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
FLAME was a randomized, parallel-group, triple-masked phase 3 trial evaluating fluorometholone 0.1% ophthalmic suspension versus artificial tears as adjunctive medical therapy for trichiasis surgery in people with trachomatous trichiasis.
| Feature | FLAME |
|---|---|
| Trial name | FLuorometholone as Adjunctive MEdical Therapy for TT Surgery (FLAME) Trial |
| Phase | Phase 3 |
| Status | Completed |
| Therapeutic area | Ophthalmology |
| Condition | Trachomatous Trichiasis (TT) |
| Allocation | Randomized |
| Design model | Parallel |
| Masking | Triple |
| Primary purpose | Treatment |
| Enrollment | 2,410 |
| Interventions | Fluorometholone 0.1% Oph Susp; Artificial Tears |
| Primary endpoint | Incidence of postoperative TT by one year, as determined by trained study team members |
| Primary endpoint time frame | 12 months |
| Primary hypothesis | Superiority |
2. Clinical Question
The central question was whether adjunctive treatment with fluorometholone 0.1% ophthalmic suspension changed the incidence of postoperative trachomatous trichiasis by one year compared with artificial tears following trichiasis surgery.
Population
People enrolled in the phase 3 trial for trachomatous trichiasis, with postoperative follow-up through the registered endpoint time frame.
Intervention
Fluorometholone 0.1% ophthalmic suspension.
Comparator
Artificial tears.
Primary question
Does adjunctive fluorometholone change the incidence of postoperative TT by one year relative to artificial tears?
3. Trial Design
Fluorometholone
- Fluorometholone 0.1% ophthalmic suspension
- Adjunctive medical therapy for TT surgery
Artificial Tears
- Artificial tears
- Comparator adjunctive therapy
The registry reports enrollment of 2,410 participants. The statistical analyses indicate that the number analyzed could differ across follow-up visits because of events such as death or loss to follow-up. This is important when interpreting longitudinal endpoint comparisons: the denominator underlying an analysis may depend on the availability of outcome information at that assessment.
4. Trial Timeline and Registry Status
| Registry milestone | Date |
|---|---|
| Study start | 2021-08-19 |
| Primary completion | 2024-11-30 |
| Registry status | Completed |
The ClinicalTrials.gov record reports 19 posted outcome measures and 18 posted statistical analyses. One primary endpoint has a formal statistical analysis posted with an estimate and confidence interval. The registry also contains analyses covering efficacy, safety/adverse outcomes, patient-reported outcomes, and visual acuity.
5. Primary Endpoint
| Endpoint | Registry definition | Time frame |
|---|---|---|
| Incidence of Postoperative TT by One Year, as Determined by Trained Study Team Members | The primary outcome measure is the postoperative recurrence of trachomatous trichiasis (TT), defined as one or more eyelashes in the upper eyelid touching the globe or evidence of epilation (lash stubs) in the upper eyelid on examination, or a history of repeat trichiasis surgery at any time in the upper eyelid during the one year follow-up period after the baseline surgery. | 12 months |
The endpoint is binary: by one year, an eye either meets the registry definition of postoperative TT recurrence or does not. The posted analysis compares artificial tears with fluorometholone using logistic regression and reports the treatment contrast as a risk difference.
6. Statistical Methodology
Logistic regression for the primary binary endpoint
The primary analysis used logistic regression for the incidence of postoperative TT. The registry reports the effect measure as a risk difference rather than an odds ratio. Thus, the principal reported quantity is the estimated difference in event probability between the two treatment groups.
For the FLAME primary analysis, the registry labels the groups as artificial tears versus fluorometholone. The reported estimate is 0.00, with a two-sided 95% confidence interval from −0.02 to 0.02.
A risk difference is an absolute rather than relative effect measure. An estimate of 0.00 means that the estimated event risks were equal on the reported risk-difference scale. It does not mean that every individual patient had the same outcome, nor does it establish that the two treatments are clinically identical.
