This preview shows page 1. Sign up to view the full content.
Unformatted text preview: Cohort Studies
Readings: Chapter 12, Oleckno Prospective Study of Serum Vitamin D and Cancer Mortality in the United States. Freedman DM et al. JNCI 2007;99:1594-602. EPID 301 Kristan Aronson March 11, 2009 Study Designs Cohort Study: Learning Objectives Understand: Source population Exposure measurement Outcome measurement Effect Measures Describe: Strengths Weaknesses Cohort Study Ancient Roman military unit of 300-600 men
A group of soldiers marching on...
Any designated group of persons who are followed or traced over a period of time
(JM Last, Dictionary of Epidemiology) Cohort Study Analytic study Also called concurrent, follow-up, incidence, longitudinal and prospective/retrospective studies Key components: Source population defined by exposure status Observation over time (commonly years) Comparison of incidence rate of disease or other outcome Design
Develop Outcome Exposed Do Not Develop Outcome Source Population At risk Develop Outcome Unexposed Do Not Develop Outcome not at risk Period of Follow-up Design: Prospective
Exposed Do Not Develop Outcome Source Population At risk Develop Outcome Unexposed Do Not Develop Outcome Develop Outcome not at risk Period of Follow-up Design: Retrospective
Develop Outcome Exposed Do Not Develop Outcome Source Population At risk Develop Outcome Unexposed Do Not Develop Outcome Period of Follow-up not at risk When to conduct a cohort study Reasonably strong rationale for suspecting association between exposure and outcome Most efficient if short time interval between exposure and outcome development Want to study several exposures/outcomes Classic example: Framingham Heart Study Began in 1948 Framingham, Massachusetts Eligibility Criteria: residents, aged 30-62 years, free from CVD Study population = 5,127 men and women Examination every 2 years, with daily surveillance of local hospital Multiple exposures including smoking, obesity, high blood pressure, physical activity, and alcohol intake Primary Outcome = CVD 1288 publications generated 1951-2004 http://www.nhlbi.nih.gov/about/framingham/ Our example:
Prospective Study of Serum Vitamin D and Cancer Mortality in the United States. Freedman DM et al. JNCI 2007;99:1594-602 Objective: to study the relationship between serum 25(OH)D and cancer mortality Data from the Third National Health and Nutrition Examination Survey (NHANES III) in the United States Representative sample of non-institutionalized population N=16,818 >=17 years old Enrolled 1988-94 followed through 2000 Source/ Target Population Population from which study sample is drawn -Eligibility Criteria (demographic characteristics,
specific time-frame for recruitment, exposure status etc.) [External population/ External validity: The degree to which the study finding is relevant to the wider population (beyond the target population)]
External Population Target Population Study Sample Source Population: Required Characteristics: Large number exposed Ability to measure exposure accurately Appropriate "unexposed" group available Participants are all free from the outcome at baseline Source Population: Selection Community Based: Representative of General Population (e.g. Framingham Heart Study) Specific Factor: Exposure (e.g. workers in particular occupation / industry) High risk group Enumerated professional groups/societies (e.g. nurses, physicians, civil servants) Geographic area/facility (e.g. school, healthcare plan, hospital clinic, military service) Cohort Study Design
Sources for Study Subjects General population sample Occupational group Members of a health plan Military, church groups, schools... Selection considerations: Representativeness Frequency of exposure Ability to follow Ability to get accurate information Logistics/access/con Cohort Selection
Internal Comparison Group Unexposed portion of the cohort Variation in amount of exposure: the lowest exposure group becomes your baseline comparison Dose-response gradient assessment possible In our example, serum vit D levels were measured and categorized into 6 groups Internal Comparison Group Selection Select source population Framingham residents Determine exposure status Questionnaires, physical examination, Lab tests Cohort divided into exposure categories Non-exposed become internal comparison Multiple levels of exposure constructed e.g. Nondrinkers, <5g/day, 5-<15 g/day, >15 g/day Common to use quantiles/tertiles and use extremes as comparison group Cohort Selection
External Comparison Group Needed when everyone is exposed, or don't have info on variability in exposure (ex: occupation) Must be as similar as possible to the exposed group on other risk factors for the outcome Multiple comparison groups can increase validity of the study, but also has cost External Comparison Group Selection If everyone in cohort exposed e.g. persons exposed to radiation after Chenobyl Workers (but then have potential healthy worker effect) Select separate comparison group General population From similar group Must be as similar as possible (e.g. age, gender, SES) to exposed group but without exposure Sometimes both internal and external comparisons Minimize healthy worker effect Accuracy of a Study Findings Validity
