This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record.
1. Trial at a Glance
RECESS was a randomized, parallel-group, phase 3 trial evaluating whether the duration of red blood cell storage affected postoperative multiple organ dysfunction and other clinical outcomes in patients undergoing cardiac surgery.
| Feature | RECESS |
|---|---|
| Trial name | Red Cell Storage Duration Study |
| Phase | Phase 3 |
| Conditions | Cardiac Surgery; Erythrocyte Transfusion |
| Allocation | Randomized |
| Design | Parallel |
| Masking | None |
| Primary purpose | Treatment |
| Enrollment | 1,481 |
| Interventions | Red blood cell units stored ≤ 10 days; red blood cell units stored ≥ 21 days |
| Start | 2010-01 |
| Primary completion | 2014-03 |
| Lead sponsor | Carelon Research |
| Registry status | Completed |
| ClinicalTrials.gov | NCT00991341 |
2. Clinical Question
The central question was whether transfusion of shorter-storage red blood cell units, stored for 10 days or less, produced a different postoperative change in multiple organ dysfunction score than transfusion of longer-storage red blood cell units, stored for 21 days or more, among evaluable patients undergoing cardiac surgery.
Population
Patients undergoing cardiac surgery who were randomized, underwent surgery within 30 days after randomization, and received at least one red blood cell transfusion between randomization and postoperative hour 96 for the evaluable analyses.
Intervention
Red blood cell units stored for ≤ 10 days.
Comparator
Red blood cell units stored for ≥ 21 days.
Primary question
Does shorter red blood cell storage duration change postoperative multiple organ dysfunction relative to longer storage duration?
3. Trial Design
Red blood cell units stored ≤ 10 days
- Biological intervention
- Compared with the longer-storage reference arm
- Primary and secondary outcomes analyzed in evaluable subjects
Red blood cell units stored ≥ 21 days
- Biological comparator
- Reference category for reported treatment effects
- Used as the reference arm in the ANCOVA and Cox analyses
The trial was randomized and parallel-group, with no masking. The statistical comparison therefore focused on the assigned storage-duration strategy rather than a blinded treatment contrast.
4. Analysis Population
The posted analyses were restricted to evaluable subjects. The registry defines an evaluable subject as a randomized subject who underwent cardiac surgery within 30 days after randomization and received at least one red blood cell transfusion between randomization and postoperative hour 96.
5. Primary Endpoint
| Endpoint | Registry definition | Time frame | Analysis |
|---|---|---|---|
| Change in the Composite Multiple Organ Dysfunction Score (MODS) From the Pre-operative Baseline | The worst post-operative values of each component of MODS were used to calculate the change in MODS. | Through post-operative day 7, hospital discharge, or death, whichever occurs first | ANCOVA, adjusting for baseline MODS |
The follow-up MODS used to calculate 7-day change was based on the worst value of each MODS component observed through postoperative day 7, hospital discharge, or death, whichever occurred first. The worst values for different components could occur on different dates. Subjects who died during this period were assigned the worst possible follow-up MODS score of 24 points, with each component set at 4.
The total MODS score ranges from 0, representing the best possible score, to 24 points, representing the worst possible score. If a subject did not die during the period but had at least one day when the Glasgow Coma Score could not be scored for the specified clinical reasons, the postoperative MODS score was set to missing and a 7-day change in MODS was not computed.
6. Primary Result
Change in MODS Through Postoperative Day 7
Adjusted mean difference
95% CI: −0.60 to 0.26 · P = 0.44
Shorter-storage red blood cell units versus longer-storage red blood cell units, with baseline MODS adjustment.
| Endpoint | Shorter-storage vs longer-storage | 95% CI | P-value | Method |
|---|---|---|---|---|
| Change in MODS through postoperative day 7, discharge, or death | Mean difference −0.17 | −0.60 to 0.26 | 0.44 | ANCOVA |
The adjusted mean difference of −0.17 MODS points means that the estimated average change in MODS was 0.17 points lower in the shorter-storage arm than in the longer-storage reference arm after adjustment for baseline MODS. The registry states that positive values are in favor of longer storage duration.
