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
PARALLAX was a randomized, quadruple-masked, phase 3 trial comparing sacubitril/valsartan with individualized medical therapy for comorbidities in patients with heart failure with preserved ejection fraction (HFpEF). The registry reports 2,572 enrolled participants and 2 primary endpoints: change from baseline in NT-proBNP at Week 12 and change from baseline in 6-minute walk distance at Week 24.
| Feature | PARALLAX |
|---|---|
| Phase | Phase 3 |
| Population | Patients with heart failure with preserved ejection fraction |
| Design | Randomized, parallel-group, quadruple-masked controlled study |
| Allocation | Randomized |
| Primary purpose | Treatment |
| Enrollment | 2,572 |
| Primary endpoints | Change from baseline in NT-proBNP at Week 12; change from baseline in 6-minute walk distance at Week 24 |
| Intervention | Sacubitril/valsartan (LCZ696) |
| Comparator | Individualized medical therapy (IMT) |
| ClinicalTrials.gov | NCT03066804 |
| Status | Completed |
2. Clinical Question
The primary clinical-statistical question was whether sacubitril/valsartan was superior to individualized medical therapy for comorbidities in reducing NT-proBNP from baseline at Week 12 and improving change from baseline in 6-minute walk distance at Week 24 in patients with HFpEF.
Population
Patients with heart failure with preserved ejection fraction.
Intervention
Sacubitril/valsartan (LCZ696).
Comparator
Individualized medical therapy for comorbidities, represented in the registry as the IMT comparator.
Primary question
Does sacubitril/valsartan produce a superior change in NT-proBNP at Week 12 and 6-minute walk distance at Week 24?
3. Trial Design
LCZ696
- Sacubitril/valsartan
- Placebo to match individualized medical therapy components
- Primary comparison against the IMT comparator
IMT Comparator
- Enalapril
- Valsartan
- Placebo to match sacubitril/valsartan
The registry lists the intervention components as sacubitril/valsartan, enalapril, valsartan, placebo to match sacubitril/valsartan, placebo to match enalapril, and placebo to match valsartan. The overall design is randomized, parallel, quadruple-masked, and treatment-focused.
4. Endpoints
| Endpoint | Registry definition | Time frame |
|---|---|---|
| NT-proBNP | Change from baseline in N-terminal Pro-brain Natriuretic Peptide (NT-proBNP) at Week 12. The endpoint was intended to demonstrate that LCZ696 was superior to individualized medical therapy for comorbidities in reducing NT-proBNP from baseline at Week 12 in patients with HFpEF. | Baseline, Week 12 |
| 6MWD | Change from baseline in 6-minute walk distance (6MWD) at Week 24. The 6-minute walk test was to be performed in accordance with the guidelines of the American Thoracic Society 2002. | Baseline, Week 24 |
Both primary endpoints are continuous outcomes, but they use different effect measures. NT-proBNP was analyzed on the log scale and reported after transformation as a ratio of geometric means, whereas 6MWD was reported as a net mean difference in meters.
5. Analysis Populations
The registry describes analysis populations separately for the two primary endpoints. This distinction is important because the estimand for each endpoint depends on which participants have the required measurements available.
| Endpoint | Analysis population | Statistical implication |
|---|---|---|
| NT-proBNP | Full Analysis Set: patients with non-missing change from baseline in log-transformed NT-proBNP at Week 12. | The mixed model used change from baseline in log-transformed NT-proBNP and then transformed the modeled result back to a geometric mean ratio. |
| 6MWD | Subset of the Full Analysis Set with a baseline 6MWD from 100 meters to 450 meters. | The reported estimate applies to this defined subset rather than automatically to every enrolled participant. |
| KCCQ CSS outcomes | Full Analysis Set with the required baseline and Week 24 KCCQ observations and non-missing covariates. | The analyses incorporated covariate adjustment and longitudinal modeling. |
The registry also identifies intention-to-treat analysis as a concept associated with the posted statistical analyses. For randomized treatment comparisons, the intention-to-treat principle preserves treatment assignment as the basis for comparison rather than redefining groups according to subsequent treatment exposure.
