1 of 79

Study designs: Cohort studies

2/15/2011

Cohort studies

1

Principles of Epidemiology for Public Health (EPID600)�

Victor J. Schoenbach, PhD home page

Department of Epidemiology�Gillings School of Global Public Health�University of North Carolina at Chapel Hill

www.unc.edu/epid600/

2 of 79

“Bottled at the source”

Italy?

France?

Germany?

Switzerland?

3 of 79

Not quite – “Bottled at the CG Roxane Source a the Mountains of Tennessee”

Welcome to the American Alps!?

4 of 79

Students examination answers

From Ann Landers "Science answers are hilarious but scary” (taken from a Popular Science article citing [real] test answers given by students in science classes)

10/8/2001

4

5 of 79

“To collect fumes of sulfur, hold a deacon over a flame in a test tube.”

“Water is composed of two gins, oxygin and hydrogin. Oxygin is pure gin. Hydrogin is gin and water.”

“Nitrogen is not found in Ireland because it is not found in a free state.”

2/11/2003

5

6 of 79

“When you smell an odorless gas, it is probably carbon monoxide.”

“The pistol of a flower is its only protection against insects.”

“Germinate: to become a naturalized German.”

“To prevent contraception, wear a condominium.”

10/8/2001

6

7 of 79

Study designs: Cohort studies

2/15/2011

Cohort studies

7

Principles of Epidemiology for Public Health (EPID600)�

Victor J. Schoenbach, PhD home page

Department of Epidemiology�Gillings School of Global Public Health�University of North Carolina at Chapel Hill

www.unc.edu/epid600/

8 of 79

Cohort studies

Intuitive approach to studying disease incidence and risk factors:

1. Start with a population at risk

2. Measure characteristics at baseline

3. Follow-up the population over time with �a) surveillance or b) re-examination

4. Compare event rates in people with and without characteristics of interest

10/1/2001

Cohort studies

8

Cohort studies, of which intervention studies are a special case, are an “intuitive” approach to studying the incidence of disease and associations between risk factors and health outcomes.

We start by identifying a “population at risk”, a group of people who are in fact at risk (as far as we know) of developing the disease. For a chronic disease, that generally means that they have not yet developed it. We then measure the characteristics of this population at baseline, i.e., the start of a period of follow-up.

We follow-up the population over time, either through (a) surveillance of routinely-detected cases (people get symptoms and go to their health care providers, who diagnose them or, sometimes, people get screened and then referred for diagnosis) or (b) re-examining people periodically or at the end of the follow-up.

We then compare the incidence of events (development of the outcome) in people with characteristics we suspect may increase risk and compare this incidence to that in people without those characteristics.

9 of 79

Cohort studies

Can be large or small

Can be long or short

Can be simple or elaborate

Can be local or multinational

For rare outcomes need many people and/or lengthy follow-up

May have to decide what characteristics to measure long in advance

7/1/2009

Cohort studies

9

Cohort studies are, of course, quite diverse. There are large cohort studies, (e.g the Nurses Health Study with nearly 100,000 participants, or even the entire population of a country, by making use of comprehensive population registries that exist in Scandinavian countries) and small ones with only a few dozen participants. Cohort studies can be long, continuing over many years (e.g., the Framingham study) or short (e.g. a cohort of Egyptian newborns followed for 1 year for rotavirus; Naficy, et al. AJE 1999;150:770-7) . Cohort studies can be simple or elaborate, local or multinational.

Since we generally need to have at least tens of cases, preferably hundreds, and sometimes thousands, if incidence is low we need to follow many people and/or follow them for a long time. Among other things, that means that we often have to decide what risk characteristics we will study long before we get the results, by which time understanding and measurement technology will have evolved.

10 of 79

Case example – Atherosclerosis Risk in Communities (ARIC) Study�

Prospective study in four U.S. communities to investigate:

1. etiology and natural history of atherosclerosis

2. etiology of clinical atherosclerotic diseases

3. variation in CVD risk factors, medical care and disease by race, sex, place, and time.

10/1/2001

Cohort studies

10

To explain some of the concepts and features of cohort studies, let us take as a case example the Atherosclerosis Risk in Communities (ARIC) Study. The ARIC study was launched at the end of the 1980’s as “the Framingham of the 1990’s” – the Framingham Study was a community-based cardiovascular disease (CVD) cohort study begun at the end of the 1940’s, the first study of this genre, and studied several thousand residents of the city of Framingham, Massachusetts for several decades. The Framingham Study demonstrated and documented many of the “classic” CVD risk factors (e.g., high blood pressure, elevated blood cholesterol, cigarette smoking, and diabetes). The ARIC study, also organized and funded by the NIH National Heart, Lung, and Blood Institute (NHLBI), was intended as a major, long-term investigation to advance understanding of CVD and coronary heart disease (CHD) etiology and epidemiology beyond what was possible with studies like Framingham.

The term “prospective study” is usually used as a synonym for “cohort study.” ARIC was designed as a prospective study to be carried out simultaneously in four communities, in order to study a more diverse population than would be possible in any single location. Its three primary objectives were to investigate 1) the etiology and natural history of atherosclerosis (the underlying condition that is responsible for most CHD and CVD), 2) the etiology of clinical atherosclerotic diseases (angina pectoris, “heart attack”, stroke), and 3) variation in CVD risk factors, medical care and disease by race, sex, place, and time.

11 of 79

Background to ARIC Study – the CVD epidemic of the 20th century

Heart disease became the leading cause of death in men and women
Major CVD cohort studies, e.g.:

Framingham, MA British Civil Servants
Tecumseh, MI Paris
Evans County, GA …
Honolulu, HI

10/4/2005

Cohort studies

11

To better understand the purpose and rationale for the ARIC study, we should consider the background leading up to its creation. The 20th century witnessed a major epidemic of cardiovascular disease (CVD), during which heart disease (generally coronary heart disease, CHD) increased in frequency to become the leading cause of death in men and women in the U.S. and most industrialized countries.