The estimate is therefore close to zero relative to the uncertainty around it. The 95% confidence interval of −0.60 to 0.26 spans both negative and positive differences, so the data are compatible with modest differences in either direction as represented by this interval. The interval describes statistical uncertainty around the estimated group difference; it does not describe the range of individual patient responses.
The P-value of 0.44 is a measure of compatibility between the observed data and the statistical null hypothesis under the specified analysis. It is not a measure of the size, clinical importance, or probability of the treatment effect. The magnitude of the estimated difference and its confidence interval provide the more direct information about effect size and precision.
Because this endpoint is continuous and baseline MODS was prespecified in the analysis, ANCOVA estimates an adjusted between-group difference rather than simply comparing unadjusted postoperative means. The interpretation also depends on the evaluable analysis population and on the way missing MODS observations were handled by the endpoint definition.
7. Secondary Endpoint Results
The registry contains posted analyses for additional clinical, physiologic, and time-to-event outcomes. These analyses use the same evaluable-subject definition unless otherwise stated below.
All-Cause Mortality
Hazard ratio for death through 28 days
95% CI: 0.48 to 1.43 · P = 0.50
Cox proportional-hazards model adjusted for baseline MODS.
A hazard ratio of 0.83 corresponds to an estimated instantaneous rate of death approximately 17% lower in the shorter-storage arm relative to the longer-storage reference arm, under the fitted Cox model. It does not mean that mortality was 17% lower at every time point or that 17% fewer participants died.
The 95% CI of 0.48 to 1.43 is wide and includes 1.00, indicating substantial uncertainty around the estimated hazard ratio. The P-value of 0.50 does not quantify the magnitude or clinical importance of the observed hazard ratio.
As with any Cox model, interpretation of a single hazard ratio relies on the proportional-hazards framework. The registry identifies Cox regression as the analysis method but does not provide a separate assessment of that assumption in the posted analysis.
Change in MODS Through 28 Days
| Endpoint | Estimate | 95% CI | P-value | Method |
|---|---|---|---|---|
| Change in MODS from pre-operative baseline through 28 days | Mean difference −0.32 | −0.82 to 0.17 | 0.20 | ANCOVA |
| Ventilation duration through day 28 | Mean difference −0.12 days | −0.62 to 0.37 | 0.75 | Kruskal-Wallis |
| Days alive and ventilator free through postoperative day 28 | Mean difference 0.21 days | −0.59 to 1.00 | 0.71 | Kruskal-Wallis |
Major Postoperative Complications
| Endpoint | Risk difference | 95% CI | P-value | Method |
|---|---|---|---|---|
| Composite of major in-hospital postoperative complications | 0.018 | −0.033 to 0.069 | 0.50 | Fisher exact test |
| Composite of major cardiac events | −0.027 | −0.090 to 0.035 | 0.40 | Fisher exact test |
| Composite of major pulmonary events | −0.018 | −0.059 to 0.023 | 0.41 | Fisher exact test |
| Any mechanical ventilation more than 48 hours post-operation | −0.015 | −0.057 to 0.027 | 0.53 | Fisher exact test |
The major in-hospital postoperative complication composite included death, stroke, myocardial infarction, renal failure, and culture-proven sepsis or septic shock. The major cardiac event composite included death, myocardial infarction, low cardiac output, ventricular tachycardia, and ventricular fibrillation. The major pulmonary event composite included any mechanical ventilation from 48 hours post-operation through day 7, hospital discharge or death, whichever came first, or pulmonary embolism.
For these binary endpoints, the risk difference was calculated as the proportion with the event in the shorter-storage arm minus the proportion with the event in the longer-storage arm. A value of zero represents equal observed event proportions between the two groups.