6. Statistical Methodology
Mixed-effects modeling
The registry reports mixed models analysis for both primary endpoints and for the continuous secondary KCCQ and SF-36 outcomes. Mixed-effects models are useful for repeated or longitudinal clinical measurements because they can represent within-patient correlation and incorporate information across the relevant measurement structure.
The exact model specification beyond the registry's reported mixed-model method should not be assumed. The key statistical feature is that observations within a patient are not treated as independent when the model explicitly represents longitudinal structure.
Log-transformation of NT-proBNP
NT-proBNP was analyzed on the logarithmic scale. The registry states that the model fit change from baseline in log-transformed NT-proBNP and produced a difference in mean change on that scale. An exponential transformation was then used to express the result as a geometric mean ratio.
This explains why the reported effect measure is a ratio rather than a difference in the original NT-proBNP units.
Longitudinal binary logistic regression
The two KCCQ deterioration/improvement endpoints were binary outcomes. The registry reports longitudinal binary logistic regression, with covariate adjustment and intention-to-treat analysis identified as associated concepts. The treatment effect was summarized as an odds ratio.
Proportional cumulative odds model
Change from baseline in NYHA functional class at Week 24 was analyzed using a proportional cumulative odds model. NYHA functional class is ordered, so a cumulative-odds approach can account for the ordered nature of the categories rather than treating them as an ordinary continuous measurement.
Confidence intervals and superiority testing
All seven posted statistical analyses report two-sided 95% confidence intervals. The analyses are identified as superiority hypotheses. A confidence interval quantifies statistical uncertainty around the estimated treatment effect; the P-value addresses evidence against the corresponding null hypothesis but does not measure the magnitude or clinical importance of the effect.
7. Primary Results: NT-proBNP at Week 12
The primary NT-proBNP analysis compared sacubitril/valsartan with individualized medical therapy using a mixed-effects model. The endpoint was change from baseline in log-transformed NT-proBNP, expressed after exponential transformation as a geometric mean ratio.
Geometric mean ratio
95% CI: 0.7987–0.8754 · P < 0.0001
Analysis time point: Week 12
| Feature | Reported result |
|---|---|
| Endpoint | Change from baseline in NT-proBNP at Week 12 |
| Comparison | Sacubitril/valsartan vs individualized medical therapy |
| Analysis | Mixed Models Analysis |
| Analysis population | Full Analysis Set with non-missing change from baseline in log-transformed NT-proBNP |
| Effect measure | Geometric mean ratio |
| Estimate | 0.8362 |
| 95% CI | 0.7987–0.8754 |
| P-value | <0.0001 |
The reported geometric mean ratio of 0.8362 means that the modeled geometric mean at Week 12, relative to baseline on the original scale, was estimated to be lower in the sacubitril/valsartan group relative to the individualized-medical-therapy group under the analysis framework used for the log-transformed outcome. Equivalently, the ratio is approximately 0.84, rather than a difference measured in the original NT-proBNP units.
The estimate does not mean that every participant experienced a 16.38% reduction, nor does it establish that an individual patient's NT-proBNP would change by exactly that amount. It is a group-level model-based comparison of geometric means.
The 95% confidence interval of 0.7987–0.8754 describes uncertainty around the estimated ratio. Because the interval is entirely below 1, the reported result is consistent with the superiority hypothesis tested for this endpoint. The P-value of <0.0001 measures evidence against the null hypothesis under the specified model; it does not measure effect size, clinical importance, or the probability that the treatment hypothesis is true.
Why a geometric mean ratio was appropriate
NT-proBNP was modeled after logarithmic transformation. A log transformation is commonly used for positively skewed biomarker measurements because multiplicative differences on the original scale become additive differences on the log scale. Exponentiating the model-based difference returns the estimate to a ratio scale that is easier to interpret as a relative change.
8. Primary Results: 6-Minute Walk Distance at Week 24
The second primary endpoint assessed change from baseline in 6-minute walk distance at Week 24. The registry reports a mixed-model analysis and a net mean difference between sacubitril/valsartan and individualized medical therapy.
