Evidence-Based Medicine

01 Overview & History of EBM

Evidence-based medicine (EBM) is the conscientious, explicit, and judicious use of current best evidence in making decisions about the care of individual patients. Rather than relying solely on pathophysiologic reasoning, clinical anecdote, or authority, EBM demands that treatment decisions be informed by the highest quality research evidence available, integrated with clinical expertise and patient values. EBM is not a rigid cookbook; it is a disciplined approach to lifelong learning and decision-making that has reshaped every field of medicine over the past three decades.

Historical Development

The conceptual roots of EBM stretch back centuries — Pierre Louis used the numerical method in 1830s Paris to show that bloodletting did not improve pneumonia survival, one of the first statistical critiques of accepted therapy. Archie Cochrane, a British epidemiologist, published Effectiveness and Efficiency in 1972, arguing that medical practice should be based on rigorous randomized trials and systematic reviews; his name now marks the Cochrane Collaboration, founded in 1993 to produce systematic reviews of healthcare interventions.

The term “evidence-based medicine” was coined by Gordon Guyatt at McMaster University in 1991, building on David Sackett’s clinical epidemiology tradition. The landmark 1992 JAMA paper by the Evidence-Based Medicine Working Group formally introduced EBM as a “new paradigm for medical practice.” Since then EBM has expanded into evidence-based nursing, evidence-based public health, and evidence-based healthcare more broadly.

Key Figures in EBM

Why EBM Matters in Modern Practice

The medical literature doubles roughly every 10–15 years; more than 1 million biomedical articles are indexed annually in PubMed. No clinician can read, let alone synthesize, this volume unaided. Without structured methods for finding, appraising, and applying evidence, practice drifts toward habit, authority, and marketing influence. Studies repeatedly show that routine care lags best evidence by 10–20 years, with many widely used interventions later found ineffective or harmful (e.g., hormone replacement for cardiovascular prevention, routine arthroscopy for knee osteoarthritis, tight glycemic control in ICU patients).

02 Defining EBM & the Three Pillars

The Sackett Definition

David Sackett’s 1996 BMJ editorial provides the canonical definition: “Evidence-based medicine is the conscientious, explicit, and judicious use of current best evidence in making decisions about the care of individual patients. The practice of evidence-based medicine means integrating individual clinical expertise with the best available external clinical evidence from systematic research.” Sackett emphasized that EBM is neither “cookbook medicine” nor a tool for cost-cutters — it is a method for integrating evidence with clinical judgment.

The Three Pillars in Detail

What EBM Is — and Is Not

03 EBM vs Traditional Medicine

Traditional (pre-EBM) medical practice was dominated by eminence-based decisions — senior authorities in a field determined standards of care, often based on pathophysiologic reasoning or uncontrolled clinical experience. While valuable, this approach produced and perpetuated many practices later shown to be ineffective or harmful.

Historical Examples of Evidence Overturning Authority

Pre-EBM vs EBM Decision-Making

Why Pathophysiologic Reasoning Alone Fails

Pathophysiologic reasoning is seductive because it is mechanistic and satisfying — if a drug lowers LDL, and high LDL causes atherosclerosis, then lowering LDL should prevent heart attacks. This chain of reasoning is correct for statins but wrong for torcetrapib, ezetimibe (initially), and many niacin trials, where LDL or HDL changes did not translate to outcome benefits. The body is a complex system with countless feedback loops, off-target effects, and unanticipated interactions. Only empirical testing in humans with patient-important outcomes can confirm whether a biologically plausible mechanism produces a real clinical benefit.

04 Step 1 — Ask (PICO/PICOT)

The 5-step EBM cycle begins with transforming a clinical information need into an answerable question. A well-constructed question is the foundation of an efficient literature search and a relevant appraisal. The most widely used framework is PICO (sometimes extended to PICOT or PICOS).

