1. Chapter 7 Design Strategies and Statistical Methods in Analytic Epidemiology

2. Objectives Define analytic epidemiology Distinguish between observational and experimental analytic epidemiologic studies Define case-control and cohort studies, and identify their distinctive features, strengths, and weakness Identify appropriate measures of association in case-control and cohort studies

3. Objectives (cont’d) Identify common measures used in epidemiology for describing cohort data Identify potential biases in case-control and cohort studies Identify ways to control for biases in case-control and cohort studies at the design and analysis levels Distinguish between effect modification and confounding

4. Analytic Epidemiology An analytic study attempts to answer why and how a health-related state or event occurred Tests specific a priori hypotheses Comparison group

5. Analytic Observational vs. Experimental Studies In analytic observational studies, researchers observe relationships between variables In analytic experimental studies, a portion of the participants are assigned the intervention Experimental studies are the focus of the next chapter

6. Analytic Observation Studies Case-control studies Cohort (prospective and retrospective) studies These studies may be exploratory or analytic (specific a prior hypothesis)

7. Case-Control Study Retro spicere means to look back

8. Case-Control Study The outcome is always identified prior to the exposure 1. Identify cases (persons experiencing a health-related state or event) 2. Identify controls (similar except not cases) 3. Investigate whether the cases are more or less likely than controls to have had past experiences, lifestyle behaviors, or exposures

9. Selection of Cases Establishing the diagnostic criteria and definition of disease is the first step in conducting a case- control study

10. Selection of Cases May consist of new cases (incidence) that show selected characteristics during a specific period, in a specified population, and a particular area Cases may also consist of existing cases at a point in time (prevalence)

11. Sources of Cases Cases come from Records from public health clinics Physician offices Health maintenance organizations Hospitals Industrial and government sources Cases should be representative of all persons with the disease

12. Selection of Cases Sampling Representation requires random selection with a sufficiently large sample size Restriction May improve validity May limit generalization

13. Selection of Controls Control subjects should look like the case subjects with the exception of not having the disease An epidemiologic assumption is that controls are representative of the general population in terms of probability of exposure and that controls have the same possibility of being selected or exposed as the cases

14. Selection of Controls General population Hospital Family, friends, relatives

15. Example

16. Exposure Status Information about exposure status may be obtained through Medical records Interviews Questionnaires Surrogates, such as spouses, siblings, or employers

17. Measuring the association between exposure and outcome variables The appropriate measure of association to use depends on the nature of the data When exposure and outcome variables are dichotomous (two-level nominal data) Odds ratio – use with case-control study Risk ratio – use with cohort study Rate ratio – use with cohort study

18. Example

19. Measuring association between exposure and disease in case-control studies

20. Bias Systematic error in the collection or interpretation of epidemiologic data Results in inaccurate (over or under) estimation of the association between exposure and disease Avoiding bias at the design stage of a study is paramount because of the difficulty identifying and accounting for it later

21. Selection Observation Recall Interviewer Types of bias in case-control studies

22. Selection Bias Cases and controls selected into the study is based in some way on the exposure The relationship between exposure and disease among participants in the study differs from what the relationship would have been among individuals in the population of interest

23. Selection Bias Berkson’s bias Prevalence-incidence bias (also called Neyman bias)

24. Observation Bias Recall bias – differential accuracy of recall between cases and controls Interviewer bias – interviewer probes cases differently than controls

25. Misclassification Differential Non-differential

26. Confounding Occurs when an extrinsic factor is associated with a disease outcome and, independent of that association, is also associated with the exposure Exposure Outcome Confounder

27. Confounding Coffee Heart Disease Smoking

28. Example OR =

29. SmokersNon-smokers Example

30. Controlling for Bias Matching – a strategy for controlling confounding at both the design and analysis levels of a study

31. Example

32. Case-Crossover Study Design Compares the exposure status of a case immediately before its occurrence with that of the same case at a prior time The case-crossover study design is especially appropriate where individual exposures are intermittent, wherein the disease occurs abruptly and the incubation period for detection and the induction period are short

33. Nested case-control study design Also called a case-cohort study A case-control study “nested” within a cohort study

34. Cohort Studies In epidemiology, it generally refers to a group of persons being studied who were born in the same year or period As time passes, the group moves through different and successive periods of life; as the group ages, changes can be seen in the health and vital statistics of the group

35. NEJ 1989;321:1285-9 We examined the mortality from breast cancer in a cohort of 31,710 women who had been treated for tuberculosis at Canadian sanatoriums between 1930 and 1952. Women exposed to ≥10 cGy of radiation had a relative risk of death from breast cancer of 1.36, compared with those exposed to <10 cGy. The risk was greatest among women who had been exposed to radiation when they were between 10 and 14 years of age…. With increasing age at first exposure, there was substantially less excess risk, and the radiation effect appeared to peak approximately 25 to 34 years after the first exposure. We conclude that the risk of breast cancer associated with radiation decreases sharply with increasing age at exposure.