When World War II ended in 1945, freeing up resources and attention for other national needs, the National Institutes of Health (NIH) was created. It’s first major epidemiologic study was a cohort study in the city of Framingham, Massaachusetts to verify evidence from various studies indicating that elevated blood pressure (hypertension), elevated blood cholesterol (hypercholesterolemia), and diabetes were contributing to the epidemic of CVD. The Framingham Study began in the 1940’s, and during the following two decades major CVD cohort studies were launched in a number of other communities (e.g., Tecumseh MI, Evans County, GA, Honolulu HI, Paris) and occupational settings (Western Electric employees, Chicago Gas employees, British Civil Servants). These studies were similar in that they enrolled people from the general population (i.e., not at special risk), gave them a baseline clinical and electrocardiographic examination, and followed those without evidence of CVD for at least several years.

12 of 79

Background to ARIC Study – the CVD epidemic of the 20th century

CVD (heart disease, stroke, hypertension, etc.) rose from the 4^th leading cause of death in 1900 to the leading cause by 1910
CVD death rates peaked in 1963 and proceeded to fall by over one-half (56%)
Death rates from coronary heart disease (CHD) and stroke fell most

2/15/2011

Cohort studies

12

The leading causes of death in the U.S. in 1900 were pneumonia, tuberculosis, diarrheal disease, and heart disease, but by 1910 heart disease had risen to first. By 1950 cardiovascular disease accounted for half of all U.S. deaths, and myocardial infarction (“heart attack”), first clinically recognized in 1912, had become the single most common cause of death.

Although findings from the Framingham Study began emerging in the early 1950’s (as did other evidence indicating that men who smoked cigarettes had a higher incidence of CVD), U.S. CVD rates continued to climb relentlessly during the 1950’s and into the 1960’s. Toward the end of the 1960’s, however, it began to appear that CVD death rates had begun to fall. During the following decade the decline continued and became more pronounced. CVD death rates fell by over half from their peak in 1963. Death rates from CHD and stroke fell the most.

13 of 79

10/1/2001

Cohort studies

13

CHD

Non-CVD

Stroke

This graph shows CVD death rates in the U.S. during the second half of the 20th century. From these data we see that CHD continued to climb steeply until 1963 and then declined just as steeply. Stroke death rates leveled off during the 1950’s and began their decline a few years before CHD. Death rates from all causes other than CVD began to decline during the 1950s and continued a moderate decline until beginning to rise again in 1980.

There are several things to take note of in interpreting these graphs. First, note that the data are age-adjusted to the 2000 standard. CVD rates are much higher in older persons, so if a population is aging – the proportion of older people is growing – then the CVD rate in that population will tend to increase. If a decline is in progress for other reasons, then the aging of the population will tend to mask that decline. So it is important to use age-adjusted rates when making comparisons of disease or death rates that are related to age. Since the rates within each age band, or stratum, are presumably not affected by changes in their number or proportion, we can compare age-specific rates (e.g., rates among 50-54 year-olds) across populations or time periods. However, since there are numerous strata to compare (40-44 years, 45-49 years, 50-54 years, etc.), it is usually more convenient to compare weighted averages of these stratum-specific rates. As long as we weight the rates from both populations in the same way, then the comparison should be valid, though there are subtleties that we will ignore for now.

14 of 79

10/1/2001

Cohort studies

14

CHD

Non-CVD

Stroke

Another point to note is the term “comparability ratios” under the graph. Diseases do not come with name tags telling us what they are. The process of diagnosing a disease is based on the medical profession’s understanding of that disease, as well as the training of and resources available to the health care provider. In reaching a diagnosis, the health care provider looks for defining signs and symptoms that enable her or him to decide among alternative diagnoses. Of course, people have different symptoms, different severity of symptoms, etc., so the clinician must always decide which symptoms, test results, and so on are of most significance. So diagnosis is based on a classification that attempts to combine “like with like” and draw important distinctions while ignoring unimportant ones. Similarly, disease classification for statistical purposes and epidemiologic research must employ some classification scheme to define the various outcomes to be studied.

The International Classification of Diseases (ICD), prepared under the auspices of the World Health Organization (WHO), is the official compendium for classifying health conditions and is generally the classification manual used for epidemiologic research on recognized diseases. However, as with any classification, it is necessary to revise it as knowledge grows and understanding evolves. So the ICD is revised every decade or so, which as you can imagine creates problems for examining changes in rates of disease or death. Comparability ratios are used to adjust for the effects of these revisions.

15 of 79

10/1/2001

Cohort studies

15

White male

Black male

Black female

White female

16 of 79

1978 National Heart, Lung and Blood Institute (NHLBI) Workshop on the Decline in Coronary Heart Disease Mortality

1. Is the decline in CVD mortality real?

2. How much of the decline reflects lower incidence (blood pressure control, smoking cessation, dietary change)?

3. How much reflects lower case fatality rate (better survival due to emergency medical services and coronary care units)?

10/1/2001

Cohort studies

16

A substantial and sustained decline in the leading cause of death is a rather important development for public health, and CVD researchers were naturally excited about it. NHLBI convened a workshop in 1978 to attempt to answer three key questions: 1) Was the decline real or due to some kind of artifact (e.g., a temporary fluctuation or due to classification of CVD deaths under some other rubric - these possibilities were soon dismissed, since the decline had persisted for over a decade, and total non-CVD mortality was stable or declining); 2) How much of the decline reflects lower incidence of CVD due to reductions in risk factors; 3) How much of the decline reflects lower case fatality due to better treatment? Needless to say, the answers to these questions had strong implications for giving credit and channeling resources.

17 of 79

Conclusion: The decline is real

Recommendations:

Need data on incidence and risk factor change in order to determine causes
NHLBI Community Cardiovascular Surveillance Program (1980-1984) developed and pilot-tested protocol for community surveillance

10/1/2001

Cohort studies

17

Workshop participants were able to conclude that the decline was real, but the existing data were not adequate to determine how much of the decline was due to each of the two components of CVD mortality – incidence and case fatality. The key gap, of course, is the lack of data on CVD incidence. Since a large proportion of CHD deaths in particular occur before the victim reaches the hospital, case fatality rates cannot be assessed solely from hospital data. A population denominator is needed.