Valid: sound or defensible; well-grounded
(Oxford dictionary) Validity in epidemiologic studies: - internal validity: the extent, within a given study that the results reflect the truth - external validity (generalizability): the extent to which the study findings can be applied to people outside the study population Internal Validity Selection Bias Information Bias Confounding Bias and confounding are types of systematic error in the design, conduct or analysis of an epidemiologic study that results in a parameter estimate that does not represent the true effect in the target population Selection Bias: Definition Error due to systematic differences in characteristics between those who are selected for (and retained in) the study and those who are not When characteristics of those chosen for the study differ systematically from those in the target population When the study and comparison groups are selected from different populations Primarily occurs during the design phase of a study Bias: Selection Bias Non-participation/Selection Must be related to both exposure and disease for selection bias to exist in cohort study Best example: Healthy worker effect Loss to follow-up If those exposed and with outcome are more likely to drop out Try to minimize/ assess impact
STRENGTH AND DIRECTION? Source Population: Selection Comparison / Unexposed group Internal External Both internal and external Eligibility Criteria Inclusion Exclusion Selection Bias: Example Loss to Follow-up Cohort study examining the association between dust exposure and chronic lung disease
Lung Dx No Lung Dx 9,000 9,500 Total Truth Dust 1,000 10,000 10,000 No Dust 500 RR = 1,000/10,000 = 2 500/10,000 Selection Bias: Loss to Followup
How will this loss to followup effect the RR?
Lung Dx Dust No Dust 1,000 (10%) 500 (2%) Lung Dx Dust 900 No Lung Dx 9,000 (1%) 9,500 (0%) No Lung Dx 8,910 9,500 Total 10,000 10,000 Total 9,810 9,990 No Dust 490 RR = 900/9810 / 490/9990 = 1.85 Loss to follow-up bias Occurs when there are a lot of losses to follow-up in an ongoing cohort study Final population may be systematically different than the one you started with regarding exposure status and outcome susceptibility Longer the study duration, more likely loss to follow-up will be a problem Often largest part of expense for a cohort study Strategies include: keep losses low (<20%) and compare characteristics of those lost to those who remain to help you interpret your results Measuring Exposure Depends on: Type of cohort Source population Characteristics of exposure Presence Duration Intensity Measuring Exposure Method: Questionnaire Existing records Laboratory tests Physical measurements Clinical examinations / procedures Repeated measures? Determining Exposure Status Medical records Employment records Birth records Questionnaires Interviews Physical examination Body specimens (blood, urine) Environmental tests Exposure Assessment
Exposure may change during the follow-up Person-years analysis can help with this Choice of method depends on logistics, completeness and accuracy Often larger studies have poorer exposure assessment trade-offs ("data thin", "data thick") Exposure Assignment:
Prospective Study of Serum Vitamin D and Cancer Mortality in the United States. Freedman DM et al. JNCI 2007;99:1594-602 Levels of serum 25(OH)D were assayed at start of study Categorical cut-points based on insufficiency values for three lower categories and on observed distribution for highest three categories Measured in November March in southern latitudes and in April October in northern latitudes due to mobile van data collection being sensitive to cold weather Any concerns? Induction Time and Latency Timing of exposure in relation to disease occurrence should be addressed Disease occurrence during the induction period should not be assigned to the exposure Latency (time before disease that is present can be detected) needs to also be considered Induction and latency times may not be known, in which case multiple induction times may need to be tested Latency Consideration:
Prospective Study of Serum Vitamin D and Cancer Mortality in the United States. Freedman DM et al. JNCI 2007;99:1594-602 Average follow up was 8.9 years Mean age at entry was about 43 years old Any concerns? Internal Validity Selection Bias Information Bias Confounding Information Bias Synonyms: measurement bias, observation bias Error in classifying subjects with regard to their exposure or outcome status OR Differential quality (accuracy) of information between compared groups Primarily occurs during the data collection phase of a study
