REACT: Complete Statistical Analysis of Roflumilast in Chronic Obstructive Pulmonary Disease

1. Trial at a Glance

REACT was a randomized, double-blind, parallel-group phase 4 trial evaluating roflumilast 500 µg versus placebo in patients with chronic obstructive pulmonary disease treated with fixed combinations of long-acting β2-agonists and inhaled glucocorticosteroid. The registry reports 1945 enrolled participants and a primary endpoint based on the rate of moderate or severe COPD exacerbations per patient per year over 52 weeks.

2. Clinical Question

The primary statistical question was whether roflumilast 500 µg reduced the rate of moderate or severe COPD exacerbations per patient per year over 52 weeks compared with placebo.

3. Trial Design

4. Primary Endpoint

The registry classifies the primary outcome unit as exacerbations per patient per year. This is fundamentally a rate outcome: the analysis accounts for the amount of time participants contribute to observation rather than treating every participant as if they contributed exactly the same follow-up time.

5. Statistical Methodology

Generalized linear regression for exacerbation rates

The primary analysis was reported as generalized linear regression and described specifically as a Poisson regression model using time in trial as the model offset. The effect measure was a rate ratio.

Time in trial as an offset

Participants can contribute different amounts of observation time. In a rate model, an offset allows the model to distinguish the number of observed events from the amount of time during which those events could occur. This is particularly important when the outcome is expressed as events per patient per year.

ANCOVA

The change from baseline in post-bronchodilator FEV1 was analyzed using ANCOVA, with treatment by time interaction included. The reported effect measure was the least-square mean difference.

MMRM

Several repeated-measures outcomes were analyzed using a repeated measurement model, normalized here as an MMRM. The registry reports unstructured covariance with restricted maximum likelihood (REML) for FVC, FEF25-75%, and FEV6, while rescue medication and COPD symptom score used compound symmetry covariance with REML. CAT total score used an unstructured covariance structure and REML.

Cox proportional-hazards models

Time-to-first-exacerbation and several other time-to-event outcomes were analyzed with Cox proportional-hazards models. The effect measure was the hazard ratio.

Wei-Lin-Weissfeld method

The time to second and time to third moderate or severe COPD exacerbation outcomes were analyzed using the Wei-Lin-Weissfeld method. These analyses extend beyond a simple time-to-first-event comparison by addressing recurrent ordered events.

6. Analysis Population

The reported efficacy analyses used participants from the intent-to-treat population, defined in the registry analysis records as all randomized participants who received at least 1 dose of study drug, analyzed according to the treatment for which they were randomized. Several outcomes additionally required data to be available for the corresponding analysis.

This distinction matters because the ITT principle preserves the randomized comparison, while missing observations and treatment exposure can still affect the precision and interpretation of individual endpoint analyses.

7. Primary Result: Rate of Moderate or Severe COPD Exacerbations

The registry reports a rate ratio of 0.868 for moderate or severe COPD exacerbations per patient per year. Because the estimate is below 1, the estimated exacerbation rate was lower with roflumilast than with placebo in the fitted model. Expressed as a simple relative interpretation, 0.868 corresponds to an estimated rate approximately 13.2% lower under the model.

8. Secondary Exacerbation Rate Results

The registry reports several additional analyses of exacerbation rates. These endpoints use different clinical definitions of exacerbation, so their rate ratios should not be treated as repeated measurements of one identical outcome.

The reported rate ratios are all below 1, but their confidence intervals and p-values differ. This is exactly why a trial should not be summarized from the direction of the point estimates alone. Different endpoint definitions can produce different statistical evidence even when they concern related clinical events.

Severe exacerbations

The registry states that severe exacerbations were analyzed using a negative binomial regression model excluding a correction for overdispersion. The reported rate ratio of 0.757 indicates a lower estimated severe-exacerbation rate in the roflumilast group relative to placebo under that model.

