This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record. Reported estimates are kept separate from statistical interpretation, which is labelled as such throughout.
1. Trial at a Glance
AWARD-10 was a double-blind, randomized, three-arm parallel-group phase 3 trial asking whether two doses of dulaglutide lowered HbA1c more than placebo over 24 weeks. Its most distinctive statistical feature is that the primary HbA1c comparison was registered under two estimands, one that uses all data including post-rescue values and one that excludes them.
Headline estimates are least-squares mean differences in HbA1c change (percentage units) under the treatment-regimen estimand.
| Feature | AWARD-10 |
|---|---|
| Registry title | A Study of Dulaglutide (LY2189265) in Participants With Type 2 Diabetes Mellitus |
| Phase | Phase 3 |
| Condition | Type 2 Diabetes Mellitus |
| Design | Randomized, parallel-group, double-masked, treatment purpose |
| Arms | 3 (dulaglutide 1.5 mg, dulaglutide 0.75 mg, placebo) |
| Interventions listed | Dulaglutide; Placebo; SGLT2 inhibitor; Metformin |
| Enrollment | 424 |
| Primary endpoints | Change from baseline in HbA1c at 24 weeks, under a treatment-regimen estimand and under an efficacy estimand |
| Study start / primary completion | 2015-11 / 2017-02 |
| Status | Completed; results posted |
| ClinicalTrials.gov | NCT02597049 |
| Lead sponsor | Eli Lilly and Company (industry) |
2. Clinical Question
The trial asked whether adding dulaglutide at either of two doses improved glycemic control, measured as the change in HbA1c from baseline to week 24, compared with placebo in people with type 2 diabetes. Because the registry lists an SGLT2 inhibitor and metformin among the interventions, and because the primary model adjusted for SGLT2 inhibitor dose and metformin use, these agents formed the background regimen against which dulaglutide and placebo were compared.
Population
Participants with type 2 diabetes mellitus receiving an SGLT2 inhibitor, with metformin use recorded as a model covariate.
Intervention
Dulaglutide 1.5 mg or dulaglutide 0.75 mg, in addition to the background SGLT2 inhibitor regimen.
Comparator
Placebo, in addition to the same background SGLT2 inhibitor regimen, under double masking.
Primary question
Is the 24-week change in HbA1c larger with each dulaglutide dose than with placebo, whether or not post-rescue data are counted?
3. Trial Design
Dulaglutide 1.5 mg
- Dulaglutide 1.5 mg
- Background SGLT2 inhibitor
- Metformin use recorded
Dulaglutide 0.75 mg
- Dulaglutide 0.75 mg
- Background SGLT2 inhibitor
- Metformin use recorded
Placebo
- Matching placebo
- Background SGLT2 inhibitor
- Metformin use recorded
4. Estimands and Analysis Populations
An estimand is a precise statement of the treatment effect a trial is trying to estimate. In diabetes trials, the key complication is rescue medication: participants whose glucose rises too high may receive additional glucose-lowering therapy, which changes their later HbA1c values. AWARD-10 registered the same HbA1c comparison under two estimands that handle rescue differently.
| Analysis | Analysis population as described in the registry |
|---|---|
| HbA1c, treatment-regimen estimand | All randomized participants who received at least one dose of study medication and had evaluable data |
| HbA1c, efficacy estimand | All randomized participants who received at least one dose of study drug and had a baseline and post-baseline value, excluding values collected after rescue medication |
| Body weight | All participants who received at least one dose of study drug |
| Fasting serum glucose | All participants who received at least one dose of study drug and had baseline and at least one post-baseline value |
| 6-point SMPG profile | All participants who received at least one dose of study drug and had a baseline SMPG value and at least one post-baseline SMPG value |
| Fasting glucagon | All participants who received at least one dose of study drug with non-missing baseline values and at least one post-baseline value at the specified time point |
| Safety (serious adverse events) | Participants at risk: 142, 141 and 140 in the three arms |
Two differences are worth noticing. First, the populations are defined by dosing and data availability rather than by randomization alone. Second, the treatment-regimen and efficacy estimands differ not only in which observations are used but also, potentially, in which participants contribute: a participant whose only post-baseline values were collected after rescue would contribute to the first analysis but not to the second.