Information Bias MISCLASSIFICATION OF EXPOSURE MISCLASSIFICATION OF OUTCOME Measurement error: Recall bias Interviewer bias Exposure suspicion bias Data entry error Measurement error : Diagnostic bias Data entry error Information Bias: Misclassification Epidemiologists recognize that measurement error often exists, the question is: Is the misclassification differential? degree of error varies between comparison groups Is the misclassification non-differential? degree of error uniform between comparison groups Why is it important know? Information Bias: Misclassification Differential misclassification can bias the study result in either direction: an overestimate or underestimate of the true effect Non-differential misclassification biases the study result toward the null value of no association when the exposure is dichotomous
Dichotomous: two categories ex. yes/no, high/low Minimizing Information Bias Study design must be carefully planned including consideration of bias Data collection must be standardized Objective data is more valid than subjective Blinding of those collecting data to study's purpose and subject's status Determining Outcome Status Medical records Disease registries Death certificates Questionnaires Interviews Physical examination Diagnostic tests Determining Outcome Status Blinded, standardized physical exam best Accuracy of medical records varies depending on clarity of diagnostic criteria Self-reports usually need validation from another source Assessment should be equally applied to exposed and unexposed groups Example: more checkups in an occupational cohort than in the general population external comparison group Outcome Measurement:
Prospective Study of Serum Vitamin D and Cancer Mortality in the United States. Freedman DM et al. JNCI 2007;99:1594-602 "Follow-up of the cohort continued from data collection until December 31, 2000, and was based on the NHANES III Linked Mortality File (with 113 underlying causes of death) which had been derived through probabilistic linkage with the National Death Index." What else might you want to know? Differential Misclassification If the exposed group is more (or less) likely to be mistakenly classified as having developed the outcome than the unexposed group Any concerns in our study? If so, what would be the direction of the bias in our effect estimate? Non-Differential Misclassification There is some degree of error in classifying the outcome, but it is equally frequent in both the exposed and unexposed groups Any concerns in our study? If so, what would be the direction of the bias in our effect estimate? Interviewer Bias Systematic differences in how data are collected across exposure groups In our example, some data on confounders were collected by interview race/ethnicity, education, smoking, calcium intake, physical activity Any concerns? Diagnostic Suspicion Bias Occurs in cohort studies when knowledge of the subjects exposure status leads to systematic differences in the procedures for diagnosing the outcome Any concerns? Measuring Outcome Method must be similar for both exposed and unexposed Outcome of interest must be clearly defined a priori Diagnostic Criteria Sub-types of disease Measuring Outcome Mortality Death certificates, autopsy report, medical records, or family Hospital records Physician records Absenteeism from school/employment Self-report Diary Telephone survey Laboratory tests Physical examination Clinical measures/procedures Overview of Cohort Studies Distinguishing feature: Cohort is defined on basis of exposure; exposure is assessed before the outcome: temporality logical Subjects are followed over time and incidence of the outcome is measured Many advantages. Considered strong evidence of possible cause-effect relationships Often expensive and time-consuming Analysis Based on Incidence Proportion If you recruit all subjects at one point in time and follow them for the same amount of time and calculate the event rate, you have measured an incidence proportion Recall that with incidence proportion estimates, you can calculate Risk Ratio or Risk Difference as the effect estimates Odds ratio is sometimes calculated for convenience (ex. when using logistic regression) Analysis Based on Incidence Rates Used when your study subjects are followed for unequal amounts of time and/or you allow subjects to move in and out of your exposure category Recall that with person-time incidence estimates, you can calculate Rate Ratio or Rate Difference as the effect estimates Survival analysis methods often used, which calculate a hazard ratio (ex. when using Cox proportional hazards regression) Analysis in our example:
Prospective Study of Serum Vitamin D and Cancer Mortality in the United States. Freedman DM et al. JNCI 2007;99:1594-602 Investigators used a form of multivariate regression that takes person-time into account ("Cox proportional hazards") They do not provide us with incidence rates in each exposure group They use the generic term "relative risk" to represent what they actually measured (hazard ratio) Analysis in our example:
Prospective Study of Serum Vitamin D and Cancer Mortality in the United States. Freedman DM et al. JNCI 2007;99:1594-602