Fisher exact test
Fisher exact testing was used for several binary outcomes, including reoperation, lid closure defect, granuloma, intraocular-pressure-related outcome, cataract surgery, and adverse events. Fisher's exact test evaluates the association between treatment group and a binary outcome without relying on the large-sample approximation used by a conventional chi-square test.
Linear regression
Linear regression was used for continuous outcomes including the number of lashes touching, pain level, health utility, and visual acuity. The registry reports these results using mean differences in final values.
Poisson regression
Poisson regression was used for the lashes endpoint, which concerns the number of upper-eyelid lashes touching the globe at one year. Count outcomes require a modeling framework that reflects their discrete nature; the registry specifically identifies Poisson regression for this endpoint.
Proportional-odds regression
The patient-reported satisfaction and cosmetic-outcome analyses used a cumulative logit regression model, normalized here as proportional odds (ordinal) regression. Although the registry's endpoint type field labels these outcomes as binary, the posted method is an ordinal cumulative-logit model. The interpretation therefore depends on the ordered response categories represented in the underlying measure.
Analysis populations and missing follow-up
For most analyses, the registry states that numbers analyzed may differ because different numbers of patients completed each follow-up visit, including because of death or loss to follow-up. The adverse-events analysis specifically notes that randomized patients with surgery were used for statistical analysis. The fellow-eye TT analysis has a more specific eligibility definition, excluding certain eyes based on prior surgery, baseline TT status, or inability to determine TT status at specified visits.
7. Primary Result: Postoperative TT by One Year
Risk difference in postoperative TT
Artificial Tears vs Fluorometholone
95% CI: −0.02 to 0.02 · P = 0.97
Logistic regression · two-sided 95% confidence interval · superiority hypothesis
| Primary endpoint | Estimate | 95% CI | P-value | Method |
|---|---|---|---|---|
| Incidence of postoperative TT by one year | RD 0.00 | −0.02 to 0.02 | 0.97 | Logistic regression |
The estimated risk difference was 0.00, meaning the estimated absolute difference in postoperative TT incidence between artificial tears and fluorometholone was zero on the reported scale. The two-sided 95% confidence interval extends from −0.02 to 0.02, so the reported estimate is compatible with a small absolute difference in either direction within that interval.
The P = 0.97 does not measure the size of the treatment effect. A p-value addresses the compatibility of the observed data with the null hypothesis under the specified statistical model; it is not a probability that the treatment has no effect and is not a measure of clinical importance.
The confidence interval is especially useful here because it places the estimated absolute difference in context. It is relatively narrow around zero compared with the range of possible risk differences, but it does not prove that the two treatments are equivalent. This was a superiority analysis, not a non-inferiority or equivalence analysis, so the interpretation should remain framed around the observed superiority comparison rather than an equivalence claim.
Because the registry notes that analyzed numbers can differ by follow-up completion, the precision of this result also depends on how much outcome information was available for the relevant analysis population. Logistic regression additionally relies on the specification of the model used to obtain the reported effect measure.
8. Secondary Efficacy Results
The registry posts three additional efficacy analyses: entropion, reoperation, and the number of lashes touching the globe at one year. These analyses use different statistical methods because their outcome structures differ.
| Outcome | Time frame | Method | Effect | 95% CI | P-value |
|---|---|---|---|---|---|
| Efficacy Measure 1 - Entropion | 1 year | Logistic regression | RD 0.02 | −0.01 to 0.04 | 0.13 |
| Efficacy Measure 2 - Reoperation | 1 year | Fisher exact test | RD 0.00 | 0.00 to 0.00 | 0.37 |
| Efficacy Measure 3 - Lashes | 1 year | Poisson regression | Mean difference −0.01 | −0.05 to 0.04 | 0.75 |
Entropion
The reported risk difference for entropion was 0.02, with a 95% confidence interval of −0.01 to 0.04 and P = 0.13. The interval spans zero, so the estimate does not distinguish a positive risk difference from a small negative one at the stated confidence level.