All-category exacerbation analyses

The hospitalisation-related analysis used negative binomial regression, whereas the other listed rate analyses used Poisson regression. The distinction is important because the statistical model is part of the definition of the reported estimate: the same clinical trial does not necessarily use one universal rate model for every exacerbation endpoint.

9. Pulmonary Function Results

Post-Bronchodilator FEV1

The reported ANCOVA estimated a least-square mean difference of 0.056 liters for change from baseline in post-bronchodilator FEV1, comparing roflumilast 500 µg with placebo. The registry specifies an ANCOVA including treatment by time interaction.

Forced Vital Capacity

FVC was analyzed with a repeated measurement model using an unstructured covariance structure and restricted maximum likelihood (REML). The reported least-square mean difference was 0.092 liters.

FEF25-75%

The registry reports an MMRM with unstructured covariance and REML. The outcome unit is liters/second.

FEV6

FEV6 was also analyzed using an MMRM with an unstructured covariance structure and REML.

10. Patient-Reported and Daily-Diary Outcomes

The rescue-medication analysis produced a negative least-square mean difference, while the reported COPD symptom-score and CAT estimates had confidence intervals spanning zero. These outcomes illustrate why the direction of an estimate should always be interpreted together with its uncertainty interval and the endpoint's measurement scale.

11. Time-to-Event Results

Time to First COPD Exacerbation — All Categories

The hazard ratio below 1 indicates a lower estimated hazard of first exacerbation in the roflumilast group under the Cox model. The 95% CI extends from 0.815 to 1.031, and the P-value is 0.1461.

Time to Second Moderate or Severe COPD Exacerbation

This endpoint was analyzed using the Wei-Lin-Weissfeld method. The reported hazard ratio was 0.790, with a 95% CI of 0.641–0.974 and P = 0.0270.

Time to Third Moderate or Severe COPD Exacerbation

The third-exacerbation analysis also used the Wei-Lin-Weissfeld method. Its confidence interval extends above 1, unlike the reported interval for the time to second exacerbation.

12. Mortality, Withdrawal, Cardiovascular Events, and Hospitalisation

Withdrawal

The reported hazard ratio for time to withdrawal was 1.529. Unlike the exacerbation rate-ratio analyses, the registry's Cox interpretation states that values below 1 represent a lower hazard for the test treatment. Thus, an HR above 1 represents a higher estimated hazard of the event in the roflumilast group under this time-to-event model.

Mortality and cardiovascular outcomes

The mortality and MACE estimates have comparatively wide confidence intervals. For mortality due to any reason, the HR was 1.025 with a 95% CI of 0.528–1.990. For mortality due to COPD exacerbation, the HR was 1.078 with a 95% CI of 0.378–3.075. For first MACE, the HR was 1.083 with a 95% CI of 0.542–2.167.

Wide intervals mean that the point estimate should not be interpreted as highly precise. In particular, an HR close to 1 with a wide interval can represent substantial uncertainty about the underlying treatment effect.

13. Number Needed to Treat to Benefit

The registry reports a secondary endpoint for the Number of Patients Needed to Treat to Avoid 1 Moderate or Severe COPD Exacerbation Derived From Exacerbation Per Patient Per Year. The reported NNTB is 9, with a two-sided 95% CI of 4–31.

14. Statistical Methods Explained

Why was a rate ratio used for the primary endpoint?

The primary endpoint counts COPD exacerbations over a period of observation and expresses the result as exacerbations per patient per year. A rate ratio compares the estimated event rates between treatment groups. This is different from comparing the proportion of patients who experienced at least one event.

Why was time in trial used as a model offset?

Participants may contribute different amounts of follow-up. The offset allows the Poisson model to account for the exposure time when estimating event rates. Without that adjustment, a participant observed for substantially less time could be treated too similarly to a participant observed for the full analysis period.

Why were Poisson and negative binomial models both used?

The registry reports Poisson regression for several exacerbation-rate analyses and negative binomial regression for severe exacerbations and the hospitalization-related analysis. These are count-data modeling approaches, but they make different assumptions about the distribution of event counts. The appropriate model depends on the structure of the observed event data and the prespecified statistical analysis.