The PICO(T) Framework

Foreground vs Background Questions

Clinicians ask two fundamentally different types of questions:

• Background questions seek general knowledge about a condition or intervention (“What causes atrial fibrillation?”). These are best answered by textbooks, review articles, or point-of-care resources (UpToDate, DynaMed).
• Foreground questions seek specific evidence to guide a clinical decision (“In patients with new-onset AF, does rhythm control with catheter ablation improve mortality compared to rate control?”). These demand systematic literature searches and critical appraisal.

05 Types of Clinical Questions

Different types of clinical questions demand different study designs for their answer. Recognizing the question type determines where to look and how to appraise what you find.

The Five Major Question Types

Less Common but Important Question Types

06 Steps 2–5 — Acquire, Appraise, Apply, Assess

Once the question is framed, the remaining steps of the EBM cycle turn evidence into action and feed back into improved practice.

The Full 5-Step EBM Cycle

Time Constraints at the Point of Care

A complete 5-step cycle — formulating a question, searching, appraising, and applying — takes hours and is impractical during a busy clinic. For real-time decisions, clinicians rely on pre-appraised resources (Cochrane reviews, UpToDate, DynaMed, BMJ Best Practice), which summarize evidence in structured, regularly updated formats. The full 5-step process is reserved for challenging questions, teaching, or when pre-appraised summaries are unavailable.

07 Databases & Point-of-Care Resources

Knowing which database to search is as important as knowing how to search it. Resources differ in scope, filtering, and the degree to which evidence has already been appraised.

Major Biomedical Databases

Pre-Appraised & Point-of-Care Resources

Gray Literature

“Gray literature” includes conference abstracts, theses, regulatory documents (FDA, EMA), unpublished trials, preprints (medRxiv, bioRxiv), and clinical trial registries (ClinicalTrials.gov, WHO ICTRP, EU Clinical Trials Register). It is essential for comprehensive systematic reviews because positive results are preferentially published — ignoring gray literature can bias conclusions.

08 Search Strategy, Boolean Logic & MeSH

Efficient searching combines controlled vocabulary (MeSH), free-text terms, Boolean operators, and filters. A good strategy balances sensitivity (capturing all relevant studies) with specificity (excluding irrelevant ones).

Boolean Operators

MeSH (Medical Subject Headings)

MeSH is the National Library of Medicine’s controlled vocabulary. Each MEDLINE-indexed article is tagged with MeSH terms by human indexers, allowing retrieval of articles on a concept regardless of exact wording. Searching “myocardial infarction”[MeSH] captures papers that say “heart attack,” “MI,” or “STEMI” without having to list every synonym.

Useful PubMed Filters

Four-Step Search Strategy

1. Identify concepts using PICO.
2. Find MeSH terms and synonyms for each concept.
3. Combine synonyms with OR, then concepts with AND.
4. Apply limits (study type, date, language) appropriate to the question.

09 The Evidence Pyramid

Not all evidence is created equal. The evidence pyramid visualizes a rough hierarchy of study designs by their ability to reduce bias in estimating treatment effects. Higher-tier designs are more likely (on average) to provide valid estimates, but a well-done cohort can outperform a flawed RCT.

Classic Evidence Pyramid (Unfiltered Information)

Haynes “5S” / “6S” Pyramid (Pre-Appraised Information)

Brian Haynes proposed an alternative hierarchy focused on how the evidence reaches the clinician, prioritizing pre-appraised sources that save time at the bedside:

Systematic Review vs Meta-Analysis — The Critical Distinction

These terms are often confused:

• A systematic review (SR) is a structured, reproducible summary of all the evidence relevant to a focused question, using explicit methods to search, select, and appraise studies.
• A meta-analysis (MA) is a statistical technique that quantitatively pools results from multiple studies into a single summary estimate.
• All meta-analyses should be embedded in a systematic review. Not all systematic reviews include a meta-analysis — when studies are too heterogeneous to pool, a qualitative (narrative) synthesis is appropriate.

10 Study Designs Overview

The choice of study design is driven by the question and by ethical and practical constraints. This section summarizes the major designs and when to use each.