Organ and Laboratory Measures
| Endpoint | Estimate | 95% CI | P-value | Method |
|---|---|---|---|---|
| Change in serum creatinine to worst postoperative value | Mean difference −0.02 mg/dL | −0.08 to 0.05 | 0.62 | ANCOVA |
| Change in troponin-I to worst postoperative value | Mean difference 1.93 ng/mL | −2.11 to 5.98 | 0.35 | ANCOVA |
| Change in bilirubin to worst postoperative value | Mean difference −0.65 mg/dL | −0.89 to −0.41 | <0.01 | ANCOVA |
| Change in lactate to worst postoperative value | Not reported | Not reported | 0.10 | Kruskal-Wallis |
For creatinine, the analysis adjusted for baseline creatinine. For troponin-I, values recorded as "too low to detect" were recoded as 0 because the median value of the minimum quantitative troponin-I value obtained among participating sites was 0.01. The posted analysis adjusted for baseline troponin-I.
For bilirubin, the estimated difference was −0.65 mg/dL, with a 95% confidence interval of −0.89 to −0.41 and P < 0.01. The registry describes this as a 0.65 mg/dL lower change in bilirubin in the shorter-storage arm after adjustment for baseline bilirubin.
The registry does not provide an estimate or confidence interval for the lactate endpoint, although a Kruskal-Wallis analysis with P = 0.10 is posted.
Postoperative Recovery Time
| Endpoint | Hazard ratio / estimate | 95% CI | P-value | Method |
|---|---|---|---|---|
| Days to first bowel movement through day 28 | HR 1.11 | 0.98 to 1.25 | 0.11 | Cox proportional-hazards model |
| Days to first solid food through day 28 | HR 1.08 | 0.96 to 1.22 | 0.22 | Cox proportional-hazards model |
For both endpoints, the registry states that the mean time to an event is estimated by the area under the survival function. For first solid food, the treatment groups were also compared adjusting for baseline MODS. The longer-storage arm was the reference category.
8. Serious Adverse Events
The registry reports serious adverse events by treatment arm as affected participants divided by the number at risk.
| Treatment arm | Participants with serious adverse events | At risk |
|---|---|---|
| Shorter-storage Red Blood Cell Units | 283 | 538 |
| Longer-storage Red Blood Cell Units | 288 | 560 |
These figures describe participants affected by serious adverse events and the corresponding number at risk in each arm. They are separate from the statistical analyses of the efficacy and physiologic endpoints above.
9. Statistical Methodology
ANCOVA for continuous outcomes
ANCOVA was used for the primary change in MODS endpoint and several secondary continuous outcomes. The central feature was adjustment for a prespecified baseline measurement, such as baseline MODS, baseline creatinine, baseline troponin-I, or baseline bilirubin.
The treatment coefficient represents the adjusted mean difference between storage-duration groups after accounting for the specified baseline measurement.
Covariate adjustment can improve precision when the baseline variable is related to the outcome. In a randomized trial, the adjustment is not intended to replace randomization; rather, it can account for prognostic baseline variation while preserving the treatment-group comparison.
Cox proportional-hazards model
Cox regression was used for all-cause mortality, days to first bowel movement, and days to first solid food. The model expresses the treatment comparison through a hazard ratio while accommodating time-to-event follow-up and censoring.
The hazard ratio is a relative, model-based measure. It is not a risk ratio, an absolute risk difference, or the proportion of patients who experience an event.
Fisher exact test
Fisher exact testing was used for binary postoperative outcomes, including major complications and prolonged mechanical ventilation. The method evaluates the treatment-group association with a binary event using the observed contingency table without relying on a large-sample approximation.
Kruskal-Wallis test
The registry identifies the Kruskal-Wallis test for ventilation duration, days alive and ventilator free, and change in lactate. This is a nonparametric method for comparing distributions across groups and is useful when an outcome may not satisfy the assumptions underlying a conventional parametric comparison.
Risk difference
For the binary outcomes, the reported risk difference was defined as the event proportion in the shorter-storage arm minus the event proportion in the longer-storage arm. A positive value therefore indicates a higher event proportion in the shorter-storage arm, whereas a negative value indicates a lower event proportion in that arm.
