Why Compare New Users of Active Treatments?
A comparison of two treatments can be difficult to interpret if the groups differ in how long they have already been treated, why treatment began, or what clinical choices were available. A new-user active-comparator protocol makes the comparison more specific: it contrasts patients initiating one clinically plausible treatment with patients initiating another, after applying the same eligibility and treatment-free lookback rules.
The design has two important features. The new-user requirement establishes a common point at which treatment starts and follow-up can begin. The active comparator restricts the comparison to another treatment that could reasonably have been chosen for the same clinical purpose. Neither feature alone removes confounding. Instead, they help define a relevant question and can reduce some differences between treatment groups that would arise in a comparison with nonuse.
Specify the Protocol Before Looking at Outcomes
The protocol should say who could start either treatment, what counts as initiation, and how far back treatment history is checked. Eligibility must be assessed using information available before treatment initiation. In particular, requiring a patient to remain event-free or continue treatment after initiation would make eligibility depend on post-baseline information and change the question.
A washout period is a prespecified lookback interval, such as 180 days, during which records must show no use of the treatment or class being excluded. State the interval’s length, which records count as evidence of use, and how incomplete history is handled. A washout can distinguish observed new initiation from recorded ongoing use, but it cannot rule out treatment before the available records begin. A longer washout may better identify recent nonuse but can reduce the eligible sample and exclude people with sparse historical data.
Define the treatment strategies as initiation of treatment A versus initiation of treatment B at time zero. Specify dose or formulation if clinically relevant to the intended contrast. For this protocol, follow-up starts on the same initiation date in either group. The question is about the consequences of initiating A rather than B; by itself, it is not a comparison of sustained use, adherence, or a future treatment-switching strategy.
Identify the indication and the two treatments that are reasonable alternatives for patients in the target population.
Specify clinical and data criteria using information determined before the qualifying treatment start.
Check the prespecified lookback history for both groups, with treatment-specific details stated explicitly.
Classify the initiating treatment and begin outcome follow-up at that initiation date for either group.
Name the target population, outcome, follow-up horizon, and effect measure, then assess whether a causal interpretation is supportable.
Worked Example: Defining a New-User Comparison
Suppose a hypothetical study asks whether initiating Drug A rather than Drug B changes the one-year risk of hospitalization among adults eligible for either medication for the same condition.
Eligibility and washout. The protocol includes adults with the condition who meet prespecified clinical criteria for both drugs and have 12 months of observable medication history. Patients with a recorded dispensing of either drug during those 12 months are excluded. The protocol defines initiation as the first recorded dispensing of A or B after eligibility is met.
Time zero and contrast. Time zero is the dispensing date that qualifies a patient as initiating A or B. The outcome is hospitalization within one year after that date. The target contrast is the one-year hospitalization risk if eligible patients initiated A versus the risk if those same eligible patients initiated B. This is a contrast of treatment initiation, not a claim that patients remain on their initial drug throughout follow-up.
Protocol checks. The same eligibility criteria and washout length apply to both groups. Eligibility is determined before the dispensing date, and outcome follow-up starts on that date in both groups. These choices avoid defining one group using a later start date or granting one group guaranteed event-free time before follow-up. However, they do not establish that the groups are comparable on prognosis: differences in severity, contraindications, or clinician preference may still confound the comparison.
Interpretation and Conditions for a Causal Contrast
Let \(E=1\) denote eligibility under the protocol, \(A\) denote the treatment initiated at time zero, and \(Y^a\) denote the outcome that would be observed under initiation of treatment \(a\), where \(a\) is A or B. For a binary outcome, the treatment-initiation risk difference in the eligible population is
Here \(Y^{A}\) and \(Y^{B}\) are potential outcomes, not the observed treatment assignments. A negative risk difference means a lower average outcome risk under initiation of A than under initiation of B in the stated population and horizon. The protocol defines this target; the observed difference between treatment groups is not automatically equal to it.
A causal interpretation from observational data requires, at minimum, consistent treatment and outcome definitions, no interference between individuals, conditional exchangeability, and positivity. With \(X\) denoting measured pre-initiation covariates, conditional exchangeability means \(Y^a\) is independent of \(A\) given \(X\) and eligibility. Positivity means that, within relevant covariate patterns among eligible patients, each treatment has a positive probability of initiation. Consistency requires that the observed outcome for a patient who initiates \(a\) equals that patient’s potential outcome under the protocol’s version of \(a\). The covariates must be measured before initiation and should capture important common causes of treatment choice and outcome.