Experimental Designs

Observational Designs

Quasi-Experimental Designs

When randomization is impossible (policy changes, system interventions), quasi-experimental designs provide stronger inference than simple before-after comparisons:

Key Concepts in Observational Analysis

11 Validity of RCTs

Critical appraisal of a therapy study asks three core questions: (1) Are the results valid? (2) What are the results? (3) Will they help me care for my patient? This section focuses on validity — whether the study design protects against bias.

Key Validity Criteria for RCTs

Randomization vs Allocation Concealment

These are distinct concepts: randomization is the process of generating a random sequence; allocation concealment is hiding that sequence from the person enrolling participants. Studies with inadequate allocation concealment overestimate treatment effects by ~30–40%.

Levels of Blinding

Intention-to-Treat vs Per-Protocol

Methods of Randomization

When an RCT Is Not Possible

Randomized trials are sometimes impossible or unethical (rare diseases, emergency conditions, long-term exposures, harms). In such cases, observational evidence must carry the inference. Key strategies to strengthen observational inference include large sample sizes, propensity score methods, instrumental variable analysis, negative controls, target trial emulation, and triangulation across independent designs.

12 Results — Effect Size, RR, ARR, NNT, CI

Once validity is established, the next appraisal question is: how large is the effect, and how precise is the estimate? Effect size measures quantify the impact; confidence intervals quantify precision.

Core Effect Measures (Dichotomous Outcomes)

Why RRR Alone Is Misleading

Press releases love relative risk reductions because they sound larger (“30% reduction”) than absolute numbers (“3 fewer events per 100 patients”). Always translate relative figures into absolute ones. A large RRR applied to a very low baseline risk yields a small ARR and a large NNT, which may not justify the cost or side effects.

Confidence Intervals & Statistical Significance

A 95% confidence interval (CI) is the range within which the true effect likely lies, repeated across hypothetical studies. If the CI for an RR or OR crosses 1.0 (or the CI for a mean difference crosses 0), the result is not statistically significant at α = 0.05. The width of the CI reflects precision — wider = less precise, narrower = more precise.

Statistical vs Clinical Significance

A statistically significant result may be clinically trivial (e.g., a 2 mmHg blood pressure difference from a huge trial), and a non-significant result may hide a clinically important effect obscured by low power. Always ask: is this difference big enough to matter to a patient?

Types of Outcomes in Trials

Power & Sample Size

Statistical power (1 − β) is the probability of detecting a true effect of a specified size. Convention requires ≥80% power. Sample size depends on expected effect size, baseline event rate, variability, α (usually 0.05), and desired power. Underpowered trials produce wide confidence intervals and risk missing real effects (type II error); they also produce inflated effect estimates when positive (“winner’s curse”).

13 Applicability & CONSORT

A valid, precise result is useless if it does not apply to your patient. Applicability (external validity, generalizability) asks whether you can extrapolate the study findings to the person in front of you.

Applicability Questions

• Is my patient similar enough to the trial population (age, sex, comorbidities, severity, ethnicity)?
• Was the treatment protocol feasible in my setting?
• Were all clinically important outcomes considered, including harms and quality of life?
• Do the benefits outweigh the harms, costs, and burdens for this patient?
• Are the patient’s values and preferences consistent with the intervention?

CONSORT — Reporting RCTs

The CONSORT (Consolidated Standards of Reporting Trials) statement is a 25-item checklist and flow diagram for transparent reporting of parallel-group RCTs. Journals increasingly require CONSORT compliance. Key elements include:

14 Diagnostic Study Validity & Biases

Diagnostic studies evaluate how well a test distinguishes patients with the disease from those without it. Critical appraisal here has its own vocabulary and pitfalls.

Validity Criteria for Diagnostic Studies

Key Biases in Diagnostic Studies

15 Sensitivity, Specificity, LRs, ROC, STARD & QUADAS-2

The 2×2 Diagnostic Table

Core Diagnostic Metrics

Likelihood Ratios & Bayesian Reasoning

Likelihood ratios convert pre-test probability into post-test probability without requiring the population prevalence. They are the most clinically useful diagnostic metric because they work at the individual patient level.

