1. 1 Two-stage sampling JF Boivin Version 14 November 2007 S:\BOIVIN\695\Winter 2007\Two-stage Sampling.ppt

2. 2 1980s-1990s: Progress in use of administrative drug databases

3. 3 Advantages Large Population-based Valid prescription data Long-time periods

4. 4 Disadvantages Missing data on certain outcomes Temporal sequence not always clear Glucocorticoids  cataracts Cataract surgery  glucocorticoids Lack of data on confounders

5. 5 NSAIDs and breast cancer

6. 6 Previous research Poor exposure data Dose Duration Self-reports Small numbers Short follow-up Inadequate control of confounding

7. 7 NSAIDs and breast cancer Cases: Saskatchewan cancer registry Controls: Saskatchewan population Drug exposure: 15 yr of computerized information Missing: - Over the counter drugs - Other confounding factors: Menarche Menopause Pregnancies Obesity

8. 8 Entire population (= truth) Obese E+ E−E− cancerno cancer Not obese E+ E−E− All E+ E−E− 2 000 40 10 000 100 2 040 10 100 200 400 10 000 600 20 000 OR=0.5 32 740 2 200 440 20 000 10 100 2 640 30 100 OR=2.5

9. 9 Obese E+ E−E− cancerno cancer Not obese E+ E−E− All E+ E−E− 2 200 440 20 000 10 100 2 640 30 100 not available computerized databases

10. 10 What to do about missing confounder data?

11. 11 Option #1 Do not conduct research on that topic

12. 12 Option #2 Cohort or case-control study without data on confounder 32 740 Obese women E+ E−E− cancerno cancer Not obese E+ E−E− All women E+ E−E− ? ? ? ? ? ? ? ? 2 200 440 20 000 10 100

13. 13 Advantages Cheaper May be scientifically reasonable for certain questions

14. 14 Option #3 Collect covariate data on a sample of the study subjects two-stage samples three-stage samples partial questionnaire case series only etc.

15. 15 Two-stage sample Sampling approaches: simple random balanced etc.

16. 16 Two-stage balanced design Obese E+ E−E− no cancer Not obese E+ E−E− All E+ E−E− cancer 227 23 125 2 23 227 125 248 250/ 32 740 (I) 2 200 440 20 000 10 000

17. 17 White JE. A two-stage design for the study of the relationship between a rare exposure and a rare disease. AJE 1982 Cain KC, Breslow NE. Logistic regression analysis and efficient design for two-stage studies. AJE 1988

18. 18 Consent for interviews Cases:49% Controls : 39% (Sharpe et al. Saskatchewan study)

19. 19 Other related sampling designs three-stage sampling partial questionnaire confounder data on cases only

20. 20 Obese E+ E−E− cancerno cancer Not obese E+ E− All E+ E− 2 200 440 20 000 10 100 2 640 30 100 medical record review computerized databases Confounded data on cases only 2 000 40 200 400 ? ? ? ?