# Confounding/ controlling for confounding. Flashcards

1
Q

At what stage(s) of the study can confounding be corrected for?

A

The design or analysis stage.

2
Q

What is a crude association?

A

An association between an exposure and an outcome before adjusting for any confounders.

3
Q

Describe what is meant by confounding.

A

An extraneous variable that is a common cause of an exposure and outcome that fully or partially explains the observed effect.

4
Q

What three conditions need to be satisfied at the data analysis stage for something to be a confounder?

A
1. Association with the study exposure in the control group.
2. Associated with the disease in the unexposed group.
3. Not on the causal pathway for exposure-disease.
5
Q

If you stratify by a potential confounder (in this case smoking) what will the H0 be?

A

ORnonsmoke = ORsmoke.

6
Q

What is a Mantel-Haenazel OR and how is it used?

A

A OR adjusted for a confounding variable. This will take into account the weighted averages of the separate OR’s of the confounding groups. This adjusted OR can then be compared to the crude OR. if there is a roughly 15% different the factor can then be counted as a confounder.

7
Q

You only need to correct for confounders if the Mantel-Haenazel ratio is 15% different from the crude OR. True or false?

A

False. You should still correct for a confounding factor as it is still an error, even if it is not technically classes as a confounder.

8
Q

If the analysis of data suggests that there is no confounding effect you should accept that there is no cofounding effect, even if literature says that something may be a biologically confounding factor. This is because the statistical analysis is more robust. True or false?

A

False. The causal definition is the most important thing to consider when deciding wether if something is a confounder.

9
Q

If the analysis of data suggests that there is a confounding effect you should accept that there is a cofounding effect, even if past literature says that something is not a biologically confounding factor. This is because the statistical analysis is more robust. True or false?

A

False. The causal definition is the most important thing to consider when deciding wether if something is a confounder.

10
Q

What is the limiting factor when using the causal definition in deciding whether something is a confounding factor?

A

Current biological knowledge.

11
Q

It is not ideal to adjust for every single factor that could be confounding. Why? (2 reasons).

A
1. Efficiency is reduced.

2. Confidence intervals become wider.

12
Q

If you adjust for a non confounder what can it produce?

A

13
Q

What sort of confounding is this?

A

Positive confounding.

14
Q

What sort of confounding is this?

A

Negative confounding.

15
Q

What is qualitative confounding?

A

When one OR is less than 1 and one is more than 1.

16
Q

If a confounder is very highly related to an exposure (‘Excessive correlation/ over adjustment) what can adjusting it also do?

A

17
Q

Computationally you can adjust for confounding, but it is complex. True or false?

A

False. There is no statistical test for confounding.

18
Q

What three ways can you control for confounding at the design stage of a study?

A
1. Randomisation (in RCTS.)
2. Restriction.
3. Matching.
19
Q

What two ways can you control for confounding at the analysis stage of a study?

A
1. Stratification.

20
Q

What is the H0 used after stratifying for a potential confounder?

A

There is no significant difference between the OR in the stratified groups.

21
Q

What is the best method to use when adjusting for confounding? Why is this the best method?

A

Randomisation (design stage). This is best as you can adjust for both known and unknown covariates.

22
Q

How does the restriction method help adjust for confounders?

A

Eligibility criteria is limited to a certain subgroup.

23
Q

What are the downsides of using restriction to adjust for confounders (2 things)?

A
1. The pool of available subjects is reduced.

2. External validity reduced.

24
Q

What is eliminated through the use of individual matching when controlling for confounders?

A

Bias comparison between cases and controls.

25
Q

What can matching for a strong confounder do?

A

Increase statistical power and with it efficacy.

26
Q

What do you risk happening if you plan to adjust at the analysis stage instead of the design stage?

A

There may be distinct distributions due to confounding due to chance. Without an overlap of distributions you can not perform adjustment (0’s will be present in the 2 by 2 table).

27
Q

What sort of confounders are normally adjusted for at the design stage?

A

Very important ones, such as age and ethnicity.

28
Q

Matching for a strong confounder can improve statistical efficacy. What can reduce it?

A

Matching for a non confounder.

29
Q

What is ‘Over matching’?

A

A phenomenon that occurs when you match for a variable that is too strongly correlated to an exposure resulting in a reduction of the effect that the exposure has on the outcome.

30
Q

‘If you have a strong indication that something is a confounder you should match for it.’ Why is this statement false?

A

As you can only practically match for a few confounders. You should hence match for the most important ones.

31
Q

When you have used matching to reduce confounding you need to use specific statistical tests to account for this. Name two examples of these tests.

A
1. McNemar test.

2. Conditional logistic regression.

32
Q

What ratio of cases:controls maximises statical power.

A

1:4.

33
Q

What is the odds ratio equation used to calculate the OR after matching?

A

b/c (exposed case and nonexposed control)/(exposed control and non exposed case).

34
Q

What are four advantages of using regression models to control for confounding?

A
1. Can study numerous cofounders.
2. Can look at large datasets.
3. Confounders can have any distribution.
4. Can include non linear terms.
35
Q

What type of bias is confounding?

A

Systematic bias.

36
Q

By accounting for confounding factors in a case control or cohort study what are you essentially trying to imitate?

A

Randomisation in a RCT.