Linear Disease-Progression Trajectories

Why Represent Untreated Progression?

A response can change over time even when no active drug is given. For example, a disease may gradually worsen, gradually improve, or remain approximately stable. A disease-progression component represents this background time trend so it is not automatically attributed to placebo or treatment. Here, the focus is the simplest changing trajectory: a response that changes by a constant absolute amount per unit of time.

The trajectory describes the expected response under the specified untreated natural-history condition. It does not, by itself, describe a placebo response or an active-drug effect. The response scale and the direction of clinical benefit must be stated, because a falling value can mean improvement for one endpoint and worsening for another.

Definition: A linear untreated disease-progression trajectory is an expected response that changes at a constant absolute rate over time. Let $t$ denote time, $B$ the expected response at the reference time $t=0$, and $k$ the slope. The trajectory is $$ \mu_{\mathrm{NH}}(t)=B+kt. $$ Here, $\mu_{\mathrm{NH}}(t)$ is the expected response under the modeled untreated natural-history condition. If response is measured in units $U$ and time in units $T$, then $B$ has units $U$ and $k$ has units $U/T$.

Meaning and Direction of the Slope

The slope $k$ is the expected change in response for each one-unit increase in time. If $k=0$, the expected response is constant. If $k>0$, the response increases; if $k<0$, it decreases. These signs describe numerical direction, not clinical benefit or harm by themselves. For a symptom score where higher values indicate worse symptoms, $k>0$ indicates worsening and $k<0$ indicates improvement. For a functioning score where higher values indicate better function, the clinical interpretation is reversed.

For any two times $t_1$ and $t_2$, the expected change is

\mu_{\mathrm{NH}}(t_2)-\mu_{\mathrm{NH}}(t_1)=k(t_2-t_1).

This equation makes clear that the slope applies to a time interval, not just to the endpoint. For example, a slope of $0.4$ score points per month implies an expected increase of $0.4\times 6=2.4$ points over six months, provided the linear approximation is appropriate throughout that interval.

Worked Example: Worsening on a Higher-Is-Worse Scale

Suppose a hypothetical untreated disease has a symptom score for which higher values indicate worse symptoms. At the reference time, the expected score is $B=12$ points, and the estimated natural-history slope is $k=0.6$ points per month. Calculate the expected score at 3 and 8 months and describe the change.

Model. The linear trajectory is $$ \mu_{\mathrm{NH}}(t)=12+0.6t, $$ where $t$ is measured in months and the response is measured in score points. The slope is positive, so the expected score worsens over time on this scale.

At 3 months: $$ \mu_{\mathrm{NH}}(3)=12+(0.6\ \text{points/month})(3\ \text{months})=12+1.8=13.8\ \text{points}. $$

At 8 months: $$ \mu_{\mathrm{NH}}(8)=12+(0.6\ \text{points/month})(8\ \text{months})=12+4.8=16.8\ \text{points}. $$

The expected increase from month 3 to month 8 is $0.6(8-3)=3.0$ points. These are model-based expected responses, not claims that every individual’s observed score will follow the line exactly.

Why the Linear Form Represents a Constant Rate

The linear form follows directly from the assumption that the expected response changes at a constant rate. In differential notation, that assumption is

\frac{d\mu_{\mathrm{NH}}(t)}{dt}=k.

Integrating this constant derivative from the reference time $0$ to time $t$ gives

\mu_{\mathrm{NH}}(t)-\mu_{\mathrm{NH}}(0)=\int_0^t k\,du=kt.

Writing $\mu_{\mathrm{NH}}(0)=B$ yields $\mu_{\mathrm{NH}}(t)=B+kt$. Equivalently, for a discrete set of equally spaced visits, the assumption means that the expected difference between successive visits is the same when the interval length is the same. It is the constancy of the absolute change—not a constant percentage change—that defines this model.

In an analysis, observed responses may vary around the expected trajectory. One simple observation model is $Y_{ij}=\mu_{\mathrm{NH}}(t_{ij})+\epsilon_{ij}$, where $Y_{ij}$ is an observed response for individual $i$ at time $t_{ij}$, and $\epsilon_{ij}$ is deviation from the expected trajectory. This equation separates the population-level trajectory from individual observations; it does not imply that the deviations are independent, identically distributed, or normally distributed. Those assumptions depend on the observation and sampling model being used.

Worked Example: Improvement on a Higher-Is-Better Scale

Consider a hypothetical untreated condition measured with a functioning score, where higher values indicate better function. At the reference time, expected function is $B=58$ points. The untreated slope is $k=1.2$ points per week. Find the expected response after 5 weeks and the expected change from week 2 to week 5.

Response at 5 weeks: $$ \mu_{\mathrm{NH}}(5)=58+(1.2\ \text{points/week})(5\ \text{weeks})=58+6=64\ \text{points}. $$

Because higher values indicate better function and $k>0$, this is an expected improving trajectory.

Change from week 2 to week 5: $$ \mu_{\mathrm{NH}}(5)-\mu_{\mathrm{NH}}(2)=k(5-2)=(1.2)(3)=3.6\ \text{points}. $$

The expected score at week 2 is $58+(1.2)(2)=60.4$ points, and $64-60.4=3.6$ points, confirming the interval-change calculation. The slope is positive, but it does not mean worsening here: the clinical direction depends on what a higher score represents.