ROC Curves

A receiver operating characteristic (ROC) curve plots sensitivity (y-axis) against 1 − specificity (x-axis) across all possible thresholds for a continuous test. The area under the curve (AUC) summarizes overall discrimination: AUC 0.5 = no better than chance, 0.7–0.8 acceptable, 0.8–0.9 excellent, >0.9 outstanding.

Reporting & Appraisal Tools

Worked 2×2 Example

Suppose a troponin assay is tested in 1,000 patients with chest pain. 100 truly have MI (10% prevalence). The test is positive in 95 of the 100 with MI and in 90 of the 900 without MI.

• Sensitivity = 95/100 = 95%
• Specificity = 810/900 = 90%
• PPV = 95/185 = 51%
• NPV = 810/815 = 99.4%
• LR+ = 0.95 / 0.10 = 9.5
• LR− = 0.05 / 0.90 = 0.056

Notice that even with excellent sensitivity and specificity, PPV is only 51% because of the 10% prevalence. If prevalence were 1% (e.g., screening asymptomatic young adults), PPV would fall below 10%.

Diagnostic Thresholds & Tradeoffs

Most continuous tests force a tradeoff between sensitivity and specificity via the cutoff. Lower cutoffs raise sensitivity (and NPV) but reduce specificity (and PPV). The optimal threshold depends on the clinical consequence of false negatives vs false positives — missing a pulmonary embolism vs a false-positive CT angiogram involve very different harms. The ROC curve displays every possible tradeoff.

16 Prognosis Studies & Cohort Design

Prognosis studies estimate the likely course of a disease — rate of recovery, progression, complications, death, or functional outcome. They answer questions patients ask most often: “What will happen to me?” The ideal design is a prospective inception cohort with complete follow-up.

Validity Criteria for Prognosis Studies

Sources of Bias in Prognosis Studies

Reporting Prognostic Results

Reporting Guidelines

The STROBE statement guides reporting of observational studies (cohort, case-control, cross-sectional). TRIPOD guides multivariable prediction model reporting, and REMARK guides prognostic tumor marker studies.

17 Harm Studies & Hill Criteria

Harm studies ask whether an exposure (drug, environmental factor, device) causes an adverse outcome. Because harms often cannot be ethically randomized, most harm evidence comes from cohorts, case-control studies, and pharmacovigilance.

Cohort vs Case-Control for Harm

Validity Criteria for Harm Studies

• Clearly identified comparison group with similar baseline risk
• Exposure and outcome measured the same way in both groups
• Sufficient follow-up
• Adjustment for confounders (including indication bias)
• Dose-response relationship where possible

Bradford Hill Criteria for Causation

Austin Bradford Hill proposed nine viewpoints in 1965 to help judge whether an observed association is causal — not a rigid checklist but a structured way to reason:

18 SR Methodology & PRISMA

A systematic review is the highest form of literature synthesis when done well. Its value depends entirely on the rigor of its methods — a poorly conducted SR is worse than a single well-done trial because its conclusions carry false authority.

Steps of a Systematic Review

1. Frame a focused question (PICO)
2. Pre-register the protocol (e.g., PROSPERO)
3. Develop an explicit search strategy (multiple databases + gray literature)
4. Screen titles, abstracts, and full texts in duplicate
5. Extract data in duplicate with standardized forms
6. Assess risk of bias in each included study (e.g., Cochrane RoB 2, ROBINS-I)
7. Synthesize results (narrative and/or meta-analysis)
8. Assess certainty of evidence (GRADE)
9. Report transparently (PRISMA)

PRISMA — Reporting Standard

The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) statement is a 27-item checklist and flow diagram for transparent SR reporting. The 2020 update emphasizes protocol registration, data sharing, and explicit handling of bias.

