Epidemiology quiz

Confounding and Effect Modification

David D. Celentano, ScD, MHS Johns Hopkins University

2

Learning Objectives

1. To identify and learn how to eliminate or reduce confounding in the design and analysis of observational epidemiologic studies ► A type of bias

2. To identify and learn how to highlight the presence of effect modification ► Not bias, possibly biology

3

Disclosure

► Design and analysis methods to identify confounding and effect modification are similar

► However, confounding is a problem to be dealt with, while effect modification may reflect biology and should be highlighted

► There’s always controversy in the teaching of epidemiologic methods about whether the two concepts should be taught side by side—to make it clear that they are distinct—or in completely separate lectures—to avoid making it seem like both are a problem

The material in this video is subject to the copyright of the owners of the material and is being provided for educational purposes under rules of fair use for registered students in this course only. No additional copies of the copyrighted work may be made or distributed.

Confounding

5

Recall: Goal of Epidemiologic Research

► Identify exposure-disease associations

► Observational epidemiologic methods cannot determine whether an association is causal

► There are always several possible explanations for observed epidemiologic associations, only one of which is cause ► First, we will focus on confounding

6

Possible Explanations for Associations

► Bias ► Selection bias ► Information bias

► Confounding

► Chance

► Cause

7

Confounding—Classic Definition

► A confounding variable (or confounder) is a third factor that is… 1. A risk factor for the disease, 2. Associated with the exposure, and 3. Not a factor in the causal pathway from exposure to disease (not a mediator)

► If any of these criteria is not satisfied, then the third factor is not a confounder

► Confounding is an error in a study ► The observed association between the exposure and disease appears stronger or

weaker than it actually is

8

Defining Confounding—1

► “…the apparent effect of the exposure is distorted because the effect of an extraneous factor is mistaken for—or mixed with—the actual exposure effect (which may be null).” — Rothman, Greenland, and Lash, Modern Epidemiology, 2008

► B causes A and C

9

Defining Confounding—2

► “…the apparent effect of the exposure is distorted because the effect of an extraneous factor is mistaken for—or mixed with—the actual exposure effect (which may be null).” — Rothman, Greenland, and Lash, Modern Epidemiology, 2008

► B causes A and C ► If B is not taken into account, then A

appears to cause C

10

Defining Confounding—3

► “…the apparent effect of the exposure is distorted because the effect of an extraneous factor is mistaken for—or mixed with—the actual exposure effect (which may be null).” — Rothman, Greenland, and Lash, Modern Epidemiology, 2008

► B causes A and C ► If B is not taken into account, then A

appears to cause C

Alcohol does not cause kidney cancer, but not taking into account smoking makes alcohol appear to be a cause of kidney cancer

11

Evaluating Criteria for Confounding—1

► Three criteria for a factor to be a confounder: (1) Factor must be a risk factor for the disease, (2) Factor must be associated with the exposure, and (3) Factor must not be in the causal pathway from exposure to disease (not a mediator)

► B is a confounder

12

Evaluating Criteria for Confounding—2

► Three criteria for a factor to be a confounder: (1) Factor must be a risk factor for the disease, (2) Factor must be associated with the exposure, and (3) Factor must not be in the causal pathway from exposure to disease (not a mediator)

► B is a confounder ► Contrast with: B is a mediator

13

Evaluating Criteria for Confounding—3

► Three criteria for a factor to be a confounder: (1) Factor must be a risk factor for the disease, (2) Factor must be associated with the exposure, and (3) Factor must not be in the causal pathway from exposure to disease (not a mediator)

► B is a confounder ► Contrast with: B is a mediator

14

Evaluating Criteria for Confounding—4

► Three criteria for a factor to be a confounder: (1) Factor must be a risk factor for the disease, (2) Factor must be associated with the exposure, and (3) Factor must not be in the causal pathway from exposure to disease (not a mediator)

► B is a confounder ► Contrast with: B is a mediator

15

Evaluating Criteria for Confounding—5

► Three criteria for a factor to be a confounder: (1) Factor must be a risk factor for the disease, (2) Factor must be associated with the exposure, and (3) Factor must not be in the causal pathway from exposure to disease (not a mediator)

