1. Lecture 9: Retrospective cohort studies and nested designs
Jeffrey E. Korte, PhD
BMTRY 747: Foundations of Epidemiology II
Department of Public Health Sciences
Medical University of South Carolina
Spring 2015

2. Why do these kinds of studies?
Retain advantages of prospective cohort study (e.g. exposure measured before disease occurs)
Eliminate disadvantages of prospective cohort study (e.g. high cost, long time commitment)
e.g. study of cancer with year latent period

3. Retrospective cohort study
Identify cohort of individuals based on their characteristics in the past
Reconstruct their subsequent disease experience up to some defined time point in the more recent past, or up to the present (or into the future)
This involves tracking them down and finding out what happened to them

4. Retrospective cohort study
Examples: Studies of
Workers (asbestos, benzene)
College students (future risk of CHD predicted by participation in college sports; future risk of back pain predicted by spinal curvature)
Patients (future risk of hysterectomy predicted by tubal sterilization; future risk of cancer predicted by DES exposure during pregnancy)

5. Should I do a retrospective cohort study or a case-control study?
Participants identified on the basis of exposure
Therefore may be the only option in the case of a rare exposure
Must have good quality historical exposure data
Case-control study
Participants chosen on the basis of disease
Therefore may be the only option in the case of a rare disease
May be the only option if no information is available on past exposures

6. Retrospective cohort study
Choosing participants
Must have large % exposed to agent of interest
Some must have high enough exposures to expect a detectable excess risk of disease
Exposure data must be reasonably accurate
Appropriate comparison group must be identifiable

7. Retrospective cohort study
Measuring exposure
Occupational studies: can use proxies for dose (length of employment) to assess dose-response
Can sometimes impute exposure dose by synthesizing data sources: e.g. study of fluoroscopic radiation of the chest for tuberculosis in relation to breast cancer: collect number of procedures from patient or medical records; obtain procedure dose, body positioning, etc. from technicians; estimate radiation dose for each procedure

8. Retrospective cohort study
Measuring disease
Death certificates
Disease registries (may allow analysis of incidence as well as mortality)
Likely to underascertain disease occurrence
If disease is underascertained in exposed group, but risk is compared to accurate population rates, then true risk factors may appear null or even protective

9. Retrospective cohort study
Measuring confounders
Often not possible
But: keep in mind that not all risk factors are confounders (must be related to both exposure and disease in order to be a confounder)
Can contact surviving cohort members to ascertain confounders or other information of interest
Compare distribution between exposed/unexposed survivors
Assess likelihood of confounding in full cohort

10. Retrospective cohort study
Follow-up
Individualized tracing methods required (in contrast to most prospective cohort studies)
National Death Index
Credit bureaus, state DMV, etc.
City directories
Contact relatives

11. Nested case-control study
Useful when:
Contemplating an expensive lab test
Have frozen stored biologic samples from prospective cohort
Select cases and (matched) controls from cohort; test only their samples
Save enormous sums of money

12. Nested case-control study

13. Nested case-control study
If sufficient number of cases has accrued:
Select controls from unaffected cohort members who are still at risk and under surveillance at the time the cases develop disease
May match controls on age, gender, time of entry into the study (in addition to length of follow-up)
Controls may later serve as cases in analysis

14. Nested case-control study
Example: Epstein-Barr virus and Hodgkin’s disease (Mueller et al 1989)
Cohort of 240,000: stored blood in serum banks
Cases identified through cancer registries and hospital records
Controls matched on serum bank, gender, race, date of index serum (+/- 6 months), date of birth (+/- 1 year)
Sera tested for antibody patterns to EBV and cytomegalovirus
Results: EBV antibodies predict Hodgkin’s; antibody changes detectable 3-5 years prior to disease

15. Nested case-control study
Relevant for designs other than stored blood: e.g. study of EMF and brain tumors/leukemia (Tynes et al 1994)
Cohort of 13,300 railway workers
39 cases of brain tumor; 52 cases of leukemia
4-5 age-matched controls for each case
Case-control design was beneficial, because the determination of exposure to electromagnetic fields was time-consuming and expensive for each participant

16. Case-cohort study Also called case-base study
All cases identified from cohort
Controls are a random sample of cohort at baseline
Use a form of Cox model to compare each case to the subset of controls still at risk
Controls are not matched on follow-up time, or other factors (age, sex, etc.)
These factors must be accounted for in analysis

17. Case-cohort study

18. Case-cohort study
Can be used to obtain rate ratio, risk ratio, and odds ratio estimates within a cohort
Can compare rate and risk estimates between exposed groups within cohort versus an external population
e.g. risk of disease in exposed can be compared to risk in unexposed; can also be compared to risk in general population

19. Should I do a case-cohort study or a nested case-control study?
Advantages of nested case-control:
Matching on time (more statistically efficient, unless everyone joins the study at the same time and has the same follow-up)
Analysis is similar to case-control study
Easier to calculate sample size
Does not require control disease ascertainment beyond the time of case data collection

20. Should I do a case-cohort study or a nested case-control study?
Advantages of case-cohort
Can use same control group for multiple case groups, and for unanticipated future studies
Flexible analysis in choice of time scale (no matching on cohort entry or follow-up time)
Quicker and easier to select controls
Subcohort represents population; can estimate risks of various things

21. Analysis overview
Nested case-control study: analyzed as case-control study
Conditional logistic regression (matched on follow-up time, perhaps other factors
Case-cohort study: analyzed as cohort study
Cox proportional hazards models for time-to-event data
Poisson for rare event/count data
Log binomial to estimate relative risk