Risk-of-Bias Tools

Cochrane Systematic Reviews

The Cochrane Collaboration produces the most methodologically rigorous systematic reviews in healthcare. Cochrane reviews follow standardized procedures documented in the Cochrane Handbook, undergo peer review, and are updated as new evidence accumulates. They are published in the Cochrane Database of Systematic Reviews (CDSR). Key features include:

• Pre-specified protocols with peer review
• Comprehensive search strategies including gray literature
• Standard risk-of-bias assessment (RoB 2)
• GRADE Summary of Findings tables
• Plain language summaries for patients
• Living review format for rapidly evolving topics

Living Systematic Reviews

A living systematic review is continuously updated as new evidence becomes available, rather than representing a snapshot in time. COVID-19 accelerated adoption of this approach — treatment guidelines needed to incorporate new RCTs within weeks rather than years. Living reviews rely on automated search alerts, streamlined screening, and modular updating.

19 Meta-Analysis & Forest Plots

Meta-analysis statistically combines results from multiple studies into a pooled effect estimate. Done well, it increases precision and power; done poorly, it produces a misleadingly confident summary of heterogeneous or biased studies.

Reading a Forest Plot

Fixed vs Random Effects

When heterogeneity is present, random-effects models are generally more appropriate because they acknowledge variability in the underlying effect. Random-effects models give more weight to small studies than fixed-effect models, which can be problematic if small studies are lower quality or biased.

Pooling Methods

Sample Forest Plot Interpretation Walk-Through

Imagine a forest plot of eight trials of a new oral anticoagulant vs warfarin for stroke prevention in atrial fibrillation. Six trials show point estimates to the left of the vertical line (favoring the new drug), one is right on the line, and one is slightly to the right. The pooled diamond sits at HR 0.82 with 95% CI 0.75–0.90, I² = 22%. What do you conclude?

• The pooled estimate shows an 18% relative reduction in stroke with the new drug.
• The 95% CI (0.75–0.90) excludes 1.0, so the result is statistically significant.
• I² of 22% indicates low heterogeneity — studies are reasonably consistent.
• The consistency across studies (most favoring the new drug) supports the pooled finding.

But before accepting the conclusion, you would check risk of bias in each trial, whether the heterogeneity has a clinical explanation (dose, population), and whether the pooled NNT (derived from baseline stroke risk in the control arm) is clinically meaningful.

20 Heterogeneity, Publication Bias & GRADE

Heterogeneity

Heterogeneity is variation between studies beyond what would be expected by chance. Sources include clinical (different populations, interventions), methodological (different designs, risk of bias), and statistical heterogeneity.

Publication Bias & Funnel Plots

Publication bias is the preferential publication of studies with positive, statistically significant, or novel results. It inflates meta-analytic effect estimates. Detection tools include:

• Funnel plot: scatter of effect size vs precision (typically 1/SE). In the absence of bias, it is symmetric about the pooled estimate. Asymmetry suggests missing small “negative” studies.
• Egger’s test: statistical test for funnel plot asymmetry.
• Trim-and-fill: imputes missing studies to estimate an adjusted pooled effect.

GRADE Certainty of Evidence

Evidence starts at High (RCTs) or Low (observational studies) and is rated up or down:

GRADE Recommendation Strength

Recommendations are dichotomized as strong (“we recommend”) or weak / conditional (“we suggest”). Strong recommendations require high-certainty evidence and benefits clearly outweighing harms for most patients. Weak recommendations reflect uncertainty or tradeoffs — shared decision-making is particularly important here.

Summary of Findings Tables

GRADE-compliant SRs present a Summary of Findings (SoF) table: for each critical outcome, it shows the risk in the comparator group, the risk in the intervention group, the relative effect, the number of participants and studies, and the GRADE certainty with reasons for downgrading. This single table is often the most valuable artifact of a systematic review.

