Epidemiology quiz

Synthesis and Practical Applications: Comparisons and Inferences

Jennifer Deal, PhD Johns Hopkins University

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.

Interpret Measures of Association

3

Have You Had at Least One Cup of Coffee Today?

4

Motivating Example: Coffee Consumption and Clinical Outcomes

► We will cover key concepts and material from: ► Lectures: Measures of Association 1 & 2; Bias; Confounding and Effect Modification,

Causal Inference ► Activities: Obesity and Colon Cancer; Analyzing Published Papers

► In order to investigate the association between coffee consumption and all-cause mortality

5

Association between Exposure and Outcomes

► Association: a relationship between an exposure and a disease ► In other words, an outcome or a condition ► Although not necessarily a causal relationship

► We measure associations when we hypothesize that an exposure temporally precedes the disease—and, thus, could be a cause of the disease

► Associations can be measured as the risk (or rate) difference or the relative risk

6

Why Care about Associations?

► Epidemiology and public health ► Asses the burden of disease ► Study natural history and prognosis of

disease ► Assess risk factors for disease (risk

factor: a factor that is causally related to a change in the risk of a relevant health process, outcome, or condition (Porta, 6th edition)

► Evaluate interventions ► Make policy ► Communicate with the public

► Why assess risk factors for disease? 1. To identify groups at high risk for

disease ● Factors associated with increased

risk may or may not be causal (risk marker)

2. To direct preventive efforts to appropriate populations ● To prevent disease, the factor must

be a cause of disease

7

Example 1: Coffee Consumption and Mortality—1

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

8

Example 1: Coffee Consumption and Mortality—2

► Population: 185,555 African Americans, Native Hawaiians, Japanese Americans, Latinos, whites aged 45 to 75 years at recruitment

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

9

Example 1: Coffee Consumption and Mortality—3

► Exposure: coffee consumption

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

10

Example 1: Coffee Consumption and Mortality—4

► Outcome: total and cause-specific death

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

11

Multiethnic Cohort Study—Study Population

“The authors describe the design and implementation of a large multiethnic cohort established to study diet and cancer in the United States. They detail the source of the subjects, sample size, questionnaire development, pilot work, and approaches to future analyses. The cohort consists of 215,251 adult men and women (age 45–75 years at baseline) living in Hawaii and in California (primarily Los Angeles County) with the following ethnic distribution: African-American (16.3%), Latino (22.0%), Japanese-American (26.4%), Native Hawaiian (6.5%), White (22.9%), and other ancestry (5.8%). From 1993 to 1996, participants entered the cohort by completing a 26-page, self-administered mail questionnaire that elicited a quantitative food frequency history, along with demographic and other information. Response rates ranged from 20% in Latinos to 49% in Japanese-Americans. As expected, both within and among ethnic groups, the questionnaire data show substantial variations in dietary intakes (nutrients as well as foods) and in the distributions of non-dietary risk factors (including smoking, alcohol consumption, obesity, and physical activity). When compared with corresponding ethnic-specific cancer incidence rates, the findings provide tentative support for several current dietary hypotheses. As sufficient numbers of cancer cases are identified through surveillance of the cohort, dietary and other hypotheses will be tested in prospective analyses.”

Source: Kolonel et al. (2000). A multiethnic cohort in Hawaii and Los Angeles: Baseline characteristics. Am J Epidemiol, 151(4), 346-357.

12

Multiethnic Cohort Study—Source Population

► Source population: driver’s licenses, census tracts with ≥65% African Americans, Health Care Financing Administration files to identify ethnic minority adults ages 45 to 75 living in Hawaii & California between 1993 and 1996

“The authors describe the design and implementation of a large multiethnic cohort established to study diet and cancer in the United States. They detail the source of the subjects, sample size, questionnaire development, pilot work, and approaches to future analyses. The cohort consists of 215,251 adult men and women (age 45–75 years at baseline) living in Hawaii and in California (primarily Los Angeles County) with the following ethnic distribution: African-American (16.3%), Latino (22.0%), Japanese-American (26.4%), Native Hawaiian (6.5%), White (22.9%), and other ancestry (5.8%). From 1993 to 1996, participants entered the cohort by completing a 26-page, self-administered mail questionnaire that elicited a quantitative food frequency history, along with demographic and other information. Response rates ranged from 20% in Latinos to 49% in Japanese-Americans. As expected, both within and among ethnic groups, the questionnaire data show substantial variations in dietary intakes (nutrients as well as foods) and in the distributions of non-dietary risk factors (including smoking, alcohol consumption, obesity, and physical activity). When compared with corresponding ethnic-specific cancer incidence rates, the findings provide tentative support for several current dietary hypotheses. As sufficient numbers of cancer cases are identified through surveillance of the cohort, dietary and other hypotheses will be tested in prospective analyses.”

Source: Kolonel et al. (2000). A multiethnic cohort in Hawaii and Los Angeles: Baseline characteristics. Am J Epidemiol, 151(4), 346-357.

