Association and Causation

Association Basics - Spotting Connections

• Association: Statistical relationship between two or more variables/events; they occur together more or less often than expected by chance.
• Types of Association:
  • Spurious/Artifactual: False association due to chance or bias (e.g., selection bias, measurement bias).
  • Indirect: Statistical association due to a common factor (confounder) linking exposure and outcome.
  • Direct/Causal: Exposure directly leads to outcome.
    • One-to-one: Factor A → Disease X.
    • Multifactorial: Factor A + Factor B → Disease X.
• Strength of Association: Measured by Relative Risk (RR), Odds Ratio (OR).
  • RR/OR = 1: No association.
  • RR/OR > 1: Positive association (risk factor).
  • RR/OR < 1: Negative association (protective factor).

⭐ Spurious association can arise from Berksonian bias, where hospital admission rates differ for exposed/unexposed or cases/controls, leading to a distorted association observed only in hospital settings.

Bradford Hill Criteria - Causation's Commandments

Framework for assessing if an observed association is likely causal. Fulfilling more criteria strengthens causal inference; not all are required. 📌 Mnemonic: "BTS CAPS CE".

• Strength of Association: Strong association (e.g., high RR/OR) is more indicative of causality.
• Consistency: Association observed repeatedly by different researchers, in diverse populations, settings, and times.
• Specificity: Exposure linked to a specific disease, not multiple outcomes. (Often the weakest criterion).
• Temporality: Exposure must precede disease onset. (This is the absolutely essential criterion).
• Biological Gradient (Dose-Response): Increased exposure (dose) correlates with increased risk or severity of disease.
• Plausibility: A biologically sensible mechanism can explain the association.
• Coherence: Association aligns with existing knowledge of the disease's natural history and biology.
• Experimental Evidence: Findings from controlled experiments (e.g., RCTs, animal studies) support the causal link.
• Analogy: Similar causal relationships are known for related exposures or diseases.

⭐ Temporality is the sine qua non for causation: the cause must precede the effect. Without this, an association cannot be deemed causal.

Bias & Confounding - Tricky Twins

• Bias: Systematic error in study design, conduct, or analysis; distorts true association. Not reduced by ↑ sample size.
  • Selection Bias: Groups differ systematically.
    • E.g., Berkson's (hospital-based), Neyman (incidence-prevalence), Healthy worker effect.
  • Information (Measurement) Bias: Errors in measuring exposure/outcome.
    • E.g., Recall bias (cases recall more), Interviewer bias (systematic questioning differences).
  • Control: Blinding, randomization, standardized protocols.
• Confounding: A third variable (confounder) distorts the exposure-outcome link.
  • Confounder:
    1. Associated with exposure.
    2. Independent risk factor for outcome.
    3. Not on causal pathway.
  • Control:
    • Design: Randomization, restriction, matching.
    • Analysis: Stratification, multivariate analysis.

⭐ Randomization is the gold standard for controlling confounding, especially for unknown confounders.

oka

Study Designs & Causality - Evidence Ladder

• Hierarchy of Evidence (↑ Causal Inference Strength):
  • Case reports/series, Ecological (hypothesis generation)
  • Cross-sectional (Snapshot, prevalence)
  • Case-Control (Retrospective, Odds Ratio)
  • Cohort (Prospective/Retrospective, Relative Risk, Incidence)
  • Randomized Controlled Trial (RCT - Intervention, experimental)
  • Systematic Review & Meta-analysis (Highest, pooled evidence)
• 

⭐ RCTs provide the strongest evidence for causality among primary study designs due to minimized confounding through randomization.

• Bradford Hill criteria are key for assessing causality, not just association.
• Strength of association (e.g., high RR/OR) is a strong criterion, but not definitive.
• Temporality (cause precedes effect) is the only essential criterion for causality.
• Dose-response relationship (↑ exposure → ↑ risk) strengthens causal inference.
• Consistency of findings across diverse studies and populations is crucial.
• Specificity (one cause → one effect) is weak for chronic diseases.
• Biological plausibility and coherence with existing knowledge support causality.