Guide series, Part 5 of 5 · What gets measured gets managed
How would Nassim Nicholas Taleb use HbA1c and Time in Range to assess Type 1 diabetes risk?
CareLink, Clarity, LibreView and Glooko all default to a 14-day report. That default is an ensemble approach to a profoundly individual disease. In a population, the 14-day average is informative. For one person whose lifetime risk is being decided, it is roughly the precision of a horoscope.
Population probability is not your probability
If you flip a coin 1,000 times, the population result is 50:50. If you flip the coin once, the population result tells you almost nothing about whether your single flip will be heads or tails. The two probabilities are not the same.
This distinction has a name. A process is ergodic if the time-average for one individual converges on the ensemble-average across many individuals. Coin flips are roughly ergodic over enough time. Type 1 diabetes is not.
For a person with T1D, your risk does not only average across “people like you.” It also accumulates across your own 50 to 70 years of glycaemic exposure, your own hypos, your own DKA episodes, your own cardiovascular trajectory. Population means are useful for designing trials and writing guidelines, and the average HbA1c hazard ratio (Bebu 2020) does apply to individuals on average. The error is treating the population mean as if it were your personal probability, when the cumulative-exposure path you are on is what actually delivers the outcome.
Nassim Taleb’s revolver example. Imagine 100 people each play one round of Russian roulette with a six-shooter holding one bullet. Roughly five out of six walk away. The ensemble survival rate is ~83%. For each individual player, that ensemble rate tells you nothing about whether they are the sixth. The two probabilities are not the same number applied to different things; they are different questions. T1D is not roulette, but the structural lesson holds: when the consequences of being on the wrong side of the distribution are catastrophic and irreversible, you have to plan for your individual path, not the population average. Taleb first used this example in Fooled by Randomness (2001) to illustrate how easily ensemble statistics mislead individuals; he returned to it in Skin in the Game (2018) to argue that this distinction is structurally invisible to people interpreting data they do not have to live with.
Most people are not “most people”
If a population has a mean of 70% TIR, the mean is by definition where most of the population sits. But “most” hides a lot. In any well-distributed population, roughly:
- One-third sit close to the mean (within a tight band)
- One-third sit meaningfully above (better-controlled, lower-glycator, or both)
- One-third sit meaningfully below (higher exposure, higher-glycator, or both)
If a clinic looks at your last 14-day TIR and compares it to “the consensus 70%,” they are doing two things at once. They are taking your short-term snapshot, and pretending it sits comfortably on a population mean that two-thirds of the population are not actually on. The 14-day window adds noise; the population comparison adds the wrong reference frame. Both errors point the same way: false reassurance, or false alarm.
Why short windows mislead in T1D specifically. One holiday, one virus, one stretch of shift work, one menstrual cycle, one course of steroids can move a 14-day TIR by 10 to 20 percentage points without anything changing about your underlying biology or your management. Three months smooths most of that out. Two weeks does not.
The 90-30-14 rule
Different windows answer different questions. Here is what each is actually for:
90 days, for risk
This is the window that matches HbA1c (red blood cell turnover ~120 days, weighted to recent weeks). It is the window that matches what the GMI formula estimates (HbA1c, which reflects roughly 90 to 120 days of glycaemic exposure). It is the window that lets you calculate an mHGI pair (HbA1c + 90-day mean glucose). It is the window that smooths out one-off events and reveals your stable glycaemic exposure. If you are deciding what target you should be on, what therapy might shift your trajectory, or what your long-term complication risk looks like, you use 90 days.
30 days, for patterns
This is the window for spotting structural patterns: morning highs, post-lunch crashes, weekend versus weekday differences, cycle-related shifts, exercise responses. Long enough to see the pattern repeat, short enough to be actionable. If you are tweaking basal rates, ratios, or AID settings, this is the window.
14 days, only after a change
14 days is for confirming that something specific has shifted: a new pump, a new sensor, a new insulin, a new schedule, a new medication. It is a short-term verification window, not a risk-assessment window. If a clinic uses 14 days as the default for everyone, every visit, that is an ensemble approach pretending to be a personal one.
Three pairs of (HbA1c + 90-day CGM mean glucose) is the minimum to estimate your mHGI properly (Part 3). Four to five pairs is better. This is not because you cannot calculate mHGI from one window; it is because your individual probability stabilises with more independent observations of your own biology.
The platform defaults are wrong by default
Look at where everyone in T1D gets their data:
| Platform | Manufacturer | Default report window | What that window is good for |
|---|---|---|---|
| CareLink | Medtronic / MiniMed | 14 days | Pattern checking after a change |
| Clarity | Dexcom | 14 days (90-day option exists) | Pattern checking after a change |
| LibreView | Abbott | 14 days | Pattern checking after a change |
| Glooko | Glooko (multi-device) | 14 days | Pattern checking after a change |
None of these defaults are wrong as such; they are calibrated for short-term clinical workflow. The problem is that the default sets the expectation. A clinician opens the report, sees a 14-day TIR number, compares it to the 70% target, and treats that comparison as a clinical signal. It isn’t. It is a noisy snapshot compared to a population statistic. Two layers of mismatch in one decision.
What good practice looks like:
- Open the 90-day view first. That is your risk view. Compare it against your personalised TIR target from the matrix in Part 4.