13

Multiethnic Cohort Study—Target Population

► Target population? “The authors describe the design and implementation of a large multiethnic cohort established to study diet and cancer in the United States. They detail the source of the subjects, sample size, questionnaire development, pilot work, and approaches to future analyses. The cohort consists of 215,251 adult men and women (age 45–75 years at baseline) living in Hawaii and in California (primarily Los Angeles County) with the following ethnic distribution: African-American (16.3%), Latino (22.0%), Japanese-American (26.4%), Native Hawaiian (6.5%), White (22.9%), and other ancestry (5.8%). From 1993 to 1996, participants entered the cohort by completing a 26-page, self-administered mail questionnaire that elicited a quantitative food frequency history, along with demographic and other information. Response rates ranged from 20% in Latinos to 49% in Japanese-Americans. As expected, both within and among ethnic groups, the questionnaire data show substantial variations in dietary intakes (nutrients as well as foods) and in the distributions of non-dietary risk factors (including smoking, alcohol consumption, obesity, and physical activity). When compared with corresponding ethnic-specific cancer incidence rates, the findings provide tentative support for several current dietary hypotheses. As sufficient numbers of cancer cases are identified through surveillance of the cohort, dietary and other hypotheses will be tested in prospective analyses.”

Source: Kolonel et al. (2000). A multiethnic cohort in Hawaii and Los Angeles: Baseline characteristics. Am J Epidemiol, 151(4), 346-357.

14

Baseline Characteristics: Coffee Consumption and Mortality

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

15

Incidence or mortality

Measures of Association

► Study design: prospective cohort

► Which measure of association did authors use for the MEC prospective cohort study evaluating association between coffee consumption and cause-specific death?

► Naming measures of association ► Difference in risks (or rates)

● Risk difference (or rate difference) ● Absolute risk ● Attributable risk

► Ratio of risks (or rates) ● Relative risk ● Risk ratio ● Cumulative incidence (or mortality) ratio ● If rates or hazards are estimate  Rate ratio  𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻 𝐻𝐻𝐻𝐻𝑟𝑟𝑟𝑟𝑟𝑟 = 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 𝑅𝑅𝑖𝑖 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑅𝑅𝑒𝑒𝑒𝑒

𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 𝑅𝑅𝑖𝑖 𝑖𝑖𝑒𝑒𝑖𝑖𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑅𝑅𝑒𝑒𝑒𝑒

16

Reminder: Interpretation of Risks—1

► Interpreting relative risks ► The direction of the association provides

information on the nature of the influence of the exposure on the disease

► The magnitude of the association provides information about the strength of the relationship between an exposure and disease

17

Reminder: Interpretation of Risks—2

► Interpreting relative risks ► The direction of the association provides

information on the nature of the influence of the exposure on the disease

► The magnitude of the association provides information about the strength of the relationship between an exposure and disease

► Interpreting relative risks ► RR = 1

● Risk in exposed = risk in nonexposed ● No association

► RR > 1 ● Risk in exposed > risk in nonexposed ● Positive association ● Causal?

► RR < 1 ● Risk in exposed < risk in nonexposed ● Negative association ● Protective?

18

Reminder: Interpretation of Risks—3

► Interpreting relative risks ► The direction of the association provides

information on the nature of the influence of the exposure on the disease

► The magnitude of the association provides information about the strength of the relationship between an exposure and disease

► Same applies to hazard ratios: ● HR > 1 → Increased risk ● HR = 1 → No association ● HR < 1 → Decreased risk

► Interpreting relative risks ► RR = 1

● Risk in exposed = risk in nonexposed ● No association

► RR > 1 ● Risk in exposed > risk in nonexposed ● Positive association ● Causal?

► RR < 1 ● Risk in exposed < risk in nonexposed ● Negative association ● Protective?

19

Hazard Ratios and 95% CIs

► Six exposure categories ► Reference: 0 cups of

coffee per day

► Model 1: Age, sex, ethnicity ► Model 2: +Smoking status,

# cigarettes, # years smoking, # years since quitting

► Model 3: +BMI, education, physical activity, alcohol consumption, energy, preexisting illness

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

20

Interpret the Hazard Ratio—1

Coffee consumption Participants, n Deaths, n Adjusted hazard ratio (95%) Model 1 Adjusted hazard

ratio (95%) Model 2 Adjusted hazard

ratio (95%) Model 3

None 30,082 9,460 1.00 (reference) 1.00 (reference) 1.00 (reference)

1-3 cups per month 13,370 4,277 1.00 (0.96-1.04) 0.98 (0.95-1.02) 1.00 (0.95-1.05)

1-6 cups per week 24,637 7,894 0.99 (0.96-1.02) 0.94 (0.91-0.97) 0.97 (0.93-1.01)

1 cup per day 57,488 19,623 0.97 (0.95-1.00) 0.88 (0.85-0.90) 0.88 (0.85-0.91)

2-3 cups per day 47,282 13,395 0.95 (0.93-0.98) 0.80 (0.78-0.83) 0.82 (0.79-0.86)

≥4 cups per day 12,996 3,748 1.11 (1.07-1.16) 0.80 (0.77-0.84) 0.82 (0.78-0.87)

P for trend - - 0.098 <0.001 <0.001

Increase per cup - - 1.00 (1.00-1.01) 0.94 (0.94-0.95) 0.95 (0.94-0.96)

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

21

Interpret the Hazard Ratio—2

Coffee consumption Participants, n Deaths, n Adjusted hazard ratio (95%) Model 1 Adjusted hazard ratio