Net mean difference in 6MWD
95% CI: −8.5267 to 3.52297 m · P = 0.4164
Analysis time point: Week 24
| Feature | Reported result |
|---|---|
| Endpoint | Change from baseline in 6-minute walk distance at Week 24 |
| Comparison | Sacubitril/valsartan vs individualized medical therapy |
| Analysis | Mixed Models Analysis |
| Analysis population | Subset of the Full Analysis Set with baseline 6MWD from 100 meters to 450 meters |
| Effect measure | Mean Difference (Net) |
| Estimate | −2.4985 meters |
| 95% CI | −8.5267 to 3.52297 meters |
| P-value | 0.4164 |
The estimated net mean difference was −2.4985 meters, meaning that the modeled change from baseline in 6MWD was estimated to be 2.4985 meters lower in the sacubitril/valsartan group than in the individualized-medical-therapy group under the reported analysis.
The estimate does not mean that every participant walked 2.4985 meters less. It is a difference between group-level modeled mean changes. It also does not by itself establish whether an individual change is clinically meaningful.
The 95% confidence interval extends from −8.5267 to 3.52297 meters. Because the interval includes zero, the data are compatible with a range of treatment differences that includes no mean difference. The P-value of 0.4164 does not provide evidence against the null hypothesis at conventional significance thresholds, but it should not be interpreted as proof that the two treatments are identical.
Importantly, this endpoint was analyzed in a defined subset of the Full Analysis Set consisting of participants with baseline 6MWD from 100 meters to 450 meters. The reported estimate therefore pertains specifically to that analysis population.
Why the confidence interval is more informative than the P-value alone
The estimate of −2.4985 meters is only one point within the uncertainty interval. The interval shows that the estimated treatment difference could plausibly lie on either side of zero under the statistical framework used. This makes the confidence interval essential for understanding the precision and direction of the estimate rather than reducing the analysis to a binary significant/not-significant conclusion.
9. Secondary Endpoint Results
The registry reports five secondary endpoint analyses: KCCQ Clinical Summary Score, two binary KCCQ deterioration/improvement measures, NYHA functional class, and SF-36 Physical Component Summary score. These analyses used either mixed-effects models, logistic regression, or a proportional cumulative odds model.
| Secondary endpoint | Method | Effect estimate | 95% CI | P-value |
|---|---|---|---|---|
| Mean change in KCCQ Clinical Summary Score at Week 24 | Mixed Models Analysis | Mean difference 0.5231 | −0.9258 to 1.9720 | 0.4791 |
| Percentage with ≥5-points deterioration in KCCQ CSS at Week 24 | Longitudinal binary logistic regression | OR 0.8993 | 0.6461–1.2518 | 0.5294 |
| Percentage with ≥5-points improvement in KCCQ CSS at Week 24 | Longitudinal binary logistic regression | OR 1.1060 | 0.8287–1.4760 | 0.4938 |
| Change from baseline in NYHA Functional Class at Week 24 | Proportional cumulative odds model | OR 0.9798 | 0.8122–1.1820 | 0.8314 |
| Change from baseline in SF-36 PCS at Week 24 | Mixed Models Analysis | Mean difference −0.1570 | −0.8093 to 0.4953 | 0.6370 |
KCCQ Clinical Summary Score
The estimated mean difference in change from baseline was 0.5231, with a 95% confidence interval of −0.9258 to 1.9720 and P = 0.4791. The confidence interval includes zero, so the reported estimate is compatible with both a small negative and a positive difference in mean change under the analysis framework.
KCCQ deterioration
For the binary endpoint defined as the percentage of patients with at least 5-points deterioration in KCCQ CSS at Week 24, the reported odds ratio was 0.8993 with a 95% CI of 0.6461–1.2518 and P = 0.5294. An odds ratio below 1 corresponds to lower estimated odds of the binary outcome in the sacubitril/valsartan group, but the confidence interval includes 1.
KCCQ improvement
For the endpoint defined as the percentage of patients with at least 5-points improvement in KCCQ CSS at Week 24, the odds ratio was 1.1060 with a 95% CI of 0.8287–1.4760 and P = 0.4938. The point estimate is above 1, but the confidence interval includes 1, indicating substantial uncertainty around the direction and magnitude of the odds ratio.
NYHA Functional Class
The proportional cumulative odds model produced an odds ratio of 0.9798, with a 95% CI of 0.8122–1.1820 and P = 0.8314. Because NYHA functional class is ordered, the cumulative-odds framework models the probability of being at or below specified category thresholds rather than treating the class as a continuous numerical score.
