1. LECTURE 3 – June 9 2006 Cohort Studies, Selection Bias Survival analysis
Dr. Dick Menzies

2. Cohort Studies – General
Prospective study: Incidence of new disease in persons who start without disease.
Follow-up period – weeks, months, years
One or more diseases can be measured
Measure exposures – at start or ongoing.
Can measure multiple exposures
Compare incidence in exposed vs unexposed groups within population – per unit of time

3. Advantages of cohort over case-control or cross-sectional designs
KEY – exposure measurement is made before disease occurs
Exposure more accurate – prospective, and often repeated
Eliminates bias in measurement of exposures:
Recall bias of patients, or observer bias in exposure assessment - with knowledge of disease status.

4. Experimental vs cohort studies
Expt studies are a form of cohort study
Same - Persons are free of disease at outset
But - Exposure is RANDOMLY ASSIGNED to some/not others
Same - Measure outcomes after exposure
Cohort study – exposures NOT assigned, but occur naturally, or are chosen purposely by subjects, or by their MD’s, etc

5. Advantages of cohort studies over experimental
Ideal to study natural history, course of disease, prognostic factors.
Etiologic research for exposures that can not be given experimentally, for ethical reasons
Smoking, asbestos, air pollution
Interventions not feasible for randomization
Diagnostic tests, complex care management
Some outcomes not well measured in trials:
Compliance by patients and MD’s,

6. Advantages of cohort studies over experimental
Total population studied.
Children, elderly, pregnancy, mentally incompetent,
Full spectrum of illness
From patients in ICU to minimal forms of disease
Often excluded in RCT – esp Pharma trials
Findings more likely to be applicable in real world
Adverse events often more accurately measured
Population based estimates of exposure effects
BUT you MUST include the full spectrum of patients as possible (No exclusions in observational studies)

7. Disadvantages
Selection bias – Persons who get exposed not same as unexposed
Surgery – who is ‘operable’ vs ‘inoperable’
Exposures that seem same, are not
Potential bias in measuring
Drop-outs – reduce power, may bias (a lot)
Outcome assessment can be biased

8. Cohort Designs
Prospective: Subjects without disease followed to determine incidence of diseases
Exposures measured at baseline, and/or concurrently.
Disease – measured during follow-up
Retrospective: Subjects first identified based on past Exposures (Hiroshima survivors, work-force)
Outcomes may then be ascertained directly, or also have already occurred
Key – exposure well defined, AND occurred well before disease (useful for diseases like cancer)

9. Cohort Populations General populations – no special exposures
Framingham study – a true general population
All persons in the community invited
Proxy general Pop’n - Nurses, Military, Company
Exposures studied are those of general pop’n.
Diet, exercise, smoking, alcohol
Exposure defined cohort
Work-force to study occupational exposures
Group of patients who received certain therapy

10. Cohorts of patients Clinical cohorts – patients with a given condition
Case series can be form of cohort study
But – must have differences in ‘exposure’
Different types, severity, causes
Potential problems in cohort studies with patients:
Referral bias – only sickest, rarest,
Lead-time bias – better facilities = earlier Dx
Multi-serial cohorts –
Cohort starts with all diabetics in 2004
New, and old = very different patients

11. Open versus Closed Cohorts
An open cohort – or dynamic cohort - is one where people can enter or leave
Examples: A workforce study that is ongoing
A city or other geographic location
A closed cohort is where all persons in the cohort are defined at entry. No one enters, members can only exit.
Eg. McGill medical school class of 2004

12. Selection Bias
Definition – selection bias occurs when there is a distortion in the estimate of effect (association) because the study or sample population is not truly representative of the underlying population in terms of the distribution of exposures and/or outcomes.
Other terms: referral bias, volunteer bias, healthy worker effect, susceptibility bias, drop-out bias
How/where in a study can this occur?

13. GROUPS LOSSES REASONS FOR LOSSES INTENDED POPULATION AVAILABLE GROUP
CANDIDATE GROUP
ELIGIBLE GROUP
QUALIFIED GROUP
ADMITTED GROUP
NOT
AVAILABLE
CANDIDATES
ELIGIBLE
EXCLUDED
NON-
RECEPTIVE
Treated at other hospitals or by other doctors
Not identified or accessible
Did not fulfill diagnostic criteria
Superimposed condition of severity, co-morbidity, co-medication, or non-compliance
Refused participation or acceptance of assigned maneuver
Figure Diagram showing successive transfers from the intended population to the group admitted to a study of therapy

14. Obtaining a representative sample
In a representative sample we hope for a sample that shows us the true underlying distribution of exposure and disease:
Truth – distribution of exposure and disease in source population
Exposed
Not Exposed
Diseased A B
Not Diseased C D
Odds Ratio = (A/B) / (C/D)
= A x D
B x C

15. Un-biased Sampling x 1 = Truth! Odds Ratio = (P1 x P4) x (A x D)
Exposed
Not Exposed
Diseased
P1A
P2B
Not Diseased
P3C
P4D
Odds Ratio = (P1 x P4) x (A x D)
(P2 x P3) (B x C)
IF (P1 x P4) THEN OR = (A x D)
(P2 x P3) (B x C)
x 1
= Truth!

