Week 8: Confounding and Effect Modification

Question 1

What is the main purpose of RCTs?

Answer 1

To test the efficacy of an intervention by ensuring similar distributions of known and unknown characteristics between treatment groups

Question 2

Why are observational studies sometimes used instead of RCTs?

Answer 2

RCTs can be expensive, time-consuming, and unethical for harmful interventions. Observational studies provide an alternative method to analyse exposures and outcomes

Question 3

Define confounding in the context of statistical analysis

Answer 3

Confounding occurs when a third variable influences both the exposure and the outcome, creating a spurious association

Question 4

What are the criteria for a variable to be a confounder?

Answer 4

• It is a risk factor for the outcome
• It is associated with the exposure
• It is not a result of the exposure

Question 5

What is stratified analysis, and why is it used?

Answer 5

Stratified analysis divides data into strata based on potential confounders to control for their effects and analyse the exposure-outcome relationship within each stratum

Question 6

What is the Mantel-Haenszel Odds Ratio?

Answer 6

A weighted average of ORs across strata, giving more weight to larger strata, used to control for confounding in categorical data

Question 7

How does logistic regression help in addressing confounding?

Answer 7

Logistic regression models the relationship between exposure and outcome while adjusting for confounding variables

Question 8

Describe the difference between effect modification and confounding

Answer 8

Effect modification occurs when a third variable alters the strength or direction of the association between the exposure and outcome, whereas confounding creates a spurious association

Question 9

What methods can address confounding in study design?

Answer 9

Restriction (e.g., limiting participants to a specific group) and matching (e.g., pairing cases and controls on confounding variables)

Question 10

When is the assumption of homogeneity for Mantel-Haenszel analysis valid?

Answer 10

When the association between exposure and outcome is valid across strata

Question 11

How is effect modification identified?

Answer 11

By stratifying data on a third variable and observing different exposure-outcome associations across strata

Question 12

Why should factors on the causal pathway not be adjusted for in analysis?

Answer 12

Adjusting for these factors can bias the estimate of the exposure-outcome relationship

Question 13

What is the key advantage of randomisation in RCTs?

Answer 13

It minimises confounding by equally distributing known and unknown characteristics between groups

Question 14

How does matching help in study design?

Answer 14

It pairs subjects with similar values of confounding variables, ensuring balanced groups for analysis

Question 15

What does a significant test of homogeneity in Mantel-Haenszel analysis suggest?

Answer 15

That there is effect modification, and separate ORs should be reported for each stratum - we wouldn’t use MH

Question 16

Why is continuous outcome analysis different, and which method is used?

Answer 16

Continuous outcomes require linear regression to assess associations and adjust for confounding

Question 17

What is the impact of crude analysis on confounding?

Answer 17

Crude analysis may overestimate or underestimate the true association by ignoring confounding variables

Question 18

Explain the term “interaction” in statistical analysis

Answer 18

Interaction (effect modification) occurs when the relationship between two variables changes across levels of a third variable

Question 19

How do you calculate Q (weighted numerator) as part of the Mantel-Haenszel OR?

Answer 19

Q = a_s x d_s / N_s + a_n x d_n / Nn
- a_s, d_s = cases and controls exposed in one stratum (e.g., smoker)
- Ns = total participants in the smoker stratum
- a_n, d_n = cases and controls in the non-smoker stratum
- Nn = total participants in the non-smoker stratum

Question 20

How do you calculate R (weighted denominator) as part of the Mantel-Haenszel OR?

Answer 20

R = b_s x c_s / N_s + b_n x c_n / N_n
- b_s, c_s = cases and controls unexposed in one stratum (e.g., smokers)
- b_n, c_n = cases and controls unexposed in the non-smoker stratum (e.g., non-smokers)

Question 21

How do you calculate Mantel-Haenszel OR?

Answer 21

MH_OR = Q/R

Question 22

What is the difference between pair and frequency matching?

Answer 22

Pair matching - pair one individual to another individual
Frequency matching - match groups of individuals

Question 23

What is the H0 for a test of homogeneity (M-H) between smokers and non-smokers?

Answer 23

H0: No difference between OR for smokers and non-smokers
If p > 0.05, we should use the M-H combined estimate - the M-H is a valid overall estimate of the association between exposure and outcome in such cases

Question 24

How else is effect modification called?

Answer 24

Interaction, heterogeneity between strata