SF-36 Physical Component Summary
The reported mean difference in change from baseline was −0.1570, with a 95% CI of −0.8093 to 0.4953 and P = 0.6370. The interval includes zero and therefore encompasses both a small negative and a small positive mean difference.
10. Statistical Methods Explained
Why was a mixed-effects model used?
Mixed-effects models are designed for longitudinal clinical data in which measurements from the same participant can be correlated. Rather than treating every observation as statistically independent, the model can represent patient-level variation and the repeated-measure structure. PARALLAX used mixed models for the two primary endpoints and for several continuous secondary outcomes.
Why was NT-proBNP analyzed on the log scale?
The registry explicitly states that the NT-proBNP analysis fitted change from baseline in log-transformed NT-proBNP. The resulting difference in mean log change was then exponentiated to obtain a geometric mean ratio. The transformation therefore changes the interpretation from an additive difference in the original biomarker scale to a multiplicative ratio.
What does a geometric mean ratio of 0.8362 mean?
A ratio of 0.8362 represents the estimated relative geometric-mean result under the reported model and transformation. It is not a raw mean difference and should not be interpreted as a difference of 0.8362 units of NT-proBNP. The ratio scale is a direct consequence of modeling the biomarker after logarithmic transformation.
What does an odds ratio of 0.8993 mean?
An odds ratio of 0.8993 means that the estimated odds of the specified binary outcome were lower in the sacubitril/valsartan group relative to individualized medical therapy by the ratio represented by the model. An odds ratio is not the same as a risk ratio or risk difference, particularly when the outcome is not rare.
Why does the 6MWD confidence interval matter?
The 6MWD estimate was −2.4985 meters, but the 95% confidence interval ranged from −8.5267 to 3.52297 meters. Looking only at the point estimate would conceal the uncertainty around it. The interval demonstrates that the estimated treatment difference is compatible with values below and above zero under the reported model.
Why are the KCCQ endpoints analyzed differently?
The continuous KCCQ Clinical Summary Score endpoint was analyzed using a mixed model, whereas the ≥5-points deterioration and ≥5-points improvement endpoints were converted into binary outcomes and analyzed using longitudinal binary logistic regression. The statistical method follows the form of the outcome: continuous measurements require a mean-based model, while binary events can be represented through odds.
Why is NYHA functional class modeled with cumulative odds?
NYHA functional class is categorical and ordered. A proportional cumulative odds model uses the ordering of those categories and evaluates cumulative probabilities across thresholds. This differs from treating NYHA class as a continuous measurement where the numerical spacing between categories would have to carry a specific quantitative interpretation.
11. Understanding the Primary Effect Measures
Geometric mean ratio
Used for the log-transformed NT-proBNP analysis. A value below 1 indicates a lower estimated geometric mean in the treatment comparison under the reported model.
Mean difference
Used for 6MWD, KCCQ CSS, and SF-36 PCS. A value below zero means the modeled mean change was lower in the sacubitril/valsartan group.
Odds ratio
Used for binary KCCQ outcomes and NYHA functional class. A value below 1 indicates lower estimated odds of the modeled outcome in the treatment comparison.
P-value
Quantifies evidence against a specified null hypothesis under the statistical model. It does not quantify treatment benefit, clinical importance, or the probability that a hypothesis is true.
The three quantities answer different questions. The estimate describes the observed model-based treatment contrast; the confidence interval describes statistical uncertainty around it; the P-value summarizes evidence against the null hypothesis. None of the three alone describes the complete clinical meaning of a treatment effect.
12. Serious Adverse Events
The registry reports serious adverse events by treatment arm. There were 186 affected participants among 1,280 at risk in the sacubitril/valsartan group and 191 among 1,284 at risk in the individualized-medical-therapy group, for a total of 377 among 2,564 at risk.
| Treatment group | Participants with serious adverse events | Participants at risk |
|---|---|---|
| Sacubitril/valsartan (LCZ696) | 186 | 1,280 |
| Individualized medical therapy (IMT) | 191 | 1,284 |
| Total | 377 | 2,564 |
These figures describe the number affected and the corresponding number at risk in each group. They are not accompanied in the registry data by a posted inferential comparison of serious adverse-event rates, so they should be read descriptively rather than as a treatment-effect estimate.
