生物統計與研究設計

Buzzwords → Dx

Study Design Hierarchy

• Meta-analysis / systematic review — synthesizes RCTs
• RCT — gold standard for causation; randomization eliminates confounding
• Prospective cohort — exposure measured before outcome
• Retrospective cohort / case-control — easier; bias-prone
• Case series / case report — hypothesis-generating

Treatment Algorithm — Reading a Trial

flowchart TD
  A[New trial publication] --> B[Population: who was included?<br>generalizability to your patient]
  B --> C[Intervention vs comparator<br>was control SOC?]
  C --> D[Outcome: PFS / OS / ORR / etc.<br>surrogate vs definitive?]
  D --> E[Effect size: HR / OR / RR + 95% CI<br>does CI exclude 1.0?]
  E --> F[Statistical significance: p-value]
  F --> G[Clinical significance: NNT, ARR<br>magnitude relevant to patient?]
  G --> H[Subgroup analyses: hypothesis-generating, not confirmatory]
  H --> I[Quality: ITT analysis? blinding? attrition?]
  I --> J[Apply to patient with shared decision-making]

Pearls / Pitfalls

• CI excluding 1.0 = statistically significant for ratios (OR, RR, HR). For absolute differences (ARR, NNT), the significance line is 0.
• OS is gold standard; PFS is surrogate. Crossover trials complicate OS interpretation (control can switch to active at progression — diluting OS difference).
• NNT < 50 is generally considered clinically meaningful; NNT depends on baseline risk + ARR.
• Sensitivity / specificity are intrinsic to test; PPV / NPV depend on prevalence.
• Likelihood ratios are prevalence-independent and shift pre-test probability.
• Subgroup analyses are hypothesis-generating; not confirmatory unless pre-specified + adjusted for multiple comparisons.
• Intention-to-treat preserves randomization — preferred primary analysis. Per-protocol introduces bias from differential dropout.
• Lead-time bias in screening — earlier diagnosis appears to extend survival without changing outcome. Length-time bias — screen catches slower-growing tumors disproportionately.
• Hazard ratio interpretation: HR 0.50 means 50 % reduction in hazard at any time point (assumes proportional hazards).
• Median survival is robust to outliers; mean is not. KM curves visualize.
• Cox regression allows multivariable adjustment for prognostic variables; reports HR with CI.
• Type I error = false positive (claiming benefit that isn't there); Type II = false negative (missing real benefit).
• p-value < 0.05 does NOT prove benefit — context, magnitude, biological plausibility, replication all matter.
• Adaptive trial designs: master protocols, basket / umbrella trials common in oncology (e.g., BATTLE-2, NCI-MATCH).
• Surrogate endpoints: ORR, CR, MRD-negativity, PFS — accelerated approval pathway in oncology, not always confirming OS.
• Cox proportional hazards assumption — violated when curves cross; use stratified or time-varying coefficients.
• Competing risks in oncology (death from non-cancer causes) — Fine-Gray model, cumulative incidence functions.

Cross-references

• Staging — IPI / R-ISS / ELN are prognostic models
• Drug Regimens — trial-derived therapies
• DLBCL — POLARIX trial design example
• MM — PERSEUS, CARTITUDE-4

相關題目

• Q-214 — Biostats — sensitivity vs specificity, PPV / NPV
• Q-215 — Biostats — hazard ratio interpretation
• Q-216 — Biostats — number needed to treat (NNT)

Sources