5. Endpoints
Primary endpoints
| Endpoint (registry wording) | Time frame | Registered analysis |
|---|---|---|
| Change From Baseline in Hemoglobin A1c (HbA1c) at 24 Weeks (Treatment-regimen Estimand) | Baseline, Week 24 | LS mean change adjusted by treatment, country, SGLT2 inhibitor dose, metformin use and treatment-by-visit interaction as fixed effects, baseline HbA1c as a covariate and participant as a random effect, via MMRM; uses all data including post-rescue data |
| Change From Baseline in the HbA1c at 24 Weeks (Efficacy Estimand) | Baseline, Week 24 | Same MMRM specification; excludes post-rescue data |
Secondary endpoints with posted statistical analyses
| Endpoint | Time frame | Unit | Method |
|---|---|---|---|
| Change From Baseline in Body Weight at 24 Weeks | Baseline, Week 24 | kg | Mixed models analysis |
| Change From Baseline in Fasting Serum Glucose (Central Laboratory) at 24 Weeks | Baseline, Week 24 | mg/dL | ANCOVA |
| Change From Baseline in 6-Point Self-Monitored Plasma Glucose (SMPG) Profile at 24 Weeks | Baseline, Week 24 | mg/dL | Mixed models analysis |
| Change From Baseline in Fasting Glucagon at 24 Weeks | Baseline, Week 24 | pmol/L | ANCOVA |
| Percentage of Participants With HbA1c <7% | 24 Weeks | Percentage of participants | Logistic regression |
6. Primary Results: Treatment-Regimen Estimand
Under the treatment-regimen estimand, both dulaglutide doses were compared with placebo using a mixed model for repeated measures. The effect measure is the difference in least-squares (LS) mean change in HbA1c from baseline to week 24, in HbA1c percentage units. All comparisons were registered as superiority tests with two-sided 95% confidence intervals.
Dulaglutide 1.5 mg vs placebo
95% CI: −0.97 to −0.61 · P <.001
LS mean difference in HbA1c change from baseline at week 24 (treatment-regimen estimand)
Dulaglutide 0.75 mg vs placebo
95% CI: −0.84 to −0.49 · P < .001
LS mean difference in HbA1c change from baseline at week 24 (treatment-regimen estimand)
What the estimate means. A difference of −0.79 means that, after adjustment for baseline HbA1c and the other model terms, the model-estimated mean HbA1c at week 24 fell by 0.79 percentage points more with dulaglutide 1.5 mg than with placebo. For dulaglutide 0.75 mg the corresponding figure is 0.66 percentage points. Because rescue medication was counted, this is the effect of being assigned to and following each regimen in practice, including any rescue therapy that followed.
What it does not mean. It is not a 79% or 66% relative reduction, and it is not the HbA1c change experienced by any individual participant. It is also not the absolute change within the dulaglutide arm; it is a between-arm difference in mean change, so part of any within-arm fall may also have occurred with placebo.
Precision. Both intervals lie entirely below zero. The 1.5 mg interval (−0.97 to −0.61) and the 0.75 mg interval (−0.84 to −0.49) are of similar width, reflecting similar arm sizes. The intervals overlap substantially, so these results alone do not establish that 1.5 mg is more effective than 0.75 mg; a direct dose-versus-dose comparison would be needed for that claim.
Why the p-value is not the effect size. P <.001 indicates that a difference this large would be very unlikely if the true mean difference were zero. It says nothing about whether 0.79 or 0.66 percentage points is clinically meaningful; that judgment rests on the estimate and its interval.
Cautions. MMRM gives valid estimates when missing data are missing at random given the observed data; that assumption cannot be verified from the data themselves. The trial made two dose comparisons under two estimands, and the registry does not describe how the type I error was controlled across these four tests, so the P-values should be read as reported rather than as individually adjusted.