- Switch to the 30-day view second. That is your pattern view. Look for repeating signatures.
- Use 14 days only when you have changed something specific and want to know if the change worked.
- Do not interpret a 14-day TIR comparison against population consensus as a clinical signal. It is barely a signal.
T1D is a Black Swan game
Taleb’s Black Swan framing applies to events that are rare, high-impact, and only obvious in retrospect. In T1D, the events that matter most are exactly this shape:
- Severe hypoglycaemia, especially nocturnal, especially driving, especially pre-DKA-rescue dosing in a stomach bug
- Diabetic ketoacidosis, especially during illness, especially during a sensor failure window, especially with new SGLT2 use
- Cardiovascular events at age 50 to 70, driven by 30 to 40 years of cumulative glycaemic exposure
- Severe non-traumatic hypoglycaemia at age 70+ as cognition begins to slow
None of these are well captured by ensemble averages. A clinic-level “average HbA1c was 58 mmol/mol last year” tells you almost nothing about who in that clinic walked into A&E with severe DKA. A clinic-level “average TIR was 65%” tells you almost nothing about who had three night-time hypos last month. The events that matter are individual, episodic, and clustered around moments of vulnerability.
Taleb’s argument in Skin in the Game (2018) is that the people interpreting the data should bear some of the consequences of being wrong. In T1D, the person bearing the lifetime cumulative-exposure consequences is the person living with it. Clinicians do carry accountability through clinical-governance routes, but they do not carry the 30-year glycaemic, hypoglycaemic and cardiovascular exposure. The structural mismatch is mostly an ergodicity problem (a 14-day snapshot is the wrong window for a non-ergodic disease) with a skin-in-the-game flavour around how target-setting decisions are made.
The implication: for individual T1D risk, do not collapse into population statistics. Use your own 90-day CGM data, your own 3 to 4 paired HbA1c values, your own glycator status, your own CGM zone. The matrix in Part 4 is one of the few clinical tools that does this; almost everything else available right now defaults back to ensemble.
What this changes for how you use your CGM data
If you are the person with T1D
- Open the 90-day view in your CGM app at least once per quarter. Note your mean glucose and your TIR.
- Pair it with your most recent lab HbA1c from the same window. That is one mHGI data point.
- Three to four of these pairs over 12 to 18 months is enough to know your glycator status.
- Compare your 90-day TIR against your personalised target from the matrix in Part 4, not against the consensus 70%.
If you are a clinician
- Default the patient’s CGM platform view to 90 days for risk reviews. Use 30 days for the pattern conversation. Reach for 14 days only when there has been a recent change worth verifying.
- Do not interpret a 14-day TIR shift as a clinical change unless you have ruled out a 14-day-window event (illness, holiday, cycle, steroids).
- For complications-risk discussions, anchor on HbA1c and 90-day mean glucose together, not on TIR alone (see Part 2 on why HbA1c is still indispensable).
- If you are pooling CGM data across patients with different sensors, the comparability problem (Part 1) compounds with the ergodicity problem. Both layers of mismatch apply.
The Via Negativa thread through this guide
Taleb’s four books form a spine that runs through everything GNL does. This guide is one expression of it.
- Fooled by Randomness (Part 5): a 14-day TIR snapshot fools you exactly the way a single cricket innings or a single stock return fools you. Use 90 days for signal.
- The Black Swan (Part 5): severe hypos, DKA, and late-life cardiovascular events are rare, high-impact, and clustered around individual vulnerability. Ensemble averages do not protect you from your own tail.
- Antifragile (Part 4): antifragility is the property of getting stronger from variability and stress, the opposite of fragility (which gets weaker). A 3 mmol/mol HbA1c gap compounds over a T1D lifetime the way a 1% savings gap compounds over a working life: small, persistent exposures define the trajectory. Removing them (Via Negativa, the subtraction move) matters more than adding interventions on top.
- Skin in the Game (Parts 4 and 5): the person interpreting your data should bear some of the consequence of being wrong. In T1D, that person is you. Your target should be set from your own biology (glycator status), your own device (CGM zone), and your own 90-day trajectory; not from a population statistic produced by people who will not carry the result.
The Via Negativa principle ties it together: what you remove matters more than what you add. Remove the 14-day default. Remove the assumption that everyone glycates the same. Remove the non-inferiority margin that flatters the headline. What remains is your individual signal, and that is what this guide exists to find. The same philosophy drives Via Negativa Health, the consultancy arm of GNL: subtraction over addition, skin in the game, and building things that get stronger from use.
This part is conceptual. The 90-30-14 rule is a working clinical heuristic supported by the way HbA1c, GMI and CGM variability behave; it is not yet a guideline. The ergodicity framing draws on Taleb (Skin in the Game, 2018) and a working GNL manuscript on ensemble vs individual probability in diabetes care. Discuss any change to how you use your CGM reports with your diabetes care team.
Concepts cited: ergodic theory (general); Taleb 2001 (Fooled by Randomness); Taleb 2007 (The Black Swan); Taleb 2012 (Antifragile); Taleb 2018 (Skin in the Game); Via Negativa (subtraction over addition); Bergenstal 2018 (GMI); Pemberton-Chalew mHGI framework (working); GNL CGM Zone framework.