Bias in Epidemiologic Studies

🔍 What Is Bias?
●​ Bias = Systematic error in estimating the association between exposure and outcome
●​ Results in an incorrect risk ratio, odds ratio, or rate ratio
●​ Introduced at either the design or analysis stage

🔁 Bias vs. Chance

Feature Random Error (Chance) Bias (Systematic Error)

Cause Natural variation Design flaw or measurement issue

Effect Random distortion Predictable distortion of association

🧠 Types of Systematic Error

1.​ Selection Bias
2.​ Information Bias
3.​ Confounding (covered separately)

📌 1. Selection Bias: Error due to procedures used to select participants or

factors that influence participation or follow-up

🔁 Common Sources:
●​ Loss to follow-up: Differential dropout by exposure or outcome status
●​ Volunteer/non-response bias: Participants differ from the target population
●​ Hospital patient bias (Berkson’s bias): Hospital-based controls may have other
conditions related to the exposure
●​ Healthy Worker Effect: Workers are generally healthier than the general population
🧪 Case-Control Example
●​ Exposure: Coffee
●​ Cases: Pancreatic cancer patients
●​ Controls: Hospital patients with GI issues (who avoid coffee)
●​ ➤ Problem: Controls have abnormally low exposure → false positive association

👥 Cohort Example
●​ Exposure: Smoking
●​ Outcome: Dementia
●​ Problem: Older participants = survivors; bias may suggest smoking is protective
because vulnerable smokers died earlier

📝 2. Information Bias:
exposure, outcome, or covariates
Distortion due to inaccurate measurement of

🧩 Sources of Information Bias:

Type Description

Misclassification Errors in classifying exposure or disease

Interviewer bias Different probing/measurement between groups

Recall/reporting bias Unequal accuracy of self-reported info

📊 Misclassification
🔹 Non-differential Misclassification
●​ Misclassification occurs equally across groups
●​ Usually biases toward the null
●​ Example: Overreporting physical activity in both groups

🔹 Differential Misclassification
●​ Misclassification differs between cases/controls or exposed/unexposed
●​ Can bias toward or away from the null
●​ Example: Exposed individuals more likely to report symptoms
🎙️ Interviewer Bias
Occurs when:

●​ Interviewers probe more in one group

●​ Reviewers interpret data differently by group​

✅ Prevention Strategies
1.​ Blind data collectors to exposure/outcome status
2.​ Use standardized protocols
3.​ Train interviewers thoroughly
4.​ Use multiple data sources when possible

🧠 Recall/Reporting Bias
●​ Occurs when recall differs between groups
●​ Cases may over-report exposures due to concern
●​ Exposed may report symptoms more accurately due to fear​

✅ Prevention Strategies
●​ Use well-designed questionnaires
●​ Allow anonymous or self-administered responses​
Confirm with records or biomarkers, if possible

✅ Best Practices to Reduce Bias

●​ Plan to minimize errors at the design stage
●​ Pilot test instruments
●​ Use blinding, objective measurements, and clear definitions​

●​ Ask:​

○​ Could bias have occurred?

○​ Would it cause distortion toward or away from the null?
○​ Is the bias likely meaningful?
 🧾 Summary Table
Type of Bias Definition Example

Selection Bias Error from how participants are Dropout differs by

selected/followed exposure/outcome

Information Bias Error from inaccurate measurement Misclassification, recall error

Interviewer Bias Different methods for different groups Probing exposed group more

Recall Bias Differential recall between groups Cases over-report past

exposures
Confounding & Random Error

🧠 Confounding
🔑 Definition
A confounder is a third variable that distorts the true association between an exposure and an
outcome.

Not due to study design. Can occur in any type of epidemiologic study.

⚠️ Confounding Effects
●​ Can overestimate or underestimate true associations
●​ Can make a harmful exposure appear protective (and vice versa)
●​ Example:
○​ True RR = 0.5 → Confounded RR = 0.8 (bias toward null)
○​ True RR = 2.5 → Confounded RR = 3.1 (bias away from null)

✅ A Variable Is a Confounder If:

1.​ It is an independent predictor of the outcome
2.​ It is associated with the exposure
3.​ It is not on the causal pathway between exposure and disease

🧪 Example: Alcohol & Lung Cancer

●​ Alcohol initially appears associated with lung cancer
●​ BUT: Alcohol drinkers are more likely to be smokers
●​ When separated by smoking status → no alcohol effect
●​ ➤ Smoking is the confounder