36. Cohort Studies Cohorts of persons placed in a group can be studied as a group, forward in time (prospectively) or backward in time (retrospectively)

37. Prospective Cohort Study The predictor variable is measured before the outcome has occurred

38. Retrospective Cohort Study A historical cohort is reconstructed with data on the predictor variable (measured in the past) and data on the outcome collected (measured in the past after some follow-up period)

39. Cohort Effect Also called generation effect The change and variation in the disease or health status of a study population as the study group moves through time Cohort effects include any exposure or influence, from environmental effects to societal changes

40. Common measures used in epidemiology for describing cohort data Cumulative incidence rate – attack rate Incidence density rate – person-time rate

41. Measures of association in cohort studies Ratio of attack rates Risk ratio Ratio of person-time rates Rate ratio

42. Example

43. Equations Based on 2x2 Table

44. Expressing RRs as Percentages We can also express these RRs as percent change RR > 1 % Increase Change = (RR – 1) × 100 RR < 1 % Decrease Change = (1 – RR) × 100

45. When does the odds ratio approximate the risk ratio? For health-related states or events that are rare (i.e., affecting less than 10% of the population), a + b can be approximated by b, and c + d can be approximated by d

46. Example Portion of Table 7-5 Commonly Used Epidemiologic Measures for Describing Cohort Data

47. Incidence rate per 100,000 persons-years of cardiovascular disease among current smokers is 399, among non-current smokers is 356, and overall is 379. The RR is 1.122—male current smokers are 1.122 times (or 12.2%) more likely than non-smokers to develop cardiovascular disease Example

48. Attributable Risk The AR is 43 (399 – 356) per 100,000 Interpretation The excess occurrence of cardiovascular disease among male smokers attributable to their smoking is 43 per 100,000

49. Attributable Risk Percentage (AR%) The AR% equals 10.9% [(1.122-1) ÷ 1.122 × 100] Interpretation If smoking causes cardiovascular disease, nearly 10.9% of cardiovascular disease in males who currently smoke is attributed to their smoking

50. Population Attributable Risk The PAR for our example is 23 (379 – 356) per 100,000 Interpretation If current smoking were eliminated from the population, we would expect the cardiovascular disease incidence rate to drop by 23 per 100,000

51. Population Attributable Risk Percentage (PAR%) The PAR% is 6.2% [(379-356) ÷ 379 × 100] Interpretation If smoking were eliminated from the population, we would expect a 6.2% decrease in the incidence rate of cardiovascular disease

52. Double Cohort Distinct from conventional cohort studies in that two distinct populations are involved with different levels of an exposure of interest Double cohorts are employed when the exposure is rare and a relatively small number of people are affected

53. Selecting the Study Cohort From population, choose those at risk of becoming a case Exclude Individuals who already have a disease outcome of interest (prevalent cases) Those not at risk (e.g., they have had an organ removed such that they cannot become a case)

54. Restriction Used to improve validity of study Restriction involves selecting cohorts with limited exposure, narrow behaviors or activities, or from a limited work environment with restricted exposures or health problems This limits generalization but often improves feasibility and focus

55. Bias in Cohort Studies Selection bias Healthy worker effect Loss to follow-up

56. Healthy Worker Effect Occurs in cohort studies when workers represent the exposed group, and a sample from the general population represents the unexposed group This is because workers tend to be healthier, on average, than the general population

57. Loss to Follow-Up A circumstance where researchers lose contact with study participants, resulting in unavailable outcome data on those people A common problem in cohort studies, increasingly so in cohorts with longer follow-up times Reasons Refusal to participate Unable to locate Unable to be interviewed Death

58. Confounding More a problem in double-cohort studies if factors (age, sex, race/ethnicity, education, etc.) differ between the populations being compared

59. Misclassification Also may result in cohort studies Differential (non-random) Non-differential (random)

60. Misclassification Differential misclassification arises if exposure classification influences differential accuracy in ascertaining outcome information Nondifferential misclassification may arise through inaccuracies in classifying exposure status of individuals, but these misclassifications occur similarly between exposed and unexposed groups

61. Controlling for bias in cohort studies Healthy worker bias may be avoided by selecting a comparison group made up of workers, only unexposed Misclassification may be minimized by refining the definition of the exposed and unexposed groups and avoiding exposure classifications that result in differential outcome ascertainment

62. Controlling for bias in cohort studies Loss to follow-up bias may be minimized by Restricting the study participants to those likely to remain in the study (e.g., excluding those with a highly fatal disease or who are likely to move out of the area) Collecting personal identifying information (e.g., the participant’s telephone number and address, as well as that of their employer and a family member) Making periodic contact, and providing incentives (e.g., cash or free medical exam)

63. Controlling for confounding in cohort studies At the study design level Restriction to avoid bias due to confounding In double-cohort studies, reduced by choosing comparison groups as alike as possible to the exposed population At the analysis level Collecting data on potential confounders at the beginning of the study makes it possible to adjust for these potential confounders at the analysis level through stratification and multiple regression techniques

64. Example

65. Effect Modification When an association between an exposure and disease outcome is modified by the level of an extrinsic factor, beyond random variation, the extrinsic variable is called an effect modifier Effect modification may occur in either cohort or case-control data

66. Examples of confounding and effect modification in case-control studies using stratified data Crude ORStratum-specific OREffect (OR c )(OR 1 )(OR 2 )Confounding Modification 2.02.02.0NoneAbsent 3.01.21.3PositiveAbsent 1.02.62.5NegativeAbsent 2.61.04.5NonePresent 3.80.82.9PositivePresent 0.82.13.9NegativePresent

67. Examples of confounding and effect modification in case-control studies using stratified data (cont’d) If OR c > OR 1 = OR 2, positive confounding If OR c < OR 1 = OR 2, negative confounding (small differences in OR 1 and OR 2 assumed explained by random error) OR 1 ≠ OR 2, effect modification present Adapted from Hennekens CH, Buring JE. Epidemiology in Medicine. Boston: Little, Brown and Company; 1987.

68. If >, positive confounding If <, negative confounding, effect modification present Note that small differences in and are likely explained by random error. Also, the same ideas apply for. Confounding and effect modification shown using ORs

69. Confounding and effect modification treated differently Control for confounding Present results from effect modification