Cancer incidence data have been collected through the network of cancer registries that in the U.S. began with the Connecticut Tumor Registry and later, the National Cancer Institute (NCI) SEER (Surveillance, Epidemiology, and End Results) network of cancer registries. But no such system existed for CVD. In addition, cancer is diagnosed based on pathologic examination of biopsied tissue, so pathology laboratories can serve as place to verify that all cases are being counted. The different types of CVD are diagnosed in different ways, and some CVD victims die from a stroke or heart attack before their disease has been diagnosed. In contrast, cancer (other than non-melanoma skin cancer) is is not as rapidly fatal as CVD can be and is nearly always treated in a hospital, so it is easier to count cases of cancer than of CVD.

To meet the need for CVD incidence data, therefore, NHLBI inaugurated a program to develop and pilot-test a system for community surveillance for CVD.

18 of 79

Atherosclerosis Risk in Communities Study� (ARIC)�

“the Framingham of the 1990’s”

Two components:

1. Community surveillance – estimate CVD� incidence

2. Cohort – validate and facilitate� interpretation of surveillance data

(See http://www.cscc.unc.edu/aric/)

10/7/2003

Cohort studies

18

19 of 79

Communities in ARIC Study�

Forsyth County, North Carolina (biracial)
Jackson, Mississippi (blacks)
Suburban Minneapolis, Minnesota
Washington County, Maryland

Defined geographical entities, well-delineated medical care referral patterns, black and white, urban and rural

10/1/2001

Cohort studies

19

20 of 79

Demographics of ARIC study communities, 1980

10/1/2001

Cohort studies

20

21 of 79

Age-adjusted mortality rates* in ARIC study communities, 1980

10/1/2001

Cohort studies

21

* per 100,000/year

22 of 79

Cohort study added to enhance ARIC community surveillance�

Cohort study – more and better data:

1. More data: provides information on risk factors and out-of-hospital medical care

2. Better data: uses standard methods for ascertaining events (surveillance relies on health care system)

10/1/2001

Cohort studies

22

23 of 79

Measure preclinical CVD (atherosclerosis) and CVD precursors

1. assess association of risk factors with both underlying and clinical diseases

2. assess value of B-mode ultrasound diagnosis in predicting clinical diseases

3. store blood in hope of discovering unsuspected precursors of CVD

10/1/2001

Cohort studies

23

A key advance over previous population-based studies was the plan to measure preclinical CVD, i.e., CVD that had not yet begun to cause symptoms. Most CVD arises as a result of the process of gradual occlusion of arteries due to the build-up of atherosclerotic plaque over prolonged periods of time. Very limited data were available on the development of atherosclerosis. All of the major CVD cohort studies of the previous decades studied people who at baseline had no signs of CVD, while realizing that a significant fraction of them would have subclinical atherosclerosis. So the ARIC study offered the possibility of using a newly-developed measurement technology – B-mode ultrasound – to measure the extent of atherosclerosis both at baseline and follow-up examinations. Since atherosclerosis in one area of the arterial tree is an indicator of atherosclerosis in other areas, B-mode ultrasound measurements of the carotid arteries (in the neck) could provide a window into the underlying disease that erupts as angina pectoris and myocardial infarction (MI, “heart attack”).

Besides using ultrasound technology to measure underlying disease and study its ability to predict clinical (symptomatic) disease, the ARIC study would store blood specimens collected at baseline, so that possible biochemical CVD precursors or risk factors not known in 1990 might be analyzed at a later date.

24 of 79

Community surveillance enhances generalizability of cohort findings

1. Cohort study – compare incidence rates and characteristics of events in residents who do and who do not participate in cohort

2. Community surveillance – compare the study communities’ CHD experience with areas in the U.S.

10/1/2001

Cohort studies

24

25 of 79

ARIC community surveillance for �hospitalized MI and CHD death in age 35-74

Hospital records with discharge diagnosis of MI or related “screening” diagnoses
Death certificates with various CHD manifestations coded as the cause of death
Interviews with physician and next-of-kin for deaths outside the hospital

2/162004

Cohort studies

25

Let’s see how the community surveillance component of ARIC was designed. First, its objective was to identify cases of hospitalized myocardial infarction (MI) and CHD death in persons age 35-74 years. Surveillance was accomplished by reviewing records for all hospitals in the ARIC communities to identify patients with a discharge diagnosis that might reflect undiagnosed CVD (note that the diagnosis at discharge may differ from the admitting diagnosis, since information gathered during the hospitalization may lead to a change in the diagnosis). Finally, for deaths occurring outside the hospital, death certificates with possible manifestations of CHD were coded by study nosologists (people trained in the authoritative classification of cause of death). Physicians and next-of-kin were interviewed to obtain additional information about deaths occurring outside of the hospital setting.

26 of 79

ARIC cohort study – 1

Different sampling scheme in each community
Map & enumerate households
Interview all eligible persons in household
Recruit 16,000 age 45-64, clinic examination (1986-1989)

10/4/2005

Cohort studies

26

27 of 79

ARIC cohort study – 2

Review medical records
Interview participants annually
Contact health care providers, family members
Re-examine every 3 years after first exam (1990-92, 1993-95)

10/1/2001

Cohort studies

27

28 of 79

ARIC cohort study – home interview

Health status, CVD risk factors
Family health status, past history of CVD, cancer or diabetes
Smoking status and amount
Current employment status
Level of education
Participant's cooperation, literacy/comprehension, interview quality

10/1/2001

Cohort studies

28

29 of 79

ARIC cohort study – clinic examination

3 l/2 hours, 2 or 3 simultaneous exams
Fasting and 12-hour abstinence (tobacco, alcohol) required prior to blood pressure and venipuncture
Sitting blood pressure must be measured before venipuncture
Interview and exam must precede the Medical Review

10/1/2001

Cohort studies

29

30 of 79

ARIC cohort study clinic exam – 1

Greet participant; determine fasting status; collect medications
Obtain informed consent
Measure sitting blood pressure
Measure weight, height, skinfolds, girths, and wrist breadth
Blood samples for lipid, hemostasis, hematology, and chemistries

10/1/2001

Cohort studies

30

31 of 79

ARIC cohort study clinic exam – 2

Snack (no caffeine or stimulants)
Obtain a digitized 12-lead ECG and 2-minute rhythm strip
Collect medical history (incl. Rose Quest.; stroke, TIA, respiratory symptoms, reproductive history) and food frequency
Brief systems review incl. neck, neurological, chest and lungs, breast (optional), heart, extremities.

