What Does an Initiation-Based Contrast Ask?
In a randomized trial, an intention-to-treat contrast compares groups according to their randomized assignment, even when some participants later stop, switch, or add treatment. In an observational study, there is no randomized assignment to use. A useful analog is to compare people who initiate a specified treatment at time zero with people who do not, and then follow both groups regardless of subsequent treatment changes.
This contrast answers a practical question: what would the outcome difference be if eligible people initiated treatment rather than did not initiate it, while later care proceeded as it naturally would under each initial choice? The later treatment changes are part of the pathway from initiation to outcome. The effect therefore includes consequences of initiation that operate through subsequent care.
Let \(A\in\{0,1\}\) denote observed initiation at time zero, \(L\) the measured baseline covariates, and \(Y\) an outcome measured at a fixed follow-up time. Write \(Y^a\) for the potential outcome under baseline initiation set to \(a\), with subsequent care allowed to follow its natural course under that initiation. For a binary outcome, common population contrasts are the risk difference \(E[Y^1]-E[Y^0]\) and risk ratio \(E[Y^1]/E[Y^0]\), when the denominator is nonzero. The target population is the population represented by the eligible participants, not automatically the full population of all patients.
The phrase “follow regardless” is substantive: participants remain in the group defined by their baseline initiation status even if they later discontinue, switch, or begin the other treatment. This is a feature of the estimand, not a claim that later treatment is irrelevant.
Assumptions for an Observational Analog
A causal interpretation requires more than recording initiation and outcomes. Let \(Y^a\) be well-defined for every eligible participant under either baseline choice. The key identification assumptions are:
- Consistency: if a participant has \(A=a\), then the observed outcome equals \(Y^a\). The treatment at initiation must be specified sufficiently clearly that the intervention represented by \(a\) is meaningful.
- Conditional exchangeability at baseline: \(Y^a\perp A\mid L\), for \(a=0,1\). Given measured baseline covariates, initiators and noninitiators must be comparable in their potential outcomes.
- Positivity: \(0<P(A=a\mid L=l)\) for each \(a\) and for covariate values \(l\) occurring in the target population.
- No interference: one participant’s initiation does not change another participant’s potential outcome, or any interference is represented in the treatment definition.
Exchangeability is a strong, generally untestable assumption. It requires adequate measurement and control of common causes of initiation and outcome, such as disease severity, prognosis, contraindications, and access to care. Positivity says that both initiation options occur across the baseline covariate patterns relevant to the target population. It does not require equal treatment probabilities, but probabilities near zero can make estimates unstable.
Identification by Baseline Standardization
For discrete \(L\), the potential-outcome mean is identified by standardizing the observed outcome mean in each initiation group to the baseline covariate distribution of the target population:
For continuous or mixed covariates, the sum is replaced by integration with respect to the target distribution of \(L\). The formula averages within-covariate outcome means over one common baseline population. It does not compare the crude outcome means of the observed initiators and noninitiators unless their baseline distributions happen to match.
To justify the result, condition first on \(L\). By conditional exchangeability, \(E[Y^a\mid L]=E[Y^a\mid A=a,L]\). Among those with \(A=a\), consistency gives \(Y=Y^a\), so the right side equals \(E[Y\mid A=a,L]\). Taking expectations over the target distribution of \(L\) yields:
Positivity ensures the conditional observed mean in the final line is defined for the covariate patterns being averaged. The argument concerns baseline exchangeability: it does not require controlling for later treatment changes to identify this treatment-policy contrast. \(\square\)
Worked Example: Standardizing Initiators and Noninitiators
Suppose half the eligible population has high baseline severity and half has low severity. The observed treatment groups differ in severity: 80% of initiators have high severity, compared with 20% of noninitiators. Assume baseline severity suffices for exchangeability and positivity holds. The observed follow-up risks, including outcomes after any later treatment changes, are:
| Baseline severity | Risk among initiators | Risk among noninitiators |
|---|---|---|
| High | 0.30 | 0.40 |
| Low | 0.10 | 0.20 |
The crude risk among initiators is \(0.80(0.30)+0.20(0.10)=0.26\). The crude risk among noninitiators is \(0.20(0.40)+0.80(0.20)=0.24\). The crude difference is \(0.26-0.24=0.02\), suggesting higher risk among initiators.
But the target population is half high severity and half low severity. Standardized risk under initiation is \(0.50(0.30)+0.50(0.10)=0.20\). Standardized risk under no initiation is \(0.50(0.40)+0.50(0.20)=0.30\). The standardized risk difference is \(0.20-0.30=-0.10\). Under the stated assumptions, initiation reduces the target-population risk by 10 percentage points. The reversal from the crude difference reflects the groups’ different baseline severity distributions.
Because the outcome risks were defined by baseline initiation group without removing people who later changed treatment, this is an initiation-based effect, not a contrast that requires continued treatment.