Reoperation
The reoperation analysis used Fisher's exact test and reported a risk difference of 0.00, with a 95% confidence interval of 0.00 to 0.00 and P = 0.37. The analysis is explicitly based on a binary outcome and uses an exact categorical-data method.
Lashes
For the number of upper-eyelid lashes touching the globe at one year, the reported mean difference in final values was −0.01, with a 95% confidence interval of −0.05 to 0.04 and P = 0.75. The registry used Poisson regression because the outcome concerns a count.
These endpoints are not interchangeable. Entropion and reoperation are binary outcomes, while lashes is a count outcome. The corresponding analyses therefore target different statistical structures. Across these three posted analyses, the confidence intervals include zero, meaning the reported estimates are compatible with no difference as well as modest differences in either direction.
The absence of a small p-value should not be converted into a statement that the treatments are identical. Conversely, an observed estimate close to zero should not be interpreted as evidence that all patients have the same outcome. Each result describes a treatment-group comparison at the specified endpoint and follow-up time.
9. Safety and Adverse Outcomes
The registry reports ten safety/adverse-outcome analyses. Nine are eye-level or participant-level binary outcomes analyzed using logistic regression or Fisher exact testing; pain level is analyzed as a continuous outcome using linear regression.
| Safety outcome | Time frame | Method | Effect | 95% CI | P-value |
|---|---|---|---|---|---|
| Corneal Opacity | 1 year | Logistic regression | RD −0.01 | −0.05 to 0.03 | 0.53 |
| Overcorrection | 1 year | Logistic regression | RD 0.00 | −0.01 to 0.00 | 0.44 |
| Eyelid Abnormalities | 1 year | Logistic regression | RD −0.01 | −0.02 to 0.00 | 0.17 |
| Lid Closure Defect | 1 year | Fisher exact test | RD 0.00 | 0.00 to 0.00 | 0.50 |
| Granuloma | 1 year | Fisher exact test | RD 0.00 | 0.00 to 0.00 | 0.73 |
| Pain Level | 1 year | Linear regression | Mean difference −0.03 | −0.14 to 0.09 | 0.63 |
| IOP in mmHg | 4 weeks | Fisher exact test | RD 0.00 | 0.00 to 0.01 | 0.58 |
| Cataract Surgery | 1 year | Fisher exact test | RD 0.00 | −0.01 to 0.00 | 0.34 |
| TT in Non-study Eye | 1 year | Logistic regression | RD 0.01 | −0.04 to 0.06 | 0.76 |
| Adverse Events | 1 year | Fisher exact test | RD 0.00 | −0.01 to 0.00 | 0.27 |
Serious adverse events
| Treatment arm | Affected | At risk |
|---|---|---|
| Artificial Tears | 83 | 1206 |
| Fluorometholone | 88 | 1204 |
The registry reports serious adverse events as 83/1206 in the artificial-tears group and 88/1204 in the fluorometholone group. These are affected/at-risk counts and should not be treated as an adjusted treatment effect estimate.
The posted safety analyses generally report risk differences close to zero, with confidence intervals that include zero. The individual estimates should nevertheless be interpreted outcome by outcome because they represent different clinical events, different time frames, and in some cases different units of analysis.
The serious-adverse-event counts provide a direct descriptive comparison of affected participants, while the registry's formal statistical analyses cover the specified adverse-event outcomes. A descriptive count and a modeled risk difference answer related but distinct questions.
10. Patient-Reported Outcomes
Three patient-reported outcomes were analyzed at one year: patient satisfaction, cosmetic outcome, and health utility. The first two used cumulative logit regression models, while health utility used linear regression.
| Outcome | Method | Effect | 95% CI | P-value |
|---|---|---|---|---|
| Patient Satisfaction | Proportional odds (ordinal) regression | RD 0.02 | 0.00 to 0.04 | 0.04 |
| Cosmetic Outcome | Proportional odds (ordinal) regression | RD 0.01 | 0.00 to 0.03 | 0.12 |
| Health Utility | Linear regression | Mean difference −0.01 | −0.22 to 0.21 | 0.96 |
Patient satisfaction
The registry describes this analysis as the difference between groups in the incidence of reporting being very satisfied with the trichiasis surgery outcome at one year. The reported risk difference was 0.02, with a 95% confidence interval of 0.00 to 0.04 and P = 0.04.