7. Primary Results: Efficacy Estimand
The efficacy estimand used the same MMRM specification but excluded values collected after rescue medication. It therefore targets the effect of the randomized treatments when taken as directed without rescue.
Dulaglutide 1.5 mg vs placebo
95% CI: −1.00 to −0.64 · P <.001
LS mean difference in HbA1c change from baseline at week 24 (efficacy estimand)
Dulaglutide 0.75 mg vs placebo
95% CI: −0.86 to −0.51 · P <.001
LS mean difference in HbA1c change from baseline at week 24 (efficacy estimand)
What the estimate means. Under the efficacy estimand, dulaglutide 1.5 mg lowered mean HbA1c by an estimated 0.82 percentage points more than placebo at week 24, and dulaglutide 0.75 mg by 0.69 percentage points more. These estimates describe a hypothetical scenario in which rescue medication is not used.
Why it is slightly larger than the treatment-regimen estimate. The efficacy estimates (−0.82 and −0.69) are a little further from zero than the treatment-regimen estimates (−0.79 and −0.66). This is the expected direction: rescue medication is more likely to be needed in the placebo arm, and counting post-rescue values tends to pull the placebo arm's HbA1c down and narrow the between-arm gap. The small size of the difference suggests that rescue did not strongly change the comparison, although the registry does not report how often rescue occurred.
What it does not mean. The efficacy estimand is not a "truer" effect; it answers a different question. Removing post-rescue data does not create a randomized comparison of participants who would never need rescue. Its validity depends on the assumption that the excluded values can be treated as missing at random given the observed history.
Precision and p-values. Both intervals again exclude zero, with widths comparable to the treatment-regimen analysis. The P-values (<.001) confirm incompatibility with a null difference but do not measure how large the benefit is.
8. Secondary Endpoint Results
The registry posts mean differences with 95% confidence intervals for four continuous secondary endpoints and P-values only for the HbA1c <7% responder endpoint. None of these were designated as confirmatory, and the registry does not describe a multiplicity procedure covering them.
Body weight
| Comparison | Estimand | Mean difference (kg) | 95% CI | P-value |
|---|---|---|---|---|
| 1.5 mg vs placebo | Treatment-regimen | −0.9 | −1.8 to −0.1 | 0.027 |
| 0.75 mg vs placebo | Treatment-regimen | −0.5 | −1.3 to 0.4 | 0.264 |
| 1.5 mg vs placebo | Efficacy | −0.8 | −1.7 to 0.0 | 0.059 |
| 0.75 mg vs placebo | Efficacy | −0.3 | −1.2 to 0.5 | 0.435 |
Body weight is the clearest example on this page of why a confidence interval is more informative than a significance label. The 1.5 mg treatment-regimen estimate (−0.9 kg) has an interval that just excludes zero (P = 0.027); the 1.5 mg efficacy estimate (−0.8 kg) is almost the same size, but its interval reaches 0.0 and P = 0.059. The two analyses tell essentially the same story about the likely effect; they fall on opposite sides of 0.05 because of small differences in the estimate and the data used. All four intervals are wide relative to their point estimates, each spanning 1.7 kg, so the weight effect is estimated imprecisely.
Fasting serum glucose (central laboratory)
| Comparison | Estimand | Mean difference (mg/dL) | 95% CI | P-value |
|---|---|---|---|---|
| 1.5 mg vs placebo | Treatment-regimen | −24.7 | −30.8 to −18.6 | < .001 |
| 0.75 mg vs placebo | Treatment-regimen | −19.6 | −25.7 to −13.5 | <0.001 |
| 1.5 mg vs placebo | Efficacy | −26.6 | −32.7 to −20.6 | <0.001 |
| 0.75 mg vs placebo | Efficacy | −20.7 | −26.7 to −14.6 | <0.001 |
Fasting serum glucose was analysed with ANCOVA rather than a repeated-measures model. The pattern mirrors the HbA1c results: larger differences for 1.5 mg than for 0.75 mg, slightly larger differences under the efficacy estimand, and intervals well away from zero.