(95%) Model 2 Adjusted hazard ratio

(95%) Model 3

None 30,082 9,460 1.00 (reference) 1.00 (reference) 1.00 (reference)

1-3 cups per month 13,370 4,277 1.00 (0.96-1.04) 0.98 (0.95-1.02) 1.00 (0.95-1.05)

1-6 cups per week 24,637 7,894 0.99 (0.96-1.02) 0.94 (0.91-0.97) 0.97 (0.93-1.01)

1 cup per day 57,488 19,623 0.97 (0.95-1.00) 0.88 (0.85-0.90) 0.88 (0.85-0.91)

2-3 cups per day 47,282 13,395 0.95 (0.93-0.98) 0.80 (0.78-0.83) 0.82 (0.79-0.86)

≥4 cups per day 12,996 3,748 1.11 (1.07-1.16) 0.80 (0.77-0.84) 0.82 (0.78-0.87)

P for trend - - 0.098 <0.001 <0.001

Increase per cup - - 1.00 (1.00-1.01) 0.94 (0.94-0.95) 0.95 (0.94-0.96)

► Adjusted HR: 0.88 ► People who drink 1 cup of coffee per day have reduced risk (HR<1) for total mortality compared to those who

do not drink any coffee (unexposed group) ► 95% confidence interval: 0.85-0.90

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

22

Is this association real? Let’s investigate bias.…

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.

Bias

2

Definitions: Bias

► “Systematic deviation of results or inferences from truth… An error in the conception, analysis, interpretation, reporting, publication, or review of data—leading to results or conclusions that are systematically (as opposed to randomly) different from the truth.”

Source: Porta, M. (Ed.). (2008). A dictionary epidemiology (6th ed.). Oxford University Press.

3

Bias and Internal Validity

► Bias can affect… ► Internal validity

● Refers to the validity of the comparison of groups within the study

● “Degree to which a study is free from bias or systematic error”

4

Non-differential Information Bias—1

► Misclassification or measurement error in the exposure and/or outcome

► Extent of the error does not differ between the groups being compared

5

Non-differential Information Bias—2

► Can you think about how misclassification or measurement error may occur in the Park et al. (2017) coffee consumption study?

6

Example of Non-differential Bias: Coffee Consumption—1

► Key features ► Cohort study ► Self-reported

exposure: coffee consumption

► Possible misclassification of the exposure?

“Our study also has limitations. As with any observational study, we cannot exclude the possibility of residual or unmeasured confounding. However, we are reassured by similar findings within different racial/ethnic populations in our cohort and in previous studies in the United States, Europe, and Japan. Also, a formal sensitivity analysis showed that the coffee–mortality association was robust to all but very strong levels of unmeasured confounding, which seems unlikely. On the other hand, self-reported coffee consumption at baseline is subject to measurement error, and consumption might have changed throughout follow-up. Among the participants who responded to the repeated QFFQ in 2003 to 2007 (n = 84 170) (average 11.0 years between measurements), the intraclass correlation coefficient between the 2 QFFQs was 0.60, which reflects potential changes over time or misclassification of coffee consumption. In the subset of participants with follow-up data on coffee consumption, the analysis using a time-dependent model yielded similar results.”

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

7

Example of Non-differential Bias: Coffee Consumption—2

► Key features ► Cohort study ► Self-reported

exposure: coffee consumption

► Possible misclassification of the exposure?

► Non-differential vs. differential? How does this affect the study’s validity?

“Our study also has limitations. As with any observational study, we cannot exclude the possibility of residual or unmeasured confounding. However, we are reassured by similar findings within different racial/ethnic populations in our cohort and in previous studies in the United States, Europe, and Japan. Also, a formal sensitivity analysis showed that the coffee–mortality association was robust to all but very strong levels of unmeasured confounding, which seems unlikely. On the other hand, self-reported coffee consumption at baseline is subject to measurement error, and consumption might have changed throughout follow-up. Among the participants who responded to the repeated QFFQ in 2003 to 2007 (n = 84 170) (average 11.0 years between measurements), the intraclass correlation coefficient between the 2 QFFQs was 0.60, which reflects potential changes over time or misclassification of coffee consumption. In the subset of participants with follow-up data on coffee consumption, the analysis using a time-dependent model yielded similar results.”

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

8

Consequence of Non-differential Information Bias

► Bias associations toward the “null” ► Association observed is weaker than the truth

► If the true OR* > 1, the biased OR* < true OR* ► For example, true OR = 2.0 and biased OR = 1.5

► If the true OR* < 1, the biased OR* > true OR* ► For example, true OR = 0.7 and biased OR = 0.9

*Same is true for the RR in a cohort study

9

Example of Non-differential Bias: Coffee Consumption—3

► Key features ► Cohort study ► Self-reported

exposure: coffee consumption

► Possible misclassification of the exposure? → non- differential bias

“Our study also has limitations. As with any observational study, we cannot exclude the possibility of residual or unmeasured confounding. However, we are reassured by similar findings within different racial/ethnic populations in our cohort and in previous studies in the United States, Europe, and Japan. Also, a formal sensitivity analysis showed that the coffee–mortality association was robust to all but very strong levels of unmeasured confounding, which seems unlikely. On the other hand, self-reported coffee consumption at baseline is subject to measurement error, and consumption might have changed throughout follow-up. Among the participants who responded to the repeated QFFQ in 2003 to 2007 (n = 84 170) (average 11.0 years between measurements), the intraclass correlation coefficient between the 2 QFFQs was 0.60, which reflects potential changes over time or misclassification of coffee consumption. In the subset of participants with follow-up data on coffee consumption, the analysis using a time-dependent model yielded similar results.”

