Bias and Confounding

Question 1

How can we get the wrong answer?

Answer 1

Systematic error (bias)
Random error (chance)

Question 2

Systematic Erros (Bias)

Answer 2

A systematic error in selection of the study sample or measurement that can lead to incorrect results.

Question 3

Selection Bias

Answer 3

Can occur when Study Sample is systematically different from Accessible Population or Target Population

Choosing who to recruit and who is analyzed for the study

Question 4

Measurement Bias

Answer 4

A systematic error in the measurement of relevant study measures (e.g., exposure, outcome).

Question 5

Confounding

Answer 5

The true relationship between an exposure and an outcome is obscured by another variable that is associated with both the exposure and the outcome.

Question 6

Oversampling outcome

Answer 6

There are more subjects with the outcome than would have been expected had the sample been drawn randomly

Question 7

Oversampling Exposure

Answer 7

There are more subjects with the exposure
than would have been expected had the sample been drawn randomly

Question 8

Selection Bias can bias

Answer 8

• measures of occurrence (prevalence, incidence)
• measures of association (OR, RR)
• Is most problematic when there is selection by both exposure and outcome

Question 9

Differential Loss to Follow-up Can Cause Selection Bias

Example

Answer 9

When patients are not showing up for the follow up of their study results so you are unable to see the true real answer.

Question 10

How do you limit selection bias?

Answer 10

Study Design
- Select a representative sample
- Do not select by both exposure and outcome
- Randomize
Study Procedures
- Recruit subjects to achieve intended sample
- Complete follow-up in those subjects

Question 11

Measurement bias can occur when

Answer 11

Actual Measurements are systematically different from Intended Variables or Target Phenomenon

Question 12

Misclassification of Outcome

Answer 12

Subjects who should have been classified
as having the outcome are misclassified as
not having the outcome

Question 13

Misclassification of Exposure

Answer 13

Subjects who should have been classified
as having the exposure are misclassified
as not having the exposure.

Question 14

Differential Misclassification

Answer 14

Misclassification occurs with different probabilities by outcome or exposure

Question 15

Measurement Bias can bias

Answer 15

• measures of occurrence
• measures of association
• Is most problematic when it is differential by both exposure and outcome

Question 16

how to limit measurement bias

Answer 16

Study design:
Use most sensitive and specific measurements available
- “Gold-standard” measurements (closest representation to the truth phenomena)
- Validated self-report measures
- Consider duplicate measurements to reduce variability
Same measurements of exposure and outcome in all subjects
- Randomization and Blinding
Study procedures:
Operations manuals/training for standard measurement

Question 17

Is confounding a bias?

Answer 17

No, it’s a phenomenon that occurs due to
associations between variables and requires careful consideration.

Question 18

Three criteria for confounding

Answer 18

1. Variable must be associated with the exposure (cause it)
2. Variable must be associated with the outcome (cause it)
3. Variable cannot be on the causal path between the exposure and the outcome

Question 19

How to Address Confounding

Answer 19

Study Design
- Restriction (breast cancer: study women more than men; age: things happen differently in kids vs adults)
- Matching
- Randomization
Study Analysis
- Stratification
- Regression

Question 20

Restriction

Answer 20

• Include only participants with some value of the confounder
• Example: Age is a confounder in association between Smoking and Death, so, exclude people under 65 in your study
• Issues
  1. Irrevocable (you can’t go back and change it)
  2. Limits external validity/generalizability

Question 21

Matching

Answer 21

• Equal numbers of participants with an important confounder
• Often employed in case-control studies
• Must be accounted for in design, study procedures, and analysis
• Issues: Irrevocable, Sometimes not feasible

Question 21

Stratification

Answer 21

• Divide the study population into groups based on the confounder
• Estimate the measures of occurrence/association separately in the stratified groups (Can use methods to create combined measures)

Question 21

Randomization

Answer 21

• If participants are randomly assigned the exposure, then the
  exposure is not associated with anything!
  1. Limits selection bias
  2. Measurement error will not affect ratio measures
• Issues:
  1. Differential loss to follow-up
  2. Generalizability/external validity issues depending on inclusion/exclusion criteria and Study Sample

Question 22

Regression

Answer 22

it allows you to look at variables in all different ways

Question 23

What should you adjust for?

Answer 23

Factors known or suspected to be associated with exposure and outcome
* Find in
literature
* Observed in
data
Factors not on causal pathway

Question 24

Confounding – Critical Review of Literature

Answer 24

1. Did they consider things that might be associated with the exposure and outcome?
2. Did they include factors on the causal pathway?
3. Were methods employed to limit confounding (e.g., restriction, matching, randomization, stratification, regression)?
4. Are there important unmeasured (or unmeasurable) confounders that could affect the interpretation of results?

Question 25

How to Address Confounding

Answer 25

Directed Acyclic Graphs (DAGs)

Question 26

DAG Rules

Answer 26

-Directed - Arrows can only go one way
-Acyclic - There cannot be a loop where following the arrows
ends back at the variable

Question 27

Systematic errors (bias) is especially problematic when

Answer 27

related to exposure AND outcome

Question 28

Directed Acyclic Graphs (DAGs) can visualize relationships between

Answer 28

exposure and outcome, confounders, and mediators