2. (c) GerstmanCh 11: Observational Designs1 Epidemiology Kept Simple Ch 11: Observational Studies

3. (c) GerstmanCh 11: Observational Designs2 Observational Designs Cross-sectional: Sample population, no follow-up of individuals  compare disease experience of exposure groups (§11.2 and §11.3) Cohort: closed population with individual follow-up over time  compare disease experience of exposure groups (§11.4) Case-control: all cases and a sample of non- cases from population  compare exposure experience (§11.5)

4. (c) GerstmanCh 11: Observational Designs3 Cross-Sectional Designs Recall distinction between longitudinal and cross-sectional observations Recall the distinction between individual and aggregate units of observation Cross-sectional design with aggregate unit of observation ≡ ecological design Cross-sectional data with individual units ≡ cross-sectional survey

5. (c) GerstmanCh 11: Observational Designs4 Example: Ecological Data Unit of observation = geographic region Exposure = Cig1930 = cigarettes per capita, 1930 Disease = Mortal = lung cancer mortality per 100,000 p-yrs, 1950

6. (c) GerstmanCh 11: Observational Designs5 r = 0.74 Example: Ecological Data

7. (c) GerstmanCh 11: Observational Designs6 % Calories from fat and cardiovascular disease mortality by country Example: Ecological Data

8. (c) GerstmanCh 11: Observational Designs7 Farr’s (1852) Faux Pas Elevation above sea level & cholera mortality: (supports miasma theory)

9. (c) GerstmanCh 11: Observational Designs8 Farr’s faux pas Confounded! Exposure = elevation Disease = cholera Confounder ≡ proximity to contaminated water sources My “bad”!

10. (c) GerstmanCh 11: Observational Designs9 Example: X-Sectional Survey SES & Mental Disorders Prevalence per 100,000 Social classPsychosisNeurosis High188349 Moderate291250 Low518114 Very low150597

11. (c) GerstmanCh 11: Observational Designs10 Biases in Hollingshead Detection bias: different diagnostic practices  artificial differences Reverse-causality bias: “Disease” causes “exposure” (e.g., psychosis causes low SES) Prevalence-incidence bias: Difference in prevalence but not incidence (e.g., more persistent diagnoses) Limitations like these stimulated development of better quality study designs

12. (c) GerstmanCh 11: Observational Designs11 §11.4 Cohort Studies Closed population Incidence 1 Incidence 0 RR or RD Recruit cohort Classify individual as exposed or non-exposed Follow exposed and non-exposed sub-cohorts to determine incidence exposed sub-cohort non-exposed sub-cohort RR or RD

13. (c) GerstmanCh 11: Observational Designs12 Cohort: Simple Example

14. (c) GerstmanCh 11: Observational Designs13 Goldberger on Pellagra Joseph Goldberger (1874–1929) demonstrated nutritional basis of pellagra when it was thought to be infectious Goldberger’s cohort analysis of cow ownership as protective factor:

15. (c) GerstmanCh 11: Observational Designs14 British Doctors Cohort Source: Doll, R., Peto, R., Wheatley, K., Gray, R., & Sutherland, I. (1994). Mortality in relation to smoking: 40 years' observations on male British doctors. British Medical Journal, 309(6959), 901-911. 80% of nonsmoker survived to age 70 50% heavy smokers survived to 70

16. (c) GerstmanCh 11: Observational Designs15 British Doctors Cohort

17. (c) GerstmanCh 11: Observational Designs16 Wade Hampton Frost First Professor of epidemiology in U.S. First Dean of US School of Public Health Bridged gap between infectious disease epi and chronic disease epi with TB studies (infectious disease with long latency) 1880 – 1938

18. (c) GerstmanCh 11: Observational Designs17 Frost’s TB Studies TB morality per 100,000 p-yrs Columns  cross-sectional rates by age (NO follow-up of individuals )

19. (c) GerstmanCh 11: Observational Designs18 Rows  cross-sectional rates by year (NO follow-up of individuals) Frost’s TB Studies TB mortality per 100,000 p-yrs

20. (c) GerstmanCh 11: Observational Designs19 Frost’s TB Studies TB morality per 100,000 p-yrs Diagonals  mimic experience of birth cohort (1870 birth cohort shaded)

21. (c) GerstmanCh 11: Observational Designs20 TB Study Cross-Sectional Rates Note shifting peak occurrence (marked by *) with age Not a true cohort: False impression of cohort experience.

22. (c) GerstmanCh 11: Observational Designs21 Frost’s Birth Cohorts True cohort perspective. Note consistent peak at in late 20s (corresponds childbearing age) Comstock, G. W. (2001). Cohort analysis: W.H. Frost's contributions to the epidemiology of tuberculosis and chronic disease. Social and Preventive Medicine, 46(1), 7-12.