10/1/2001

Cohort studies

31

32 of 79

ARIC cohort study clinic exam – 3

Digitized spirometric measurements of timed pulmonary function (FVC, FEV1).
B-mode ultrasound scans for wall measurements in carotids and a popliteal artery
Supine brachial and ankle blood pressure; heart rate and blood pressure changes as participant arises

(www.cscc.unc.edu/aric/visit/General_Description_and_Study_Management.1_1.pdf)

9/25/2002

Cohort studies

32

33 of 79

Central laboratories & Coordinating Center

Central lipid laboratory
2 ECG reading ctrs (Dalhousie, U of Minn)
Pulmonary function center
Ultrasound reading center
Study coordinating center (data monitoring, data mgmt, quality control, data analysis)

10/1/2001

Cohort studies

33

34 of 79

ARIC committees and subcommittees

Steering Committee
Laboratory and Sample Processing
Ultrasound Subcommittee
Risk Factors and Clinic Operations
Sampling, Recruitment, and Follow-Up
Criteria and Diagnoses
Morbidity and Mortality Classification

10/1/2001

Cohort studies

34

35 of 79

ARIC committees and subcommittees

10/1/2001

Cohort studies

35

36 of 79

Anger proneness predicts coronary heart disease risk��Prospective analysis from the Atherosclerosis Risk in Communities (ARIC) Study��Janice E. Williams, Catherine C. Paton, Ilene C. Siegler, Marsha L. Eigenbrodt, F. Javier Nieto, and Herman A. Tyroler� �Circulation 2000;101:2034

10/1/2001

Cohort studies

36

The ARIC study was a sophisticated, multi-site, elaborate cohort study. Now that we have illustrated its rationale, design, and conduct, we turn to a sample analysis of ARIC data. Data analysis and manuscript preparation using ARIC data are also governed by committees. Generally an individual or group of investigators propose a study, obtains approval, and then submits requests for data analysis. A writing group is formed for each paper and its progress is tracked. The particular analysis we will examine today was led by Janice Williams, who was then a postdoctoral fellow in the UNC Department of Epidemiology. The “writing group” included investigators at Duke and Johns Hopkins as well as UNC-CH. The topic – probably not one that was a central theme in ARIC but one that was made possible by the inclusion of a psychological scale in one of the interviews – is the relation of anger proneness to coronary heart disease risk. The case study for this topic is based on the second article from these analyses.

37 of 79

Background

Persons “with trait anger have rage and fury more often, more intensely, and with longer-lasting episodes.”
Studies have linked trait anger with CHD risk factors.
Studies have found associations between CHD and suppressed anger and difficulties with controlling anger.

10/1/2001

Cohort studies

37

38 of 79

Study population

14,348 participants (92.9% of baseline) returned to the ARIC visit 2 (1990-92) exam
Exclusions for this study:�1,140 with clinically manifest CHD (incl ECG)� 38 with ethnicity other than black or white� 40 with missing data on hypertension� 144 with incomplete anger questionnaire
12,986 participants available for this analysis

10/1/2001

Cohort studies

38

39 of 79

Spielberger’s 10-item trait anger scale

(1=never, 2=sometimes, 3=often, 4=almost always)

1. I am quick tempered.

2. I have a fiery temper.

3. I am a hotheaded person.

4. I get angry when I am slowed down by others’ mistakes.

5. I feel annoyed when I am not given recognition for doing good work.

10/1/2001

Cohort studies

39

40 of 79

Spielberger’s 10-item trait anger scale

6. I fly off the handle.

7. When I get angry, I say nasty things.

8. It makes me furious when I am criticized in front of others.

9. When I get frustrated, I feel like hitting someone.

10. I feel infuriated when I do a good job and get a poor evaluation.

10/1/2001

Cohort studies

40

41 of 79

Measures

Age, Gender, Race/ethnicity, Education
Alcohol, Cigarette smoking
Waist-to-hip ratio
Diabetes (fasting serum glucose >= 140 mg/dL or history of diabetes, insulin, or diabetes medication)
Plasma LDL & HDL cholesterol

9/30/2003

Cohort studies

41

42 of 79

Hypertension

Blood pressure measured as average of 3 sitting measurements with a random-zero sphygmomanometer, after 5 min. rest period
Hypertension if any of the following

Diastolic pressure >= 90 mm Hg
Systolic >= 140 mm Hg
Use within past 2 weeks of hypotensive medication

10/1/2001

Cohort studies

42

One of the most important of these other risk factors was hypertension, since elevated blood pressure is a key risk factor for CHD. In addition, blood pressure is strongly affected by the experience of anger. Blood pressure varies from moment to moment and in response to emotional stimuli, including emotions aroused when a health care provider measures someone’s blood pressure. The phenomenon of “white coat hypertension” refers to observation that some people’s blood pressure rises to hypertensive levels when their blood pressure is being measured by a clinician but stays at normal levels when measured in other circumstances (e.g., by the patient her/himself).

To reduce the influence of random variation in blood pressure measurements, for the study blood pressure was measured three times, with a five-minute rest period between each time and then averaged. A random-zero sphygmomanometer (blood pressure measurement instrument) was used to avoid the influence of “digit preferences” (in this case, the reduce the inclination to round to multiples of 5).

