Research Methodology and Evidence-Based Emergency Medicine
What the board tests
Whether you can appraise a paper, calculate diagnostic and therapeutic metrics, recognise bias, and decide if a result applies to your patient.
Design and bias
- Hierarchy: systematic review and meta-analysis, randomised controlled trial, cohort, case-control, cross-sectional, case series.
- Randomisation with allocation concealment controls selection bias; blinding controls measurement and performance bias; intention-to-treat analysis preserves randomisation.
- Diagnostic accuracy studies must avoid spectrum bias by enrolling consecutive, appropriate patients; check for verification bias and a consistent reference standard.
- Reporting standards: CONSORT for trials, STARD for diagnostic studies, PRISMA for reviews, GRADE for strength of recommendations.
- Clinical decision rules need prospective validation in a different setting before adoption; a derivation study alone is hypothesis-generating.
Calculations
- Sensitivity = TP/(TP+FN); specificity = TN/(TN+FP); PPV and NPV depend on prevalence.
- Likelihood ratio positive = sensitivity/(1-specificity); a value above 10 meaningfully raises probability.
- ARR = control event rate minus treatment event rate; NNT = 1/ARR. Report confidence intervals and absolute, not just relative, effects.
Worked example
A new troponin assay is tested in 1000 patients with chest pain; prevalence of MI is 10%. Sensitivity 90% and specificity 80% give TP 90, FN 10, FP 180, TN 720. PPV = 90/(90+180) = 33%; NPV = 720/(720+10) = 99%. LR+ = 0.9/0.2 = 4.5, LR- = 0.1/0.8 = 0.125. So a negative result is genuinely useful for rule-out, but a positive result needs confirmation, because only one in three positives has MI at this prevalence. Reasoning: PPV falls as prevalence falls, which is why high-sensitivity assays are used to rule out rather than rule in.
Traps
Quoting PPV without prevalence; equating statistical significance with clinical importance; computing NNT from relative risk reduction; ignoring composite or surrogate endpoints; trusting a subgroup or post-hoc analysis; applying a derivation rule that was never externally validated.