A 1.9-million-person health cohort maps its own biases
Our Future Health now has the scale to study uncommon conditions, but its first deep profile shows why volunteer cohorts cannot be read as a census of Britain.
The 60-second version
A baseline analysis of more than 1.9 million Our Future Health participants shows enormous research capacity alongside measurable volunteer-selection bias.
Key points
- Disease and medication patterns often aligned with other cohorts, supporting the resource's use for comparative research.
- Women, older adults and less-deprived neighborhoods were overrepresented relative to the UK population.
- Different case definitions and time windows mean absolute disease rates cannot be compared mechanically.
- The cohort is best suited to carefully adjusted comparisons, prospective follow-up and externally replicated findings.
Verdict. The database is a powerful research platform, but 1.9 million volunteers are not a census and should not be treated as one.
ScaleThe cohort is already large enough to change study design
Our Future Health aims to recruit five million UK-resident adults. More than 2.5 million had enrolled when this analysis was prepared, and baseline phenotype data were available for more than 1.9 million. The paper combines questionnaires, clinical measurements, location, diagnoses, medication, hospital visits, cancer records and causes of death.
That scale matters because researchers can form large case groups even for some uncommon conditions. The authors report, for example, 668 participants with myasthenia gravis and 172 with cystic fibrosis. Large absolute counts can improve statistical power, but they do not by themselves remove selection or measurement bias.
ChecksMany patterns reproduced across other datasets
Across 109 self-reported conditions, relative prevalence patterns correlated with the UK Biobank at r = 0.784. Associations with known clinical correlates also reproduced across the two cohorts at about r = 0.80. Medication use and cancer prevalence generally followed expected age gradients.
These are consistency checks, not certificates of representativeness. A high correlation can mean conditions are ranked similarly while their absolute rates remain different. The comparison sources also use different case definitions, coding systems and time windows.
| What the paper supports | The cohort contains large case groups and reproduces many familiar relative patterns of disease and medication use. |
|---|---|
| What it does not support | It cannot be treated as a census or used uncritically to estimate every national prevalence and incidence rate. |
| What comes next | Linked records, repeat questionnaires, bias analyses and replication can test how health outcomes change over time. |
BiasWho volunteers shapes the picture
The cohort broadly reflected several UK population patterns, but important groups were underrepresented proportionally. Women made up 57% of participants. Younger adults and all but one minority ethnic group were underrepresented, although the database's size still produced large absolute numbers in some of those groups.
Socioeconomic selection was clearer. People from the most deprived fifth of neighborhoods accounted for 13% of participants, versus 20% nationally. The least deprived fifth accounted for 27%, versus 20%. Access, time, trust, health awareness and the geography of recruitment can all influence participation.
A huge volunteer sample can sharpen comparisons without becoming a miniature copy of the country.
MeasurementThe denominator is not the only source of uncertainty
Some self-reported mental health conditions, including depression and anxiety, appeared more common than in national comparisons, while lung cancer prevalence was lower. Those differences may arise from cohort age, survival, recruitment, healthcare use or the way a condition is recorded.
A lifetime self-reported diagnosis is not directly comparable with active disease measured at one point in time. Hospital coding, questionnaires and registry linkage each see a different slice of health. The authors therefore caution against using the current cohort to calculate generalizable national prevalence or incidence for all conditions.
UseHow to read future findings from this resource
- Ask the right question: within-cohort comparisons and prospective associations are different from national head counts.
- Inspect selection: check age, sex, ethnicity, deprivation and recruitment geography for the analysis at hand.
- Align definitions: compare like with like across questionnaires, hospital codes, registries and time windows.
- Test robustness: use weighting, sensitivity analyses and external replication before generalizing.
- Protect participants: individual-level data require an approved research application and are not openly downloadable.
- Disclose interests: the programme has public, charity and industry funding; several authors reported advisory links to the cohorts.
TakeawayTreat scale as an opportunity, not a guarantee
Our Future Health can make previously underpowered questions testable and support long-term follow-up at unusual scale. Its strongest studies will be those that model who entered the cohort, define outcomes carefully and verify conclusions elsewhere. The practical rule is simple: use the database as a research engine, not as a national scoreboard.