🔄 Confounder vs. Intermediate

●​ Confounder: A third variable that distorts the relationship
 ●​ Intermediate: Lies on the causal pathway
○​ Example: Running → ↑ Lung Capacity → Faster Speed
○​ Lung capacity is not a confounder, but an intermediate

📚 How to Identify Potential Confounders

●​ Know your subject area
●​ Do a literature review
●​ Use lists of historical confounders (e.g., age, sex, race/ethnicity)

🛠️ Controlling Confounding
📐 Design Phase
1.​ Randomization
○​ Equal chance of exposure assignment
○​ Balances known and unknown confounders
○​ Used in Randomized controlled trials (RCTs)​

2.​ Restriction
○​ Limit study to one level of the confounder (e.g., only men, only smokers)
○​ Reduces generalizability​

3.​ Matching
○​ Select comparison groups with similar levels of the confounder
○​ Used in cohort and case-control studies

🧮 Analysis Phase
1.​ Standardization: Adjust rates to a standard population
2.​ Stratified Analysis: Analyze within strata of the confounder
3.​ Multivariate Analysis: Adjust for multiple confounders simultaneously

📊 Assessing Confounding
 ●​ Use this formula:​

●​ If <10%, likely little confounding

●​ If >10%, confounding may be present

⚠️ Residual Confounding
●​ Confounding left over after adjustment
●​ Possible reasons:
○​ Confounder not measured
○​ Measurement error
○​ Broad categories (e.g., wide age groups)
●​ ➤ Acknowledge in the discussion section of your paper

🎲 Random Error:
association
Errors due to chance that result in inaccurate estimates of

🧬 Sources
●​ Measurement error
●​ Sampling variability
●​ Occurs in all studies

🔄 Effects of Random Error

●​ May suggest an association that doesn’t exist
●​ May mask a real association
●​ Described using probability-based statistics

🧪 Sampling Variability
●​ We use samples to estimate the truth about a population
●​ Variability can occur because:
○​ Samples differ from the population
 ○​ Small samples = more error
●​ Example: Anecdotal evidence (e.g., older men gamble more) may be misleading

📉 Evaluating Random Error

🧮 Two Statistical Tools:
1.​ Confidence Intervals (CIs)
○​ Range of values within which the true value likely falls
2.​ Hypothesis Testing / p-values
○​ p < 0.05 often considered “statistically significant”

💊 Example: Clinical Significance ≠ Statistical Significance

●​ Finasteride study showed a 0.3 point improvement in symptoms
●​ Statistically significant
●​ But not clinically meaningful (need ≥ 3 point change to feel a difference)

✅ Summary: Error Types in Epidemiology

Type Description Source

Confounding Third variable distorts exposure-outcome Systematic

Selection Bias Error in participant selection Systematic

Information Bias Error in measurement of exposure/outcome Systematic

Random Error Chance-related distortion Random

​
Causal Inference in Epidemiology

🔎 What Is a Cause?
📖 Definitions
●​ Merriam-Webster: Something that brings about a result
●​ Kenneth Rothman: An event or characteristic without which disease would not
occur
●​ Mervyn Susser: Something that makes a difference

⚙️ Characteristics of Causes
●​ Can be host or environmental factors
●​ Can be positive (e.g., smoking) or negative (e.g., lack of exercise)
●​ Causative exposures: air pollution, toxins
●​ Preventive exposures: exercise, healthy diet, vitamins

🔁 Three Essential Attributes of a Cause

1.​ Association: Exposure and outcome must co-occur
2.​ Time Order: Cause must occur before the effect
3.​ Direction: One-way relationship (A → B, not B → A)
○​ E.g., smoking causes low birth weight, but low birth weight doesn’t cause
smoking

⚠️ Risk Factors ≠ Causes

Characteristic High-Risk Group Low-Risk Group

Place of Birth North America/Europe Asia, Africa

Socioeconomic Status High Low

Marital Status Never Married Ever Married

 Not all risk factors are direct causes.

🧬 Historical Models of Causation

Era Theory

Ancient Divine punishment or body humors

Pre-modern Miasma theory (bad air)

Modern Germ theory (microorganisms)

Contemporary Web of causation & SCM

🕸️ Web of Causation
●​ Many interconnected factors contribute to disease
●​ Used especially for chronic diseases
●​ Example (MI): Includes stress, diet, smoking, inactivity, genetics

🥧 Sufficient Cause Model (SCM)

🔑 Key Concepts
●​ Sufficient cause = a full “pie” of component causes that inevitably leads to disease
●​ Component causes = individual “pie pieces”
●​ Necessary cause = appears in every sufficient cause (e.g., HIV virus for AIDS)​

✅ Application
●​ Prevent disease by removing just one pie piece
●​ SCM explains how different combinations can cause the same disease

🔍 Association vs. Causation

 Just because two things are associated doesn’t mean one causes the other.

