Selection bias
Selection bias is systematic error caused by how units enter, remain in or are excluded from an analyzed sample relative to its target population.
Start with the target: a failure rate among currently observed drives and lifetime reliability among all installed drives ask different questions. A restricted sample need not be biased for its stated conditional target. Transferring that estimate to a different population requires justified observation and sampling assumptions. [1] [2]
Survivor-only inclusion can omit failures before observation starts. Left-truncation describes that entry pattern; right-censoring retains a known unit with incomplete follow-up after entry. Neither term is a synonym for selection bias. Correct delayed-entry risk sets require assumptions about entry, and cannot identify survival before the earliest supported entry age from those observations alone. [2]
More records and narrower conventional confidence intervals do not remove an unmodelled selection mechanism. Record eligibility, entry times, exit reasons and missing-data exclusions; compare retained and excluded outcomes and exposure before extrapolating. A selection-adjusted analysis still needs defensible assumptions. [1]
Relations
Citations
- CDC Field Epidemiology Manual, chapter 8: confidence intervals and epidemiologic interpretation
- Schisterman et al. (2013): Accuracy Loss Due to Selection Bias in Cohort Studies with Left Truncation (doi:10.1111/ppe.12073)
Sources & caveats
- The methodological sources concern epidemiologic inference; the drive examples apply those distinctions without estimating bias in a storage fleet.