► B is a confounder ► Contrast with: B is a mediator

16

Evaluating Criteria for Confounding—6

► Three criteria for a factor to be a confounder: (1) Factor must be a risk factor for the disease, (2) Factor must be associated with the exposure, and (3) Factor must not be in the causal pathway from exposure to disease (not a mediator)

► B is a confounder ► Contrast with: B is a mediator

17

Testing the Criteria for Confounding

Note: We may need additional information beyond what is in this figure to determine if a criterion is met

► Walk through the criteria ✔ (1) Smoking is a risk factor for kidney cancer ✔ (2) Smokers tend to drink alcohol ✔ (3) Smoking is not a consequence of alcohol drinking on the

path to kidney cancer (smoking is not a mediator)

► Yes, smoking meets all three criteria for being a confounder of the alcohol-kidney cancer association

18

The Classic Example of Confounding— 1

► Does coffee drinking cause pancreatic cancer?

19

The Classic Example of Confounding— 2

► Does coffee drinking cause pancreatic cancer? ► Truth is that coffee drinking does NOT cause pancreatic cancer

20

The Classic Example of Confounding— 3

► Does coffee drinking cause pancreatic cancer? ► Truth is that coffee drinking does NOT cause pancreatic cancer

21

The Classic Example of Confounding— 4

► Does coffee drinking cause pancreatic cancer? ► Truth is that coffee drinking does NOT cause pancreatic cancer

22

The Classic Example of Confounding— 5

► Does coffee drinking cause pancreatic cancer? ► Truth is that coffee drinking does NOT cause pancreatic cancer ► If smoking is not taken into account, then the observed

association between coffee and pancreatic cancer is confounded by cigarette smoking

23

Confounding Is Specific to Research Question

► Depending on the research question, the same factor can be… ► An exposure ► A third factor that is not a confounder ► A third factor that is a confounder ► A mediator ► An effect modifier (later in this lecture) ► An outcome

► Must consider each research question and all of its factors carefully

24

Research Question 1

► Is alcohol drinking a risk factor for lung cancer?

► For Research Question 1, is cigarette smoking a confounder?

25

Is Confounding Present in Research Q 1?—1

26

Is Confounding Present in Research Q 1?—2

27

Is Confounding Present in Research Q 1?—3

28

Is Confounding Present in Research Q 1?—4

29

Research Question 2

► In Research Question 2, the exposure and the possible confounder are switched from Research Question 1

► Is smoking a risk factor for lung cancer?

► For Research Question 1, is cigarette smoking a confounder?

30

Is Confounding Present in Research Q 2?—1

31

Is Confounding Present in Research Q 2?—2

► Don’t need to assess the second or third criterion because the first one isn’t satisfied (i.e., the third factor is not a risk factor for the outcome)

32

Is Confounding Present in Research Q 2?—3

► Don’t need to assess the second or third criterion because the first one isn’t satisfied (i.e., the third factor is not a risk factor for the outcome)

33

Is Confounding Present in Research Q 2?—4

► Observed = True: Strong Association

► Don’t need to assess the second or third criterion because the first one isn’t satisfied (i.e., the third factor is not a risk factor for the outcome)

► Alcohol drinking is not a confounder when addressing Research Question 2

The material in this video is subject to the copyright of the owners of the material and is being provided for educational purposes under rules of fair use for registered students in this course only. No additional copies of the copyrighted work may be made or distributed.

How to Evaluate and Minimize Confounding

2

Ways of Evaluating Confounding

► So far, we’ve discussed a conceptual way to determine whether confounding is likely present

► There’s another way: empirical ► Empirical Method 1: Practical assessment of the presence of confounding using 2x2

tables ► Empirical Method 2: Use of statistical models to assess the presence of confounding

3

Use the Data from Your Study to Assess the Presence of Confounding— 1

a) Calculate crude RR

4

Use the Data from Your Study to Assess the Presence of Confounding— 2

b) Stratify by potential confounder and calculate stratum-specific RRs

5

Continued

► Compare crude and stratum-specific RRs ► If they differ by more than 10 to 20%, then confounding is likely present