Blood pressure is a continuous variable – it can take on any of a range of values. But in this study the investigators wanted to categorize participants as “normotensive” or “hypertensive”, so they formulated a definition of “hypertension” for this purpose. Hypertension was coded if the diastolic pressure (the lower number) was 90 mm Hg or greater, the systolic pressure was 140 mm Hg or greater, or the participate had used hypotensive medication with the past two weeks.

43 of 79

Follow-up for events

Participants were followed from date of their first clinic reexamination in ARIC (1990-92) through December 31, 1995
Median 53 months, maximum 72 months
Abstraction of death certificates and hospital discharge records

10/1/2001

Cohort studies

43

44 of 79

“Incident CHD event”

1. acute myocardial infarction (MI) or fatal CHD (“hard” events)

2. cardiac revascularization procedure (percutaneous transluminal coronary angioplasty or coronary artery bypass graft surgery)

3. silent MI

10/1/2001

Cohort studies

44

45 of 79

Table 1. Distribution of Population Characteristics by Level of Trait Anger: ARIC Study, 1990 to 1992

10/1/2001

Cohort studies

45

The first data we typically want to know for a cohort study are the baseline characteristics of the cohort, by exposure status, as well as the distribution of exposure. Table 1 shows that the vast majority of the cohort have low or moderate anger scores – only 7.7% have high anger scores. Because of the size of this cohort, the number of participants with high anger levels (1,000) is adequate for analysis. However, the relative infrequency of high anger levels means that unless high anger is a very strong risk factor for CHD, this level of anger can at most contribute to a modest fraction of CHD cases.

We also see from the table that anger is related to younger age (mean age decreases with increasing anger scores), male gender (no surprise there), cigarette smoking (quite strongly, since only 18% of persons with low anger scores are current smokers, compared to 31% of those with high anger scores), and alcohol consumption. The relations of anger scores to education and to race are not so clear, since the percentages are not monotonic (i.e., consistently decreasing with higher anger scores or consistently increasing). Note that these are all prevalence data.

46 of 79

Table 1. Distribution of Population Characteristics by Level of Trait Anger: ARIC Study, 1990 to 1992

10/1/2001

Cohort studies

46

Additional data in Table 1 shows an association between trait anger and waist-to-hip ratio, a measure of fat distribution that is related to CHD risk. Other relations are less persuasive, either because it is the moderate anger category that has the highest or lowest value or because the differences in means are quite small. The p-values provide an indication of how readily differences of the magnitude observed would result from a completely random process. Even though it appears that the percentages of diabetics would be somewhat unlikely to occur by chance alone, the fact that the moderate anger category has a lower prevalence of diabetes than the low category makes one suspect that the pattern is nevertheless a chance finding. Alternatively, it could be that only high trait anger is related to diabetes, with the difference in prevalence between low (11.4%) and moderate (10.2%) anger levels being due to random variability.

47 of 79

From table 2. Hazard Ratios (95% CI) for association between trait anger and all CHD – Normotensives

10/1/2001

Cohort studies

47

The data for this study were analyzed and presented separately for normotensive participants and hypertensive participants, with hypertension defined as previously shown. Table 2 shows hazard ratios for the association between trait anger and all CHD events (i.e., both “hard” and “soft” events) in normotensives. The table had 95% confidence intervals, but these have been omitted from the slide to simplify it. (We will present confidence intervals in a future lecture.) The hazard is a term that statisticians and demographers use to refer to the same concept as incidence density or incidence rate.

From Table 2 we see that the majority of participants who had a CHD event (i.e., MI, CHD death, coronary bypass surgery, or electrocardiographic evidence of an MI but no symptoms) had moderate anger levels. That is not in itself surprising, since the moderate trait anger group is the largest, at least among the normotensives, shown here. But note that there are only 27 high anger participants with a CHD event, a fairly small number for an epidemiologic study. From the slide we see that the hazard ratio – the ratio of CHD incidence in an exposure group to the CHD incidence in the low trait anger group – is higher for moderate anger participants and highest for high anger participants. This observation holds for hazard ratios adjusted only for age differences among the anger groups as well as for hazard ratios adjusted for all of the CHD risk factors listed earlier.

48 of 79

From table 2. Hazard Ratios (95% CI) for association between trait anger and all CHD – Normotensives

2/15/2011

Cohort studies

48

Incidentally, I’m sure that there was a reason why epidemiologists felt the need to introduce the term “incidence density” when the biostatisticians were already using the term “hazard”, but please do not ask me what the reason was. That reminds me of a joke about epidemiologists and biostatisticians, but I had better not try to tell it here – remind me sometime, and I’ll send an e-mail.

Let’s practice our calculation of measures of incidence by calculating the cumulative incidence (CI) of a CHD event for participants in the three anger groups. Ideally all participants would have been followed for the same length of time, which we know was not the case. But let us ignore that for the moment and calculate the CI that each group experienced during the existing follow-up.

49 of 79

From table 2. Hazard Ratios (95% CI) for association between trait anger and all CHD – Normotensives

2/15/2011

Cohort studies

49

50 of 79

From table 2. Hazard Ratios (95% CI) for association between trait anger and all CHD – Normotensives

2/15/2011

Cohort studies

50

51 of 79

From table 2. Hazard Ratios (95% CI) for association between trait anger and all CHD – Normotensives

2/15/2011

Cohort studies

51

Did you get these values in your calculations? The calculation is simply number of persons with a CHD event divided by the person-time, repeated for each anger group.

