Selection Bias

Confounding refers to variables that influence both the dependent and independent variables.
Key Terms:
- Adjusted for: When confounders are accounted for in the analysis.
- Controlled for: Similar to adjusted for, but often implies using specific methodologies to isolate the effect of the confounder.
Sources of Confounding: Often identified in tables or supplementary materials in research articles.
Examples from recent AJPH articles: Political science and public health cross-sections.

Study focused on employed adults aged 18-64 years in the United States during 2021-2022.
Confounding Variables Included:
- Age
- Sex
- Race/Ethnicity
- Country of Birth
- Family Income
- Family Size
- Marital Status
- Education
- Region
- Insurance Status
- Occupation
- Industry of Employment
- Survey Year
- Potential Effect Modifier: Hours worked per week
Statistical Methodology: Adjusted using quasibinomial logistic regression with a logit link.

Focused on general responders to the World Trade Center disaster.
Adjusted Models for:
- Sociodemographic characteristics (age, sex, race/ethnicity, education, marital status)
- Body mass index
- Census tract-level socioeconomic variables (e.g., percentage below poverty line)
Statistical Methodology: Cox proportional hazard ratio models to analyze associations of WTC exposure on outcomes.
Sample Size: 47,795.
- Metrics: Results reported per interquartile range increments, which were 176 hours and 68 days of exposure.

Study time frame: January 2022.
Factors Considered:
- Age, sex, race, education, service branch, and military service history.
Table Data: Adjusted odds ratios (OR) presented for completion of COVID-19 vaccination.
- Example Entries:
- Males (Reference):
  - Adjusted OR: 0.86 (95% CI: 0.84, 0.88)
- Females:
  - Adjusted OR: 1.19 (95% CI: 1.15, 1.23)
- Age breakdown detailed from <20 to ≥45, observing significant odds ratios differing by demographic factors.

Importance of effectively finding, reading, synthesizing, and critiquing literature as essential skills in public health.
Core competencies including applying epidemiologic principles for population health understanding.
Students are equipped to evaluate evidence quality and limitations for future public health solution proposals.

Main Goal: Differentiate between common types of selection bias and their impacts on study results.

Selection Bias: A systematic error in study design/conduct resulting in an inaccurate estimate of the association between exposure and outcome.

Information Bias: Distortion in exposure, outcome, or covariates measurement.
Selection Bias: Error during the participant selection process.
Consequences of random and systematic error on study validity.

Assessing Influence:
- Are selection probabilities influenced by exposure and outcome?
- What are the differences between groups in relation to exposure/outcome?
- How were participants recruited?

Participant Selection Procedures:
- Voluntary participation, illness, migration, or refusal can introduce bias.
Loss to Follow-up: Affects cohort studies, with implications for data integrity based on participant retention.
Sampling Procedures: High probabilities of differing inclusion rates can produce systematic discrepancies.

Healthy Worker Effect: More likely for healthier individuals to meet study criteria compared to those unemployed or disabled.
Publicity Bias: Individuals may self-identify for studies post-publicity due to perceived relevance.
Selection bias can skew results towards or away from the null hypothesis and is often exacerbated in retrospective studies.

Appropriate Participant Selection: Ensure groups stem from the same source population and utilize effective recruitment strategies.
Minimizing Attrition: Keep in communication with participants and track reasons for non-responses.

Review the provided handout regarding design improvement for studying selection bias, to discuss as a class.