► Don’t report the confounded RR—remember, the estimate is wrong! ► Unless trying to document the fact of confounding ► Otherwise, report the stratum-specific RRs (or use statistical methods to take the

confounder into account—more in a moment)

6

Empirical Method 2: Statistical Models

► Use a statistical model, such as logistic regression or Cox proportional hazards regression, to estimate the association between the exposure and the disease before and after including the third factor C in the model

► Compare RR before adjusting for the confounder to the RR after adjusting for the confounder ► If the RRs differ by more than 10 to 20%, then confounding is likely present

7

Is Smoking Associated with MI?

► Smoking is a risk factor for myocardial infarction (MI). Smokers tend to drink. Determine whether smoking is a confounder of the alcohol-MI association.

Practical Assessment of the Presence of Confounding (Empirical Method 1)

Adapted from Schlesselman. (1982).

8

Is Alcohol Drinking Associated with MI?—1

► Is alcohol drinking associated with MI? ► No, the un-confounded, stratum-

specific ORs = 1.0

► Is smoking a confounder of the alcohol-MI association? ► Yes, the stratum-specific ORs differ

from the crude OR

9

Is Alcohol Drinking Associated with MI?—2

► In this example, the stratum-specific ORs are equal to 1.0 ► Do the stratum-specific ORs (or

RRs) have to be equal to 1.0 to rule in confounding?

► No, they just have to be equal to each other and different from the crude OR (RR)

10

Methods to Minimize Confounding

► In the study design ► Restriction ► Matching

► In the analysis ► Stratify (or restrict to one stratum) ► Adjust

11

Minimize Confounding by Study Design—1

► Restriction ► Only enroll participants

● Without the possible confounder, or ● With the possible confounder, or ● With only one level of the possible confounder

► For example, if age is the confounder, then only enroll a narrow age range

► Easy to use restriction, just need to carefully screen participants for study entry

12

Minimize Confounding by Study Design—2

► Matching ► In a case-control study, match the cases and the controls on the possible confounder

● Individual matching—for every case enrolled with the third factor, enroll a control also with the third factor, and for every case enrolled without the third factor, enroll a control also without the third factor

● Frequency matching—determine the prevalence of the third factor in the cases and then enroll controls to achieve the same prevalence [to avoid other biases, best to ascertain controls as the cases are being ascertained]

13

Minimize Confounding by Study Design—3

► Matching—We can match in a cohort study! ► In a cohort study, match the exposed and the unexposed on the possible confounder

● For each exposed participant enrolled with the third factor, enroll an unexposed participant also with the third factor, and for every unexposed participant enrolled without the third factor, enroll an unexposed participant without the third factor

● Not done very often—in large cohort studies, usually enroll all eligible individuals and deal with possible confounding in the analysis

14

Minimize Confounding by Study Design—4

► Individual matching in a case-control study ► 1 control per case ► >1 control per case

► Individual and frequency matching can be done for several factors ► Often in case-control studies we match controls to cases on age, gender, and race ► Typically, we recruit matched controls close in time to when the case is enrolled; we

don’t wait until all of the cases have been enrolled to recruit controls

15

Minimize Confounding in the Analysis

► Stratification (or restrict to one stratum) ► Perform the statistical analysis separately within strata of the third factor ► Obtain and report stratum-specific RRs

► Adjustment ► Weighted average of the stratum-specific RRs ► Multivariable modeling (next term Biostatistics)

16

Residual Confounding

► Despite controlling for confounding, some confounding can remain because the confounder isn’t measured perfectly or the method used to control for confounding is inadequate ► The adjusted RR still has some bias

► No foolproof method to identify or quantify residual confounding ► Consider how well the confounder has been measured or modeled

The material in this video is subject to the copyright of the owners of the material and is being provided for educational purposes under rules of fair use for registered students in this course only. No additional copies of the copyrighted work may be made or distributed.

Effect Modification

2

What Is Effect Modification?

► In contrast to confounding, effect modification is not an error in the design or analysis of the study

► Effect modification is present when the association between an exposure and outcome differs between categories of a third factor (another exposure). It may be due to the biological interaction between the two exposures.

