L2 - Measurement Error, Bias & Confounding

Measuring Exposures

• First step in any public-health study: decide how an exposure will be operationalised and captured.
  • Questionnaire
  • Self-administered vs. interviewer-administered.
  • Consider differences in response patterns, data quality, feasibility, scalability.
  • Biological sampling
  • Blood, urine, hair.
  • Radiology / imaging.
  • Physical examination (e.g. height, weight).
  • Abstraction of previous medical records.
  • Environmental sampling (air, water, soil, noise, etc.).
  • Biomarkers & genetic markers
  • Example: School of Public Health pharmacogenomic screening group (Prof. Paul McKay).
  • Cheek-swab panel → tests multiple genetic variants to predict medication response.
• Core research question: “Is there an association? Is there a possible cause-and-effect?”
  • To answer, we must judge whether findings are due to ▸ bias ▸ confounding ▸ chance ▸ a true causal relationship.

Measurement Error & Imprecision

• Virtually all measurements include error.
• Mathematical representation:
  • \text{Measured score}=\text{True score}\pm \text{Error}
  • Error contributes to \sigma (standard deviation) of the measure.
• Two overarching categories of error:
  1. Bias (systematic error) – consistent deviation in one direction.
  2. Random error / imprecision – non-systematic fluctuation around the true value.

Bias: Definition & Importance

• Bias = systematic error that leads to incorrect estimate of the exposure–outcome relationship.
  • Causes the difference between study result and true population value.
  • Repeating the study under identical biased conditions will reproduce the wrong answer.
• Risk-of-bias assessment / critical appraisal = formal process of identifying & minimising these errors.
• Eliminating bias protects against:
  • Inappropriate influence of funders.
  • Co-interventions & contamination of results.
  • Selective reporting of subgroups.
  • Baseline imbalances in important factors.

Taxonomy of Bias Covered in the Course

• Selection bias
  • Sub-types: ascertainment bias, detection bias.
  • Occurs when selected participants are not representative of the target population.
  • Example: Choosing a new-treatment group that is inherently healthier → overestimates treatment benefit.
• Recall bias
  • Archetypal in case–control studies.
  • Participants with an event (e.g. stroke) may recall past exposures more (or less) accurately than controls.
  • Not intentional; arises from differential memory.
• Measurement (information) bias
  • Systematic errors in how data are collected or instruments are calibrated.
• Publication bias
  • Studies with significant or positive findings are more likely to be published.

Confounding

• Literal meaning: “mixed together.”
• Formal definition: A non-causal association between exposure (E) and outcome (O) produced by a third variable (C) that is related to both E and O.
  • C must not be an intermediate step on the causal pathway.

Coffee–Heart-Disease Example

• Observation: Coffee drinkers appear to have higher heart-disease rates than non-drinkers.
• Hypothesis: “Coffee causes heart disease.”
• Investigation reveals third variable smoking:
  • Smokers drink more coffee and smoking causes heart disease.
  • Association E\rightarrow O was distorted: estimate contained contribution from both coffee and smoking.
• After removing/adjusting for smokers → no coffee–heart-disease association, aligning with wider literature.

Strategies to Control Confounding

Design Stage (pre-data-collection)

• Randomisation
  • Equal distribution of both known & unknown confounders across groups.
• Restriction
  • Exclude participants with the confounder (e.g. analyse only non-smokers).
• Matching
  • Select controls with identical values of confounders as cases.
  • Downsides: expensive, time-consuming, and matched variables can no longer be analysed as exposures.

Analysis Stage (post-data-collection)

• Stratification
  • Separate analysis within levels of the confounder (e.g. smokers vs. non-smokers).
• Multivariable statistical modelling
  • Include multiple covariates in regression to adjust simultaneously.
  • Allows exploration of numerous potential risk factors while preserving the “full picture.”
  • Example: Late antenatal care among Aboriginal infants investigated with up to 9 risk factors; modelling allowed stepwise identification of the dominant contributors without discarding information.

Practical & Ethical Implications

• Good study design prevents erroneous public-health recommendations.
• Failure to recognise bias or confounding can:
  • Waste resources.
  • Lead to ineffective or harmful interventions.
  • Undermine trust in research.
• Continuous cycle: Design ↔ Measurement ↔ Analysis ↔ Interpretation must address error at every step.

Key Take-Home Equations & Concepts

• Measurement model: \text{Observed}=\text{True}\pm \varepsilon
• Bias ≠ random error; elimination of one does not automatically remove the other.
• Confounder criteria:
  1. Associated with the exposure.
  2. Independent risk factor for the outcome.
  3. Not an intermediate variable in the causal pathway.
• Control options summary:
  • \text{Design stage}\rightarrow {\text{randomisation, restriction, matching}}
  • \text{Analysis stage}\rightarrow {\text{stratification, multivariable models}}

Connections to Previous & Future Lectures

• Builds on foundational principles of epidemiologic study design (cohort, case–control, RCT).
• Will dovetail with upcoming sessions on statistical inference, validity vs. reliability, and causal diagrams (DAGs).
• Lays ethical groundwork for transparent reporting (CONSORT, STROBE) and systematic-review methodology.