Use Hill’s guidelines as a framework to evaluate whether an observed association is likely to

be causal.

📋 Hill’s Guidelines for Assessing Causality

Guideline Description

Strength Stronger associations are more likely causal

Consistency Repeated in different studies/populations

Specificity One cause → one effect (rare in chronic diseases)

Temporality Cause must precede effect (required)

Biological Gradient Dose-response relationship

Plausibility Fits with known biology

Coherence Aligns with existing knowledge

Experiment Evidence from intervention studies

Analogy Similar exposures cause similar outcomes

⚠️ These are guidelines, not strict rules. They support, but do not prove,
causation.

🧠 Quote from Sir Austin Bradford Hill (1965)

“All scientific work is incomplete... That does not confer upon us a freedom to
ignore the knowledge we already have, or to postpone the action that it appears to
demand.”

📌 Summary
●​ Causality is complex and often cannot be proven definitively
●​ Use SCM and Hill’s Guidelines to evaluate evidence
●​ Public health decisions often rely on best available evidence—not perfect certainty

​
Reading & Critiquing Epidemiologic Studies

🧭 Overview: Critical Review Framework

A structured critique of epidemiologic studies includes three main phases:

1.​ A. Collection of Data

2.​ B. Analysis of Data
3.​ C. Interpretation of Data

🧪 A. Collection of Data
A.1: Study Context

●​ What prompted the study?

●​ Was it informed by previous literature or research gaps?

A.2: Study Objectives

●​ What hypotheses are being tested?

A.3: Primary Exposure

●​ What is the main exposure being evaluated?

●​ Was it accurately measured?
○​ Distinguish conceptual (theoretical) vs. operational (measured) definitions
○​ Be aware of misclassification bias

A.4: Primary Outcome

●​ What is the primary health outcome of interest?

●​ Was it accurately measured?
○​ Again, consider conceptual vs. operational definitions
○​ Consider possible misclassification

A.5: Study Design

●​ Identify the type of study:

○​ Experimental​
 ○​ Cohort (prospective or retrospective)
○​ Case-control
○​ Cross-sectional
○​ Ecological

A.6: Study Base & Participants

●​ What population was studied over what period?

●​ How were participants selected?
●​ What was the sample size and the exposed:unexposed ratio?
○​ Affects statistical power, comparability, and generalizability

A.7: Selection Bias

●​ Did refusal, non-response, or loss to follow-up differ by exposure and disease?

●​ How likely is bias from selection issues?

A.8: Information Bias

●​ Was there potential for:

○​ Recall bias
○​ Interviewer bias
○​ Misclassification?

A.9: Confounding (Pre-analysis)

●​ How was confounding minimized before analysis?

○​ Methods:
■​ Randomization
■​ Restriction
■​ Matching
■​ Careful data collection on known confounders
■​ Use of comparable groups from the same source population

📊 B. Analysis of Data
B.1: Control of Confounding (During Analysis)

●​ Were appropriate methods used?

○​ Stratified analysis
○​ Multivariate analysis
○​ Standardization
B.2: Measures of Association

●​ What was reported?

○​ Risk/rate ratio
○​ Risk/rate difference
○​ Odds ratio

B.3: Measures of Statistical Stability

●​ Were p-values or confidence intervals reported?

○​ P-values → hypothesis testing
○​ Confidence intervals → precision & estimation

🧠 C. Interpretation of Data
C.1: Major Results

●​ What were the key numbers and findings?

C.2: Bias & Confounding

●​ Were results affected by:

○​ Selection bias
○​ Information bias
○​ Confounding?
●​ Consider magnitude and direction of potential bias

C.3: Misclassification

●​ Did non-differential misclassification affect results?

○​ How large was the potential impact?

C.4: Study Limitations

●​ Did the authors acknowledge major limitations?

●​ Did they discuss how those limitations might influence the findings?

C.5: Conclusions

●​ Were they justified by the data?

●​ Were the implications consistent with the study results?
C.6: Generalizability

●​ Can results be applied to a broader population?

○​ To whom do these findings apply—everyone or only a specific group?

📌 Summary: What Makes a Good Epidemiologic Study?

✔ Clearly defined and relevant research question​
✔ Appropriate study design​
✔ Valid and reliable measurement of exposure and outcome​
✔ Proper control of bias and confounding​
✔ Correct and transparent analysis​
✔ Reasonable interpretation and conclusion based on evidence