23. (c) GerstmanCh 11: Observational Designs22 Example: Historical Cohort Historical info on exposure to aniline dyes (from work records) were used to compile exposed and non-exposed worker cohorts Retrospective data from death certificates on bladder cancer occurrence Result: bladder cancer occurrence was 100 times as frequent in aniline-exposed cohort Figure shows induction time between exposure onset and bladder CA occurrence

24. (c) GerstmanCh 11: Observational Designs23 Case-Control Studies Identify population cases Randomly select non-cases (“controls”) Compare exposure histories in cases & controls Population All cases Sample non-cases Exposure histories Odds Ratio

25. (c) GerstmanCh 11: Observational Designs24 Case-Control CasesControls Exposed A1A1 B1B1 Non- exposed A0A0 B0B0 M1M1 M2M2 Cross-tabulate disease and exposure status of cases and controls Calculate:

26. (c) GerstmanCh 11: Observational Designs25 Historical Example: Levin Mort Levin (1904 -1995) 1950 case-control studiesMort Levin (1904 -1995 Population: Roswell Park Cancer Institute (Buffalo, NY) 236 lung cancer cases 481 non-cases (nonmalignant lung conditions) OR estimate of 2.29 suggests smokers had 2.29× the risk of non- smokers Heavy smoker Lung CA+ Lung CA− Exposed+156221 Exposed−80260 Total236481

27. (c) GerstmanCh 11: Observational Designs26 Interpretation of the Odds Ratio When the disease is rare, interpret the OR as if it were an RR The illustrative OR of 9.3 suggests that tampon users had 9.3 times the risk as non-tampon users [The suspected brand of tampon has since been removed from the market]

28. (c) GerstmanCh 11: Observational Designs27 Multiple Levels of Exposure Historical Example: (Wynder & Graham, 1950, p. 212) Smoking status CasesControls Chain 12364 Excessive 18698 Heavy 213274 Moderate 61147 Light 1482 Non-smoker 8115 Total 605780 Table 1 CaseCntl Light smoke1482 Non-smoker8115 OR 1 = (14)(115)/(82)(8) = 2.5 Table 2CaseCntl Moderate61148 Non-smoker8115 OR 2 =(61)(115)/(147)(8) = 6.0 Table 3CaseCntl Heavy213274 Non-sm.8115 OR 3 = (213)(115)/(274)(8) =11.2 Exposure may be measured at various levels. In this historical example, smoking is classified into 6 levels. To analyze the table, break-up it up into five separate 2- by-2 tables with each table referencing the nonexposed group as follows: Table 4CaseCntl Excessive18698 Non-smoker8115 OR 4 = (186)(115)/(98)(8) = 27.3 Table 5 CaseCntl Chain12364 Non-smoker8115 OR 5 = (123)(115)/(64)(8) =27.6

29. (c) GerstmanCh 11: Observational Designs28 Longitudinal Designs

30. (c) GerstmanCh 11: Observational Designs29 Longitudinal Designs TrialCohortCase-Control ExperimentalObservational Assign exposureClassify exposureSelect cases and non-cases Calculate RD and/or RR Calculate OR Study multiple outcomes Study multiple exposures ProspectiveProspective or retrospective Retrospective

31. (c) GerstmanCh 11: Observational Designs30 End of HS 261 Presentation

32. (c) GerstmanCh 11: Observational Designs31 Why the OR is an estimate of the RR There are two justification for using the OR as an estimate of the RR between an exposure and outcome The classical justification: Cornfield, J. (1951). A method of estimating comparative rates from clinical data. Application to cancer of the lung, breast, and cervix. Journal of the National Cancer Institute, 11, 1269-1275. The modern justification: Miettinen, O. (1976). Estimability and estimation in case-referent studies. American Journal of Epidemiology, 103, 226-235.

33. (c) GerstmanCh 11: Observational Designs32 Incidence Density Sampling: Miettinen’s Modern Justification of the Case-Control OR Imagine 5 people followed for disease D. Every time a case occurs in the source population, select at random a control from the same source population For example, At time t1, D occurs in person 1. That time, select at random a non-cases to serve as a control. Note: person #2 becomes a case later on, but can still as a control at time t1. Miettinen, O. (1976). Estimability and estimation in case-referent studies. American Journal of Epidemiology, 103(2), 226-235.

34. (c) GerstmanCh 11: Observational Designs33 The Ecological Fallacy (aggregation bias) The ecological fallacy occurs when an association seen in aggregate does not hold for individuals The illustrative example on p. 195 discusses the historically important demonstration of a negative ecological association between high foreign birth rates and illiteracy rate (r = −0.62). However, when data are disaggregated, there is a positive association. The reason? High immigration states had better public education Optional

35. (c) GerstmanCh 11: Observational Designs34 Logic of the Ecological Ecological measurements in conjunction with individual level measurements (multi-level designs) can elucidate interdependence of upstream and downstream factors Types of aggregate-level risk factors (Susser, 1994) –Integral variables – factors that effect all community members (e.g., the local economy) –Contextual variables – summary of individual attributes (e.g., % of calories from fat) –Contagion variables – a property that involves a group outcome (e.g., prevalence of HIV effects risk of exposure) Optional

36. (c) GerstmanCh 11: Observational Designs35 Goldberger’s Field Study of Food Intake see pp. 200 - 201 for details Optional Caloric Intake by Food Group

37. (c) GerstmanCh 11: Observational Designs36 Small Sample Size Formula For the Odds Ratio (Optional) Some statisticians recommend adding ½ to each cell before calculating the odds ratio, esp. when some cells have very few counts (This is known as the “small sample odds ratio formula”) For the illustrative data:

38. (c) GerstmanCh 11: Observational Designs37 Matched-Pairs Matching is employed to help adjust for confounding e.g., matching on age and sex will adjust for these factors Each pair now represents a single observation Cross-tabulate pairs to determine odds ratio Control E+ Control E− Case E+ tu Case E− vw Odds ratio formula for matched pairs

39. (c) GerstmanCh 11: Observational Designs38 Example (Matched Pairs) Control E+Control E− Case E+530 Case E−105 Exposure triples risk