Bias and Confounding Lecture Notes

BIAS AND CONFOUNDING

• Objectives of the Lecture
  • Describe bias and confounders in observational and interventional studies.
  • Identify sources of bias and confounders.
  • Discuss methods to reduce bias in study designs.

Key Concepts

Observed Associations

• The observed association in quantitative studies can arise from:
  • Bias (Systematic Error): Errors in study design that create a false association between exposure and disease.
  • Confounding Bias: A distortion in the association between exposure and disease due to a third variable (confounder).
  • Variance (Random Error): Observed associations caused by chance.
• These are alternative explanations for observed associations.

Validity of a Study

Internal Validity

• Refers to whether a study is designed and conducted properly to yield accurate results.
  • Depends on:
  • Randomization
  • Measurement and assessment of variables
  • Participant dropouts
  • Overall methodology

External Validity

• The extent to which study results can be generalized to the larger population.
  • Depends on:
  • Recruitment and selection of participants
  • Eligibility criteria
  • Sample size
  • Study setting

Errors in Study Design

• Validity Issues: How well the study measures what it's intended to measure.
• Errors: Uncertainties in data estimates and inferences.
  • Bias: Systematic deviations from the true value leading to a statistical bias.
  • Variance: Random errors caused by mistakes from various sources (respondents, interviewers).

Understanding Bias

• Defined as a systematic error leading to incorrect estimates of association.
• Can occur at any research stage and is classified by occurrence stage and direction.
• Examples:
  • Failure to calibrate equipment
  • Systematic shifts from true values, usually clustered away from them.

Detecting Bias

• Investigative steps include:
  • Identifying Sources: Determines strength and magnitude of bias.
  • Estimating Magnitude: A large bias can falsely weaken associations.
  • Assessing Direction: Bias can underestimate or overestimate associations.

Examples of Bias Impact

1. Example 1: Relative risk of cancer = 1.9 (actual) vs. biased estimate = 1.4 (underestimated).
2. Example 2: Treatment group cancer risk = 0.4 (actual) vs. biased estimate = 0.7 (underestimated).
3. Example 3: Exposure risk = 2.0 (actual) vs. biased estimate = 2.6 (overestimated).
4. Example 4: Treatment risk = 0.5 (actual) vs. biased estimate = 0.3 (overestimated).

Types of Bias

• Bias is categorized based on recruitment methods, participation factors, and data collection distortions.
• Selection Bias:
  • Types include:
  • Sampling bias
  • Survivor bias
  • Non-response bias
  • Attrition bias
  • Volunteer bias
  • Under-coverage bias
• Information Bias:
  • Types include:
  • Recall bias
  • Interviewer bias
  • Misclassification bias
  • Performance bias
  • Regression to the mean

Selection Bias

• Results from how study participants are selected; leads to false associations.
• More pronounced in case-control and retrospective cohort studies.
• Key Features:
  • Occurs when eligibility differs based on exposure/outcome.
  • Difficult to correct once identified.

Avoiding Selection Bias

• Use consistent criteria for case/control selection.
• Aim for high participation rates and improve recruitment retention.
• Adopt various participant tracing methods.

Information Bias

• Results from differences in information accuracy among study groups.
• Examples and types include:
  • Recall Bias: Differential recall ability in a case-control context.
  • Interviewer Bias: Bias arising from the interviewer's knowledge of participant disease status.

Avoiding Information Bias

• Use blinding techniques to prevent bias in information collection and classification.
• Design clear and unbiased questionnaires.

Confounding

• Occurs when additional exposures affect both the disease and the exposure of interest.
• Independent Risk Factors: Related to both the disease of study and the exposure.
• Controlling for Confounding:
  • Can be done in design/analysis phase through various methods such as randomization, stratification, and multivariate analysis.