Interpreting Treatment-by-Time Interaction Models

Why the Treatment Coefficient Needs a Time Reference

When outcomes change during enrollment, a single treatment contrast may not describe every participant’s enrollment period. For example, prognosis may improve or worsen over calendar time, and the difference between treatment arms may itself change. A treatment-by-time interaction represents that changing contrast. Its interpretation depends on the units and zero point used for time.

The key is to distinguish three quantities: the treatment contrast at a particular enrollment time, the rate at which that contrast changes over time, and a summary of contrasts averaged over a specified distribution of enrollment times. A treatment main-effect coefficient in an interaction model is the first of these, but only at the time value coded as zero.

Definition: Let $A_i=1$ denote assignment to treatment and $A_i=0$ assignment to control. Let $t_i$ be participant $i$'s enrollment time, measured in a stated unit and possibly centered at a reference time. A linear treatment-by-time model for a continuous outcome is $$ E(Y_i\mid A_i,t_i)=\beta_0+\beta_A A_i+\beta_t t_i+\beta_{At}A_it_i. $$ Here $\beta_0$ is the control-arm mean at $t_i=0$; $\beta_t$ is the control-arm change in mean outcome per time unit; $\beta_A$ is the treatment-minus-control contrast at $t_i=0$; and $\beta_{At}$ is the change in that contrast per time unit.

The corresponding model-based treatment contrast at time $t$ is

\Delta(t)=E(Y\mid A=1,t)-E(Y\mid A=0,t)=\beta_A+\beta_{At}t.

The control-arm time slope is $\beta_t$, while the treatment-arm time slope is $\beta_t+\beta_{At}$. Thus $\beta_{At}$ is the difference between those slopes. If it is positive, the treatment-minus-control contrast becomes more positive as time increases; whether that means outcomes are improving or worsening depends on the outcome scale.

Assumptions and Scope of Interpretation

The equation defines a conditional-mean model, not a guarantee that the data follow it. It assumes that, over the time range being interpreted, the mean outcome in each arm is adequately represented by a straight line in enrollment time. The interaction is linear too: each additional time unit changes the treatment contrast by the same amount. The model also assumes that the specified treatment coding, time scale, and outcome units are used consistently.

A fitted contrast is initially a model-based comparison of expected outcomes. A causal interpretation additionally requires appropriate conditions for the trial: treatment assignment must be randomized, outcomes must correspond to the assigned treatment under the consistency assumption, and each treatment must have positive assignment probability in the enrollment-time range being summarized. In an adaptive trial, assignment can depend on recorded trial history. If that history affects outcomes and the allocation rule, a causal analysis may need to account for the relevant history as well as time; time adjustment alone does not automatically remove every such dependence. Missing outcomes and departures from the allocation procedure can also undermine interpretation.

The model is useful only over times with meaningful data support in both arms. A fitted contrast far beyond observed enrollment times is an extrapolation, not a direct summary of observed trial evidence. This tutorial concerns coefficient and contrast interpretation; it does not supply an uncertainty calculation for those quantities.

Read the Contrast Before Reading the Main Effect

For any chosen time, substitute that time into $\Delta(t)$. This gives the treatment-minus-control contrast in outcome units. A negative value means the treatment mean is lower; a positive value means it is higher. Whether either direction is desirable depends on the outcome. The main-effect coefficient $\beta_A$ is simply $\Delta(0)$, so choosing a meaningful time origin makes it directly interpretable.

Worked Example: Contrasts at Early and Late Enrollment Times

Suppose, hypothetically, an outcome is a symptom score in points, with lower scores indicating better health. Enrollment time is measured in months from month 6, so $t=0$ is month 6; month 2 corresponds to $t=-4$, and month 10 to $t=4$. A fitted model is

\widehat{E}(Y\mid A,t)=50-4A+0.6t+0.4At.

At month 6, the treatment-minus-control contrast is $-4+0.4(0)=-4$ points. At month 2, it is

\widehat{\Delta}(-4)=-4+0.4(-4)=-4-1.6=-5.6\text{ points}.

At month 10, it is

\widehat{\Delta}(4)=-4+0.4(4)=-4+1.6=-2.4\text{ points}.