Serum 25(OH)D (nmol/L)* Total cancer mortality RR 95% CI <50 50 < 62.5 62.5 < 80 80 < 100 100 <120 >=120 1.0 1.22 1.02 1.00 0.92 1.49 0.91, 1.64 0.69, 1.50 0.71, 1.40 0.58, 1.46 0.85, 2.64 p-value (trend test) = 0.65 *adjusted for age, race/ethnicity, sex, smoking What do these results tell us? Overview of Observational Cohort Studies Distinguishing feature: exposure is assessed before the outcome. Subjects are followed over time and incidence of the outcome is measured Many advantages. Considered strongest observational evidence. Often (but not always) expensive and time-consuming. Internal Validity Selection Bias Information Bias Confounding Confounding A key concept in epidemiology We need to assure ourselves that the differences we are seeing are due to the factor we are interested in (risk factor) and not some other factor that is associated with the disease and the risk factor. In the example we've just seen, age is a confounding factor for studying the differences in cancer rates between the 1940 and 1980 US populations Confounding
Definitions Positive confounding: crude effect is more extreme than adjusted effect Example: Crude RR = 5.0 Adjusted = 3.5 Negative confounding: crude effect is less extreme than adjusted effect Example: Crude RR = 3.5 Adjusted = 5.0 Useful to think about direction of effect of a particular confounder if it was not considered in the study Controlling Confounding at the Design Stage Restriction: limit study subjects to one category of the potential confounder But can interfere with external validity (generalizability) Matching: study groups have similar distribution of the potential confounder Most often used in case-control studies we'll revisit when we cover that design Random allocation: assignment of exposure status Can ensure balance of known and unknown confounders Controlling Confounding at the Analysis Stage
Requires collecting data on potential confounders usually defined as known risk factors for the disease Stratification Standardization (as in our age-adjustment example) Mantel-Haenszel procedure (weighted average of stratum-specific RR) Multivariate analysis Simultaneously control for more than one confounder Use of various mathematical models depending on the nature of the data Confounding Latin confundere: to mix together Definition: An exposure between and exposure and outcome are distorted by an extraneous third variable (referred to as the confounding factor) Recall previous lectures on age standardization and causation Bias: Confounding "Distortion of the estimated effect of an exposure on an outcome, caused by the presence of an extraneous factor associated both with the exposure and the outcome" (Last) Must try to eliminate/control Measurement methods same as exposure Take into account in analysis Confounding
For a variable to be a confounder it must be .....
1. Associated with exposure 4. Distribution varies between study group and comparison group Confounder 2. Independent risk factor for the outcome Exposure
3. Not an intermediate factor on the causal chain Outcome Bias: Confounding e.g. Coffee, Caffeine and Cardiovascular Disease in Men (Grobbee et al NEJM 1990: 323 1026-32) Prospective cohort of 45,589 men in USA Adjusted RR = 1.04 (0.74 to 1.46) Discussion: Coffee drinkers twice as likely to smoke ? Confounding in previous studies Minimizing Confounding
Design Stage Restriction: limit study subjects to one category of the potential
confounder Problems with this strategy? Matching: study groups have similar distribution of the potential confounder Matched pairs each case matched with one or more controls by value of confounder Frequency matching controls picked to have same distribution of confounder that is observed in the cases Most often used in case-control studies Random allocation: assignment of exposure status in randomized clinical trials Benefits? Problems? Controlling Confounding
Analysis Stage Stratified analysis Analogous to examination of specific rates Study parameter calculated for each strata of the confounder Multivariate analysis Can simultaneously control for more than one confounder Use of various mathematical models depending on the nature of the data Data Analysis Effect Measures Incidence Relative Risk Attributable Risk Statistical Measures Confidence Intervals p-value 2 x 2 Table Analysis
Disease Exposed Unexposed No Disease A C A+C B D B+D A+B C+D N Incidence rate in exposed = A/A+B Incidence rate in unexposed = C/C+D Relative Risk = (A/A+B)/(C/C+D) Relative Risk (RR) e.g. study of the association between smoking and lung cancer Lung Cancer Yes Smoked Cigarettes Yes No Total 432 27 459 No 229,355 142,078 371,433 Total 229,787 142,105 371,892 Incidence rate in the exposed = 432/229,787 = 188/100,000 Incidence rate in unexposed = 27/142,105 = 19 / 100,000 RR = (188/100,000)/(19/100,000) = 9.9 Persons who smoke cigarettes are nearly 10 times as likely to develop lung cancer Accuracy of a Study Findings External Validity Synonym: generalizability Definition: The degree to which the study finding is relevant to the wider population (beyond the target population) External validity can only be judged after you are assured you have an internally valid result External Population Target Population Study Sample External Validity: Example
To whom would you generalize the results of a study that evaluates the efficacy of a specific treatment in elderly Caucasian men with coronary heart disease?