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

10

Example of Non-differential Bias: Coffee Consumption—4

► Key features ► Cohort study ► Self-reported

exposure: coffee consumption

► Possible misclassification of the exposure?

► Did study mention other types of bias?

“Our study also has limitations. As with any observational study, we cannot exclude the possibility of residual or unmeasured confounding. However, we are reassured by similar findings within different racial/ethnic populations in our cohort and in previous studies in the United States, Europe, and Japan. Also, a formal sensitivity analysis showed that the coffee–mortality association was robust to all but very strong levels of unmeasured confounding, which seems unlikely. On the other hand, self-reported coffee consumption at baseline is subject to measurement error, and consumption might have changed throughout follow-up. Among the participants who responded to the repeated QFFQ in 2003 to 2007 (n = 84 170) (average 11.0 years between measurements), the intraclass correlation coefficient between the 2 QFFQs was 0.60, which reflects potential changes over time or misclassification of coffee consumption. In the subset of participants with follow-up data on coffee consumption, the analysis using a time-dependent model yielded similar results.”

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

11

Selection Bias (Always a Differential Bias)

► Distortion of a measure of association from what would be observed in the target population due to the choice (that is, selection) of participants ► Error occurring in the design phase of the study

► Distortion in the association means the measure of association is too large or too small ► Sometimes the direction of the association is even distorted

12

Example of Selection Bias: Coffee Consumption and Pancreatic Cancer

Coffee and Cancer of the Pancreas “Abstract: We questioned 369 patients with histologically proved cancer of the pancreas and 644 control patients about their use of tobacco, alcohol, tea, and coffee. There was a weak positive association between pancreatic cancer and cigarette smoking, but we found no association with use of cigars, pipe tobacco, alcoholic beverages, or tea. A strong association between coffee consumption and pancreatic cancer was evident in both sexes. The association was not affected by controlling for cigarette use. For the sexes combined, there was a significant dose-

response relation (P ∼ 0.001); after adjustment for cigarette smoking, the relative risk associated with drinking up to two cups of coffee per day was 1.8 (95 per cent confidence limits, 1.0 to 3.0), and that with three or more cups per day was 2.7 (1.6 to 4.7). This association should be evaluated with other data; if it reflects a causal relation between coffee drinking and pancreatic cancer, coffee use might account for a substantial proportion of the cases of this disease in the United States. (N Engl J Med. 1981; 304:630–3.)”

Source: MacMahon et al. (1981). Coffee and cancer of the pancreas. N Engl J Med, 304(11), 630-633. https://doi.org/10.1056/NEJM198103123041102

13

How Selection Bias Was Introduced in Pancreatic Cancer and Coffee Consumption Example—1

Study: Pancreatic cancer and coffee consumption (MacMahon et al., 1981) Design: Case-control study Cases: Patients with histologic diagnosis of pancreatic cancer Controls: Other patients under care of same physician as the cases—patients with diseases known to be associated with smoking or alcohol consumption excluded

14

How Selection Bias Was Introduced in Pancreatic Cancer and Coffee Consumption Example—2

Study: Pancreatic cancer and coffee consumption (MacMahon et al., 1981) Design: Case-control study Cases: Patients with histologic diagnosis of pancreatic cancer Controls: Other patients under care of same physician as the cases—patients with diseases known to be associated with smoking or alcohol consumption excluded

15

How Selection Bias Was Introduced in Pancreatic Cancer and Coffee Consumption Example—3

Study: Pancreatic cancer and coffee consumption (MacMahon et al., 1981) Design: Case-control study Cases: Patients with histologic diagnosis of pancreatic cancer Controls: Other patients under care of same physician as the cases—patients with diseases known to be associated with smoking or alcohol consumption excluded

Thus, eligibility criteria of controls distorted exposure prevalence in the controls relative to the source population and resulted in selection bias

16

Selection Bias Here? Multiethnic Cohort Study

► Source population: driver’s licenses, census tracts with ≥65% African Americans, Health Care Financing Administration files to identify ethnic minority adults ages 45 to 75 living in Hawaii and California between 1993 and 1996

Source: Kolonel et al. (2000). A multiethnic cohort in Hawaii and Los Angeles: Baseline characteristics. Am J Epidemiol, 151(4), 346-357.

17

Selection vs. Selection Bias—Important Notes about Selection

► “Selection” of the study population from a source population ≠ selection bias—internal validity

► For a study to be internally valid, it should be free of selection (and information) bias

► Sometimes a selected study population can help improve the internal validity of the study ► E.g., nurses participating in the Nurse’s Health Study ► But often not representative of the target population

► If not representative, the results may be internally valid, but not generalizable to the target population—external validity

18

Another Example of Selection: Clinicopathological Evaluation of Chronic Traumatic Encephalopathy in Players of American Football

► “Key Points ► Question: What are the neuropathological and clinical features of a case series of

deceased players of American football neuropathologically diagnosed as having chronic traumatic encephalopathy (CTE)?