The fitted contrast is negative at all three times, so the treatment arm has a lower predicted symptom score at each of these times. Because lower is better, those predicted contrasts favor treatment. The interaction coefficient says the contrast becomes 0.4 points more positive per month: the predicted treatment advantage is smaller later in this example. The control slope is 0.6 points per month, and the treatment slope is $0.6+0.4=1.0$ point per month.

Changing the Time Origin Changes Coefficients, Not Fitted Contrasts

Centering time means subtracting a reference time from every recorded time. It changes where $t=0$ lies. Consequently, it changes the treatment main-effect coefficient to the contrast at the new zero point, but it does not change the fitted contrast at any actual calendar time.

Result: If the original time is $t$, and a new time variable is $u=t-c$, then $t=u+c$. The treatment contrast can be written in either coordinate: $$ \beta_A+\beta_{At}t=(\beta_A+c\beta_{At})+\beta_{At}u. $$ The new treatment main effect is $\beta_A+c\beta_{At}$, the contrast at the original time $t=c$. The interaction coefficient remains $\beta_{At}$.

This result follows by direct substitution. Replacing $t$ with $u+c$ gives $\beta_A+\beta_{At}(u+c)$, which expands to $(\beta_A+c\beta_{At})+\beta_{At}u$. No fitted contrast has changed; only its coordinate description has changed.

Worked Example: Re-centering at a Different Enrollment Month

Suppose, hypothetically, $Y$ is a score in points, with higher scores indicating worse symptoms. Let $u$ be months from month 4. The fitted model is

\widehat{E}(Y\mid A,u)=12+2A+0.5u+0.3Au.

The treatment contrast is $2+0.3u$ points, so at month 4 ($u=0$) the fitted contrast is 2 points. Now let $t$ be months from the start of enrollment. Since month 4 is four months after the start, $u=t-4$. Substituting:

$$ 2+0.3u=2+0.3(t-4)=0.8+0.3t. $$

In the new coding, the treatment main effect is 0.8 points, the contrast at the start of enrollment. At month 8, $t=8$, so the contrast is $0.8+0.3(8)=3.2$ points. In the original coding, month 8 gives $u=4$, and the contrast is $2+0.3(4)=3.2$ points as well. Since higher scores are worse, both calculations indicate a higher predicted score in the treatment arm at month 8. The coefficient changed with the time origin; the prediction did not.

Summarizing Contrasts Over Enrollment Time

A coefficient at one reference time is not automatically the overall treatment effect for the trial. To obtain an average contrast, specify a target distribution of enrollment times and average the fitted time-specific contrasts over that distribution. For a finite set of target times $t_j$ with weights $w_j$, where $w_j\geq 0$ and $\sum_j w_j=1$, the standardized contrast is

\Delta_{\mathrm{target}}=\sum_j w_j\Delta(t_j) =\beta_A+\beta_{At}\sum_jw_jt_j.

Thus, under this linear interaction model, the standardized contrast equals the treatment contrast at the target distribution’s mean enrollment time. The target weights matter: weighting participants equally, enrollment periods equally, or a prespecified population distribution can produce different summaries. State the target rather than calling one of them simply “the” average effect.

Worked Example: Standardizing to a Target Enrollment-Time Distribution

Suppose, hypothetically, a model for an outcome measured in points is

\widehat{\Delta}(t)=1.5+0.25t,

where $t$ is months from the start of enrollment. Consider a target population assigning equal weight to enrollment times $t=1,3,5$ months. Its mean time is $(1+3+5)/3=3$ months. The three fitted contrasts are

\widehat{\Delta}(1)=1.5+0.25(1)=1.75,\quad \widehat{\Delta}(3)=1.5+0.25(3)=2.25,\quad \widehat{\Delta}(5)=1.5+0.25(5)=2.75.

Their equally weighted average is $(1.75+2.25+2.75)/3=2.25$ points. Equivalently, substituting the target mean time gives $1.5+0.25(3)=2.25$ points. If lower scores are better, a positive treatment-minus-control contrast of 2.25 points means the treatment group’s predicted score is higher, and therefore worse, on average for this hypothetical target population. The average contrast favors control by 2.25 points.