Equivalent Identification by Baseline Weighting
The same mean can be written using inverse-probability weights based on baseline treatment probabilities. Define \(e_a(L)=P(A=a\mid L)\). Under the assumptions above:
where \(I(A=a)\) equals 1 when \(A=a\) and 0 otherwise. To derive this identity, condition on \(L\). Consistency gives \(E[I(A=a)Y\mid L]=P(A=a\mid L)E[Y\mid A=a,L]\); dividing by \(e_a(L)\) gives \(E[Y\mid A=a,L]=E[Y^a\mid L]\), by exchangeability. Averaging over \(L\) proves the identity. Positivity makes the denominator nonzero. \(\square\)
In a sample, this identity motivates a weighted mean: people with an uncommon observed initiation choice for their baseline covariates receive more weight. Standardization and weighting are alternative ways to address baseline differences; both depend on the same causal assumptions. They can also be combined in doubly robust estimators, but combining methods does not remove the need to define the target population, exposure, and outcome correctly.
Worked Example: Checking the Weighted Risk Calculation
Suppose 60% of the target population has low severity and 40% has high severity. Initiation probabilities are 0.50 for low-severity participants and 0.25 for high-severity participants. The observed risks are 0.12 and 0.30 among initiators with low and high severity, respectively; among noninitiators they are 0.20 and 0.40.
For initiators, the baseline weights are \(1/0.50=2\) in the low-severity group and \(1/0.25=4\) in the high-severity group. The expected weighted outcome contribution is:
For noninitiators, the probabilities are 0.50 for low severity and 0.75 for high severity, giving weights 2 and \(1/0.75\). Their weighted outcome contribution is:
Thus the weighted risk difference is \(0.192-0.280=-0.088\), or an 8.8-percentage-point lower risk under initiation. The same answer follows by standardization: \(0.60(0.12)+0.40(0.30)=0.192\) and \(0.60(0.20)+0.40(0.40)=0.280\). Agreement checks the arithmetic and illustrates that the two formulas identify the same means.
Why Later Treatment Changes Stay in the Contrast
A frequent error is to adjust for, stratify by, or exclude participants according to treatment received after baseline while still describing the result as the initiation effect. Later treatment can be affected by initial treatment and can itself affect the outcome. Conditioning on it can remove part of the effect the treatment-policy estimand is meant to include; in some settings it can also introduce bias through common causes of later treatment and outcome.
The distinction can be seen in a small example. Suppose initiation is randomized, so there is no baseline confounding. Let \(M=1\) indicate a later treatment change. The probability of a change is 0.20 after initiation and 0.60 after no initiation. Suppose outcome risk is 0.40 when \(M=1\) and 0.10 when \(M=0\), regardless of initial group. Then:
The initiation-based risk difference is \(0.16-0.28=-0.12\). Yet within either value of \(M\), the outcome risks are identical across initial groups, so a comparison that adjusts away \(M\) yields no initiation contrast in this example. That adjusted comparison has removed the pathway through treatment change; it does not estimate the same total treatment-policy effect.
This does not mean later treatment is never relevant to analysis. It means the estimand must decide how it is handled. An effect under sustained treatment strategies is a different causal question. The intention-to-treat analog defined here does not require participants to sustain their initial treatment, and it should not be interpreted as the effect of doing so.
Diagnostics, Alternatives, and Limits
Useful diagnostics focus on whether the observational comparison can support its assumptions. Examine baseline covariate distributions and estimated treatment probabilities; very small probabilities signal weak overlap and potentially extreme weights. Check whether important prognostic factors and reasons for treatment choice were measured before time zero, not after initiation. Confirm that initiation, follow-up, and outcome ascertainment are defined consistently for both groups. If follow-up or outcome measurement differs by group, the observed contrast may also reflect selection or measurement differences.
Standardization uses an outcome model for \(E[Y\mid A,L]\); weighting uses a model for \(P(A\mid L)\). Standardization can be sensitive to outcome-model misspecification, while weighting can be sensitive to propensity-model misspecification and extreme weights. Their agreement is a useful diagnostic, not proof of exchangeability. Neither method detects unmeasured confounding from the observed data alone. Trimming or restricting to a region of overlap can improve stability, but changes the target population and therefore can change the estimand.
The main alternative is not simply another estimator: it is a different estimand, such as an effect under sustained treatment strategies. That question asks what outcomes would be under specified ongoing treatment behavior, rather than under baseline initiation followed by usual subsequent care. The two effects can differ even in the same population. A careful report should state the baseline treatment contrast, follow-up outcome, target population, and how subsequent treatment changes are treated.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A participant initiates treatment at baseline, stops it after two months, and has the outcome at one year. Which baseline group contributes this participant to in an initiation-based analysis?
- State the baseline exchangeability assumption using \(Y^a\), \(A\), and \(L\).
- In one sentence, distinguish an initiation-based treatment-policy effect from a sustained-treatment effect.
- What does a propensity score near zero for initiation among a common covariate pattern imply for baseline weighting?
Solutions. (1) The participant remains in the initiation group, because groups are defined by baseline initiation and follow-up continues regardless of later treatment. (2) For each \(a\in\{0,1\}\), \(Y^a\perp A\mid L\): conditional on measured baseline covariates, initiation is independent of the potential outcome. (3) The initiation-based effect sets baseline initiation and lets later care evolve naturally; a sustained-treatment effect specifies continued treatment behavior after baseline. (4) The weight \(1/P(A=1\mid L)\) is very large, so the estimate can be unstable and sensitive to a few observations; the near-zero probability also indicates weak positivity for that covariate pattern.