Lecture #21 - Confounding Pt1

Question 1

Definition of confounding

Answer 1

A mixing or muddling of effects when the relation we’re interested in is confused by the effect of so,thing else - the confounder

A 3rd factor messing with your associations

(It’s causing your association to be less accurate or close to the truth)

Question 2

Remember - analytic epi’s principle is that the results are only valid if…..

Answer 2

……groups are equivalent for anything else associated with outcome.

Question 3

Why do we randomise?

How do we know randomisation worked?

Go read the bottom page of P23 and don’t rate this card a 5 until you have and understood it

Answer 3

To minimise the effect of confounders (also why we age standardise).

Look at table 1

Question 4

Three properties of potential confounder:

1. Independently associated with _______
   - A ____/______ factor for the outcome by itself
   - explain this
2. Independently associated with _____
   - Different ______ of people with potential confounder across _____ ______
   - explain this
   - textbook says (controls)
3. Not on the causal pathway
   - Not the mechanism by which the _____ _____ the risk of _____
   - explain this

Answer 4

1. Independently associated with outcome
   - A risk/protective factor for the outcome by itself
   - doesn’t matter if you have exposure, it will increase or decrease risk of outcome so it’s independent of exposure (‘independently associated’)
2. Independently associated with exposure
   - Different proportions of people with potential confounder across study groups
   - Aren’t the same proportion of those factors in both groups so like smokers more likely to be heavy alcohol consumers (so smoking - the confounder - is more prevalent in the heavy consumer study group
   - Textbook says it needs to be associated with the exposure in the controls so if you take away the outcome, it should still be related to the exposed controls.
3. Not on the causal pathway
   - Not the mechanism by which the exposure affects the risk of outcome
   - So like, think as, “Will this exposure change my status of this potential confounder which will then lead to my outcome?” So like, being male is confounder and drinking alcohol won’t change you to being male which will then lead to liver cancer. But if you think having a high error rate in cell multiplication is a confounder for alcohol and liver cancer - it isn’t because drinking alcohol increases high error rate which then leads to liver cancer (I’m making this up but I hope you get the idea).
   - Or like for measuring the protective relationship of physical activity and heart attack - low blood pressure is related to both exposure and outcome BUT it lies on the causal pathway (i.e. physical activity means you lower your blood pressure and THEN reduce risk of heart attack)

Question 5

What are the four types of affects confounding can have on the estimate of the association?

Answer 5

1. Overestimates a true association (takes it further from the null)
2. Underestimate true association (takes it closer to null)
3. Changes direction of a true association (Simpson’s paradox) - so risk factor becomes protective
4. Get appearance of an association when there isn’t one (not overestimating) - going from null to something else.

Question 6

How do you deal with potential confounders in a study:

1. Plan _____
2. Collect info on ____ ______ ______
3. Use literature to what? Which criteria
4. Collect info on factors strongly what? Which criteria?
5. If you don’t measure it - what could happen?

Answer 6

1. Plan ahead; need to do something about this in design time, can’t do much about it after study is done
2. Collect info on all potential confounders
3. Use literature to identify known and suspected risk factors for outcome (criteria 1). So go through stuff you think would be risk/protective factors for outcome
4. Collect information on factors strongly associated with exposure, regardless if known it’s a risk factor for outcome (criteria 2) i.e. collect info on things known to be associated with exposure
5. If you don’t measure it - difficult to do anything about it later

Question 7

Controlling confounding in study design

1. What are the three methods of minimising confounding in study?
2. What do all three attempt to do?
3. Try to control for confounding in a study design through these three by manipulating what?

Answer 7

1. Randomisation, restriction and matching
2. All attempt to make groups being compared alike with regard to potential confounder(s) - all three try to make groups have same proportion of confounders (so same proportion of people with potential confounder in each group)
3. Try to control confounding in a a study design through these three things by manipulating the people who come into the study (who comes in and how we manipulate them)

Question 8

Randomisation:

1. What uses randomisation (only one thing)?
2. What is the strength of randomisation?
3. What does it work best with?
4. Needs what?
5. Needs also what?
6. In your own words, explain how randomisation controls for confounding

Answer 8

1. RCT
2. Applies to known and unknown confounders?
3. Needs large sample size for it to work best
4. Needs equipoise (genuine uncertainty else unethical)
5. Needs intention-to-treat (can be a problem if loss to follow up/missing data) - breaks randomisation if not
6. Basically, you grab a bunch of people and you randomly assign them in groups e.g. say that these people will drink heaps of alcohol and these won’t. Since you’ve randomly assigned them, the proportion of smokers is (or should be) equal in both groups so you can’t have the potential confounder having an effect on the association since both groups have the same proportion of people with the confounder. So like, you know how smoking increases the likelihood of lung cancer - yeah, if you had heaps of the heavy drinkers who smoked and barely any of the non drinkers who didn’t, it would seem like alcohol increases risk heaps (but it’s actually the smoking confounder that’s at play here). So if you equally split the confounders in the the two groups, it can’t ave more effect on one group than the other.

Question 9

Restriction:

1. What do you do?
2. What is a stratum?
3. What’s its advantage?
4. But what can it reduce?
5. What can it also reduce?
6. Potential for _____ _____ with imprecisely measured(or broadly defined) confounders
7. Usually only what?

Answer 9

1. Restrict sample to one strata of potential confounder
   e. g. just include the males in study so no difference in two groups of M/F or with alcohol and lung cancer - only assess the smokers (restrict to one stratum)
2. In potential confounders, you have different strata e.g. sex is confounder but male and female are the strata.
3. Easy and can be applied to all study designs (unlike randomisation which is only RCT).
4. Reduce generalisability (if only study males, can’t apply findings to females too)
5. Reduce number of potential participants (number of people you recruit decreases)
6. Residual confounding - Left over confounding because haven’t properly controlled for it. E.g. you’re including people in your non-smokers group who just gave up three days ago- so it can still affect association
7. Usually only one potential confounder because otherwise you get a very specific group and so you don’t get relevant findings that are generalisable

Question 10

Matching:

1. What do you do?
2. What kind of study is it mainly used in?
3. What are the two types - explain them
4. What kind of confounders is it useful for measuring?
5. What can it improve?
6. But is it hard and what can it limit?
7. Need special ____ ____ for individual matching

Answer 10

1. Choose people for the comparison (control) group who have the same values of the potential confounders as the people in the exposed group (cohort studies) or case group (case-control)
2. Case-control
3. Individual = 1:1 match one case with one control
   Frequency matching = e.g. 1:4 match one case with four controls but the proportion of the confounder is the same in each case e.g. if out of 20 people, you have 5 males then if you recruit 80 people, you have 20 males. So the proportion of the potential confounder stays the same in both the case group and the control group
4. Useful for difficult to measure/complex confounders (e.g. might select controls from the same neighbourhood to match the environment)
5. Can improve efficiency of case-control studies with small numbers
   - you can get away with smaller number of people in your study and still find the association if it exists
6. Can be difficult (can’t always find exactly the same type of person) and it can limit the number of potential participants
7. Need special matched analysis for individual matching