What does a rate ratio of 0.868 mean?

It means the estimated rate under roflumilast was 0.868 times the estimated rate under placebo in the reported primary model. A rate ratio below 1 indicates a lower estimated event rate. It does not mean that the treatment caused exactly the same percentage reduction in each individual participant's risk.

Why is a hazard ratio different from a rate ratio?

A rate ratio compares event rates over observed person-time, whereas a hazard ratio compares instantaneous event hazards in a time-to-event model. They can both be below 1 without representing the same quantity. In REACT, the primary endpoint uses a rate ratio, while time-to-event endpoints such as time to first exacerbation use hazard ratios.

Why does the primary confidence interval matter?

The primary rate ratio is 0.868 with a 95% CI of 0.753–1.002. The point estimate alone suggests a lower estimated rate, but the confidence interval communicates uncertainty and extends slightly above the null value of 1. The P-value of 0.0529 provides another representation of the evidence against the null hypothesis under the specified testing framework.

What does MMRM add to repeated measurements?

Repeated measurements from the same participant are correlated. An MMRM models those repeated observations jointly rather than treating them as independent observations. REACT reports different covariance structures and REML estimation for different longitudinal outcomes, reflecting the fact that the correlation structure is part of the statistical model.

15. Statistical Interpretation of the Primary Analysis

The statistical story of the primary endpoint is therefore more informative when all three quantities are considered together: the rate ratio describes the estimated relative effect, the confidence interval describes uncertainty around that estimate, and the p-value describes the strength of evidence against the null under the specified testing framework.

16. Safety

The registry reports serious adverse events by treatment arm using affected participants over participants at risk.

The reported serious-adverse-event figures are 249/968 for roflumilast 500 µg and 285/967 for placebo. These are descriptive safety counts as reported in the registry by the registry data. The ClinicalTrials.gov record does not include a formal statistical comparison of serious adverse events, so no comparative p-value or confidence interval is assigned here.

17. Censoring and Time-to-Event Interpretation

Several REACT outcomes are time-to-event endpoints measured in days, including time to first exacerbation, second and third exacerbations, mortality, withdrawal, MACE, and hospitalization. Time-to-event analysis is useful because participants can have different amounts of observed follow-up and because the timing of an event contains information beyond whether an event occurred.

The REACT registry data explicitly note that some participants extended treatment beyond 52 weeks and were included in certain analyses. This matters because the stated analysis time frame and the actual information contributing to a time-to-event analysis are not always identical for every participant.

18. Recurrent Exacerbation Analysis

The time to second and time to third moderate or severe COPD exacerbation endpoints illustrate a different statistical question from time to first exacerbation. Once a participant has experienced one exacerbation, subsequent events can provide additional information about recurrent disease activity.

The registry used the Wei-Lin-Weissfeld method for both the second- and third-exacerbation endpoints. The reported estimates were HR 0.790 for the second exacerbation and HR 0.749 for the third exacerbation.

These results should not be interpreted as if they were simply two independent replications of the primary endpoint. They address different positions in the sequence of recurrent events and use a method specifically identified by the registry for those repeated-event questions.

19. Longitudinal Modeling: Why the MMRM Results Are Different From ANCOVA

REACT uses both ANCOVA and MMRM for continuous outcomes. The distinction is instructive.

The least-square mean difference is therefore not necessarily a simple observed mean difference. It is a model-based comparison adjusted according to the structure of the specified analysis.

20. Multiplicity and Multiple Secondary Endpoints

The trial data contain one registered primary endpoint and numerous secondary analyses, including multiple definitions of exacerbation rate, pulmonary-function outcomes, patient-reported outcomes, recurrent exacerbation endpoints, withdrawal, mortality, MACE, hospitalization, and NNTB.

This is particularly important when interpreting a collection of related endpoints. Several nominally small p-values can coexist with other nonsignificant results without implying that every endpoint has identical evidentiary status. The role of each endpoint in the prespecified testing hierarchy matters.

