Evidence: what we have shown so far

How we grade claims

Shown Found again in independent data.
Preliminary Computed once, not yet checked elsewhere.
Open The question being tested now.

What the map shows, and what it does not

Each element of the map, with the claim it carries and the claim it does not.
Element	It shows	It does not show
A colour	Risk of death from any cause at that point, relative to the median US adult, from a model fitted on NHANES.	A diagnosis, or any one person's chance of dying.
A curve's shape	Where risk rises or falls as a marker changes, across the population.	Why it does. Cause, reverse causation and confounding are not separated.
An arrow	The average change one study measured in a marker, drawn from where the marker sits.	What that intervention would do for a person, or that moving the marker moves the risk.
A stable point	A region people drift towards and stay near when measured years apart.	A target to aim for.
A grade	Whether a claim was found again in independent data (Shown), computed once (Preliminary), or is being tested (Open).	How large or how useful an effect is.

Test register

Of the 40 published biomarker–mortality patterns we could check in NHANES:

Shown
Published patterns show up again in a national sample.
34 of 40 directions match in NHANES (51,662 adults, validation sample); 2 run the other way, 4 show no clear pattern.¹ Not registered in advance (validation). By biomarker
Shown
The same risk per standard deviation holds in other countries.
hs-CRP, creatinine and HbA1c agree within 0.02 between the US and Costa Rica; see the heatmap below. Not registered in advance (validation).
Preliminary
How fast do people's measurements change after 50?
Gait speed and grip strength fall measurably within the same people in HRS 2006–2022.² Exploratory. Ageing
Preliminary
Which single measurement adds most to age and sex?
Grip strength and urine albumin lead.³ Exploratory. Ranking
Running
Does within-person variability across body systems predict death?
Registered in advance at osf.io/kzfs4. The HRS analysis is running; the result will be posted here, null or not.
Open
Do shared curves plus a few personal numbers predict individual paths better than baseline risk?
To be registered before data arrive. Awaiting repeated blood data: this is the question the project exists to answer.

The register as a table

Each row is one analysis. Registered means the plan was published before we looked at the outcome.
Test	Registered in advance	Status	Result
Do published patterns between a biomarker and risk of death show up again in a national sample?	No (validation)	Shown	34 of 40 directions match in NHANES (51,662 adults, validation sample); 2 run the other way, 4 show no clear pattern.¹ By biomarker
Does the same risk per standard deviation hold in other countries?	No (validation)	Shown	hs-CRP, creatinine and HbA1c agree within 0.02 between the US and Costa Rica; see the table below.
How fast do people's measurements change after 50? (HRS 2006–2022)	No (exploratory)	Preliminary	Gait speed and grip strength fall measurably within the same people.² Ageing
Which single measurement adds most to age and sex?	No (exploratory)	Preliminary	Grip strength and urine albumin lead.³ Ranking
Does within-person variability across body systems predict death? (HRS)	Yes, osf.io/kzfs4	Analysis running	To be posted here, null or not
Do shared curves plus a few personal numbers predict individual paths better than baseline risk?	To be registered before data arrive	Open: awaiting repeated blood data	–

LDL and non-HDL cholesterol run opposite. Low LDL in older adults often reflects existing illness, so the raw pattern looks reversed. ↩
Rates per decade are left out until we state which standard deviation they use. ↩
Held out: 30% of NHANES 1999–2018 participants, not used in fitting. Measure: gain in how well the model ranks who died first, among people in the same 10-year age band (Harrell's concordance). ↩

The same risks in three countries' data

Darker means a larger effect per standard deviation in whichever direction is harmful (0.58 for grip strength is as strong as 1.72 would be). Cells that agree across columns mean the pattern holds outside the US. Hover or focus a cell for the 95% interval. Adjusted for age and sex; associations, not predictions of what changing a marker would do.

The same numbers as a table

Hazard ratio per standard deviation and 95% interval, by cohort.
Biomarker	NHANES · US, 20–84	CRELES · Costa Rica, 60+	HRS · US, 50+
Maximal Grip Strength	0.47 0.42–0.53	–	0.58 0.55–0.61
High-Sensitivity C-Reactive Protein	1.30 1.26–1.33	1.30 1.19–1.43	–
Serum Creatinine	1.25 1.22–1.28	1.27 1.16–1.39	–
Total Cholesterol	0.91 0.89–0.94	0.79 0.69–0.90	–
Resting Heart Rate	1.21 1.19–1.24	–	1.20 1.17–1.23
Glycated Hemoglobin	1.18 1.16–1.21	1.20 1.10–1.31	–
Fasting Plasma Glucose	1.18 1.15–1.21	1.12 1.02–1.22	–
Waist-to-Height Ratio	1.09 1.06–1.12	–	1.18 1.14–1.22
Waist Circumference	1.09 1.06–1.12	1.14 1.00–1.31	1.17 1.13–1.22
Serum Triglycerides	1.06 1.02–1.10	0.86 0.75–0.99	–
Systolic Blood Pressure	1.08 1.05–1.10	0.89 0.78–1.00	1.06 1.04–1.09
Pulse Pressure	1.11 1.08–1.13	–	1.09 1.07–1.12
Body Mass Index	1.03 1.00–1.06	1.10 0.95–1.28	1.04 1.01–1.08
Standing Height	1.01 0.97–1.04	–	0.91 0.88–0.95
Diastolic Blood Pressure	0.96 0.94–0.98	0.95 0.85–1.07	1.00 0.98–1.03
Mean Arterial Pressure	1.01 0.99–1.03	–	1.03 1.01–1.06
Body Weight	1.03 1.00–1.06	–	1.00 0.97–1.04

Systolic blood pressure: in older adults, low blood pressure often reflects existing illness, which is why the Costa Rica figure runs the other way.

The models behind the map: Models. Questions we want help with: Open problems.

Data sources

NHANES 1999–2018 with linked mortality to 2019
RAND HRS 1992–2022
Wisconsin Longitudinal Study 1993–2011
ELSA HCAP 2018 and 2023
CRELES waves 1–3
MIDUS 3

How we handle study data