Confounding and Stratification Flashcards Preview

Epidemiology > Confounding and Stratification > Flashcards

Flashcards in Confounding and Stratification Deck (13)
Loading flashcards...

What is prevalence ratio?


Total number of people with disease/total population in exposed v. unexposed.


What is risk ratio?


New disease/total population in exposed v. unexposed.

Compare cumulative incidence.


What is incidence rate ratio?


New disease/person years in exposed versus unexposed.

Compare incidence rate.


What is odds ratio?


Odds of disease in exposed v unexposed.

Compare odds.


What is the counterfactual?


Groups of unexposed and exposed identical to each other except for exposure of interest (same time, same place, same people).

This is impossible so it is the counterfactual.

What would the outcome have been for an exposed person if , counter to fact, an exposed person had been unexposed but everything else was the same?


What is a confounder?

  • Differences between exposed and unexposed that make it difficult to make valid comparisons between them.
  • All the things that interfere with ideal of comparing counterfactual to reality.
  • Associated with both exposure and outcome but not on causal pathway.

What is the consequence of a confounder?

  • The desired measure of association between exposure and outcome will reflect associations between other variables and the exposure and outcome.
  • Measure of association will be biased.
  • Magnitude or direction affected or both - do not know the extent to which it is.

How do we assess confounding?


1) See if potential confounder associated with exposure.
2) See if potential confounder associated with outcome.
3) Does magnitude or direction of association for exposure and outcome differ across strata of confounder versus overall? (Ratios that are the same or similar to each other but different from crude-then confounding present)


How do we see if a variable is associated with exposure or outcome?


Determine if the measure of association in strata of a third variable differs by more than 10%:


Calculate prevalence ratio for a cohort study and see if different from 1 by more than 10%.
Smoking and lung cancer. Sex as confounder.

F v M smokers prevalence.
20% of women smoke and 80% of men
.2/.8 = .25
(men 4 times more likely to smoke)
Different from 1 by at least 10%? Yes.

F v M lung cancer pxs.
33% of women lung cancer and 16% of men

.33/.16 = 2.1
(half as likely to have lung cancer in men)
Different from 1 by 10%

If the prevalence ratio is different from 1 by more then 10% then there if confounding.


How do we see if magnitude or direction for exposure and outcomes differ across strata of confounder?


Determine if the measure of association are the same or similar to each other but different from the crude:


Calculate risk ratios by separating the lung cancer and smoking into females and males.

If overall RR is .9:

F RR is 1.53 and M is 1.53.
(Recall that this is through comparing cumulative incidence of those with and without disease in groups)

Since the risk ratios are the same as each other but different from the crude, they are different in direction and magnitude

You should see this with a confounder in strata.


What if risk ratios are different in the strata (but similar)?


You want to present only one ADJUSTED RR.

Mantel-Haenszel method to pool measures of association to see if there is effect modification.


Example of Mantel-Haenszel method.


Combines measures of association across strata to adjust for confounder. (Can use for all measures of association). OR 5 steps:

1) For each 2*2 table of the strata, multiply number of exposed CASES times number of UNexposed CONTROL and divide by total number of people.
2) Sum quantities of the two tables (the different strata).
3) Multiply number of UNexposed CASES times exposed CONTROLS and then divide by total number of people.
4) Sum that quantity (across the strata)
5) Divide the first sum by the second sum:

Ex: Case-control of cancer and coffee.

Crude OR = (8882)/(6268) = 1.71

Think about associations: smokers is associated with both.

When you divide this association by smokers and non-smokers, you see OR for smokers as 1 and non-smokers as 1.06.

The MH OR is 1.01 so that pooled OR is closer to smokers than non-smokers (bc more people in smoking strata).

Provides a less biased answer

For RR:

1) ][Exposed cases (total controls)] / [total in that table]
2) Do that for each table and SUM
3) [(Unexposed controls)(total controls)/total in that table]
Do that for each table and SUM
1) Quantity 1/2


How do you calculate percentage difference?


To calculate the percentage difference between two numbers, a and b, perform the following calculations:

Find the absolute difference between two numbers: |a - b|
Find the average of those two numbers: (a + b) / 2
Divide the difference by the average: |a - b| / ((a + b) / 2)
Express the result as percentages by multiplying it by 100