How do these numbers compare to the CI’s we obtained earlier? Note that the CI’s are expressed as proportions, whereas the ID’s are expressed per 1,000 py. If we express the CI’s as cases per 1,000 persons, we see that the numerical value of each CI is 4.3-4.4 times the numerical value of the corresponding ID (e.g., 17 per 1,000 / 3.9 per 1,000 py = 4.36 years, approximately the length of the follow-up period. This makes sense, since if the incidence of CHD events is just under 4 per 1,000 persons each year, then in a median of 4.4 years of follow-up about 17 people should have a CHD event, which out of 3,110 people (see previous slide) is a CI of 0.017. When the incidence rate is low, the cumulative incidence will approximately equal the incidence rate multiplied by the follow-up time. The reason the equivalence holds is that when the incidence rate is small, the size of the population at risk does not meaningfully decrease. If the incidence rate is high, then the number of people at risk declines from one year to the next, so the same incidence rate will produce fewer cases in each successive year. The CI (cumulative incidence) will therefore be lower than the product of the incidence rate and follow-up time.

Now, how about the hazard ratios (IDR’s)? Let’s calculate them.

52 of 79

From table 2. Hazard Ratios (95% CI) for association between trait anger and all CHD – Normotensives

10/23/2012

Cohort studies

52

These were easy – we just divide one incidence density (a.k.a. [= also known as] hazard) by the value for the low anger group, 3.9 / 1,000 py. The results of our calculations differ only slightly from the age-adjusted hazard ratios in the table. We should have expected such a close correspondence since, after all, the mean ages of the three anger groups differed only slightly, so age adjustment would not make much difference. An exercise that I recommend to strengthen your understanding of adjustment is to explain why the multivariable-adjusted hazard ratio for the high anger group, in particular, should be lower than the crude hazard ratio (of course, in this case the ID’s were calculated using my estimates of person-years, so they are only approximations; but if we had the actual data, the multivariable-adjusted ratio should indeed be below the hazard crude ratio for the high anger group).

53 of 79

From table 2. Hazard Ratios (95% CI) for association between trait anger and all CHD – Normotensives

9/25/2002

Cohort studies

53

Even if one has to estimate some numbers, recalculating the figures in a paper and comparing the results to what the authors present is a valuable practice to make sure that one is indeed understanding the data – and occasionally to find errors that have slipped by the authors’ proofreading. I highly recommend doing so, both in EPID600 and when you want to use the results of a published paper or are reviewing a manuscript or a colleague’s work.

One final calculation is the incidence density difference (IDD), the difference in incidence rates between each of the higher anger groups and the lowest anger group. This is another way of comparing the incidence rates across the three anger groups. Whether we compare the incidence rates by taking ratios or by taking differences, of course, we see that indeed CHD incidence is higher among those with moderate trait anger and highest among those with high trait anger, with a difference of 5.8 cases per 1,000 py from the lowest to the highest anger group, a 150% increase.

54 of 79

Figure 1. CHD event-free survival probabilities among normotensive individuals by trait anger scores

10/1/2001

Cohort studies

54

High

Moderate

Low

Figure 1 presents Kaplan-Meier survival curves that compare the incidence of CHD in the three anger groups at each follow-up time. Note that follow-up time here is expressed relative to the visit 2 exam when this cohort was defined. Since participants had their visit 2 exams and their visit 3 exams at different times, their individual follow-up times would begin and end at different calendar times. However, most participants were followed for about the same length of time, something that is not always the case.

A survival curve is approximately the inverse of a graph of cumulative incidence at each point in time. A survival curve that consists of a horizontal line at the 1.00 level (the top) corresponds to a CI of 0. If we rotated the above graphs around a horizontal line at 0.5, we would see three cumulative incidence curves rising from the bottom left corner. (Note though that the vertical axis in this figure begins at 0.92 – the CI’s here are only a few percent). The Kaplan-Meier method is more sophisticated than the CI calculations we have done because it adjusts for the loss of people from the cohort at each point, e.g., due to non-CHD death.

What we see from the graph is that the curve for the high anger group drops below the other two curves almost from the beginning, and the gap continues to widen for some time. The greater CHD incidence for the high trait anger group is present at the start; the greater incidence for the moderate anger group emerges gradually.

55 of 79

From table 2. Hazard Ratios (95% CI) for association between trait anger and “hard” CHD – Normotensives

10/6/2009

Cohort studies

55

56 of 79

Relating risk factors to health outcomes – questions

Is this health condition associated with this exposure?

Association not = causation but may reflect it

How strongly are these two factors related?

Strong association more likely causal

How much of a disease can be attributed to a causative factor?

1/9/2007

Cohort studies

56

Now I would like to rehearse some of the basic concepts in defining and quantifying associations, both in regard to strength of association and the concept of “attributable risk”, or impact. This material more or less parallels the treatment in the chapter “Relating risk factors to health outcomes” in the Evolving Text.

There are three questions that arise in attempting to relate risk factors to health outcomes. The most basic one, of course, is whether or not a health condition is associated with an exposure. Association is not equivalent to causation, of course, but if there is no association between the two factors, than a causal relation must not exist.

If there is an association, then we usually ask how strong is it. Generally, stronger associations are more likely to be causal. It is much easier to come up with alternative explanations for weak associations than for strong ones.

Finally, if we are persuaded that a factor is a cause of a health outcome, or we at least want to entertain that possibility, we can ask the question how much of the outcome may be attributable to the causative factor.

57 of 79

What is an association?

Factors are associated if:

the distribution of one factor is different for different values of another.
knowing the value of one factor gives information about the distribution of the other.

9/27/2004

Cohort studies

57

58 of 79

Example – oral contraceptives and CHD

9/20/2000

58

59 of 79

Example – oral contraceptives and CHD (positive association)

9/27/2004

59

30% (30/100) of controls OC, overall

60% (30/50) of CHD cases used OC

If we compare the proportions (or percentages) of OC users for the women with CHD and the control group, we see that the percentage of OC users is greater among the women who developed CHD than it is in the control group. So the distribution of OC use (in this case, the distribution is simply the percentage who used OC and its inverse, the percentage who did not) differs according to the value of another factor (CHD). Knowing the value of CHD (i.e., CHD present or CHD absent) gives us information about the distribution of OC use. If we were betting that a randomly selected woman was an OC user, we would want different odds if we knew that she had CHD than if we knew she had not.

