)(5)}=50\sqrt{1.8}\approx67.08\) units.
The models give different intermediate predictions even though each matches the baseline and week-10 response. The exponential trajectory rises more slowly at first and then more rapidly: its initial absolute rate is \(kY_0=(0.058778)(50)=2.939\) units per week, while its rate at week 10 is \((0.058778)(90)=5.290\) units per week. Two endpoints alone cannot establish which functional form is better.
| Feature | Linear trajectory | Exponential trajectory |
|---|---|---|
| Constant quantity | Absolute change per unit time | Fractional change per unit time |
| Rate parameter units | Response units per time | Inverse time |
| Response at time zero | \(Y_0\) | \(Y_0\) |
| Typical shape when increasing | Straight line | Curves upward when \(Y_0>0\) and \(k>0\) |
| Important limitation | Can eventually predict negative values during decline | With \(Y_0>0\), cannot reach or cross zero in finite time |
Assumptions, Diagnostics, and Limitations
The exponential trajectory is a candidate model for an untreated or otherwise specified background response; it is not automatically a complete model for what is observed in a clinical trial. The response may also reflect baseline, placebo response, measurement error, and active treatment. In particular, a trend in a control arm should not be called disease progression without considering the other components of response.
Useful checks include plotting observed responses over time with the predicted curve and examining residuals—the observed values minus the model predictions—against time and predicted response. Systematic curvature in residuals, a rate that appears to change across intervals, or persistent underprediction at one end of the time range suggests that a constant proportional rate may be inadequate. Replicate measurements and uncertainty matter: a rate calculated from two measurements summarizes those endpoints under the model, but does not by itself verify the model or show how uncertain the rate is.
An exponential trajectory can produce implausibly large or small long-term predictions because its proportional rate continues indefinitely. A simple linear model may be preferable over a restricted interval if a constant absolute rate is scientifically defensible and predictions remain reasonable. Neither model should be extrapolated beyond the time range where its assumptions are credible. The subsequent treatment of saturating trajectories addresses a different shape; the key comparison here is the distinction between constant absolute and constant proportional change.
For model interpretation, always state the response scale, direction of change, time unit, and interval over which the rate is assumed constant. A positive \(k\) means that \(Y(t)\) increases; whether that means improvement or worsening depends on what \(Y\) measures. A negative \(k\) means that the response declines, not necessarily that disease severity declines.
Check Your Understanding
Try each question, then check the worked solutions below.
- A positive response has \(k=0.0300\ \text{day}^{-1}\). What is its doubling time?
- A response declines from 40 units at time zero to 25 units after 5 weeks. Under an exponential assumption, calculate \(k\) and predict the response at week 10.
- In one sentence, distinguish the parameter in a linear trajectory from the rate constant in an exponential trajectory.
- Why do two positive endpoint measurements not establish that exponential progression is preferable to linear progression?
Solutions. (1) \(t_{\mathrm{double}}=\ln(2)/0.0300=23.10\) days. (2) \(k=\ln(25/40)/5=\ln(0.625)/5=-0.09400\ \text{week}^{-1}\); at week 10, \(Y(10)=40e^{(-0.09400)(10)}=40(0.625)^2=15.625\), approximately 15.63 units. (3) A linear model’s slope is a constant absolute change in response units per time, whereas \(k\) is a constant fractional change per time and the exponential model’s absolute rate is \(kY(t)\). (4) Both models can be parameterized to match the same two endpoints but make different intermediate predictions, so the endpoints alone cannot assess their relative fit or the plausibility of their assumptions.