21 Clinical Practice Guidelines & AGREE II

Clinical practice guidelines (CPGs) are “statements that include recommendations intended to optimize patient care, informed by a systematic review of evidence and an assessment of the benefits and harms of alternative care options” (IOM, 2011). Well-made guidelines translate vast evidence into actionable recommendations; poorly made guidelines perpetuate opinion dressed as evidence.

Types of Guidelines

Guideline Development Process

1. Scope and key questions (PICO)
2. Multidisciplinary panel with declared conflicts of interest
3. Systematic reviews of the evidence
4. Rating of evidence certainty (GRADE)
5. Drafting recommendations with explicit rationale
6. External review and public comment
7. Publication with a planned update cycle

AGREE II — Guideline Quality Appraisal

The AGREE II (Appraisal of Guidelines for Research and Evaluation) tool assesses guideline methodology across six domains:

Major Evidence / Recommendation Grading Systems

22 Shared Decision-Making

Shared decision-making (SDM) is the process in which clinicians and patients make decisions together using the best available evidence, accounting for the patient’s values, preferences, and circumstances. It is the clinical operationalization of the third pillar of EBM.

When SDM Is Essential

• Preference-sensitive decisions (e.g., PSA screening, prostatectomy vs active surveillance)
• Close tradeoffs between benefits and harms
• Weak / conditional recommendations under GRADE
• Decisions with major quality-of-life implications
• Screening tests (over-diagnosis tradeoffs)

Three Talk Model (Elwyn et al.)

Tools for SDM

Framing Effects in Risk Communication

Identical data can lead to different decisions depending on how they are presented — a cognitive bias called framing. A 90% survival rate sounds better than a 10% mortality rate, though they are arithmetically identical. Best practices include:

• Present both positive and negative framings when possible
• Use consistent denominators (e.g., “3 in 100” and “97 in 100”)
• Provide absolute numbers, not only relative ones
• Use visual aids (icon arrays, bar charts)
• Avoid loaded terminology (“aggressive cancer” vs “slow-growing”)
• Describe a reasonable time horizon (“over the next 10 years”)

Barriers to SDM Implementation

23 Special Topics — RWE, Big Data, ML, NMA

Real-World Evidence (RWE)

Real-world evidence is clinical evidence derived from routine care data — EHRs, claims databases, registries, wearables. The FDA formally incorporates RWE in post-marketing drug and device decisions. RWE can answer questions RCTs cannot: long-term safety, use in populations excluded from trials, comparative effectiveness in everyday practice.

Big Data & Machine Learning in Medicine

Appraising ML prediction models uses TRIPOD-AI / TRIPOD+AI, extending TRIPOD for AI-based models. Key concerns include dataset shift, algorithmic bias, lack of transparency, and the need for prospective validation before deployment.

Network Meta-Analysis (NMA)

Network meta-analysis pools direct and indirect comparisons to compare multiple interventions simultaneously, even when head-to-head trials are unavailable. For example, NMA can rank several antidepressants using a network of trials in which some drugs have never been directly compared. Key assumptions include transitivity (the indirectly compared populations are similar enough to allow inference) and consistency (direct and indirect evidence agree).

Individual Patient Data Meta-Analysis (IPD-MA)

IPD-MA obtains individual participant data from each included trial rather than aggregate summary data. It is the gold standard for pooled analysis because it allows standardized re-analysis, time-to-event modeling, and subgroup analyses not possible with published summaries. The tradeoff is significant time and resource cost.

Target Trial Emulation

Target trial emulation is a modern framework in which observational data are analyzed as if they were an RCT, with the hypothetical (target) trial specified in advance. This discipline forces researchers to define eligibility, treatment strategies, assignment procedures, follow-up start, outcomes, and analysis plans up front — reducing many common observational biases such as immortal time bias and prevalent-user bias. It has become a standard approach in comparative effectiveness research using EHR and claims data.