Hispanic women with CHD? Caucasian women with CHD? Hispanic men with CHD? Middle-aged caucasian men with CHD? Summary: Validity Accuracy of a study depends on valid and precise findings Bias and confounding are types of systematic error that threaten the internal validity of epidemiological studies Two major categories of bias: selection and information Confounding is systematic error in the effect estimate that results from mixing of the effects of two or more variables External Validity: The applicability of the results to the target population Internal and external validity are not all-or-none, black-and-white, present-or-absent dimensions of an epidemiological study Advantages of cohort study Temporality: exposure is known to precede outcome Useful for studying rare exposures Can measure incidence of outcome in exposed and unexposed Can study multiple outcomes Minimizes recall bias Advantages relative to C/C studies
Selection bias minimized - systemic difference in the relationship under study between those selected for the study and those not. For this to occur bias in selection has to be related to both exposure and outcome. Information bias differential recall/reporting/recording/abstraction of exposure information according to exposure and outcome status Measurement of exposure easier avoids having to recall exposures many years in the past. Biomarkers of exposure can be taken at temporally relevant time. Disadvantages Tends to be expensive and labor intensive Large sample size needed Long follow-up period usually needed Inefficient to study rare or delayed outcomes Summary: Cohort Study Compare incidence of disease (outcome) between exposed and unexposed groups Conducted over time Allows inferences regarding possible causal association Nurses Health Study Established in 1976 Eligibility Criteria: married registered nurses, aged 3055 years, resident in 11 most populous states in USA Study population = 121,700 women Questionnaire every 2 years Multiple exposures: oral contraceptives, estrogen therapy, smoking, nutrition & diet, Quality of Life Multiple Outcomes: Cancer, CHD NHS II began in 1989 Registered Nurses 25-42 years 116,686 enrolled http://www.channing.harvard.edu/nhs/ The Whitehall Study Began in 1967 18,000 men working in British Civil Service Social and occupational influence on physical and mental health, and mortality Whitehall II http://www.ucl.ac.uk/whitehallII/history.htm 1985-2004 10,308 non-industrial civil servants aged between 35 and 55 Retrospective Cohort 10% Labour force sample: Overview (Howe, Newcombe, Fair, Lindsay; Aronson et
al.) Objective: develop a monitoring system to detect previously unsuspected associations between occupations and specific causes of death National employment surveys of about 10% of the Canadian labour force between 1965 and 1971 linked with records on deaths occurring from 1965 to 1991 10% Labour force sample: Methods Report in 1999 (Aronson et al.) based on 457,224 working men and 242,196 working women Exposure defined as employment in an occupation for at least 1 year Link to National Mortality Database Personal identifiers available in both databases: first name, middle name, surname, complete date of birth, sex, and usually SIN Codes for three ICD versions were converted into 70 classifications for cause of death: investigated in relation to about 670 occupational titles 10% Labour force sample: Results Average duration of follow-up: 26 years Men: > 11 million person-yrs; Women: > 6 million person-yrs 116,000 deaths (32,500 cancer) among 457,224 men 26,800 deaths (8,134 cancer) among 242,196 women To avoid publication bias, all (about 26,000) findings published
Aronson KJ, Howe GR, Carpenter M, Fair ME. Occupational Surveillance in Canada: CauseSpecific Mortality Among Workers, 1965-1991. Ottawa: Statistics Canada (CD-ROM Product 84-546-XCB), 2000. Design retrospective cohort study of fire fighters: Steps to consider
Objective Source/ Target Population Cohort selection/ minimize selection biases Define exposure(s) and outcome(s) / minimize measurement biases Length of follow-up Consider confounders ...
View Full Document
This note was uploaded on 02/23/2012 for the course EPID 301 taught by Professor Richardson&aronson during the Spring '09 term at Queens University.
- Spring '09