1. “Personality, Socioeconomic Status, and All-Cause Mortality in the United States” - Chapman BP et al. Journal Club 02/24/11

2. Monte Carlo Sensitivity Analysis (MCSA) but first...

3. Introducing nonrandom error Phillips, 2003

4. How we deal with bias Jurek et al. 2006, Orsini 2007 1)Ignore biases 2)Mention potential biases (in passing) 3)Qualitatively address the effect of bias 4)Quantitatively address the effect of bias (sensitivity analysis)

5. Why should we care? Participation rates in epidemiologic studies are falling (selection bias?) Reviewers almost certainly ask about different forms of bias Better understand uncertainty behind your findings Galea et al. 2007, Curtin et al. 2005

6. Types of quantitative sensitivity analysis 1) Deterministic e.g. externally-adjusted estimates 2) Probabilistic e.g. MCSA

7. Deterministic SA Rothman et al. 2008

8. Problems with Deterministic SA Fail to discriminate among the different scenarios in terms of their likelihood Difficult to summarize results Difficult to examine effects of multiple biases Orsini 2007

9. An example of MCSA Analysis of a case-control study to determine whether case status (lung cancer) is associated with exposure to asbestos. However, there is no measure of smoking duration!

10. An example of MCSA Observed OR = 3.0 Lung Cancer Asbestos 15080 5080

11. An example of MCSA Let’s look at uncertainty arising from: 1) Unmeasured confounding due to smoking 2) Exposure misclassification With MCSA we can simultaneously examine bias arising from these sources...

12. 1) Unmeasured confounding and MCSA What we need to generate: Prevalence of smoking among asbestos exposed (P e ) and asbestos non-exposed (P ne ) Association (OR s ) between smoking and case status Phillips 2003, Steenland et al. 2004

13. 1) Unmeasured confounding and MCSA Start by generating uniform distributions for P e and P ne P e is bounded: [0.5, 0.8] P ne is bounded: [0.2, 0.5] Generate 10,000 random numbers for each

14. 1) Unmeasured confounding and MCSA

15. Generate a normal distribution of odds ratios betweens smoking (confounder) and lung cancer (outcome) Consult literature for distribution parameters Generate 10,000 random numbers

16. 1) Unmeasured confounding and MCSA

17. 2) Exposure misclassification and MCSA What we need to generate: Sensitivity and specificity distributions for cases and controls Let’s assume recall bias has occurred (differential misclassification) Cases remember true exposure more than controls Greenland et al. 2008

18. 2) Exposure misclassification and MCSA Generate normal distributions Cases: Mean sensitivity = 0.95 Mean specificity = 0.75 Controls: Mean sensitivity = 0.75 Mean specificity = 0.75 Bound values: [0,1] Generate 10,000 random numbers for each

19. Bringing it together... Algorithm uses formulas for external adjustment method (see Rothman book 3 rd ed) Correct for biases in reverse order of data generation process Pick a random number from each of the above distributions and back-calculate a new OR Repeat this 250,000 times Phillips 2003, Steenland et al. 2004, Greenland et al. 2008

20. Median OR: 1.55 Middle 5% OR’s: (1.52 – 1.57) Middle 90% OR’s: (1.13 – 2.06) Range: (0.70 – 3.26)

21. How to implement it

22. On to the paper!

23. Socioeconomic Status and Mortality Pappas et al. 1993

24. Personality and Mortality Roberts et al. 2007 Linked with mortality risk: Optimism Neuroticism Hostility Trust Conscientiousness Cynical distrust Rationality Extraversion Creativity Agreeableness Trait Anxiety

25. Purpose of the study To examine degree to which SES and personality are mutually confounded risks in predicting all-cause mortality among US adults Two possibilities: 1)SES and personality are clustered mortality risk factors 2)SES and personality are independent mortality risk factors

26. Midlife Development in the United States (MIDUS) study English-speaking adults aged 25-74 Random digit dialing starting in 1995 Study population

27. 6,063 contacted 4,244 complete telephone interview 2,998 with complete data 70% 3,692 return mail survey 71% 87% 81%

28. Participants contacted for 10-year follow-up in 2004-2005 Names of subjects lost to follow-up submitted to NDI All-cause mortality Methods: Mortality status

29. Methods: SES factor analysis INCOME TOTAL ASSETS EDUCATION OCCUPATIONAL PRESTIGE

30. Methods: SES factor analysis INCOME TOTAL ASSETS OCCUPATIONAL PRESTIGE EDUCATION Continuous Factor Scores

31. Methods: Personality Midlife Development Inventory 30 Likert scale items Factor analysis used to separate: 1)Agreeableness 2)Openness 3)Neuroticism 4)Extraversion 5)Conscientiousness Lachman 1997

32. Methods: Covariates Demographic factors: Age Sex Ethnicity Behavioral risk factors: Smoking Heavy drinking BMI Physical activity

33. Methods: Analysis Primary analysis: Stepwise logistic regression Adjusted population attributable fractions Secondary analysis: Interactions among personality domains Mortality risk associated with “traits” within each personality domain

34. Methods: Analysis Sensitivity/Error analysis: Change in estimate in SES for all 32 combos of personality domains Multiple imputation: missing data bias Simulation extrapolation: random measurement error in health behaviors MCSA: unmeasured confounding, selection bias, nonrandom error in personality or SES measurement

35. Results

36. SES effect is reduced: 20% Neuroticism effect is reduced: 8%

37. Discussion: Key findings Support for both correlated risk model and independent risk models Low SES is mortality risk factor (no surprise) High Neuroticism is mortality risk factor (no surprise) Agreeableness X Conscientiousness interaction Health behaviors explain substantial amount of SES and personality effects

38. Thoughts on the strengths and weaknesses of this study? Significance of this study?