10. Statistical Methods Explained
Why was ANCOVA used for the primary endpoint?
The primary outcome was a continuous change in MODS, and the registry specifies adjustment for baseline MODS. ANCOVA is a natural framework for estimating an adjusted mean difference while accounting for baseline differences in the prognostic measurement. The resulting treatment estimate is interpreted on the original MODS-point scale.
What does the primary mean difference of −0.17 mean?
The estimate means that the adjusted mean change in MODS was 0.17 points lower in the shorter-storage group than in the longer-storage reference group. Because the registry defines positive values as being in favor of longer storage duration, the negative estimate points in the opposite direction. The confidence interval, however, extends from −0.60 to 0.26.
Why does the 95% confidence interval matter?
The interval describes the statistical precision of the estimated treatment difference under the specified model and sampling framework. For the primary endpoint, the interval crosses zero, so the observed estimate should not be interpreted as establishing a precise direction or magnitude of a treatment effect.
What does a hazard ratio of 0.83 mean for mortality?
Under the fitted Cox model, an HR of 0.83 represents an estimated instantaneous death rate approximately 17% lower in the shorter-storage arm than in the longer-storage reference arm. It does not mean that the probability of death was exactly 17% lower for every patient or at every follow-up time.
Why are risk differences useful for the complication endpoints?
A risk difference puts the treatment comparison on an absolute probability scale: it describes the difference in event proportions between the two randomized groups. This is directly interpretable as an absolute difference in observed event probability over the specified endpoint window, unlike a relative measure such as a hazard ratio.
Why does Fisher exact testing appear for several endpoints?
The major complication endpoints are binary: each evaluable participant either experienced the defined event or did not. Fisher exact testing provides an exact categorical comparison that is particularly appropriate when event counts may be limited or when the assumptions of large-sample chi-square approximations are less suitable.
11. Interpreting the Secondary Results as a Set
The posted analyses cover several distinct domains: organ dysfunction, mortality, postoperative complications, respiratory outcomes, laboratory measures, and recovery milestones. Because these endpoints are measured on different scales and answer different clinical questions, their estimates should not be collapsed into a single summary statistic.
| Domain | Endpoint | Effect estimate | P-value |
|---|---|---|---|
| Organ dysfunction | 28-day change in MODS | Mean difference −0.32 | 0.20 |
| Mortality | 28-day all-cause mortality | HR 0.83 | 0.50 |
| In-hospital complications | Major postoperative complications | RD 0.018 | 0.50 |
| Cardiac complications | Major cardiac events | RD −0.027 | 0.40 |
| Pulmonary complications | Major pulmonary events | RD −0.018 | 0.41 |
| Ventilation | Ventilation duration | Mean difference −0.12 days | 0.75 |
| Renal laboratory measure | Change in creatinine | Mean difference −0.02 mg/dL | 0.62 |
| Cardiac laboratory measure | Change in troponin-I | Mean difference 1.93 ng/mL | 0.35 |
| Hepatic laboratory measure | Change in bilirubin | Mean difference −0.65 mg/dL | <0.01 |
| Gastrointestinal recovery | Days to first bowel movement | HR 1.11 | 0.11 |
| Nutrition recovery | Days to first solid food | HR 1.08 | 0.22 |
| Ventilator-free time | Days alive and ventilator free | Mean difference 0.21 days | 0.71 |
| Prolonged ventilation | Mechanical ventilation >48 hours | RD −0.015 | 0.53 |
| Metabolic measure | Change in lactate | Estimate not reported | 0.10 |
The bilirubin analysis is the one listed secondary endpoint with a confidence interval entirely below zero and a P-value below 0.01. That finding should still be interpreted in the context of the number of secondary endpoints examined and the absence, in the information reported here, of a stated multiplicity-adjustment framework for this collection of secondary outcomes.
12. Multiple Endpoints and P-values
The trial contains one primary endpoint analysis and numerous secondary endpoint analyses. This distinction matters because each additional hypothesis test creates another opportunity for a small P-value to occur by chance when many hypotheses are examined.
