--- module: 066-02 language: en chapter: 66 title: "Evidence-Based Medicine, Causation, Bias, and Interpreting Clinical Research" module_title: "Causal inference, effect interpretation, diagnostic evidence, and patient-level application" source_sha256: d88cf1fec87958422c451cce06056462ec01689a969082f64a5a48accf76e65c --- # Causal inference, effects, diagnosis, and application ## Four questions and focused framing ### Valid, how large, applicable, worth the burden ### Prestige, p value, or guideline answers none alone ### Treatment question includes a time horizon ### Harm: timing, dose, latency, competing causes ### Vague questions invite surrogate answers ## Trial conduct and reporting ### Conceal sequence so recruiters cannot manipulate ### Objective outcomes not automatically immune to bias ### Loss to follow-up can destroy balance ### Per-protocol answers an adherence-conditioned question ### Check registry and protocol for post-hoc changes #### Early stopping locks in chance effects ### Funding alone is not grounds for rejection ### Small trials may be imbalanced #### Baseline significance test is not the solution ### Cluster correlation ignored narrows intervals ## Observational studies and causal diagrams ### Essential for prognosis, rare harm, long latency ### Confounder influences exposure and outcome ### Adjustment cannot erase unmeasured confounders ### Channeling: sicker patients get one drug #### New-user active-comparator design aligns start ### Mediator adjustment removes part of the effect ### Collider conditioning creates false association #### Selection into hospital or study as collider ### Adjust by causal question, not availability ## Bias and missing data ### Selection depends on exposure and prognosis ### Recall: cases reconstruct exposure differently ### Misclassification dilutes or exaggerates ### Complete-case analysis assumes ignorable missingness #### Reasons and sensitivity analysis needed ## Effect measures and uncertainty ### Odds ratio approximates risk only when rare ### Hazard ratio is an instantaneous rate ratio ### Number needed to treat keeps population and time ### Interval width reflects information #### Narrow around trivial effect is unimportant ### Non-inferiority needs a justified margin #### Non-adherence biases toward similarity ### P value measures surprise under a null model ### Multiplicity inflates false positives #### Strict correction can hide a true signal ### Bayesian prior and model must be transparent ## Composite, competing, and surrogate outcomes ### Composite may be driven by least important part ### Competing death prevents later events #### Kaplan-Meier censoring overestimates incidence ### Surrogate must predict patient-important effects #### Non-causal marker, off-target harm, timing ### Disease improvement with worse mortality ## Diagnostic and prediction evidence ### Blinded reading, independent verification ### Partial verification inflates accuracy ### Incorporation: index test within reference ### Referral-centre results may not transfer ### Good ranking can coexist with overprediction ### Internal validation estimates optimism ### External validation tests transport ## Heterogeneity and systematic reviews ### Credible effect modification criteria ### Do not cut continuous modifiers arbitrarily ### Meta-analysis cannot repair biased studies ### Random effects do not unify questions ### Prediction interval for a new setting ### Funnel asymmetry has causes besides publication ### Test robustness with low-bias studies only ## Harms and external validity ### Registries give denominators but confounding ### Spontaneous reports find signals, not incidence ### Deliberate rechallenge may be unethical ### Exclusion creates uncertainty, not contraindication ### Relative effects may transport better ### Cost, travel, caregiver burden shape effectiveness ## Certainty, decision, and time-limited trials ### Certainty is separate from magnitude ### Natural frequencies with and without treatment #### Relative framing alone can manipulate ### Decision table at the patient's baseline risk ### Weak evidence, high urgency: clinicians still act #### Target, timeframe, monitoring, stopping rule ### Regression to the mean mimics improvement #### One patient's response is not population proof ### Record evidence, limits, assumptions, triggers ### Transparency of judgement raises safety