Assumptions and Scope

A linear natural-history trajectory is most defensible when the expected response is approximately straight over the time window of interest. Its key assumptions and interpretation are:

Constant absolute rate: the expected response changes by the same amount per unit time over the modeled interval.
Defined reference and units: $B$ is the expected response at $t=0$, and the time origin and units are meaningful and consistently applied.
Untreated condition is specified: the trajectory represents natural history under the stated condition, not an unspecified combination of natural history, placebo, and active treatment.
Response scale is consistent: the same measurement and scoring interpretation apply across the time interval.
Predictions stay within a sensible range: extrapolated values should be clinically and physically plausible for the response being modeled.

The equation alone does not establish that these assumptions hold. In particular, a good fit over a short interval does not guarantee that the same slope remains appropriate over a longer period. A straight line can be a useful local approximation even when the underlying disease course is more complex.

Worked Example: Distinguishing Stable Response from Progression

Suppose a hypothetical untreated biomarker has an expected value of $B=4.5$ mg/L at $t=0$, with a slope of $k=-0.15$ mg/L per week. Calculate the expected value after 6 weeks and determine whether the response is improving or worsening if lower biomarker values indicate less disease activity.

Calculation: $$ \mu_{\mathrm{NH}}(6)=4.5+(-0.15\ \text{mg/L/week})(6\ \text{weeks})=4.5-0.90=3.60\ \text{mg/L}. $$

The expected value decreases by $0.90$ mg/L over six weeks. Given that lower values indicate less disease activity, the direction is improving. The slope’s units are mg/L per week; the predicted response retains the units mg/L.

If the slope instead were $0$ mg/L per week, the model would predict $4.5$ mg/L at every time. That is a stable expected response, not evidence that individual measurements would all equal exactly $4.5$ mg/L.

Diagnostics, Alternatives, and Limitations

A linear trajectory can be assessed by plotting observed responses against time and comparing the average pattern with the fitted straight line. If residuals—the observed responses minus the fitted expected responses—show a systematic curve, such as sustained upward bends or downward bends, the constant-slope assumption may be inadequate. Repeated runs of residuals predominantly on one side of zero can also indicate a missed time pattern. Large individual deviations alone do not prove that the mean trajectory is nonlinear; measurement variability and individual heterogeneity can produce scatter around a suitable mean trend.

Check that time is expressed in the same units used for the slope. A slope reported per week cannot be multiplied directly by a number of months without converting the time interval. Also check the response scale and the clinical direction before labeling a positive or negative slope as favorable. Finally, inspect predicted values across the full interval: a line that eventually predicts a negative count or exceeds a bounded score’s maximum is a warning against uncritical extrapolation.

A stable natural-history model, in which the expected response does not change with time, is the special case $k=0$. The linear model allows a constant nonzero trend. When the observed trend bends or its rate changes over time, a nonlinear trajectory may be more appropriate; exponential progression and saturating trajectories are distinct alternatives. Choosing among these forms requires adequate observations across time and a defensible scientific rationale, not merely a preference for a more complicated equation.

The main limitation of a linear trajectory is its constant-rate assumption. It does not represent acceleration, slowing, a plateau, or a turning point. Estimates of $B$ and $k$ can also be difficult to interpret if the time origin is not clear, or if observations cover too narrow a time range to distinguish a trend from noise. Predictions outside the observed interval are especially uncertain because the assumed rate is being extended beyond the data.

Common pitfall: “The slope is positive, so the disease is worsening.” This reasoning mistakes numerical direction for clinical direction. A positive slope means the response value increases with time. It indicates worsening only when higher values correspond to worse disease; if higher values indicate better function, the same positive slope indicates improvement.

Check Your Understanding

Answer each question using the response scale and time units stated. Brief solutions follow the questions.

A hypothetical symptom score is higher when symptoms are worse. If $B=9$ points and $k=0.5$ points per month, what is the expected response after 4 months, and what is its clinical direction?
A higher-is-better function score has $B=70$ points and $k=-2$ points per week. What is the expected change from week 1 to week 4, and how should it be interpreted?
What does $k=0$ mean in the linear trajectory, and does it require every observed response to be constant?
A model has a slope of $0.3$ mg/L per week. Why is multiplying it by 12 months invalid without a time conversion?
Residuals systematically curve upward over time. Which linear-model assumption should be reconsidered?

Solutions. (1) $\mu_{\mathrm{NH}}(4)=9+(0.5)(4)=11$ points. The score increases by 2 points, indicating worsening because higher scores mean worse symptoms. (2) The expected change is $(-2)(4-1)=-6$ points. The score is expected to fall by 6 points; because higher function scores are better, this indicates worsening function. (3) $k=0$ means the expected response is constant over time. Individual observations can still vary around that expected value. (4) The slope is per week, whereas 12 is expressed in months; convert months to weeks or convert the slope to units per month before multiplying. (5) The assumption of a constant absolute rate, and hence a straight expected trajectory, may be inadequate.