Brainscape's Knowledge GenomeTM

Browse over 1 million classes created by top students, professors, publishers, and experts.


See full index

EPIB 601 Fundamentals > Confounding > Flashcards

Confounding Flashcards

1
Q

Under what conditions does randomization produce exchangeability?

A
  • Perfect compliance
  • No bias
  • No loss to follow-up
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

Why is confounding an issue in every observational study?

A

Unlike in a RCT, we have no reason to expect that the exposed and unexposed groups will perfectly substitute for the target population under counterfactual treatment and control conditions.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

When is confounding present?

A

When our substitute imperfectly represents what our target would have been like under the counterfactual condition

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

What is a confounder?

A

A variable that at least partly explains why confounding is present

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

What are 5 sources of statistical association?

A
  1. Random error
  2. Reverse causation
  3. Selection bias (by conditioning on collider)
  4. Confounders
  5. Causal effect
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

3 common strategies for identifying confounders?

A
  1. Stepwise selection procedures: test covariate-disease association to see which variables to include (e.g. include in multivariate covariates that were sig in bivariate analyses)
  2. Compare adjusted and unadjusted estimates for the main exposure-disease association (if relative change after adjustment is greater than 10%, variable is selected)
  3. Traditional rules of confounding
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

What is the problem with stepwise procedures to identify confounders?

A

Arbitrary definitions of thresholds, leading to bias, overfitting, exaggerated p-values

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

What is the problem with change in estimate procedures to identify confounders?

A

It wrongly assumes that any variable that changes the exposure-outcome association should be adjusted for (could adjust for mediator)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

What is the problem with the traditional epidemiologic definition of a confounder?

A

Can lead us to incorrectly control for consequences of the exposure and induce bias

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

What do DAGs encode, and what do they not encode?

A

Encode: causal relations
Not: magnitude/type of statistical associations

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

4 ways to block the backdoor path?

A
  1. Restrict
  2. Stratify
  3. Match
  4. Adjust
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

When is controlling for a surrogate confounder sufficient?

A

When it is strongly associated with the unmeasured confounder (when they are perfectly correlated, you can remove all of the confounding, else only part of it)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

Is it ok to condition on the consequence (descendent) of a collider?

A

No - also induces a spurious association, to the extent that the collider and the descendent of a collider are strongly associated

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

What is positive confounding?

A

When the magnitude of the unadjusted compared to the adjusted is exaggerated
(confounding resulted in bigger effect than what really was)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

What is negative confounding?

A

When the magnitude of the unadjusted compared to the adjusted association is underestimated
(confounded resulted in smaller effect than what really was)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

What is qualitative confounding?

A

An extreme case when confounding results in an inversion of the direction of the association

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
17
Q

In a DAG, when is there a positive confounding?

A

When the association between the confounder and the exposure/outcome are either both positive or both negative

18
Q

In a DAG, when is there a negative confounding?

A

When the association between the confounder and the exposure/outcome are of different signs

19
Q

The magnitude of confounding depends on:

A
  1. The strength of the confounder-exposure association

2. The strength of the confounder-outcome association

20
Q

If the crude RR is 2, and the confounder-exposure correlation is perfect, what is the confounder-outcome association?

A

2 (doubles)

21
Q

What is residual confounding?

A

When adjustment does not completely remove the confounding effect of a given variable

22
Q

What are 3 sources of residual confounding?

A
  1. Misclassification (e.g., imperfect proxy)
  2. Improper modeling (e.g., never/ever vs pack-years)
  3. Unmeasured confounding
23
Q

3 ways to handle confounding in the design of the study

A
  1. Randomization
  2. Restriction
  3. Matching
24
Q

Why does randomization help in handling confounding?

A

Successful randomization produces groups that are similar with respect to both measured and unmeasured factors
(removes arrow from potential confounders to exposure/treatment)

25
Q

What is the feasibility of randomization?

A
  • Limited due to cost/ethics

- Not a guarantee of absence of confounding due to random errors (but unlikely in large samples)

26
Q

Why does restriction help in handling confounding?

A

If everyone in the study has the same value for a particular characteristic, then there can be no imbalances in the distribution of that characteristic and it cannot be a confounder

27
Q

What is the feasibility/limitations of restriction?

A
  • Almost always possible
  • Can be done both at the study design and analysis phase
  • But if small n, less precision
  • Less generalizability
  • Cannot evaluate association at different levels of confounder (e.g., if only non-smokers, can’t evaluate effect/magnitude of smoking)
  • If categories not sufficiently narrow, there may be residual confounding
  • Not possible if confounder and exposure are perfectly correlated
28
Q

What are concordant pairs in matching?

A

Both/neither the case and control are exposed

D+E+, D-E+
D+E-, D-E-

29
Q

What are discordant pairs in matching?

A

The case and control have different exposures

D+E+, D-E-
D+E-, D-E+

30
Q

How can we prevent selection bias in matching?

A

By computing a matched-pair OR and comparing it to the result that ignores the matched structure

31
Q

What is the formula for matched OR (mOR)?

A

(D+E+,D-E-)/(D+E-,D-E+)
or
concordant pairs/discordant pairs

32
Q

3 analytic methods to handle confounding? Limitation?

A
  1. Standardization
  2. Stratification
  3. Regression adjustment

To the extent that they were measured properly and modeled correctly

33
Q

What are the steps to stratification?

A
  1. Calculate crude measure of association
  2. Stratify by confounder
  3. Calculate stratum-specific associations
  4. Compare stratum-specific measures of associations
34
Q

What if the stratum-specific estimates are heterogenous?

A
  • We do not pool them

- We look for effect modification

35
Q

What if the stratum-specific estimates are homogeneous?

A
  • We can pool the stratum-specific estimates with the Mantel-Haenszel method
  • If crude and adjusted associations are different, confounding is likely
36
Q

What is the main principle behind Mantel-Haenszel estimates?

A

The largest weight is given to the stratum with the lowest variance

37
Q

3 limitations of stratification methods of adjustment

A
  1. Only good for binary/categorical covariates
  2. Cannot adjust for multiple covariates (data becomes sparse)
  3. Pooling estimates make the assumption that the measure of association is constant across the strata of the confounder, thus possibly resulting in residual confounding
38
Q

When is a measure of association collapsible?

A

when the crude measure of association doesn’t change if we adjust for a variable that is not a confounder

39
Q

What measures of association are collapsible and which are not?

A

Yes: RR
No: OR, IDR (for OR, crude measure may be closer to null than pooled/adjusted OR, thus resulting in a suggestion of a confounding when there isn’t one)

40
Q

Difference between bias and random error?

A

Bias results from systematic error, we are worried about how it affects the accuracy of our measures

Random error is variability that is not readily explained, and we are worried about how susceptible our findings are to this error

41
Q

What does it mean to have a 95% CI?

A

If the data collection/analysis were replicated infinite times, with no bias, the CI would contain the true value of the measure 95% of the time
- Presumption: no systematic difference, only variability due to chance (never a plausible condition in epi, so CI is more of a general guide to the amount of random error in the data)

42
Q

Definition of p value

A

The probability of the observed results, plus more extreme results, it the null hypothesis were true

EPIB 601 Fundamentals (6 decks)

Brainscape helps you reach your goals faster, through stronger study habits.
© 2024 Bold Learning Solutions. Terms and Conditions