Primary endpoint
The primary analysis was the change in MODS through postoperative day 7, hospital discharge, or death, analyzed using ANCOVA.
Secondary endpoints
The registry reports 14 additional statistical analyses covering mortality, organ dysfunction, complications, laboratory measures, recovery, and ventilation outcomes.
The presence of a P-value below 0.01 for bilirubin should therefore be distinguished from the primary endpoint result. A nominal P-value is not itself a statement about multiplicity-adjusted statistical significance unless an appropriate error-control procedure has been specified.
13. Missing Data and Endpoint Construction
The primary MODS endpoint contains explicit rules for determining when the outcome is available. The follow-up score uses the worst observed value of each component through the specified postoperative window. Death during that period is assigned the worst possible MODS score of 24.
The Glasgow Coma Score rule is particularly important. When the score could not be assessed on at least one day for the specified reasons and the subject did not die during the period, the postoperative MODS score was set to missing and the 7-day change was not computed.
This means that the endpoint is not simply the difference between two routinely observed measurements. Its construction combines worst-case component values, a death rule, and an explicit missing-outcome rule. Understanding those rules is necessary before interpreting the primary mean difference.
14. Sensitivity and Data-Availability Considerations
The registry identifies only four pediatric subjects as having been enrolled. This is a very small subgroup of the overall study population and limits the ability to characterize pediatric-specific effects from these data.
The ALT issue illustrates why a posted registry analysis can contain an outcome for which no numerical treatment estimate is displayed. The absence of a posted result for that endpoint reflects the confidentiality issue described in the registry rather than evidence of a particular treatment effect.
15. Limitations
- Analysis population: the posted statistical analyses are restricted to evaluable subjects rather than simply all 1,481 enrolled participants.
- Primary endpoint construction: MODS uses worst postoperative component values, a death assignment of 24 points, and a specific missingness rule for unscorable Glasgow Coma Scores.
- Secondary endpoint multiplicity: numerous secondary outcomes were analyzed. A nominal P-value should not automatically be interpreted as multiplicity-adjusted evidence.
- Different outcome scales: mean differences, hazard ratios, and risk differences answer different questions and cannot be compared directly by their numerical magnitude.
- Time-to-event assumptions: Cox-model hazard ratios depend on the proportional-hazards framework. The registry identifies the Cox model but does not provide a separate proportional-hazards diagnostic in the posted analysis information.
- Incomplete numerical reporting: the lactate analysis has a posted P-value but no estimate or confidence interval, and the ALT outcome has no posted result because of confidentiality protection.
- Pediatric representation: only four pediatric subjects were enrolled, limiting interpretation of that population.
16. Why This Trial Matters Statistically
RECESS is a useful statistical teaching case because the intervention is biological rather than a conventional drug-versus-placebo comparison, while the analysis still requires the same core principles of randomized clinical-trial inference: preserving the randomized comparison, defining the analysis population, specifying endpoint construction, adjusting for baseline prognostic variables where appropriate, and matching the statistical method to the outcome type.
| Concept | How it appears in RECESS |
|---|---|
| Randomization | Patients were assigned in a randomized parallel-group design to shorter- or longer-storage red blood cell units. |
| Covariate adjustment | ANCOVA analyses adjusted for baseline MODS or the relevant baseline laboratory value. |
| Continuous endpoints | MODS, creatinine, troponin-I, bilirubin, ventilation duration, and ventilator-free days were analyzed as continuous outcomes. |
| Time-to-event endpoints | Mortality, days to first bowel movement, and days to first solid food used Cox proportional-hazards models. |
| Binary endpoints | Major postoperative, cardiac, pulmonary, and prolonged-ventilation outcomes used Fisher exact testing and risk differences. |
| Nonparametric analysis | Kruskal-Wallis testing was used for ventilation duration, ventilator-free days, and lactate change. |
| Confidence intervals | Most posted effect estimates include two-sided 95% confidence intervals. |
| Endpoint construction | The primary MODS endpoint incorporates worst postoperative component values and explicit rules for death and missing neurologic assessment. |
17. Overall Statistical Interpretation
The primary analysis estimated an adjusted mean difference of −0.17 MODS points between shorter- and longer-storage red blood cell units, with a two-sided 95% confidence interval of −0.60 to 0.26 and P = 0.44. The interval includes both directions of a modest difference around the observed estimate.
