L1 - Measurement & Errors in Medical Research – Detailed Lecture Notes

Learning Objectives

• Understand challenges in planning medical & public-health research.
• Distinguish types of measurement error, bias, & confounding.
• Clarify the difference between \text{validity} (accuracy) and \text{reliability} (precision).

Core Epidemiological Vocabulary

• Determinants / Exposures / Treatments
  • Statistical notation: independent variable X in an X!\rightarrow!Y model.
  • Examples: sun exposure, use of sun-beds, helmet use.
• Health-related states & events / Outcomes / Diseases
  • Statistical notation: dependent variable Y.
  • Can be continuous, categorical, or binary; not only yes/no.
• Event vs. Disease
  • Event = discrete (e.g., stroke, intracerebral haemorrhage).
  • Disease = chronic/long-term (e.g., melanoma).
• Confounder
  • A third variable associated with both X and Y, distorting their true relationship.

Ethical & Feasibility Principles

• Every research question requests time, samples, or risk from participants → respect their contribution.
• Avoid overly vulnerable populations unless justified.
• Principle of "first, do no harm" when selecting exposures or interventions.

Building a Research Question

• Generic causal template:
  • “Exposure X is associated with an increase/decrease in risk of outcome Y.”
• Practical brainstorming cues:
  • What is happening in the population?
  • \text{Incidence? Prevalence?}
  • Which exposures & outcomes appear important or modifiable?
• Example:
  • X = wearing a motorcycle helmet.
  • Y = death from head trauma.
  • Hypothesis: Helmet use ↓ risk of death.

Causation vs. Association

• Holy Grail = causal inference (“cause → effect”).
• Measured with effect sizes such as relative risk RR or odds ratio OR.
  • RR = \dfrac{Incidence{exposed}}{Incidence{unexposed}}
• Other factors (confounders) may blur the X!\rightarrow!Y link; must be identified & controlled.

Descriptive vs. Analytical Epidemiology

• Descriptive (hypothesis generation)
  • Uses “person, place, time” framework.
  • Answers: Who? Where? When?
• Analytical (hypothesis testing)
  • Employs formal study designs (cohort, case-control, RCT, etc.).
  • Requires biostatistical input: sample size, power, adjustment for confounders.

Qualities of a Worthwhile Exposure Variable

• Influences health in a modifiable way.
• Measurable with acceptable error.
• Differentiates populations (variation exists).
• Generates testable hypotheses.
• Enables prevention/control strategies.
• Feasible financially & logistically.

Qualities of a Worthwhile Outcome Variable

• Significant impact on individuals/public health (burden, severity, costs).
• Clearly defined causal pathway or risk factors.
• Measurable with valid, reliable instruments.
• If rare, still investigable via efficient designs (e.g., case-control).

Sources of Epidemiologic Hypotheses

• Individual cases or case series.
• Descriptive data (registries, surveillance).
• Data mining (“dredging”) of large datasets.
• Laboratory/animal studies suggesting human parallels.
• Analogies from other exposure–outcome pairs.
• Common sense & everyday observations.
• Conversations with colleagues, clinicians, patients, families.
• Media (newspapers, TV) highlighting emerging issues.

Practical Search Strategy for Hypotheses

• Read primary literature & systematic reviews.
• Examine public datasets for trends/anomalies.
• Conduct stakeholder interviews (patients, clinicians).
• Stay alert to news reports of outbreaks or controversies.

Recurring Methodological Challenges

• Chronic-disease etiology often multi-factorial & time-lagged.
• One exposure can yield multiple outcomes; one outcome can stem from multiple exposures.
• Different individuals may arrive at the same disease via distinct causal pathways (gene–environment interplay, etc.).
• Observational nature of most epidemiologic work → limited control over exposure assignment.

Observational vs. Experimental Designs

• Observational studies: cohort, case-control, cross-sectional.
  • Do not manipulate exposure; rely on “inadvertent” or naturally occurring exposure patterns (e.g., workplace, geography).
• Experimental study (clinical trial)
  • Only major exception where humans are deliberately assigned an exposure/intervention, under strict ethical oversight.

Role of Biological Plausibility

• Statistical associations strengthened when consistent with known pathophysiology.
• Requires interdisciplinary understanding of biology, pathology, toxicology.

Example Walk-Through: Skin Cancer Project

• Exposure candidates:
  • Sun exposure dosage.
  • Artificial tanning (sun-beds).
• Outcomes:
  • Melanoma incidence.
  • Mortality in severe cases.
• Design considerations:
  • If melanoma is rare in target population, case-control design efficient.
  • Need accurate exposure assessment (UV dose, tanning frequency).
  • Potential confounders: skin type, sunscreen use, family history.

Example Walk-Through: Helmet Campaign

• Research underpinning public-health ad:
  • Statement: “I won’t wear a helmet—it makes me look stupid.”
  • Visual: injured youth, older caregiver hands feeding them → emotional resonance.
• Data used: prior studies quantifying risk reduction in head trauma fatalities with helmet use.
• Ethical/communication lesson: translating epidemiologic evidence into persuasive health-promotion messages.

Measurement Error, Bias & Confounding (Preview)

• Measurement error: deviation between observed & true value (random vs. systematic).
• Bias: systematic deviation of study results from truth (selection, information, publication, etc.).
• Confounding: mixing of effects—third variable distorts X-Y relation; addressed via design (randomisation, restriction, matching) or analysis (stratification, regression).

Validity vs. Reliability

• Validity (accuracy): how close a measurement is to truth.
• Reliability (precision): how consistently the measurement can be repeated.
• Interaction: a tool can be reliable but invalid; optimal instruments aim for both.

Key Take-Home Messages

• Formulating a precise, feasible research question is foundational; respect participants’ contribution.
• Always weigh ethical, financial, and logistical constraints when selecting exposures & outcomes.
• Use descriptive epi to observe patterns, then analytical epi to test causal hypotheses.
• Collaboration with biostatisticians maximises study efficiency & validity.
• Epidemiology seeks population-level associations; individual causation usually indeterminable.
• Observational designs dominate, but well-designed trials offer the strongest causal inference within ethical bounds.