6-point self-monitored plasma glucose (SMPG) profile
| Time point in profile | 1.5 mg vs placebo (mg/dL) | 0.75 mg vs placebo (mg/dL) |
|---|---|---|
| Pre-morning meal | −19.7 (−25.7 to −13.8); P < .001 | −15.2 (−21.2 to −9.2); P < .001 |
| 2-hour postprandial (morning) | −24.5 (−33.8 to −15.3); P < .001 | −21.1 (−30.3 to −11.8); P < .001 |
| Pre-midday meal | −18.3 (−26.4 to −10.2); P < .001 | −14.3 (−22.5 to −6.2); P < .001 |
| 2-hour postprandial after midday meal | −19.0 (−28.8 to −9.3); P < .001 | −12.8 (−22.5 to −3.1); P = 0.010 |
| Pre-evening meal | −22.7 (−30.7 to −14.7); P < .001 | −22.5 (−30.7 to −14.4); P < .001 |
| 2-hour postprandial after evening meal | −22.1 (−31.4 to −12.9); P < .001 | −16.7 (−26.0 to −7.4); P < .001 |
The SMPG profile generates twelve separate comparisons from a single endpoint. Each is a point on a within-day glucose curve measured in the same participants, so the tests are strongly correlated rather than independent confirmations. Postprandial intervals are generally wider than pre-meal intervals, consistent with the greater variability of post-meal glucose readings.
Fasting glucagon
| Comparison | Estimand | Mean difference (pmol/L) | 95% CI | P-value |
|---|---|---|---|---|
| 1.5 mg vs placebo | Treatment-regimen | −1.2 | −2.3 to −0.1 | 0.032 |
| 0.75 mg vs placebo | Treatment-regimen | −0.6 | −1.7 to 0.5 | 0.273 |
| 1.5 mg vs placebo | Efficacy | −1.3 | −2.4 to −0.2 | 0.023 |
| 0.75 mg vs placebo | Efficacy | −0.6 | −1.7 to 0.6 | 0.320 |
For fasting glucagon, the 1.5 mg intervals sit just below zero while the 0.75 mg intervals include it. With many secondary tests and no stated multiplicity adjustment, borderline results such as P = 0.032 and P = 0.023 are best treated as descriptive signals.
Percentage of participants with HbA1c <7% at 24 weeks
| Comparison | Estimand | Method | P-value |
|---|---|---|---|
| 1.5 mg vs placebo | Treatment-regimen | Logistic regression | < .001 |
| 0.75 mg vs placebo | Treatment-regimen | Logistic regression | < .001 |
| 1.5 mg vs placebo | Efficacy | Logistic regression | < .001 |
| 0.75 mg vs placebo | Efficacy | Logistic regression | < .001 |
For the responder endpoint the registry posts P-values from logistic regression but no odds ratio or confidence interval. The P-values indicate that the proportions reaching HbA1c <7% differed from placebo, but without an effect estimate the size of that difference cannot be judged from the posted analyses. The registry's population description for this endpoint reads "All participants who received at least one dose of Dulaglutide with HbA1c <7% at Week 24", which, taken literally, would describe only responders; it is best read as a registry wording issue rather than as the analysed denominator.
9. Safety: Serious Adverse Events
| Arm | Participants with serious adverse events / at risk |
|---|---|
| Dulaglutide 1.5 mg | 5 / 142 |
| Dulaglutide 0.75 mg | 3 / 141 |
| Placebo | 5 / 140 |
Serious adverse events were infrequent in all three arms. With counts of 5, 3 and 5, the trial had very little power to detect differences in serious adverse event rates; the absence of an apparent imbalance is not evidence of equivalence. No formal statistical comparison of serious adverse events is posted in the registry, which is typical for safety data in a trial of this size.
10. Statistical Methodology
Mixed model for repeated measures (MMRM)
The primary HbA1c analyses, body weight and the SMPG profile used a mixed-model approach. For HbA1c, the registry specifies fixed effects for treatment, country, SGLT2 inhibitor dose, metformin use and treatment-by-visit interaction, baseline HbA1c as a covariate, and participant as a random effect. The treatment effect at week 24 is read from the treatment-by-visit term at that visit.
where i indexes participants, j indexes visits, HbA1ci0 is the baseline value, bi is the participant random effect that links repeated measurements from the same person, and εij is residual error. The reported estimate is the difference in LS means between arms at week 24.