Mendelian Randomization

Mendelian randomization uses genetic variants as instrumental variables to infer causal effects of an exposure on an outcome. Because alleles are randomly assigned at conception (Mendel’s laws), they are largely independent of confounders that plague standard observational analyses. This approach has been used to clarify the causal role of LDL cholesterol, C-reactive protein, and body mass index in cardiovascular disease. Limitations include pleiotropy (variants affecting multiple pathways) and the need for large genetic datasets.

24 Common Pitfalls in EBM

Even carefully conducted studies can mislead. The following pitfalls are among the most common reasons that apparently strong evidence fails to replicate or translate.

Statistical vs Clinical Significance (Revisited)

A p-value tells you whether an effect is likely different from zero, not whether it matters. Very large trials can produce statistically significant differences of trivial magnitude; small trials can miss clinically important effects. Always look at effect size, confidence interval, and minimum clinically important difference.

Surrogate Outcomes

Surrogate outcomes substitute a lab value or imaging finding for a patient-important outcome (e.g., LDL for cardiovascular death, HbA1c for microvascular complications, bone mineral density for fractures). They are convenient but dangerous: the CAST trial (suppressing PVCs) and rosiglitazone (lowering glucose while increasing MI) are cautionary examples. Surrogates must be validated against hard outcomes before being trusted for decision-making.

Composite Endpoints

Composite endpoints combine multiple outcomes (e.g., “MACE” = death + MI + stroke + revascularization) into a single variable to boost power. They are valid only when component outcomes are of similar importance and affected in the same direction. Beware when a composite is driven by the least important component (e.g., hospitalization) while mortality is unchanged.

Subgroup & Post-Hoc Analyses

Subgroup analyses test whether an effect varies across patient subsets. They are hypothesis-generating, not confirmatory. Pre-specified subgroups with a plausible interaction test are more trustworthy than post-hoc analyses (e.g., the infamous ISIS-2 subgroup showing no benefit of aspirin for patients born under certain astrological signs — used as a parody of subgroup analysis).

p-Hacking, HARKing & the Garden of Forking Paths

White-Coat Science & Reverse Causation

White-coat science refers to overconfident interpretation of weak evidence clothed in the authority of expertise. Reverse causation occurs when the outcome causes the exposure rather than vice versa (e.g., low cholesterol in sick patients is often the result of cachexia rather than a cause of death).

Number Needed to Treat Traps

• NNT depends on baseline risk — quote NNT only for a specific population
• NNT depends on time horizon — always specify (e.g., NNT 50 over 5 years)
• NNT must be accompanied by NNH to give a fair picture of tradeoffs
• NNT from meta-analyses may be misleading when heterogeneity is high

The Reproducibility Crisis

Large replication projects (Open Science Collaboration in psychology, the Reproducibility Project: Cancer Biology) have shown that a substantial fraction of published findings do not replicate when the same studies are repeated. Causes include low statistical power, flexible analysis, publication bias, p-hacking, and inadequate description of methods. EBM’s response has included pre-registration, data sharing, methodologic transparency (CONSORT, STROBE, PRISMA), and emphasis on systematic reviews over single studies.

Ethical Considerations in EBM

25 Reference — Checklists & Tables

This section consolidates high-yield reference material — critical appraisal checklists, study design comparisons, GRADE summaries, and common database features.

Critical Appraisal Quick Checklist — Therapy

Critical Appraisal Quick Checklist — Diagnosis

Critical Appraisal Quick Checklist — Systematic Review

Study Design Comparison

GRADE Summary Table

Reporting Guidelines by Study Type

Common Databases — At a Glance

26 High-Yield Review

This final section distills the most commonly tested and clinically essential EBM concepts.

The EBM Mindset in One Paragraph

Practicing EBM means approaching every clinical decision with humility about uncertainty, discipline about evidence, and respect for patient values. It means framing answerable questions, searching efficiently, appraising critically, quantifying benefit and harm, and integrating evidence with clinical judgment at the bedside. It is not about memorizing trials — it is a durable set of skills that outlasts any individual fact and keeps practice aligned with the best available knowledge as medicine evolves.

Core Definitions to Master

Rapid-Fire Clinical Pearls