► Findings: In a convenience sample of 202 deceased players of American football from a brain donation program, CTE was neuropathologically diagnosed in 177 players across all levels of play (87%), including 110 of 111 former National Football League players (99%).

► Meaning: In a convenience sample of deceased players of American football, a high proportion showed pathological evidence of CTE, suggesting that CTE may be related to prior participation in football.”

Source: Mez et al. (2017). Clinicopathological evaluation of chronic traumatic encephalopathy in players of American football. JAMA, 318(4), 360-70. https://doi.org/10.1001/jama.2017.8334

19

Example of Selection: CTE in Football Players

► Study: CTE in Football Players

► Design: case series

► Cases: American football players at any level of play

► Can you think of potential sources of bias?

This study had several limitations. First, a major limitation is ascertainment bias associated with participation in this brain donation program. Although the criteria for participation were based on exposure to repetitive head trauma rather than on clinical signs of brain trauma, public awareness of a possible link between repetitive head trauma and CTE may have motivated players and their families with symptoms and signs of brain injury to participate in this research. Therefore, caution must be used in interpreting the high frequency of CTE in this sample, and estimates of prevalence cannot be concluded or implied from this sample. Second, the VA-BU-CLF brain bank is not representative of the overall population of former players of American football; most players of American football have played only on youth or high school teams, but the majority of the brain bank donors in this study played at the college or professional level. Additionally, selection into brain banks is associated with dementia status, depression status, marital status, age, sex, race, and education.36 Third, this study lacked a comparison group that is representative of all individuals exposed to American football at the college or professional level, precluding estimation of the risk of participation in football and neuropathological outcomes.

Source: Mez et al. (2017). Clinicopathological evaluation of chronic traumatic encephalopathy in players of American football. JAMA, 318(4), 360-70. https://doi.org/10.1001/jama.2017.8334

20

For Your Interest: NPR Discussion on CTE in Football Players

► “All the brains studied were donated, she says. ‘Families don't donate brains of their loved ones unless they’re concerned about the person. So all the players in this study, on some level, were symptomatic. That leaves you with a very skewed population.’”

Source: NPR website. Accessed January 30, 2019, at https://www.npr.org/2017/07/25/539198429/study-cte-found-in-nearly-all-donated-nfl-player-brains

21

Selection Bias Summary: How Does Selection Bias Occur?

► Eligibility criteria ► Poor inappropriate eligibility criteria

► Feasibility ► Poor sampling frame (for example, controls not from the same source population as

cases)

► Differential participation ► Non-participation ► Participant non-response ► Losses to follow-up and dropout

► Different study designs = Different biases!

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 and Effect Modification

2

On Confounding and Effect Modification

► 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

3

Confounding

► “… 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

4

Confounder: Classic Definition

► A confounding variable or confounder is a third factor that is… ► A risk factor for the disease ► Associated with the exposure and ► Not a factor is causal pathway from exposure to disease (that is, 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

5

Smoking as an Example of Confounding—1

► Key points 1. Smoking associated with coffee

consumption (exposure) 2. Smoking risk factor for total

mortality (outcome) 3. Smoking not a mediator

(consequence of coffee consumption) on pathway to mortality

► Thus, smoking is a confounder for the coffee–mortality association

6

Smoking as an Example of Confounding—2

► Key points 1. Smoking associated with coffee consumption

(exposure) 2. Smoking risk factor for total mortality

(outcome) 3. Smoking not a mediator (consequence of

coffee consumption) on pathway to mortality ► Thus, smoking is a confounder for the coffee–

mortality association ► How did authors address this?

7

Guidance for Addressing Bias in Observational Studies: Strobe Statement

STROBE Statement—checklist items that should be included in reports of observational studies

Item Item No.

Recommendation

Bias 9 Describe any efforts to address potential sources of bias.

Study size 10 Explain how the study size was arrived at.

Quantitative variables

11 Explain how quantitative valuables were handled in the analyses. If applicable, describe which groupings were chosen and why.

Statistical methods 12 (a) Describe all statistical methods, including those used to control for confounding

(b) Describe any methods used to examine subgroups and interactions

(c) Explain how missing data were addressed (d) Cohort study—If applicable, explain how loss to follow-up was

addressed Case-control study—If applicable, explain how matching of cases and controls was addressed Cross-sectional study—If applicable, describe analytical methods taking account of sampling strategy

STROBE = Strengthening the Reporting of Observational Studies in Epidemiology Source: STROBE Statement website. Accessed January 30, 2019, at https://www.strobe-statement.org/fileadmin/Strobe/uploads/checklists/STROBE_checklist_v4_combined_PlosMedicine.pdf

8

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

9

Adjusting for Confounding Variables

► Adjusted for confounding effects of smoking in regression models (Models 2 and 3)

► What other variables were considered confounders? Why?

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

10

Identifying Other Confounding Variables—1

► Examine Table 1. Are there any other differences in baseline characteristics across exposure groups?

