Evidence hierarchy (strongest to weakest):

• Level 1a — Systematic review of RCTs
• Level 1b — Individual RCT (with narrow CI)
• Level 2a — Systematic review of cohort studies
• Level 2b — Individual cohort study / low-quality RCT
• Level 3a — Systematic review of case-control studies
• Level 3b — Individual case-control study
• Level 4 — Case series / case report
• Level 5 — Expert opinion

RCT: Gold standard for testing treatments. Randomises participants to intervention vs control. Minimises bias through randomisation and blinding.

Cohort study: Follows groups over time. Can be prospective or retrospective. Good for prognosis and identifying risk factors. Cannot prove causation but can show strong associations.

Case-control study: Starts with the outcome and looks back for exposure. Good for rare diseases because you don't need to follow thousands of people. Quicker and cheaper than cohort studies.

Cross-sectional study: A snapshot at one point in time. Good for prevalence. Cannot show causation or temporal relationships.

The 2x2 table:

Sensitivity = TP / (TP + FN) — how good the test is at detecting disease. A highly sensitive test rules OUT disease when negative (SnNOut).

Specificity = TN / (TN + FP) — how good the test is at confirming health. A highly specific test rules IN disease when positive (SpPIn).

PPV = TP / (TP + FP) — affected by prevalence. Higher prevalence leads to higher PPV.

NPV = TN / (TN + FN) — affected by prevalence. Lower prevalence leads to higher NPV.

ARR (Absolute Risk Reduction) = control event rate − treatment event rate.

NNT (Number Needed to Treat) = 1 / ARR. Always uses absolute risk reduction, never relative.

ARI (Absolute Risk Increase) = treatment adverse event rate − control adverse event rate.

NNH (Number Needed to Harm) = 1 / ARI. If NNH < NNT, the drug harms more than it helps.

AUC (Area Under the Curve) measures overall test accuracy:

• 0.5 = useless (follows the diagonal — no better than chance)
• 0.7–0.8 = acceptable
• 0.8–0.9 = excellent
• >0.9 = outstanding
• 1.0 = perfect test

The top-left corner represents a perfect test. The diagonal line represents a test no better than chance. If the exam shows two ROC curves, the one with the larger AUC is the better test.

Power = 1 − β = the probability of detecting a true difference if one exists. Conventionally set at 80% (0.8).

α (alpha) = significance level (usually 0.05) = probability of Type I error (false positive — finding a difference that doesn't exist).

β (beta) = probability of Type II error (false negative — missing a real effect).

Power calculation components:

• Effect size (must be clinically meaningful)
• Alpha level
• Desired power
• Variance in the data

How each component affects sample size:

• Bigger effect size = smaller sample needed
• Smaller alpha = larger sample needed
• Higher power = larger sample needed
• More variance = larger sample needed

Power calculations are performed before the study to determine sample size.

Lowering a diagnostic threshold (e.g. troponin cut-off) increases sensitivity but increases false positives (Type I error).

Concealed allocation: Sealed envelope given BEFORE the trial starts. The patient doesn't know their group assignment. Prevents selection bias at enrolment.

Randomisation: Assigning participants to groups AFTER enrolment. Ensures groups are comparable at baseline.

Blinding:

• Single-blind: patient doesn't know their group
• Double-blind: patient + researcher don't know
• Triple-blind: patient + researcher + assessor don't know

Blinding reduces observer bias and placebo effects.

Surrogate endpoint: A proxy marker used instead of the actual clinical outcome. E.g. DVT on Doppler as a surrogate for PE in pregnancy (where CTPA radiation is a concern).

Kappa coefficient: Measures inter-rater reliability. 0 = no agreement beyond chance. 0.5 = moderate agreement. 1.0 = perfect agreement.

Demographic table: Compares baseline characteristics between groups using p-values. You WANT p > 0.05 — this means the groups are comparable.

• Selection bias: Systematic differences in who is recruited. Groups are not representative of the target population.
• Recall bias: Participants with disease remember exposures differently from those without. Common in case-control studies.
• Observer/detection bias: Assessors measure outcomes differently based on knowledge of group allocation. Prevented by blinding.
• Attrition bias: Systematic differences in dropouts between groups. If sicker patients drop out of the treatment arm, results look falsely good.
• Publication bias: Positive studies are more likely to be published than negative ones. Detected with funnel plots.
• Lead-time bias: Earlier detection appears to increase survival time without actually changing outcomes. Common in screening studies.
• Hawthorne effect: Participants change behaviour because they know they're being observed, not because of the intervention itself.

Confounding: A third variable affects both the exposure and the outcome, creating a spurious association. Controlled by randomisation, matching, stratification, or multivariate analysis.

The Jadad score (Oxford quality scoring system) assesses RCT quality on a scale of 0–5:

• Was the study described as randomised? (+1)
• Was the randomisation method appropriate? (+1) — e.g. computer-generated, NOT alternate allocation
• Was the study described as double-blind? (+1)
• Was the blinding method appropriate? (+1) — e.g. identical placebo
• Were withdrawals and dropouts described? (+1) — numbers and reasons in each group

Deductions: −1 if randomisation method inappropriate. −1 if blinding method inappropriate.

Score ≤2 = low quality. Score ≥3 = high quality.

Other quality assessment tools:

• CONSORT: Reporting standard for RCTs
• PRISMA: Reporting standard for systematic reviews
• GRADE: Overall evidence quality rating system
• Newcastle-Ottawa: Quality assessment for cohort and case-control studies

P-value <0.05 = statistically significant by convention. This does NOT mean clinically important — a huge trial can find a tiny, meaningless difference that reaches statistical significance.

Confidence interval: The range where the true value likely lies. If the 95% CI crosses 1.0 (for ratios like RR/OR) or 0 (for differences) — the result is not significant, regardless of the p-value.

• Narrow CI = precise estimate
• Wide CI = imprecise estimate (often from small sample)

Heterogeneity (I²) in meta-analysis: >50% is significant — you should question whether the studies should be combined at all.

Intention-to-treat (ITT) vs per-protocol: ITT analyses everyone enrolled (even dropouts). It's more conservative and avoids attrition bias. Per-protocol only analyses those who completed the study — risks overestimating treatment effect.

ROC curve: See Section 3 (Diagnostic Formulae) above for full detail including AUC values and illustration.

Audit: Measures current practice against an existing STANDARD. Uses the PDSA (Plan-Do-Study-Act) cycle. Must re-audit to close the loop. Does NOT need ethics approval.

Research: Generates NEW knowledge. Tests a hypothesis. NEEDS ethics approval.

QI methodologies:

• EBCD (Experience-Based Co-Design): Uses patient and staff experiences to drive improvement
• Driver diagrams: Primary aim → primary drivers → secondary drivers → change ideas
• 5 Whys / 3 Whys: Root cause analysis — keep asking "why?" until you reach the underlying cause
• Ishikawa (fishbone) diagram: Categorises causes into groups (people, process, equipment, environment, etc.)
• Pareto analysis (80/20 rule): 80% of problems come from 20% of causes. Fix the top causes first for maximum impact.
• Process mapping: Visualising the patient journey to identify bottlenecks and waste
• Run charts / SPC charts: Tracking improvement over time. Distinguishes special cause variation (something changed) from common cause variation (normal fluctuation).
• Gantt chart: Project timeline showing tasks, duration, and dependencies

You do NOT need 100% of data for PDSA — just enough, collected sequentially, to test your change idea.