3

What Does Effect Modification Look Like?

► The association between exposure E and disease D differs in the presence and absence of modifier M

4

Classic Example of Effect Modification—1

► Phenylketonuria is an inherited metabolic condition ► The mutation results in an inability to metabolize phenylalanine, an amino acid

(protein building block) ► Children with the condition experience severe neurocognitive developmental

difficulties ► Public health action: Test all children at birth and, if positive, feed them a diet that

does not contain phenylalanine

5

Phenylketonuria (PKU)—1

6

Phenylketonuria (PKU)—2

7

Phenylketonuria (PKU)—3

8

Phenylketonuria (PKU)—4

9

Phenylketonuria (PKU)—5

10

Phenylketonuria (PKU)—6

11

Phenylketonuria (PKU)—7

12

Classic Example of Effect Modification—2

View as 2x2 tables

13

Consider an Example of Effect Modification

14

Effect Modification Example

► What would we see if effect modification is present? ► A different RR for the association between the exposure (pro-carcinogen) and the

outcome (cancer) within levels of the third factor (enzyme) ► Comparing those with and without exposure to the pro-carcinogen

15

Classic Examples of Effect Modification

► Smoking, alcohol drinking, and head and neck cancer

► Smoking, asbestos, and mesothelioma

► Aflatoxin, hepatitis infection, and liver cancer

► Obesity, menopause status, and breast cancer

► Smoking, oral contraceptives, and blood clots

16

Stratification to Identify Effect Modification—1

1. Calculate the RR for the exposure-disease association within each stratum of the possible effect modifier

2. Compare stratum-specific RRs to each other ► If the same → no effect modification ► If different → effect modification

● How different? That’s the art of epidemiology!

17

Stratification to Identify Effect Modification—2

3. Divide the exposure-outcome data into two or more groups—or strata (singular, stratum)—defined by level of a third factor ► All members of the stratum have the same level of the third factor

4. The third factor can be… ► Demographic characteristics (age, race/ethnicity) ► Other kinds of factors—modifiable and non-modifiable (genetic variation, medical,

occupational, environmental)

18

A Recent Example of Effect Modification

► Can men with higher genetic risk make lifestyle changes to ameliorate their genetic risk?

Source: Loeb, S., Peskoe, S. B., Joshu, C. E., et al. (2015). Do environmental factors modify the genetic risk of prostate cancer? Cancer Epidemiol Biomarkers Prev, 24(1), 213-220. https://doi.org/10.1158/1055-9965.EPI-14-0786-T

19

Association between Higher Genetic Risk and Prostate Cancer Adjusted for Age and Education, Stratified by Environmental Factors

► 20 validated prostate cancer risk SNPs

► Bold font: evidence of effect modification between/among strata of modifiable factors ► Aspirin ► Ibuprofen ► Vegetable intake ► Selenium

supplement

Source: Loeb, S., Peskoe, S. B., Joshu, C. E., et al. (2015). Do environmental factors modify the genetic risk of prostate cancer? Cancer Epidemiol Biomarkers Prev, 24(1), 213-220. https://doi.org/10.1158/1055-9965.EPI-14-0786-T

20

Quantitative vs. Qualitative Effect Modification—1

21

Quantitative vs. Qualitative Effect Modification—2

The material in this video is subject to the copyright of the owners of the material and is being provided for educational purposes under rules of fair use for registered students in this course only. No additional copies of the copyrighted work may be made or distributed.

Interaction

2

What Is Interaction?

► Effect modification occurs because of biological interaction between two factors in the production of disease

► Statistical interaction occurs when the risk of the outcome in people who have two factors is greater or lesser than would be expected based on the independent risks of the outcome in people with one, but not both factors ► Greater—positive interaction (synergy) ► Lesser—negative interaction (antagonism)

3

Additive or Multiplicative?

► In the absence of interaction, when two factors act together, based on biology, they can follow… ► An additive model—the independent risks from the two factors add to each other ► A multiplicative model—the independent risks from the two factors multiply each

other

4

No Interaction: Risks Are Additive—1

► Additive model: incidences for participants exposed to neither, one, or both risk factors

Source: Gordis. (5th ed.).