The standardization identity follows directly from these conditions. First average the potential outcome over the covariate distribution among eligible patients. Exchangeability replaces the conditional potential-outcome mean with the observed outcome mean among initiators of \(a\) with the same covariates; consistency links that observed outcome to \(Y^a\):
For binary outcomes, this is an average of conditional risks over the eligible population’s covariate distribution. The outer expectation is appropriate whether \(X\) is continuous, discrete, or mixed. This identity does not itself guarantee an unbiased estimate: unmeasured confounding, poor covariate overlap, or misspecified estimation can undermine its use.
Worked Example: Standardizing Risks to the Eligible Population
Consider hypothetical data with two baseline severity groups. Among eligible patients, 60% have lower severity and 40% have higher severity. The observed one-year outcome risks are 4% for A initiators and 5% for B initiators in the lower-severity group; in the higher-severity group, the risks are 9% for A and 10% for B. Assume for this calculation that conditional exchangeability, consistency, and positivity hold.
For A, standardize the two conditional risks to the eligible population’s 60%/40% severity distribution:
The standardized one-year risk under A is 6.0%. For B:
The standardized risk difference is \(0.060-0.070=-0.010\), or −1.0 percentage point. Under the stated assumptions, this estimates a one-percentage-point lower one-year risk under initiation of A than under initiation of B for the eligible population. It is not a 1% relative reduction, and without the assumptions it is only a standardized association.
Choosing the Comparator and Checking the Design
An active comparator is useful when it reflects a real treatment decision: patients receiving either option should plausibly have been considered for the other. This can reduce differences in indication and healthcare contact compared with a no-treatment group. It does not ensure clinical equipoise for every patient, remove contraindication-related differences, or control unmeasured confounding.
A comparison with nonuse may answer a different question and can include people who were not candidates for treatment or had different healthcare-seeking patterns. A prevalent-user comparison includes patients who began treatment earlier and remained on it until observation; it conditions on prior treatment and survival or continuation up to study entry. Neither alternative is inherently invalid, but each targets a different contrast and has different sources of selection.
Before analysis, look for signs that the protocol is not working as intended:
- Unequal or vague washout rules: one group may contain more continuing users, or “new use” may mean different things across groups.
- Misaligned time zero: starting follow-up at eligibility for one group but at initiation for the other can create unequal opportunities for outcomes to occur.
- Ineligible comparator users: if patients receiving B would not plausibly qualify for A, the groups may represent different clinical decisions.
- Limited overlap: if some covariate patterns almost always receive one treatment, positivity is doubtful and adjusted comparisons rely heavily on extrapolation.
- Incomplete treatment history: a recorded washout may misclassify prior use when enrollment or pharmacy records do not cover the full lookback.
- Baseline variables measured after initiation: these may already be affected by treatment and should not be treated as pre-treatment confounders.
Inspect group sizes, baseline covariate distributions, treatment-history coverage, and the availability of both options across clinically important patient profiles. Large imbalances do not prove that causal estimation is impossible, but they identify where adjustment and interpretation need particular care. If key eligibility or washout information is unavailable, report that limitation rather than implying the protocol has verified it.
Worked Example: Revising a Weak Protocol
A hypothetical analyst proposes comparing patients taking Drug A on January 1 with patients who first start Drug B during the following six months. The analyst measures outcomes beginning January 1 for both groups and calls everyone with no dispensing in the prior 30 days a new user.
The groups do not share a treatment-initiation time: many B users must remain observable until a later start date, while A users are already taking treatment at the beginning of follow-up. The comparison also mixes prevalent A use with new B use, and a 30-day gap may fail to distinguish recent initiation from ongoing or recently interrupted use.
A clearer protocol would require eligibility before initiation, apply the same prespecified washout rule to both treatments, define initiation consistently from the available records, and start each person’s follow-up on the qualifying A or B initiation date. The comparator should be clinically plausible for the same eligible population. These revisions clarify the target treatment-initiation contrast; they do not, by themselves, resolve confounding.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- What does a 180-day washout establish, and what does it not establish?
- Why should follow-up begin at treatment initiation for both groups?
- In a 70%/30% target covariate mixture, the conditional risks under A are 2% and 8%. What is the standardized risk under A?
- Does an active comparator eliminate confounding by indication? Explain.
- What does a risk difference of −2 percentage points mean if causal assumptions hold?
Solutions. (1) It establishes no recorded use under the specified rule during the preceding 180 days, assuming adequate records; it does not establish lifetime nonuse or rule out unrecorded use. (2) It aligns the start of treatment classification and outcome follow-up, avoiding unequal pre-follow-up time. (3) \((0.70)(0.02)+(0.30)(0.08)=0.014+0.024=0.038\), so the standardized risk is 3.8%. (4) No. It may make the clinical decision more comparable, but differences in indication, severity, or other causes of treatment choice can remain. (5) In the specified eligible population and follow-up horizon, the average risk under initiation of A is two percentage points lower than under initiation of B; it is an absolute risk difference, not a two-percent relative reduction.