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

11

Identifying Other Confounding Variables—2

► Examine Table 1. Are there any other differences in baseline characteristics across exposure groups?

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

12

Adjusted for Confounders, What Does This Mean?—1

► Reduced risk for total mortality observed even after adjusting for confounders

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

13

Adjusted for Confounders, What Does This Mean?—2

► Reduced risk for total mortality observed even after adjusting for confounders

► Possibility for confounding of other unmeasured variables

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

14

Adjusted for Confounders, What Does This Mean?—3

► Reduced risk for total mortality observed even after adjusting for confounders

► Possibility for confounding of other unmeasured variables

► Never truly “free” of bias…

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

15

Unmeasured and 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 ► In other words, 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

16

Confounding Summary: Confounding Is Research Question Specific

► 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 outcome

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

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

17

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 ► It results due to the biological interaction between two factors

18

Back to the MEC Study—1

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

19

Back to the MEC Study—2

► No evidence that the association between coffee drinking and mortality differs by race/ethnicity

Source: Park et al. (2017). Ann Intern Med, 167(4), 228-235. https://doi.org/10.7326/M16-2472

20

Another Coffee Consumption Example

Source: Gunter et al. (2017). Ann Intern Med, 167(4), 236-247. http://dx.doi.org/10.7326/M16-2945

21

Testing for Interactions: Coffee Consumption in Europe

► Interaction tested in analysis phase

► Subgroup analysis on smoking status, BMI, physical activity…

► Interaction terms included in regression models

► Further stratified by sex

“The association between coffee consumption and mortality was further assessed across subgroups based on smoking status, body mass index, physical activity, alcohol intake, red and processed meat consumption, and fruit and vegetable consumption. Interaction terms (multiplicative scale) between these variables and coffee intake were included in separate models; the statistical significance of the cross-product terms was evaluated using the likelihood ratio test. Similar analyses examined associations according to follow-up categories (<5, 5 to <10, and ≥10 years). Heterogeneity across countries was explored using a meta-analytic approach.”

Source: Gunter et al. (2017). Ann Intern Med, 167(4), 236-247. http://dx.doi.org/10.7326/M16-2945

22

Stratifying to Identify Effect Modification

► Calculate the RR for the exposure–disease association within each stratum of the possible effect modifier

► 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.

23

Results: Subgroup Analyses—1

Source: Gunter et al. (2017). Ann Intern Med, 167(4), 236-247. http://dx.doi.org/10.7326/M16-2945

24

Results: Subgroup Analyses—2

Source: Gunter et al. (2017). Ann Intern Med, 167(4), 236-247. http://dx.doi.org/10.7326/M16-2945

25

Topic for Discussion: Effect Modification by Sex, Gender?—1

“Reporting Sex, Gender, or Both in Clinical Research? Virtually every clinical research report includes basic demographic characteristics about the study participants, such as age, and how many participants were male/men or female/women. Some research articles refer to this latter variable as sex, others refer to it as gender. As one of the first pieces of data reported, the importance of including sex appears undisputed. But what does the sex-gender category really entail, and how should it be reported? With emerging evidence that both sex and gender have an effect, for instance, on how an individual selects, responds to, metabolizes, and adheres to a particular drug regimen,1 there is an ethical and scientific imperative to report to whom research results apply. This Viewpoint explains the contexts in which sex and gender are relevant and provides suggestions for improving reporting of this characteristic.”

Source: Clayton & Tannenbaum. (2016). JAMA, 316(18), 1863-4. https://doi.org/10.1001/jama.2016.16405

26

Topic for Discussion: Effect Modification by Sex, Gender?—2

“Reporting Sex, Gender, or Both in Clinical Research? Virtually every clinical research report includes basic demographic characteristics about the study participants, such as age, and how many participants were male/men or female/women. Some research articles refer to this latter variable as sex, others refer to it as gender. As one of the first pieces of data reported, the importance of including sex appears undisputed. But what does the sex-gender category really entail, and how should it be reported? With emerging evidence that both sex and gender have an effect, for instance, on how an individual selects, responds to, metabolizes, and adheres to a particular drug regimen,1 there is an ethical and scientific imperative to report to whom research results apply. This Viewpoint explains the contexts in which sex and gender are relevant and provides suggestions for improving reporting of this characteristic.”

“To the Editor Drs Clayton and Tannenbaum1 asked whether sex, gender, or both should be reported in clinical research. They do not consider that neither should be reported. Sex or gender is routinely included as a covariate or stratification variable in statistical analyses, even in the absence of a priori hypotheses, risking false attribution. Given the close attention to such associations in society, the risks of false associations can be serious.”

Source: Wexler, B. E. (2017). [Letter to the editor]. JAMA, 317(9), 974. https://doi.org/10.1001/jama.2017.0139

27

Topic for Discussion: Effect Modification by Sex, Gender?—3

“Reporting Sex, Gender, or Both in Clinical Research? Virtually every clinical research report includes basic demographic characteristics about the study participants, such as age, and how many participants were male/men or female/women. Some research articles refer to this latter variable as sex, others refer to it as gender. As one of the first pieces of data reported, the importance of including sex appears undisputed. But what does the sex-gender category really entail, and how should it be reported? With emerging evidence that both sex and gender have an effect, for instance, on how an individual selects, responds to, metabolizes, and adheres to a particular drug regimen,1 there is an ethical and scientific imperative to report to whom research results apply. This Viewpoint explains the contexts in which sex and gender are relevant and provides suggestions for improving reporting of this characteristic.”