13. How to Interpret the Two Primary Endpoints Together
PARALLAX illustrates why a multi-endpoint trial should not be reduced to a single P-value. The two primary endpoints address different dimensions of disease and function: a biomarker outcome at Week 12 and a functional walking outcome at Week 24.
| Primary endpoint | Effect measure | Estimate | 95% CI | P-value |
|---|---|---|---|---|
| NT-proBNP at Week 12 | Geometric mean ratio | 0.8362 | 0.7987–0.8754 | <0.0001 |
| 6MWD at Week 24 | Mean difference | −2.4985 m | −8.5267 to 3.52297 m | 0.4164 |
The statistical behavior of the endpoints is different. The NT-proBNP analysis produced a ratio whose confidence interval lies below 1, whereas the 6MWD analysis produced a mean difference whose confidence interval crosses zero. These are different effect scales, so their numerical estimates should not be compared directly. The appropriate interpretation is endpoint-specific.
14. Multiplicity and Multiple Endpoints
The registry identifies two primary endpoints and five posted secondary endpoint analyses. Multiple endpoint testing raises an important statistical-design question: when several hypotheses are examined, the overall probability of obtaining at least one apparently positive result can exceed the nominal type I error rate unless the analysis plan specifies an appropriate multiplicity strategy.
The ClinicalTrials.gov record identifies the hypotheses as superiority hypotheses and provides the individual confidence intervals and P-values, but it does not provide a detailed multiplicity-adjustment strategy for the complete family of primary and secondary analyses. The individual P-values should therefore be interpreted in the context of the registered endpoint structure rather than treated as independent evidence streams.
This issue is especially important when reading secondary endpoints. A P-value for one secondary outcome does not automatically carry the same confirmatory interpretation as a prespecified primary endpoint in a multiplicity-controlled hierarchy.
15. Covariate Adjustment and Intention-to-Treat Analysis
Covariate adjustment is explicitly identified as an analysis concept for the KCCQ secondary endpoints. Adjustment can improve precision when relevant baseline variables explain part of the outcome variation. The key point is that adjustment changes the statistical model used to estimate the treatment contrast; it does not replace randomization.
Intention-to-treat analysis is also identified in the posted analyses. Under an intention-to-treat framework, randomized participants remain associated with their randomized treatment group for the treatment comparison. This preserves the causal structure created by randomization and avoids redefining treatment groups based on post-randomization behavior.
16. Confidence Intervals: Reading the Primary Results
The 95% CI of 0.7987–0.8754 surrounds the point estimate of 0.8362 on the geometric mean ratio scale. The entire interval is below 1, the null value for a ratio comparison. This indicates that the reported treatment contrast was consistently below the null value across the interval under the specified model.
The 95% CI of −8.5267 to 3.52297 meters surrounds the estimated mean difference of −2.4985 meters. Because zero lies within the interval, the interval includes both negative and positive treatment differences.
For odds-ratio endpoints, 1 is the null value. The reported confidence intervals for KCCQ deterioration, KCCQ improvement, and NYHA functional class all include 1: 0.6461–1.2518, 0.8287–1.4760, and 0.8122–1.1820, respectively.
17. What the P-values Do — and Do Not — Mean
The primary NT-proBNP analysis has a P-value of <0.0001, while the primary 6MWD analysis has a P-value of 0.4164. These values summarize evidence against their respective null hypotheses under the specified statistical models.
A P-value is not the probability that the null hypothesis is true. It is also not a measure of the size of the treatment effect. The NT-proBNP estimate of 0.8362 and its confidence interval are required to understand the magnitude and precision of the treatment contrast; the P-value alone cannot supply that information.
Likewise, the 6MWD P-value of 0.4164 should not be translated into a probability that the treatments are equivalent. The confidence interval is important because it describes the range of treatment differences compatible with the statistical framework and observed data.
18. Statistical Interpretation of the Primary Findings
NT-proBNP
The reported geometric mean ratio was 0.8362, with a 95% CI of 0.7987–0.8754 and P < 0.0001. The estimate was below the ratio null value of 1.
6MWD
The reported mean difference was −2.4985 meters, with a 95% CI of −8.5267 to 3.52297 meters and P = 0.4164. The interval included the mean-difference null value of zero.