The estimated difference was positive on the reported risk-difference scale, with the artificial-tears group listed first in the comparison. The confidence interval reaches zero at its lower boundary. The p-value of 0.04 is below 0.05, but that threshold alone does not establish the magnitude or clinical importance of the difference.
This result also illustrates why endpoint hierarchy matters. Patient satisfaction is a secondary outcome rather than the registered primary endpoint. A nominally small p-value for a secondary analysis should not automatically be interpreted in the same confirmatory framework as a prespecified primary endpoint, particularly when multiple secondary outcomes are examined.
Cosmetic outcome
The cosmetic-outcome analysis reported a risk difference of 0.01, with a 95% confidence interval of 0.00 to 0.03 and P = 0.12. The registry describes the endpoint as the difference between groups in reporting being very satisfied with the cosmetic outcome at one year.
Health utility
Health utility was analyzed with linear regression. The reported mean difference in final values was −0.01, with a 95% confidence interval of −0.22 to 0.21 and P = 0.96. The registry describes this as the mean difference between groups on severity index at one year.
11. Visual Acuity
Mean difference in visual acuity
95% CI: −0.04 to 0.03 · P = 0.78
Linear regression · 1-year assessment
The visual-acuity analysis used final values measured in logMAR. The estimated mean difference was −0.01, with a 95% confidence interval from −0.04 to 0.03. The p-value was 0.78.
The confidence interval spans zero, so the reported estimate is compatible with a small difference in either direction. The estimate is a group-level mean difference, not a statement about an individual patient's visual acuity.
Because the outcome is continuous, the relevant effect measure is a difference in means rather than a risk difference. This is why the registry uses linear regression for visual acuity and reports the result in logMAR units.
12. Statistical Methods Explained
Why was logistic regression used for the primary endpoint?
The primary outcome is binary: postoperative TT is defined according to the registry's clinical criteria over the one-year follow-up period. Logistic regression is designed for binary outcomes and can model the relationship between treatment assignment and the probability of the outcome. The registry reports the resulting treatment comparison as a risk difference.
What does a risk difference of 0.00 mean?
A risk difference of 0.00 means that the estimated absolute probability of the outcome was the same in the two groups on the reported scale. It does not mean that every participant had the same outcome or that the treatments are mathematically identical.
Why is the confidence interval important for the primary result?
The 95% confidence interval of −0.02 to 0.02 describes statistical uncertainty around the estimated risk difference under the analysis framework. It shows that the estimate is centered at zero and places an upper and lower bound on the range of treatment differences compatible with the data at the stated confidence level. It does not describe the range of individual patient responses.
Why was Fisher's exact test used for some outcomes?
Fisher's exact test is a categorical-data method appropriate for comparing binary outcomes between two groups, particularly when the observed cell counts may be small. The registry uses it for several binary safety and efficacy outcomes, including reoperation and granuloma.
Why was Poisson regression used for lashes?
The lashes endpoint concerns the number of upper-eyelid lashes touching the globe. A count outcome has a discrete structure, so a count-regression framework is appropriate. The registry specifically identifies Poisson regression for this endpoint and reports the treatment contrast as a mean difference in final values.
What does proportional-odds regression contribute to the patient-reported outcomes?
The registry identifies a cumulative logit regression model for patient satisfaction and cosmetic outcome, normalized here as proportional odds (ordinal) regression. This framework is designed for ordered response categories and models cumulative probabilities across those categories. The interpretation depends on the ordering and coding of the underlying patient-reported response measure.
