--- module: 066-01 language: en chapter: 66 title: "Evidence-Based Medicine, Causation, Bias, and Interpreting Clinical Research" module_title: "Foundations" source_sha256: 2e04606571cd0fe096b5f623d0ef844b94b61a0927a5410afbb3d56041d7d2b7 --- # Evidence-based medicine foundations ## Orientation and question framing ### Research plus expertise, values, biology, context #### Not obeying a hierarchy without judgement ### Methods before results ### Absence of evidence is not evidence of no effect ### Treatment: population, intervention, comparator, outcomes ### Diagnosis: condition, test, reference, setting ### Prioritise patient-experienced outcomes ### Biomarkers and composites may mislead #### Surrogate must reliably change clinical outcome ## Study designs ### Randomisation balances prognostic factors on average ### Concealment prevents prediction before enrolment ### Blinding reduces differential care and assessment ### Intention-to-treat preserves randomisation #### Per-protocol loses it, vulnerable to selection ### Crossover, cluster, factorial, pragmatic trials ### Case-control: efficient for rare disease #### Vulnerable to selection and recall ### Meta-analysis adds precision, not unbiasedness #### Pooling incomparable studies misleads precisely ## Random error and estimation ### Confidence interval expresses precision #### Ninety-five per cent of repeated intervals ### P value assumes the null model is true #### Not probability the null is true ### Significance depends on sample size ### Number needed to treat varies with baseline risk ### Non-significance is not equivalence #### Non-inferiority needs prespecified margins ## Bias within studies ### Selection: groups differ in prognosis ### Performance: unequal co-interventions ### Detection: outcome measurement differs ### Attrition and selective reporting ### Random error lowers precision, systematic shifts ### Recall and interviewer bias ### Confounding creates non-causal association #### Adjustment cannot remove unmeasured confounding #### Adjusting mediator or collider adds bias ## Causation ### Temporality, consistency, dose-response, mechanism #### No checklist proves causation ### Randomisation best supports intervention causality ### Triangulation across differently biased designs ### Biomarker correlation does not prove targeting helps ### Reverse causation from early disease ### Immortal time favours the exposed ## Diagnostic evidence ### Representative patients, complete verification ### Partial verification, incorporation, review bias ### Sensitivity and specificity vary with spectrum ### Predictive values vary with prevalence ### Area under curve measures discrimination only #### Adds little if management does not change ### Prediction models need external validation #### Calibration compares predicted with observed ## Prognosis and survival ### Inception cohort, follow-up, objective outcomes ### Kaplan-Meier assumes comparable censored risk ### Hazard ratio is not a risk ratio ### Ignoring competing events overestimates incidence ### Group prognosis is not individual destiny ## Harms and pharmacovigilance ### Trials too small or short for rare harms ### Spontaneous reports cannot give incidence #### Reporting and denominator unknown ### Temporality, dechallenge, rechallenge, dose ### Channeling: high-risk patients get one drug #### New-user and active-comparator designs ## Applicability and research integrity ### Does the patient resemble participants ### Exclusion raises uncertainty, not ineffectiveness ### Absolute effects depend on baseline risk ### Guidelines add values and may lag ### Credible subgroups: prespecified, tested interaction ### Positive novel studies more likely published ### Multiple analyses inflate false positives ### Disclosed conflict does not void sound methods ## From evidence to decision ### Certainty is not size of benefit ### Absolute natural frequencies, same denominator ### Include harms, burden, and no intervention ### Shared decision-making on what matters ### Weak evidence, high urgency: time-limited trial #### Document assumptions and stopping rules ### Act, measure, learn, and revise