Table 1.
Examples of large datasets with clustering
| Dataset | IMPACT29 | EPIC30 | CPRD31 | MIMIC-III32 |
|---|---|---|---|---|
| Population | Patients with a head injury | Volunteers agreeing to participate | Patients attending primary care practices in the UK | Patients admitted to the Beth Israel Deaconess Medical Center (Boston, MA, USA) |
| Data source | IPD from multiple studies | IPD from a prospective multicentre study | Linked database with EHR data | Hospital database with EHR data |
| Total sample size | 11 022 | 519 978 | 11 299 221 | 38 597 |
| No of clusters | 15 studies | 23 centres | 674 general practices | 5 critical care units |
| Heterogeneity in study designs | Phase 3 clinical trials; observational cohort studies | Observational cohort studies; nested case-control studies | Not applicable | Not applicable |
| Heterogeneity in included populations | Data collection from 1984 to 1997; data from high, low, and middle income countries; variable severity of brain injury | Participant enrolment from 1992 to 2000; data from 10 European countries; heterogeneity in participant recruitment schemes | Data collection from 1987 to present; data from England, Wales, Scotland, and Northern Ireland | Data collection from 2001 to 2012; variable patient ethnic group and social status, among other factors |
| Heterogeneity in data quality | Variable classification for head injuries; variable time points for outcome assessment | Lack of standardised procedures across cohorts; heterogeneity in dietary assessment methods; heterogeneity in anthropometric measurement methods; heterogeneity in questionnaires across countries | Selective linkage with other databases; large variation in data recording between practices; variable frequency of data recording by age, sex, and underlying morbidity; informative missingness of patient characteristics; non-standardised definitions of diagnoses and outcomes; possible variation in extent of misclassification between diseases | Different critical care information systems in place during data collection; protected health information removed from free text fields |
| Heterogeneity in level of care | Variability in level of local care; clear improvement of treatment standards over time | Not applicable | Not applicable | Variable efforts to health prevention owing to variability in health insurance programmes among patients33 |
IPD=individual participant data; EHR=electronic health records data; IMPACT=International Mission for Prognosis And Clinical Trial; EPIC=European Prospective Investigation into Cancer and Nutrition; CPRD=Clinical Practice Research Datalink; MIMC-III=Medical Information Mart for Intensive Care III.