13. Interpreting P-values in FLAME
| Outcome | P-value | What the p-value does not tell you |
|---|---|---|
| Primary postoperative TT | 0.97 | It does not quantify the size of the treatment difference. |
| Entropion | 0.13 | It does not prove that there is no difference. |
| Reoperation | 0.37 | It does not measure clinical importance. |
| Patient satisfaction | 0.04 | It does not establish the magnitude or clinical importance of the secondary-endpoint effect. |
| Visual acuity | 0.78 | It does not mean the two treatment groups are identical. |
The p-value should be interpreted together with the effect estimate, confidence interval, endpoint definition, analysis population, and role of the endpoint in the trial. A threshold such as 0.05 is not a substitute for evaluating the size and precision of an effect.
14. Multiplicity and the Endpoint Structure
The registry identifies one primary endpoint and 18 posted statistical analyses. The analyses span multiple efficacy, safety, patient-reported, and additional outcomes. This creates an important distinction between the primary hypothesis and the many secondary comparisons.
Primary endpoint
Postoperative TT by one year was the registered primary outcome and had a formal logistic-regression analysis with an estimate, confidence interval, and p-value.
Secondary analyses
The registry also reports efficacy, safety/adverse, patient-reported, and visual-acuity outcomes, each with its own analysis method and effect estimate.
When many outcomes are examined, a nominal p-value of 0.04 for one secondary outcome should be distinguished from a primary endpoint result. The number of comparisons increases the opportunity for apparently notable findings to occur by chance if each is judged independently at the same threshold.
15. Missing Follow-up and Analysis Populations
The registry repeatedly notes that numbers analyzed may differ because different numbers of patients completed each follow-up visit, including because of death or loss to follow-up. This matters because an endpoint measured at one year cannot necessarily be analyzed using exactly the full enrollment count if some participants lack an evaluable one-year outcome.
The appropriate interpretation therefore depends on the analysis population actually contributing data to each endpoint. The registry also gives a specific definition for the non-study-eye TT analysis: certain fellow eyes are excluded based on previous upper-eyelid TT surgery, baseline TT diagnosis, or inability to determine TT status at the week 4, month 6, and month 12 visits.
16. Cross-Endpoint Statistical Perspective
| Outcome structure | Examples in FLAME | Posted method | Effect measure |
|---|---|---|---|
| Binary | Postoperative TT, entropion, reoperation, corneal opacity | Logistic regression or Fisher exact test | Risk difference |
| Count | Lashes touching | Poisson regression | Mean difference |
| Continuous | Pain level, health utility, visual acuity | Linear regression | Mean difference |
| Ordered patient-reported response | Patient satisfaction, cosmetic outcome | Cumulative logit / proportional odds regression | Risk difference as reported |
This range of methods makes FLAME a useful statistical teaching example. The trial does not have a single universal analysis method because the endpoints are structurally different. The appropriate model follows the scale and distribution of the outcome being analyzed.
17. What the Primary Confidence Interval Says
The primary risk difference was 0.00 with a 95% confidence interval of −0.02 to 0.02. The interval is centered on zero and places the estimated absolute treatment difference within two percentage points in either direction on the reported risk-difference scale.
This is different from saying that the probability of no treatment effect is 95%, or that there is a 95% probability that the true effect lies inside this particular interval. The conventional confidence-interval interpretation is tied to the repeated-sampling procedure used to construct the interval.
18. Effect Measure: Risk Difference Versus Relative Measures
The primary and most secondary binary analyses report risk difference. This is an absolute effect measure. For example, an RD of 0.02 corresponds to a two-percentage-point difference in event probability when expressed on the same probability scale.
Risk difference is especially useful when the clinical question concerns absolute changes in event frequency. Relative measures such as risk ratios or odds ratios answer different questions and can produce very different numerical impressions when baseline event probabilities are low or high.
This is a direct interpretation of the risk-difference scale. It does not imply a particular relative risk, odds ratio, or number needed to treat.
19. Safety: Serious Adverse Events by Arm
Serious adverse events
Artificial Tears vs Fluorometholone
Affected participants / participants at risk
The reported serious-adverse-event counts are similar in absolute magnitude between the two randomized arms: 83 affected participants among 1,206 at risk in the artificial-tears arm and 88 among 1,204 at risk in the fluorometholone arm.