21. What the Rate Ratio Does — and Does Not — Mean

22. Important Limitations and Interpretation Issues

• Primary endpoint definition: The primary outcome is an event rate per patient per year, so it should not be reduced to a binary event proportion.
• Confidence interval crosses the null: The primary 95% CI is 0.753–1.002, which extends slightly above a rate ratio of 1.
• Multiple secondary endpoints: The registry reports numerous secondary outcomes. The ClinicalTrials.gov record does not provide a complete multiplicity-adjustment framework for all posted analyses.
• Different statistical models: Poisson regression, negative binomial regression, ANCOVA, MMRM, Cox proportional-hazards models, and the Wei-Lin-Weissfeld method are used for different endpoint structures. Their estimates are not interchangeable.
• Repeated-event interpretation: Second- and third-exacerbation analyses address recurrent events rather than simply repeating the time-to-first-event analysis.
• Time-to-event assumptions: Cox hazard ratios are model-based and require appropriate interpretation within the proportional-hazards framework.
• Analysis-population wording: Several efficacy analyses use an ITT population defined as randomized participants who received at least 1 dose, with some outcomes also requiring available data.
• Extended follow-up: Some time-to-event analyses explicitly include participants who extended treatment beyond 52 weeks, so the information contributing to those analyses can extend beyond the nominal 52-week frame.
• Safety comparison: The ClinicalTrials.gov record provides serious-adverse-event counts by arm but do not provide a formal comparative statistical analysis.
• Generalizability: The registry population is specifically described as COPD patients treated with fixed combinations of LABA and ICS; extrapolation to populations outside that description requires separate evidence.

23. Why This Trial Matters Statistically

REACT is a useful teaching case because it combines several common clinical-trial data structures in one randomized study. The primary endpoint is a recurrent-event rate, while the secondary program includes continuous repeated measures, time-to-event outcomes, recurrent-event timing, patient-reported measures, safety outcomes, and a derived NNTB.

24. Clinical Interpretation vs Statistical Interpretation

The appropriate statistical reading is therefore not a single number. The primary endpoint provides the prespecified main comparison, while the secondary endpoints provide additional information about exacerbation severity, pulmonary function, repeated outcomes, patient-reported measures, and other clinical events. Each must be interpreted according to its own estimand, scale, model, confidence interval, and role in the trial.

25. Trial Timeline

26. Data Architecture of the REACT Analysis

One reason REACT is statistically instructive is that the trial does not force every clinical outcome into the same mathematical framework. Instead, the analysis method follows the structure of the endpoint.

This endpoint-specific architecture is an important principle in clinical-trial statistics: the outcome's data-generating structure should determine the analysis rather than applying one generic test to every variable.

27. Related Tutorials

Learn more about the methods used in this trial:

28. Related Statistical Calculators

29. Sources

• ClinicalTrials.gov: REACT, NCT01329029. Official registry record containing the trial design, endpoint definitions, posted outcome measures, statistical analyses, and safety data used for this page.
• PubMed: PMID 29797235.
• PubMed: PMID 29763572.
• PubMed: PMID 28679611.
• PubMed: PMID 25684586.
• PubMed: PMID 22791991.

30. Record Summary

REACT provides a broad statistical case study in randomized clinical-trial analysis. Its primary endpoint is an exacerbation rate analyzed with Poisson regression and a time-in-trial offset, producing a rate ratio of 0.868 with a two-sided 95% CI of 0.753–1.002 and P = 0.0529. The same trial also reports ANCOVA for FEV1, MMRM for several repeated pulmonary and patient-reported outcomes, Cox models for multiple time-to-event endpoints, and Wei-Lin-Weissfeld analyses for recurrent exacerbation timing.

The most important statistical lesson is that these estimates answer different questions. A rate ratio describes relative event frequency over patient-time; a hazard ratio describes relative event hazard over time; and a least-square mean difference describes a model-based difference in a continuous outcome. Confidence intervals provide the uncertainty surrounding each estimate, while p-values address hypothesis testing rather than effect magnitude.