“To the Editor In a Viewpoint, Drs Clayton and Tannebaum1 highlighted the importance of appropriately reporting sex and gender in clinical research. However, their recommendations to report sex (male, female) and gender (man, woman) fall short of the realities of patients’ genetic, biological, and lived experiences and the effects on their health and well-being. The 2011 Institute of Medicine report on lesbian, gay, bisexual, and transgender health2 recommended that sexual orientation and gender identity be documented. In 2016, the National Institutes of Health strategic plan underscored the need to better understand sexual and gender minorities beyond binary categories of male or female and man or woman, especially how they interact with other characteristics such as race, ethnicity, and socioeconomic status;…”

Source: Streed & Makadon. (2017). [Letter to the editor]. JAMA, 317(9), 974-5. https://dx.doi.org/10.1001/jama.2017.0145

28

Effect Modification: Summary

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

► 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 or synergy ► Lesser: negative interaction or antagonism

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.

Causal Inference

2

Conclusions about Coffee Consumption

► Coffee drinking is associated with reduced risk for death in 10 European countries (Gunter et al., 2017)

► Higher consumption of coffee is associated with lower risk for death in African Americans, Japanese Americans, Latinos, and whites (Park et al., 2017)

► Both studies adjusted for confounding in analysis

► Evaluated effect modification in subgroup analyses and tested for heterogeneity

3

Source of Confusion?—1

4

Source of Confusion?—2

5

Source of Confusion?—3

6

Source of Confusion?—4

7

Is the Association Real?

► Possible explanations for associations ► Cause ► Chance ► Bias ► Confounding

8

Assessing Causality (Overview)

► Assessing causality is neither simple nor straightforward

► Requires judgment based on the totality of evidence ► Result of any single study is only a part of the process of assessing causality

► Use causal inference

9

Domain of epidemiologic

research

Approach to Assess Causality: The Process of Causal Inference

► Develop evidence ► Observations of

people ● Individuals ● Groups

► Human experiments

► Other research

► Synthesize ► Systematic

reviews ► Meta-analysis ► Other evidence

► Evaluate ► Expert judgment ► Causal criteria

► Causal inference

10

Guidelines to Consider When Assessing Causality: 1964 Surgeon General’s Report and Causal “Criteria”

► Statistical methods cannot establish proof of a causal relationship in an association. The causal significance of an association is a matter of judgment which goes beyond any statement of statistical probability. To judge or evaluate the causal significance of the association between the attribute or agent and the disease, or effect upon health, a number of criteria must be utilized, no one of which is an all-sufficient basis for judgment. These criteria include: a) The consistency of the association b) The strength of the association c) The specificity of the association d) The temporal relationship of the association e) The coherence of the association

Source: 1964 Surgeon General’s Report. Cover of Smoking and Health by US Public Health Service, Office of the Surgeon General. Public domain. Accessed at Wikimedia January 31, 2019.

11

Guidelines to Consider When Assessing Causality: Austin Bradford Hill (1897–1991)

► With Richard Doll, first to demonstrate the connection of smoking with lung cancer ► 1950 case-control study ► British Doctors’ Study

► Developed theories of association and causation still used today to characterize causal relationships between exposure and disease

Photo by unknown. Creative Commons BY 4.0. Accessed January 31, 2019, at Wikipedia.

12

Application to Coffee Consumption Example?—1

Hill’s criteria Reasoning/evidence

Consistency? (Are findings from studies conducted using different designs in different populations similar?)

Strength? (How big is the measure of association?)

Specificity? (is the factor only associated with endpoint of interest?)

Temporality? (Was the factor experienced before endpoint occurred?)

13

Application to Coffee Consumption Example?—2

Hill’s criteria Reasoning/evidence

Consistency? (Are findings from studies conducted using different designs in different populations similar?)

Consistency observed in more recent studies Depends on outcome

Strength? (How big is the measure of association?)

Specificity? (is the factor only associated with endpoint of interest?)

Temporality? (Was the factor experienced before endpoint occurred?)

14

Application to Coffee Consumption Example?—3

Hill’s criteria Reasoning/evidence

Consistency? (Are findings from studies conducted using different designs in different populations similar?)

Consistency observed in more recent studies Depends on outcome

Strength? (How big is the measure of association?) Small to moderately strong associations: reported hazard ratios ranged between 0.82 and 0.92 for USC and European studies

Specificity? (is the factor only associated with endpoint of interest?)

Temporality? (Was the factor experienced before endpoint occurred?)

15

Application to Coffee Consumption Example?—4

Hill’s criteria Reasoning/evidence

Consistency? (Are findings from studies conducted using different designs in different populations similar?)

Consistency observed in more recent studies Depends on outcome

Strength? (How big is the measure of association?) Small to moderately strong associations: reported hazard ratios ranged between 0.82 and 0.92 for USC and European studies

Specificity? (is the factor only associated with endpoint of interest?)