Different scales
The primary endpoints cannot be ranked by comparing 0.8362 with −2.4985. The first is a ratio; the second is a difference measured in meters.
Different time points
NT-proBNP was evaluated at Week 12, whereas 6MWD was evaluated at Week 24. Each estimate therefore refers to its specified endpoint and time frame.
19. Limitations
- Different analysis populations: the NT-proBNP endpoint used the Full Analysis Set with non-missing log-transformed change, whereas the 6MWD analysis was restricted to participants with baseline 6MWD from 100 meters to 450 meters.
- Different effect measures: NT-proBNP was reported as a geometric mean ratio, while 6MWD and other continuous outcomes were reported as mean differences and binary outcomes as odds ratios.
- Endpoint multiplicity: the trial has two primary endpoints and multiple secondary analyses. Interpretation of individual P-values should account for the broader endpoint structure.
- Model dependence: mixed-effects, logistic, and cumulative-odds estimates depend on their respective statistical model assumptions.
- Log-transformation: the NT-proBNP result is not a raw-scale mean difference because the analysis was performed on the log-transformed outcome and then exponentiated.
- Binary outcome interpretation: odds ratios should not automatically be interpreted as risk ratios or absolute risk differences.
- Safety interpretation: the serious adverse-event counts are descriptive in the reported information and do not include a corresponding inferential comparison.
- Clinical meaning: statistical evidence and clinical importance are related but distinct questions. A statistically precise estimate does not by itself determine whether a change is clinically meaningful.
20. Why This Trial Matters Statistically
PARALLAX is a useful teaching case because it places several common clinical-trial methods within a single randomized cardiovascular study. The primary endpoints require different statistical representations, while the secondary endpoints demonstrate how the form of the outcome determines the analysis method.
| Concept | How it appears in PARALLAX |
|---|---|
| Randomization | Randomized parallel-group comparison of sacubitril/valsartan with individualized medical therapy |
| Blinding | Quadruple-masked design |
| Superiority testing | Both primary analyses were registered as superiority hypotheses |
| Mixed-effects model | Used for both primary continuous endpoints and several continuous secondary outcomes |
| Log transformation | NT-proBNP was modeled after log transformation |
| Geometric mean ratio | Reported effect measure for the NT-proBNP primary endpoint |
| Mean difference | Reported for 6MWD, KCCQ CSS, and SF-36 PCS |
| Logistic regression | Used for binary KCCQ deterioration and improvement outcomes |
| Odds ratio | Reported for binary KCCQ endpoints and NYHA functional class |
| Cumulative odds modeling | Used for the ordered NYHA functional-class endpoint |
| Confidence intervals | All posted statistical analyses report two-sided 95% confidence intervals |
| Intention-to-treat | Identified as an analysis concept in the posted statistical analyses |
| Covariate adjustment | Identified for the KCCQ secondary analyses |
21. Related Tutorials
Learn more about the methods used in this trial:
22. Related Calculators
23. Sources
- ClinicalTrials.gov: NCT03066804 — PARALLAX.
- Linked publication: PubMed record for PMID 34783839.
Continue through the Clinical Biostats knowledge graph
Connect the endpoints and methods used in PARALLAX to deeper statistical tutorials, calculators, and related clinical-trial analyses.
24. Record Summary
PARALLAX provides a useful example of how a randomized clinical trial can use different statistical models for different endpoint types. The primary NT-proBNP endpoint used a mixed-effects model on a log-transformed outcome and reported a geometric mean ratio of 0.8362 with a 95% CI of 0.7987–0.8754 and P < 0.0001. The primary 6MWD endpoint used a mixed-effects model and reported a net mean difference of −2.4985 meters with a 95% CI of −8.5267 to 3.52297 meters and P = 0.4164.
The secondary analyses further illustrate the distinction between continuous and categorical outcomes: KCCQ CSS and SF-36 PCS were analyzed using mixed models, KCCQ deterioration and improvement used longitudinal binary logistic regression, and NYHA functional class used a proportional cumulative odds model. Across these analyses, the most informative interpretation comes from considering the effect estimate, its confidence interval, the analysis population, the outcome scale, and the statistical model together rather than relying on P-values alone.