Association is reciprocal, so we can also describe the association in this table by talking about how the percentage of women who developed CHD differs by OC history. Note, however, that describing an association in a table is not the same as describing incidences or risks. For example, it is most unlikely that half of the women who used OC (30/60) developed CHD. Most likely these data (had I not made them up) would come from a study that began with CHD cases and a sample of non-cases, which is why such a large proportion of the women had developed CHD. This note will be more understandable after the next module, Case-control studies.

60 of 79

Example – oral contraceptives and breast cancer

9/27/2000

60

61 of 79

Example – oral contraceptives and breast cancer (no association)

9/27/2004

61

30% (30/100) of noncases used OC

30% (15/50) of cases used OC

62 of 79

Measures of association

Can compare incidences (rate or proportion), prevalences
Look at differences (e.g., “incidence difference”) (retains units)
Look at ratios (e.g., “incidence ratio”) (no units)

6/9/2002

Cohort studies

62

63 of 79

Translating measures of association

If incidence ratio for runners / non-runners = 3.0:

“Incidence in runners was 3 times that in non-runners.”
“Incidence in runners was 3 times as great as in non-runners.”
“Incidence in runners was 200% greater than incidence in non-runners.”� [(3.0 – 1.0) / 1.0 = 200%]

2/21/2006

Cohort studies

63

64 of 79

Translating measures of association

“Incidence in runners was 3 times greater than incidence in non-runners” is ambiguous

Does it mean incidence ratio = 3.0 ?
Does it mean incidence ratio = 4.0 ?

10/1/2001

Cohort studies

64

65 of 79

Translating measures of association

If incidence for runners / non-runners = 0.30:

“Incidence in runners was 0.30 times that in non-runners.”
“Incidence in runners was 30% of that in non-runners.”
“Incidence in runners was 70% lower [or “less”] than incidence in non-runners.”� [(1.0 – 0.30) / 1.0 = 0.70 = 70%]

10/1/2001

Cohort studies

65

66 of 79

Translating measures of association

Or, can say “Incidence in non-runners was 3.3 times as great as incidence in runners”.

10/1/2001

Cohort studies

66

67 of 79

Measures of impact

Concept of attributable risk

How much of a disease can be attributed to a causative factor?
What is the potential benefit from intervening to modify the factor?

Important for

Public health policy
Legal liability
Clinical/individual decisions

2/21/2006

Cohort studies

67

68 of 79

Example questions

Now that I am 35 years old, my CHD risk from taking oral contraceptives is twice as great as when I was 25. But how much more risk do I have due to taking the pill?
How much of the risk of heterosexual transmission of HIV might be eliminated through eliminating bacterial sexually transmitted diseases?

10/1/2001

Cohort studies

68

69 of 79

Example questions

How many cases of asthma are due to ambient sulfur dioxide?
What proportion of motor vehicular deaths can be prevented by mandatory seat belt use.
What proportion of perinatal HIV transmission has been prevented through the use of prenatal, intrapartum, and neonatal zidovudine (AZT)?

10/1/2001

Cohort studies

69

70 of 79

Simplifying assumptions

1. “Exposure” either causes or prevents the outcome, but not both (no “two-edged swords”)

2. “Exposed” and “unexposed” groups are alike in all other respects (no confounding)

3. No other causes “compete” with the exposure

10/1/2001

Cohort studies

70

In order to proceed it is helpful to assume away some complexities. First is the conceptual problem when an exposure, such as air bags, can cause harm as well as benefit. We will limit our discussion to exposures that are either harmful or preventive, but not both. If that is not satisfying we can consider only the net effect (excess of harm or excess of benefit).

Second, we assume that there are no differences between exposed and unexposed persons that we need to take account of (e.g., age). Such other causes of the outcome could be confounding factors, so we assume them away (we will take up confounding in a later module).

Finally, we assume that no other causes of the outcome “compete” with the exposure we are examining. After all, if we remove the exposure of concern only to have the people develop the outcome due to some other exposure, then the question of attributing responsibility for an outcome is a more complex one.

It will also simplify matters if we treat incidence risk and rate interchangeably, which we can do if the outcome is rare.

71 of 79

Several concepts

Concepts

“Absolute” versus “relative”
Exposed versus total population
Disease caused, disease prevented

10/1/2001

Cohort studies

71

Having cleared the arena, we can now play with three concepts. The first one refers to whether we are taking an “absolute” perspective (i.e., looking at an amount of risk or a caseload) or a “relative” perspective (i.e., looking at a proportion of the existing caseload or risk).

The second concept relates to our frame of reference – are we considering only cases among people exposed to the factor of concern or cases in general. Cases in people who are not exposed to a factor cannot be attributed to it (unless there is “second-hand” exposure, as with environmental tobacco smoke). So the amount of risk or proportion of risk attributable to an exposure will necessarily be lower in the total population than in people exposed to the factor.

The third concept relates to whether we are interested in the amount of disease for which a harmful exposure is responsible or the amount of potential disease that a preventive measure can avoid.

72 of 79

Many terms, many meanings

E.g., “attributable risk” can mean:

Risk difference
Population attributable risk percent
Concept of assessing impact

10/1/2001

Cohort studies

72

73 of 79

“Absolute” perspective

How much risk?

In exposed persons:

risk difference (R₁– R₀)

In the total population:

(R₁– R₀) x exposure prevalence (p₁)

How many cases?

(R₁– R₀) x # of exposed persons (n₁)

2/15/2011

Cohort studies

73

Here are several measures of impact (“attributable risk”) based on an absolute perspective. If we want to answer the question, “Among people exposed to X, how much risk do they incur as a result of that exposure”, we can estimate that amount by taking the difference in their estimated incidences. If we can regard the incidence in the unexposed as indicating what the risk in the exposed would be without their exposure (remember the concept of a substitute population?), then the difference between incidence in exposed persons (R₁) and incidence in unexposed persons (R₀) should estimate the amount of risk that the exposure is responsible for. In this approach we consider R₀ as the baseline risk (e.g., CI), so any increment above that is presumably due to the exposure.