Least important causal guideline for chronic diseases

Temporality? (Was the factor experienced before endpoint occurred?)

16

Application to Coffee Consumption Example?—5

Hill’s criteria Reasoning/evidence

Consistency? (Are findings from studies conducted using different designs in different populations similar?)

Consistency observed in more recent studies Depends on outcome

Strength? (How big is the measure of association?) Small to moderately strong associations: reported hazard ratios ranged between 0.82 and 0.92 for USC and European studies

Specificity? (is the factor only associated with endpoint of interest?)

Least important causal guideline for chronic diseases

Temporality? (Was the factor experienced before endpoint occurred?)

Maybe: recent studies attempt to assess • Prospective cohort designs • Evaluate reverse causality (European study)

17

Hill’s Guidelines and Coffee Consumption (continued)—1

Hill’s criteria Reasoning/evidence

Biological gradient? (Does the magnitude of the measure of association decrease with increasing extent of factor?)

Plausibility and coherence? (Is the association supported or refuted by the contemporary biologic literature?)

Experiment? (Does the magnitude of the measure of association decrease—or increase—after changing the factor status?)

Analogy? (Have association between similar factors and endpoints been observed?)

18

Hill’s Guidelines and Coffee Consumption (continued)—2

Hill’s criteria Reasoning/evidence

Biological gradient? (Does the magnitude of the measure of association decrease with increasing extent of factor?)

As exposure to coffee increases (e.g., from 1 cup to 4 cups daily), greater magnitude in measure of association is observed

Plausibility and coherence? (Is the association supported or refuted by the contemporary biologic literature?)

Experiment? (Does the magnitude of the measure of association decrease—or increase—after changing the factor status?)

Analogy? (Have associations between similar factors and endpoints been observed?)

19

Hill’s Guidelines and Coffee Consumption (continued)—3

Hill’s criteria Reasoning/evidence

Biological gradient? (Does the magnitude of the measure of association decrease with increasing extent of factor?)

As exposure to coffee increases (e.g., from 1 cup to 4 cups daily), greater magnitude in measure of association is observed

Plausibility and coherence? (Is the association supported or refuted by the contemporary biologic literature?)

• Reported to have fibrogenic effects on hepatic stellate cells lowering proliferation, stimulating apoptosis, and inhibiting adhesion

• Reduced fat accumulation, oxidative stress, and liver inflammation

Experiment? (Does the magnitude of the measure of association decrease—or increase—after changing the factor status?)

Analogy? (Have associations between similar factors and endpoints been observed?)

20

Hill’s Guidelines and Coffee Consumption (continued)—4

Hill’s criteria Reasoning/evidence

Biological gradient? (Does the magnitude of the measure of association decrease with increasing extent of factor?)

As exposure to coffee increases (e.g., from 1 cup to 4 cups daily), greater magnitude in measure of association is observed

Plausibility and coherence? (Is the association supported or refuted by the contemporary biologic literature?)

• Reported to have fibrogenic effects on hepatic stellate cells lowering proliferation, stimulating apoptosis, and inhibiting adhesion

• Reduced fat accumulation, oxidative stress, and liver inflammation

Experiment? (Does the magnitude of the measure of association decrease—or increase—after changing the factor status?)

Unknown Lower coffee categories associated with slightly increased risk (not significant) in USC study

Analogy? (Have associations between similar factors and endpoints been observed?)

21

Hill’s Guidelines and Coffee Consumption (continued)—5

Hill’s criteria Reasoning/evidence

Biological gradient? (Does the magnitude of the measure of association decrease with increasing extent of factor?)

As exposure to coffee increases (e.g., from 1 cup to 4 cups daily), greater magnitude in measure of association is observed

Plausibility and coherence? (Is the association supported or refuted by the contemporary biologic literature?)

• Reported to have fibrogenic effects on hepatic stellate cells lowering proliferation, stimulating apoptosis, and inhibiting adhesion

• Reduced fat accumulation, oxidative stress, and liver inflammation

Experiment? (Does the magnitude of the measure of association decrease [or increase] after changing the factor status?)

Unknown Lower coffee categories associated with slightly increased risk (not significant) in USC study

Analogy? (Have associations between similar factors and endpoints been observed?)

Maybe: • Chocolate, which contains caffeine, has been to shown to

improve mortality • Similar observations for green tea

22

Final Words

► “We’re not at the point where we can say with full confidence that it’s protective,” said Dr. Eliseo Guallar, a professor at the Johns Hopkins University Bloomberg School of Health, who co-authored an accompanying editorial. “But the basic idea is that we are increasingly reassured that coffee is not harmful. As doctors, we don’t have to tell people to be worried about drinking coffee anymore. Now we can tell people to drink their coffee and be happy.”

Source: Guallar, E., Blasco-Colmenares, E., Arking, D. E., & Zhao, D. (2017). Ann Intern Med, 167(4), 283-4. https://doi.org/10.7326/M17-1503

23

Lessons Learned

► Using published studies of coffee consumption and all-cause mortality, synthesized and applied epidemiologic knowledge in order to… ► Interpret measures of association (hazard ratios) ► Identify sources of bias and how they might impact our measure of association ► Compare and contrast confounding and effect modification ► Describe the process of causal inference