5

No Interaction: Risks Are Additive—2

► Additive model: incidences for participants exposed to neither, one, or both risk factors

► 𝟗𝟗 − 𝟑𝟑 + 𝟏𝟏𝟏𝟏 − 𝟑𝟑 + 𝟑𝟑 = 𝟐𝟐𝟏𝟏

Source: Gordis. (5th ed.).

6

No Interaction: Risks Are Additive—3

► Additive model: incidences for participants exposed to neither, one, or both risk factors

► 𝟗𝟗 − 𝟑𝟑 + 𝟏𝟏𝟏𝟏 − 𝟑𝟑 + 𝟑𝟑 = 𝟐𝟐𝟏𝟏

Source: Gordis. (5th ed.).

7

No Interaction: Risks Are Multiplicative— 1

► Multiplicative model: incidences for participants exposed to neither, one, or both risk factors

8

No Interaction: Risks Are Multiplicative—2

► Incidences and relative risks for participants exposed to neither, one, or both risk factors:

9

No Interaction: Risks Are Multiplicative—3

► Incidences and relative risks for participants exposed to neither, one, or both risk factors:

10

No Interaction: Risks Are Multiplicative—4

► Incidences and relative risks for participants exposed to neither, one, or both risk factors:

11

Additive and Multiplicative Interaction

► Interaction occurs when the risk of the outcome in people who have two factors is greater or lesser than would be expected based on the independent risks of the outcome in people with one, but not both factors

12

Interaction: Additive Model

► Additive model: incidences for participants exposed to neither, one, or both risk factors

► Under the additive model, expected rate = 9 − 3 + 15 − 3 + 3 = 21 𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂 50

► 50 > 21, so interaction is present

13

Interaction: Multiplicative Model

► Multiplicative model: RRs for participants exposed to both risk factors is higher than expected based on the independent RRs

► Under the multiplicative model, expected RR=15, observed RR=40

► 40 > 3.0 × 5.0, so interaction is present

14

Additive Effect Modification—1

► Among younger men… ► RDY = (10/1,000 PY) −

(5/1,000 PY) = 5 per 1,000 person years

► Among older men… ► RDO = (100/1,000 PY) −

(50/1,000 PY) = 50 per 1,000 person years

Possible evidence for effect modification on the additive scale

15

Evaluating Effect Modification—1

► Evaluation of effect modification is optimally based on prior epidemiologic or biological evidence

► Pose primary research question ► Does A cause B? (with specification of person, place, time)

► Next, pose research questions about effect modification ► Does the association between A and B vary by levels of C?

16

Evaluating Effect Modification—2

► Effect modification should be estimated

► The finding of effect modification may be the most important result of a study

► Effect modification should be understood and reported (carefully) ► Don’t over or underplay the finding of effect modification

17

Let’s Practice! Example 1 (Question)

► Which of the following statements describes the role of X in this study? 1. Factor X is a mediator of the association between E and D 2. Factor X is an effect modifier of the association between E and D 3. Factor X is not a confounder of the association between E and D 4. Factor X is a confounder of the association between E and D

18

Let’s Practice! Example 1 (Answer)

► Which of the following statements describes the role of X in this study? 1. Factor X is a mediator of the association between E and D 2. Factor X is an effect modifier of the association between E and D 3. Factor X is not a confounder of the association between E and D 4. Factor X is a confounder of the association between E and D

19

Let’s Practice! Example 2 (Question)

► Which of the following statements describes the role of X in this study? 1. Factor X is a mediator of the association between E and D 2. Factor X is an effect modifier of the association between E and D 3. Factor X is not a confounder of the association between E and D 4. Factor X is a confounder of the association between E and D

20

Let’s Practice! Example 2 (Answer)

► Which of the following statements describes the role of X in this study? 1. Factor X is a mediator of the association between E and D 2. Factor X is an effect modifier of the association between E and D 3. Factor X is not a confounder of the association between E and D 4. Factor X is a confounder of the association between E and D