The 28-day all-cause mortality analysis produced an HR of 0.83 with a 95% CI of 0.48 to 1.43 and P = 0.50. The interval is compatible with a range of relative mortality hazards on either side of the reference value of 1.
The posted secondary analyses span multiple outcome domains. Most estimates have confidence intervals that include the corresponding null value. The bilirubin analysis is distinct, with an adjusted mean difference of −0.65 mg/dL, 95% CI −0.89 to −0.41, and P < 0.01.
The statistical story is therefore multidimensional. The primary MODS endpoint is a baseline-adjusted continuous outcome; mortality and recovery milestones are time-to-event outcomes; complications are binary outcomes; and several laboratory measures are continuous outcomes. Each requires interpretation on its own scale, with confidence intervals providing information about precision and P-values providing hypothesis-testing information rather than effect-size magnitude.
18. Important Interpretation Questions
Does a P-value of 0.44 prove that the two storage strategies are identical?
No. The P-value does not establish equality. The primary confidence interval ranges from −0.60 to 0.26 MODS points, meaning the observed data are compatible with a range of differences around zero under the specified analysis. Demonstrating equivalence or non-inferiority would require a prespecified equivalence or non-inferiority framework and margin.
Does the bilirubin result establish a clinical benefit?
The posted analysis estimates a statistically detectable difference in change in bilirubin, with a 95% confidence interval entirely below zero and P < 0.01. Statistical evidence for a difference in a laboratory endpoint is distinct from evidence about patient-important clinical outcomes such as mortality or major complications.
Can the mortality HR of 0.83 be converted directly into a percentage reduction in 28-day mortality?
No. A hazard ratio is not a risk ratio. It summarizes a relative instantaneous event rate under the Cox model. Converting it directly into a percentage difference in cumulative 28-day mortality would require additional information about the event-time distribution.
Why should the evaluable population be emphasized?
Because the posted analyses are explicitly restricted to subjects meeting the evaluability criteria. The treatment effect therefore describes the analyzed evaluable population rather than automatically describing every enrolled participant.
Why does the primary MODS endpoint use the worst postoperative values?
The registry definition is designed to capture the maximum degree of postoperative organ dysfunction during the specified period. Different MODS components may reach their worst values on different days, so the resulting score represents the worst observed postoperative component profile rather than a single-day snapshot.
19. Related Tutorials
Learn more about the methods used in this trial:
20. Related Calculators
21. Sources
- ClinicalTrials.gov: RECESS — NCT00991341.
- PubMed: PMID 28333704.
- PubMed: PMID 25853746.
- PubMed: PMID 23874777.
- PubMed: PMID 22531338.
- PubMed: PMID 20655807.
Continue with Clinical Biostats
Connect the trial's endpoints and statistical methods to deeper tutorials and practical statistical tools.
22. Record Summary
RECESS provides a useful example of how a randomized clinical trial can combine several statistical frameworks around a common biological question. The primary endpoint used baseline-adjusted ANCOVA to compare change in a composite organ-dysfunction score, while mortality and postoperative recovery milestones were analyzed with Cox proportional-hazards models, binary complications with Fisher exact testing and risk differences, and selected continuous outcomes with Kruskal-Wallis testing. The primary estimate was a mean difference of −0.17 MODS points, with a 95% confidence interval of −0.60 to 0.26 and P = 0.44. Secondary analyses provide additional information across mortality, complications, ventilation, laboratory measures, and recovery endpoints, with the bilirubin analysis reporting a mean difference of −0.65 mg/dL and a 95% confidence interval of −0.89 to −0.41.