These counts should be kept separate from the posted formal analyses of individual adverse outcomes. Serious adverse events are a broad safety category, whereas outcomes such as corneal opacity, granuloma, cataract surgery, and adverse events represent specific registry-defined measures.
20. Limitations
- Follow-up availability: the registry states that numbers analyzed may differ across visits because of death or loss to follow-up.
- Multiple endpoints: the registry contains one primary endpoint and numerous secondary analyses, so isolated nominal p-values require context.
- Different outcome structures: binary, count, continuous, and ordinal outcomes require different statistical models and effect measures.
- Risk-difference interpretation: risk differences are absolute effects and should not be interpreted as relative risks or odds ratios.
- Secondary endpoint interpretation: a result from a secondary analysis does not automatically carry the same confirmatory status as the registered primary endpoint.
- Model dependence: regression-based estimates depend on the specified statistical model and its assumptions.
- Ordinal-model interpretation: patient satisfaction and cosmetic outcomes were analyzed with cumulative logit models even though the registry's endpoint type field labels them as binary; interpretation should follow the underlying ordered response structure.
- Eye-level versus participant-level outcomes: the registry uses different units of analysis, including study eyes, fellow eyes, and participants. These should not be treated as interchangeable denominators.
21. Why This Trial Matters Statistically
FLAME is a useful example of how a randomized clinical trial can combine several statistical frameworks within one protocol. Its primary endpoint is binary and analyzed using logistic regression, while the posted secondary analyses extend across exact categorical testing, count regression, linear models, and cumulative-logit models.
| Concept | How it appears in FLAME |
|---|---|
| Randomization | Randomized, parallel-group phase 3 design. |
| Blinding | Triple-masked design. |
| Binary outcomes | Postoperative TT and multiple safety outcomes. |
| Logistic regression | Primary postoperative TT and several binary secondary outcomes. |
| Fisher exact test | Reoperation and several safety/adverse outcomes. |
| Poisson regression | Number of upper-eyelid lashes touching the globe. |
| Linear regression | Pain level, health utility, and visual acuity. |
| Proportional odds regression | Patient satisfaction and cosmetic outcome. |
| Risk difference | Primary and most binary secondary treatment contrasts. |
| Confidence intervals | Reported around the primary and secondary effect estimates. |
| Missing follow-up | Registry notes that analyzed numbers can differ by follow-up completion. |
| Multiple outcomes | 19 outcome measures and 18 posted statistical analyses. |
22. Related Tutorials
Learn more about the methods used in this trial:
23. Related Calculators
24. Sources
- ClinicalTrials.gov: NCT04149210 — FLAME.
- PubMed: PMID 41539314.
- PubMed: PMID 40737343.
- PubMed: PMID 39656933.
- PubMed: PMID 39006430.
Continue through the Clinical Biostats knowledge graph
Explore the statistical methods represented in FLAME through focused tutorials and quantitative tools.
25. Record Summary
FLAME provides a useful example of a phase 3 randomized clinical trial in which the primary endpoint is an absolute binary outcome: postoperative trachomatous trichiasis by one year. The primary analysis used logistic regression and reported a risk difference of 0.00 with a two-sided 95% confidence interval of −0.02 to 0.02 and P = 0.97.
The broader statistical record is more diverse than the primary endpoint alone. Secondary analyses use logistic regression, Fisher exact testing, Poisson regression, linear regression, and cumulative logit regression. The registry reports risk differences and mean differences across efficacy, safety, patient-reported, and visual-acuity outcomes. This makes the trial a useful illustration of a central principle in clinical biostatistics: The statistical method should follow the structure of the endpoint and the scientific question being asked.
The primary result should be interpreted through its effect estimate and confidence interval rather than the p-value alone. The secondary results require additional attention to endpoint hierarchy and multiplicity, while the safety results require attention to the distinction between participant-level and eye-level outcomes. Finally, the registry's repeated note that analyzed numbers may differ by follow-up visit is important for understanding the precision and population underlying each reported comparison.