16. Biased Sampling If sample all of A (P1=1) but only half of B (P2 =0.5)
Exposed
Not Exposed
Diseased
P1A
P2B
Not Diseased
P3C
P4D
If sample all of A (P1=1) but only half of B (P2 =0.5)
And 1/3 of C and D (P3=0.33, P4=0.33)
Odds Ratio = (P1 x P4) = (1x.33) = 2 x (A x D)
(P2 x P3) = (.5X.33) (B x C)
IF (P1 x P4)=2 THEN ORestimated = 2X ORTrue
(P2 x P3)

17. Example – Biased sampling
We are planning a case control study of spicy foods and peptic ulcer disease
Cases = endoscopy proven peptic ulcer disease
Controls = elective inguinal hernia repair at the same hospital
The truth: no relationship i.e. the odds ratio = 1
The problem – physician at this hospital strongly believe spicy foods is an important risk factor for peptic ulcer disease.
Therefore they tend to refer patients for endoscopy more often if they had a diet of spicy foods

18. Biased sampling (cont’d)
TRUTH:
Spicy Foods
No Spicy Foods
Odds
Truth
Cases 25 75
25/75
Controls
Total 50
150
1.0
Biased sample
37.5
25/37.5
112.5
2.0

19. Example: biased sampling
So, 100% of patients with peptic ulcer disease AND history of spicy foods have endoscopy
But only half of those with peptic ulcer, but WITHOUT history of spicy food are in fact diagnosed – (they do not have endoscopy, so they are missed)
Estimated association will be twice what is correct.

20. To achieve Un-biased Sampling
To achieve un-biased sampling the easiest is:
P1= P2=P3=P4
This means the proportion sampled from each group is the same, i.e., 10% are sampled from each of the groups
However if P1 is higher than P2 this can be okay as long as P4 is also increased more than P3

21. Volunteer Bias Participants in a study are different from refuseniks
Mortality of non-participants in the Framingham study
Subjects with exposure and the outcome are more (or less) likely to participate
Eg HIV infection and homosexuality – in Africa
Disease and occupational exposures, particularly for self-reported exposures, and compensable illnesses.

22. Susceptibility bias
Persons allocated to one form of treatment, or who who self-select to certain exposures are more, or less susceptible to develop health outcomes of interest.
Eg Cancer patients who have surgery vs medical or radiotherapy only. Surgical patients often appear to do better.

23. Healthy worker effect An important bias – found in work-force studies
Reflects medical screening (military, mining)
Or, physical requirements of job
Results in better health status initially than general population, or certain control pop’n
Strongly affects results in cross-sectional studies
Reduces risk or delays occurrence of health outcomes of interest.
Also occurs in smokers “healthy smoker effect”
Lung function in adolescent smokers > non-smokers

24. Example of healthy smoker effect

25. Selection Bias in Cohort Studies – Dropouts
Losses to follow up occur in all cohort studies
Reduce power, and dilute results
Problematic if more drop-outs in one exposure group
REALLY important if drop-out is due to development of disease

26. Selection Bias in Cohort Studies – Dropouts
Example:
study of incidence of diabetes in obese persons.
Truth: IRR = 3.0
Losses – 33% in diabetes/obesity group (death/other)
5% losses in all other groups
(P1 x P4) does not = 1
(P2 x P3)

27. Selection Bias from Dropouts - Example
At onset
Dropped
No DM
Out
Diabetes
Detected at end with diabetes
Obese
227 10 9 18
Not Obese
773 35 3 30
Incidence (biased):
In obese – 18/208 = 8.7%
In non-obese – 30/735 = 4.1%
Biased incidence rate ratio – 8.7%/4.1% = 2.1

28. Drop-outs from a work-force - impact
An occupational exposure causes health effects quickly in a susceptible sub-group.
They leave the work-force (quit) quickly.
Examples:
Allergy to lab animals in researchers
Asthma in Grain workers
Cross-sectional studies – no susceptibles left
Cohorts – Can miss when setting up cohort.
Outcomes occur in small number of new workers (power problem)

29. Controlling Selection Bias
Control in design - Most important is prevention
Recruitment – high % in all groups
Same %recruitment in exposed/not exposed
Close follow-up to prevent dropouts
Assess in analysis
Compare participants to non-participants
Sub-groups of non-participant
Compare dropouts with those who remained
Sensitivity analysis – best case/ worst case to assess impact of selection biases

30. Cohort Studies – Exposure Assessments
Prospective - Measure one or more exposures at start
Specific: cholesterol, obesity, smoking, blood pressure.
Proxies: occupation, housing
Measure once, or repeatedly to account for changes in exposure over time (obesity, smoking, BP).
Retrospective
Exposure based upon past events
These are rarely quantified
Proxies used (job description, distance from blast)
Sometimes records (transfusions, dust levels)