Model Checks and Common Alternatives

Interpretation should be accompanied by checks that the model is describing a relevant range of the trial. Plot or tabulate observed outcome summaries by arm and enrollment time, and compare them with the fitted arm-specific trends. Check that the time variable’s unit and zero point are explicit, that treatment is coded as intended, and that the interaction sign agrees with contrasts calculated at selected times. Look for curvature or abrupt changes that a straight-line trend cannot represent. Also inspect whether both arms have observations throughout the target time range; a model can calculate a contrast where the data provide little direct comparison.

The main alternatives answer somewhat different questions:

Approach	What it represents	Main interpretive caution
Model without an interaction	A constant treatment contrast across enrollment time, with a common time trend in both arms.	It imposes no change in the contrast; it is not evidence that the contrast truly is constant.
Linear treatment-by-time interaction	A contrast that changes at a constant rate per time unit.	The main effect is only the contrast at time zero; linearity may be inadequate.
Time-stratified contrasts and standardization	Contrasts within defined time groups, combined using stated target weights.	Results depend on the time groups and target weights; sparse groups may be unstable.
Flexible time adjustment	Potentially curved trends or more complex changes in the contrast.	Greater flexibility requires enough information and does not remove the need to define a target summary.

These approaches are not interchangeable by default. A model without an interaction can be a useful simpler summary if a constant contrast is a defensible assumption. Time-stratified or flexible approaches may better represent nonlinearity, but their summaries still require an explicit target distribution. No choice of model compensates for poor overlap or an unclear estimand.

Common Mistakes and Limitations

Calling $\beta_A$ the overall treatment effect. In an interaction model, it is the contrast only at $t=0$. Center time at a meaningful reference point or calculate the contrast at the time of interest.
Reading the interaction as the treatment effect. $\beta_{At}$ is the change in the treatment contrast per time unit, not the contrast itself. Calculate $\beta_A+\beta_{At}t$.
Ignoring outcome direction. A positive treatment-minus-control contrast means higher outcomes under treatment, not necessarily a benefit. Interpret the sign using the outcome scale.
Changing time units without changing the interpretation. A per-month interaction and a per-year interaction have different numerical values. Convert the scale explicitly.
Assuming a fitted interaction establishes causation or proves biological effect modification. The model is conditional on its assumptions and the trial’s assignment and outcome processes. A coefficient alone does not establish either claim.

The linear interaction is a compact description, not a universal pattern. It can miss curvature, thresholds, or changes that are not smooth. A single standardized contrast can also conceal meaningful variation across time. Report the reference time, interaction scale, selected time-specific contrasts, and the target distribution used for any average; these details make the result interpretable without treating one coefficient as a complete account.

Key takeaway: In a treatment-by-time model, the main effect is the treatment-minus-control contrast at the coded time zero, and the interaction is the change in that contrast per time unit. Re-centering changes the main-effect coefficient but not the fitted contrasts. An average contrast is meaningful only after its target enrollment-time distribution is specified.

Check Your Understanding

Answer each question by identifying the relevant time reference and keeping the treatment-minus-control sign convention explicit.

A fitted model has $\widehat{\Delta}(t)=-1+0.2t$, where $t$ is months from enrollment start. What is the contrast at month 5, and how does it change per month?
If the contrast is $\beta_A+\beta_{At}t$, what does $\beta_A$ mean when time is centered at month 7?
A target population has mean enrollment time 4 months, and the model is $\widehat{\Delta}(t)=0.5-0.3t$. Find the standardized contrast. If lower outcomes are better, which arm has the lower predicted outcome on average?
A colleague says that changing the time origin changed the treatment effect because the treatment main-effect coefficient changed. Explain the error.

Solutions. (1) At month 5, $\widehat{\Delta}(5)=-1+0.2(5)=0$ outcome units. The contrast becomes 0.2 units more positive per month. (2) $\beta_A$ is the treatment-minus-control contrast at month 7, because centered time equals zero there. (3) The standardized contrast is $0.5-0.3(4)=-0.7$ outcome units. Treatment is predicted to have the lower outcome, so the contrast favors treatment when lower is better. (4) Re-centering changes which time point the main-effect coefficient describes. The fitted contrast at each actual enrollment time remains the same, as substitution into either parameterization confirms.