If instead of restricting our attention to the exposed group only, we consider the entire population, then this attributable risk estimate must be multiplied by the prevalence (p₁) of the exposure in the population. Even if an agent has a powerful effect in the people exposed to it, on a population level that effect will be muted if only a small proportion of people are in fact exposed.

We might also be interested in the projected number of cases due to the exposure (or the potential number of cases that could be avoided if we could eliminate the exposure before the disease process began). That estimate is simply the attributable risk multiplied by the number of people.

74 of 79

Relative perspective

What proportion of the risk is attributable?�(What proportion of cases could be eliminated?)

In exposed persons: (R₁– R₀) / R₁ = (RR–1) / RR

(Relative strength of association)

In the population: (R – R₀) / R

(Strength of association and prevalence)

2/15/2011

Cohort studies

74

The relative perspective, in contrast, expresses the amount of risk or number of cases attributable as a proportion of the total risk or total number of cases. So here the question is what proportion of risk or cases could be eliminated if we eliminated the exposure before the disease process began.

Again, we can estimate this proportion for the exposed group only or for the total population. Among the exposed group, the proportion of cases attributable, using the same logic of a substitute population, is directly related to the relative strength of association. In the slide, the proportion of incidence attributable to the exposure is estimated by expressing the difference in incidence relative to the incidence in the exposed (“what proportion of the incidence in the exposed is attributable to their exposure”). An algebraically equivalent expression is shown to the right, where RR stands for the risk ratio. (Recall that RR = R₁/R₀)

The proportion of all (exposed and unexposed) cases attributable, however, will also reflect the proportion of the population that is exposed, i.e., the prevalence of exposure. To estimate this proportion we can take the amount of risk attributable to the exposure (overall incidence less the background incidence) and divide it by the overall incidence. There is an algebraically equivalent formula but it did not fit on the slide.

75 of 79

How much risk? What %?�How many cases? What %?

2/15/2011

Cohort studies

75

76 of 79

How much risk? What %?�How many cases? What %?

2/15/2011

Cohort studies

76

Making our simplifying assumptions, we can suppose that if the exposed group were not in fact exposed (the counterfactual) then the risk in the exposed group and in the total cohort would have been the same as what we observed in the unexposed, 0.01 or 10 per 1,000. In that case, the amount of incidence attributable to exposure would be 0.04 – 0.01 = 0.03 in the exposed group and 0.02 – 0.01 = 0.01 in the total cohort. These absolute incidence differences can also be expressed in terms of attributable cases by multiplying by the number of people in the exposed (1,000) by 0.03 or the number in the total cohort (3,000) by 0.01, yielding 30 cases. Cases occurring in the unexposed can not, by definition, be attributed to the exposure, so all attributable cases in the total cohort are those in the exposed group. The proportion of incidence attributable in the entire cohort is (0.02 – 0.01) / 0.02 = 0.50, or 50% or, equivalently, 30 attributable cases / 60 total cases.

77 of 79

Attributable risk diagram

2/15/2011

Cohort studies

77

Attributable cases

A number of years ago Ray Greenberg and I came up with the diagram above as a way of illustrating the concept of attributable risk and deriving the formulas for it. Let R represent risk in exposed (subscript 1) or unexposed (0) persons and n represent the number of people exposed (subscript 1) or unexposed (0).

If the risk in the unexposed is demarcated by the horizontal line just above R₀, then the area in the left-most rectangle represents the number of unexposed cases, R₀ x n₀. The number of exposed cases is, similarly, R₁ x n₁, the area of the taller rectangle, on the right. If the exposure had no effect, then we expect the exposed group to have the same risk as the unexposed group, so that the number of cases would be that risk (R₀) times the number of exposed persons (n₁). On that basis, then cross-hatched rectangle in the upper right represents the attributable cases, those that occurred above and beyond the background risk.

We can now express the attributable cases as a proportion of all exposed cases (R₁ x n₁) or as a proportion of all cases (R₀ x n₀) + (R₁ x n₁), yielding, respectively, the attributable risk proportion and the population attributable risk proportion. With algebra it is possible to derive the various formulas from this simple foundation (if you are interested, the algebra is in the Evolving Text).

78 of 79

Prevented fraction diagram

2/15/2011

Cohort studies

78

Prevented (potential) cases

A similar diagram can be used to illustrate the prevented fraction, which is used when we have a preventive exposure instead of a harmful one. The prevented fraction is the proportion of actual plus potential cases that were prevented by the exposure. Again, this fraction can be estimated for the group that is exposed to the preventive factor or for the entire population (why do we not estimate the prevented fraction for the group unexposed to the preventive factor?)

Note here that since the exposure is preventive, the risk in the unexposed is greater than the risk in the exposed. Note also that since, by definition, the prevented cases did not actually occur, the numerator for the preventive fraction consists of potential cases, and the denominator consists of potential cases as well as actual cases.

So the prevented fraction among those exposed to the preventive exposure is simply the prevented cases divided by the prevented plus actual cases in the exposed = (R₀ – R₁) n₁ / (R₀ n₁), which also equals (R₀ – R₁) / R₀ and 1 – RR, where RR is the risk ratio (e.g., CIR).

The prevented fraction in the population is those same prevented cases divided by all actual and potential cases, (R₀ – R₁) n₁ / (R₀ n), where n is the total population, n₀ + n₁. This can also be written as (1 – RR) x prevalence.

79 of 79

Empréstimo de amigos

O sujeito chega para o amigo e dispara:

– Ô João, me empresta mil reais?

– Não posso, só tenho setecentos.

– Não tem problema, você fica me devendo trezentos.

De Alfredo Inácio Junior, São Paulo en Bom Humor Nosso E Dos Leitores”, Almanaque Brasil de Cultura Popular. Maio 2001;3(26) (almanaquebrasil@uol.com.br). Exemplar de quem viaja TAM.

10/1/2001

Cohort studies

79