Analysis of covariance (ANCOVA)
Fasting serum glucose and fasting glucagon were analysed with ANCOVA, which models the week-24 change as a function of treatment and baseline value (with other covariates as specified in the protocol). Unlike MMRM, ANCOVA uses a single post-baseline measurement per participant, so participants without a value at the analysis time point contribute through whatever imputation or last-value convention the protocol specified; the registry does not describe that convention.
Logistic regression
The binary HbA1c <7% endpoint was analysed with logistic regression, which models the log-odds of reaching the target as a function of treatment and covariates. The natural effect measure is an odds ratio, which is not posted.
Covariate adjustment and stratification factors
Country, SGLT2 inhibitor dose and metformin use appear as fixed effects in the primary model. Adjusting for factors that predict the outcome improves precision and, when those factors were used to balance randomization, keeps the analysis consistent with the design. The registry does not state whether randomization itself was stratified by these factors.
Two-sided superiority testing
Every posted analysis is labelled as a superiority test with a two-sided 95% confidence interval. For a superiority comparison, the null hypothesis is a mean difference of zero, and a two-sided interval that excludes zero corresponds to a two-sided P-value below 0.05.
Missing data
MMRM uses all available visits for each participant and yields unbiased estimates if data are missing at random, meaning the probability of a missing value depends only on observed information. Under the efficacy estimand, post-rescue values are deliberately set aside and are therefore handled as missing in this way.
11. Multiplicity and Endpoint Hierarchy
AWARD-10 generated many tests: two dose comparisons for each of two primary estimands, plus dozens of secondary comparisons. The registry does not describe a gatekeeping order, alpha split between doses, or other multiplicity procedure. Readers should therefore distinguish carefully between the roles of the analyses.
| Analysis | Role | Interpretation |
|---|---|---|
| HbA1c, treatment-regimen estimand | Primary | Main comparison of each dose with placebo |
| HbA1c, efficacy estimand | Primary | Same comparison under a different handling of rescue; supports rather than duplicates the first |
| Body weight, fasting serum glucose, SMPG, glucagon | Secondary | Supportive; P-values near 0.05 should be read descriptively |
| HbA1c <7% | Secondary | P-values only; effect size not posted |
| Serious adverse events | Safety | Descriptive counts |
12. Statistical Methods Explained
Why was a mixed model for repeated measures used for HbA1c?
HbA1c was measured at several visits in each participant. MMRM uses all of those measurements, accounts for the correlation between visits in the same person through the participant random effect, and estimates the week-24 difference without discarding participants who miss a later visit. This makes it more efficient and less biased under missing-at-random assumptions than an analysis restricted to completers.
What does an LS mean difference of −0.79 actually mean?
It is the model-estimated difference between arms in the average HbA1c change from baseline to week 24, after adjusting for baseline HbA1c, country, SGLT2 inhibitor dose and metformin use. The negative sign means greater HbA1c lowering with dulaglutide 1.5 mg. It is expressed in percentage units of HbA1c, so it is an absolute difference of 0.79 percentage points, not a 79% or relative reduction.
Why are there two primary endpoints for the same measurement?
Because the two estimands answer different questions. The treatment-regimen estimand asks what happens when participants are assigned to a regimen and rescue is used as needed; the efficacy estimand asks what the drug does when taken as directed without rescue. Reporting both lets readers see whether the conclusion depends on how rescue is handled. In AWARD-10, the estimates differ by only 0.03 percentage points for each dose, so it does not.
Why did some secondary endpoints use ANCOVA rather than MMRM?
ANCOVA analyses a single post-baseline value adjusted for baseline and is often used for endpoints measured at fewer visits, such as a central-laboratory fasting glucose or a glucagon assay at a specified time point. MMRM is preferred when repeated measurements are available and the pattern over visits matters.
