Levels Research · Member biomarker outcomes
Most members who started out of range improved their blood markers.
Among Levels members who began with an out-of-range biomarker and completed a follow-up blood panel, 81% improved HbA1c (n=894), 79% improved fasting insulin (n=164), and 84% improved triglycerides (n=370). These are observational results from a self-selected, participant-funded cohort with no control arm.
- Robert Lustig, MD, MSL
- Clinical review · Levels advisor
- Published 24 August 2026
- Data collection closed 2025
- practice-outcomes-v1
- Analysis version
Study population
62,236
consented members
Levels IRB-approved longitudinal study · 3.5 years · 1.15B glucose readings
- HbA1c cohort
- n = 894
- Fasting insulin cohort
- n = 164
- Triglycerides cohort
- n = 370
Cohorts are defined per marker by baseline value. A member may appear in more than one.
Headline findings
Three markers, measured against a defined baseline.
Each cohort below starts from an out-of-range value. “Improved” means the follow-up result came in lower than baseline, so every member counted moved in the right direction.
HbA1c
n = 894
81%
improved from a baseline above 5.7%
- Mean change
- −0.44 pts
- Fell ≥ 0.6 pts
- ~12%
Fasting insulin
n = 164
79%
improved from a baseline above 18.4 µIU/mL
- Mean change
- −11.9 µIU/mL
- Wider cohort (> 10)
- 73% · n = 492
Triglycerides
n = 370
84%
improved from a baseline above 150 mg/dL
- Mean change
- Not reported
- Of those, fell ≥ 30 mg/dL
- 73%
We do not report confidence intervals.
These are descriptive proportions from a single-arm observational cohort, not estimates from a randomized comparison. Intervals would imply a study design this analysis does not have. We give the cohort size for every figure so you can judge precision yourself.
Results table
Every figure, with the cohort it came from.
| Marker | Baseline criterion | Cohort n | % improved | Mean change | Distribution detail |
|---|---|---|---|---|---|
| HbA1c | > 5.7% (prediabetes range) | 894 | 81% | −0.44 pts | ~12% fell by 0.6 pts or more |
| Fasting insulin | > 18.4 µIU/mL | 164 | 79% | −11.9 µIU/mL | Marked insulin resistance at baseline |
| Fasting insulin | > 10 µIU/mL | 492 | 73% | −4.3 µIU/mL | Wider elevated-insulin cohort; includes the row above |
| Triglycerides | > 150 mg/dL | 370 | 84% | Not reported | 73% of those who improved fell by 30 mg/dL or more |
Methodology
How these numbers were produced.
Version practice-outcomes-v1 identifies this analysis of the closed study dataset. The data behind it never changes. If we revise the selection rules or the definition of improvement, the version increments and the changelog records what changed.
Who is included
Consented members of the Levels IRB-approved longitudinal study who ordered at least two Levels Labs blood panels, and whose first panel fell outside the reference range for the marker in question. Levels does not randomize, assign, or select anyone. Members opt in and pay for their own testing.
Baseline and follow-up
Baseline is the first qualifying panel after joining. Follow-up is the most recent panel at least 90 days later. Where a member has more than two panels, we compare only the first and the most recent. We never pick an intermediate draw to land a better endpoint.
What counts as improved
Any follow-up value lower than baseline, for markers where lower is better. There is no minimum change threshold, so a member who fell 0.1 points counts the same as one who fell 1.0. That is why every percentage on this page carries its mean change and distribution detail.
Suppression and privacy
We report a figure only where the cohort reaches the minimum aggregate size of 11. Every result is a de-identified aggregate. Nothing on this page names an individual member or lets you work one out.
A closed dataset
Data collection ended in 2025, so these figures are a manually versioned snapshot of the closed study dataset. They are not computed from or kept in sync with customer outcomes reports. If we ever revise the analysis, the version increments and the previous version stays live at a dated URL.
Limitations
What these numbers cannot tell you.
This is an observational analysis of a self-selected population. It shows what happened to members who chose Levels and tested twice. It does not establish that Levels caused the change.
No control arm
Nothing here is a comparison
There is no untreated group measured alongside this cohort. Some of the change would likely have occurred anyway through regression to the mean, seasonal variation, or care received elsewhere.
Self-selection
Members chose to be here
Participants paid for their own membership and testing. They were, on average, more motivated and more resourced than a general or employer-wide population, and results should not be extrapolated to one.
Engagement bias
Only re-testers are counted
A member needs a second blood panel to appear in any cohort. Members who disengaged early are absent, and there is no reason to assume their results resemble those shown here.
This report summarizes changes observed during program participation. It is not a controlled study and does not establish causation. Results are not a prediction of individual outcomes and are not medical advice.
Comparison to literature
Set against the published record.
Comparators are population and natural-history data, not matched controls. They show which way an untreated or general population tends to move. They do not support a causal claim about Levels.
| Comparison | Levels cohort | Published comparator | Citation |
|---|---|---|---|
| HbA1c trend without intervention | −0.44 pts mean (n=894) | US population HbA1c distribution and drift in adults with prediabetes | NHANES, Huang 2023 |