Why does a body-weight interval ending at 0.0 give P = 0.059?
The posted limits are rounded to one decimal. An upper limit shown as 0.0 means the unrounded limit is very close to zero and, given P = 0.059, slightly above it. The result sits at the boundary of conventional significance, which is exactly why the estimate (−0.8 kg) and interval are more informative than a significant/not-significant label.
Why does the HbA1c <7% endpoint have a P-value but no effect size?
The registry posted only the logistic regression P-values. Logistic regression produces an odds ratio with a confidence interval; without it, a reader can conclude only that the proportions reaching the target differ from placebo, not by how much.
13. Important Limitations and Interpretation Issues
- Modified analysis populations: efficacy analyses required at least one dose and evaluable data, and one randomized participant who was never dosed was excluded; the analyses are therefore not strictly all-randomized.
- Missing-at-random assumption: both MMRM and the exclusion of post-rescue data rely on the assumption that missing values depend only on observed information.
- Rescue frequency not reported in the posted analyses: the similarity of the two estimands is reassuring, but the proportion of participants rescued in each arm is not shown alongside the estimates.
- Multiplicity: two doses, two estimands and many secondary endpoints were tested without a multiplicity procedure described in the registry; borderline secondary P-values should be read descriptively.
- No arm-level summaries in this analysis: between-arm differences are shown without the within-arm LS mean changes, so the absolute amount of HbA1c lowering in each arm is not presented here.
- Dose comparison: both doses were compared with placebo, not with each other; overlapping intervals mean a dose-response conclusion requires a direct comparison.
- Duration: the primary time point was 24 weeks, which does not address durability of glycemic control or long-term safety.
- Safety precision: with 13 participants with serious adverse events in total, the trial could not detect modest differences in serious adverse event rates.
- Generalizability: results apply to people with type 2 diabetes on a background SGLT2 inhibitor regimen, with or without metformin, meeting the trial's eligibility criteria.
14. Why This Trial Matters Statistically
AWARD-10 is a compact teaching case for how modern diabetes trials handle rescue medication and repeated measurements, and for why reading intervals matters more than counting significant P-values.
| Concept | How it appears in AWARD-10 |
|---|---|
| Estimands | The same HbA1c comparison registered under treatment-regimen and efficacy estimands |
| Mixed-effects model | MMRM with participant random effect for HbA1c, body weight and SMPG |
| ANCOVA | Baseline-adjusted analysis of fasting serum glucose and fasting glucagon |
| Logistic regression | Analysis of the proportion reaching HbA1c <7% |
| Confidence intervals | Show the precision of each mean difference and reveal borderline results (e.g., weight interval ending at 0.0) |
| P-values | Primary P-values <.001; secondary P-values near 0.05 illustrate why significance labels are fragile |
| Multiple comparisons | Two doses, two estimands and many secondary endpoints |
| Analysis populations | Dosed participants with evaluable data; one randomized participant not dosed |
15. Related Tutorials
Learn more about the methods used in this trial:
16. Related Calculators
17. Sources
- ClinicalTrials.gov registry record: NCT02597049 — A Study of Dulaglutide (LY2189265) in Participants With Type 2 Diabetes Mellitus.
- PubMed: PMID 29483060.
- PubMed: PMID 36894938.
Explore the methods behind the results
Each trial analysis connects its endpoints and methods to tutorials that explain them and calculators that let you apply them.
18. Summary
AWARD-10 compared dulaglutide 1.5 mg and 0.75 mg with placebo in people with type 2 diabetes on a background SGLT2 inhibitor regimen. Under both the treatment-regimen and efficacy estimands, each dose produced a greater 24-week HbA1c reduction than placebo, with LS mean differences of −0.79 and −0.66 (treatment-regimen) and −0.82 and −0.69 (efficacy), all with 95% confidence intervals excluding zero. Secondary glycemic endpoints pointed in the same direction, while body-weight and glucagon differences were smaller and less precisely estimated. The most useful reading combines the estimand being targeted, the size and precision of each mean difference, the assumptions behind the mixed model, and the number of comparisons made.