31. Pitfalls in exposure assessments
Observer bias – disease ascertained at same time
Blind observers to study hypothesis
Standardized protocols
Are all exposures the same?
Complications of pleural tap at MGH/RVH >> MCI
Did you forget something?
Hard to go back to the start of cohort
Measure everything, freeze the rest
Add measures as new things reported

32. Cohort Studies – Outcome Assessments
Baseline – ensure cohort members free of disease.
Easy if prospective, harder if retrospective
Outcomes measured periodically
Through questionnaire, exam, labs (direct)
Through health service utilization (databases)
Through vital statistics (databases)
Case definition key for outcome assessments
Diagnosis of milder disease common problem

33. Pitfalls in outcome assessments
Ascertainment bias – if patients with Factor X are more likely to have testing to detect outcome.
Standardized protocols, blinding to exposures
Observer bias – patients with Factor X more likely to be diagnosed with outcome of interest
Common with more subjective tests – eg CXR
Solution – independent reviewers, blinded to exposure status (Factor X)
Lead time bias – earlier diagnosis makes survival look better

34. Lead-time bias - example

35. Cohort Studies – Measures of Incidence
Incidence rate (simplest) =
number developing disease
Total number who entered cohort
per unit of time
Cumulative incidence =
Over total follow-up period

36. Measuring Incidence in Cohort Studies How to handle drop outs etc..?
Drop-outs from loss to follow-up, death other causes, or withdraw consent are common
Up to 50% in long term cohorts
Include or exclude from analysis?
Simple incidence measures - excludes
Need to allow variable length of follow up
And count people who enter after the first year

37. Incidence Density (ID)
Counts person-time (person-years/months)
Starts count when person enters cohort
Each year of follow-up added up
Patient
Exposed
Enter in year
Stop in Year
Years of FU
Disease occurrence 1
YES 3 2 NO 12 10 8 4 11
ID in Exposed = 1 event in 12 person years
ID in Unexposed = 1 event in 18 person years

38. Cohort studies – Measure of Association: Risk Ratios, or Incidence rate ratios
Summary measure of association in Cohort Studies
Formula for Incidence rate ratio (IRR) =
Incidence of disease in persons with exposure
Incidence of disease in persons without exposure
Ndisease/Nexposed per unit time
Ndisease/Nunexposed per unit time
* Note – in IRR there is no unit of time. This assumes the amount of time was similar for those with and without disease and those exposed and unexposed or

39. Calculation of Risk Ratio - example
Cohort at inception: 1,000 people without diabetes
Prevalence of obesity at inception = 22.7%
Outcome: Incidence of diabetes in a population
Exposure - obesity at inception of cohort
Follow-up - six years
Overall incidence of diabetes = 1% per year
Cumulative Incidence = 6%
Risk = cumulative incidence

40. Risk Ratio Calculation - Example
Number with exposure
Developed Diabetes
Cumulative Incidence rate
Obese
227 27
27/227
Non Obese
773 33
33/773
Total
1,000 60
Ratio of Incidence = risk ratio = 27/227 / 33/773
= 12 / 4 =

41. Incidence Density Ratio
Patient
Exposed
Follow up Years
Disease 1
YES 2 NO 10 3 8 4
Incidence rate ratio = (1/2) / (1/2) = 1
Density method = (0/2 years) + (1/10 years)
(0/8 years) + (1/10 years)
Incidence density ratio = (1/12) (1/18)
= 1.5

42. Incidence Rate Difference
A patient asks “How much will my risk of heart attack go down if I take this new drug (B), instead of old one (A)?”
Answer using incidence rate difference
Incidence with Drug A - Incidence with Drug B
= 0.5%/year – 0.3%/year = 0.2%/year, or, a 40% reduction
Same answer using Incidence rate ratio:
= Incidence with Drug B = 0.3% = 0.6, or, a 40% reduction
Incidence with Drug A %

43. Attributable risk
“How many lung cancers are due to air pollution in Montreal?” Same as “What is attributable risk?”
Attributable risk = IRR x Prevalence of exposure
Increases with higher IRR
Or if exposure more common
Diabetes vs Silicosis and TB
Diabetes: IRR = 3.5, Prevalence = 3%
Silicosis: IRR = 12, Prevalence = 0.1%
Attrib risk for Diabetes >> than for Silicosis

44. Cohort Studies – Survival Analysis
Analysis of time to event
Accounts for variable length of follow up.
Advantage if time to event affected by exposure.
Can find important differences in treatments even overall survival same:
Cancer treatment A increases survival at two years
But five year mortality is same as treatment B.
Treatment A - preferred by most patients!

45. Important differences found using Survival analysis

46. Types of Survival Analysis
Simplest – Direct
Kaplan-Meier – still pretty simple. Calculates cumulative proportion free of outcome (survived) at each point in time when that outcome occurs. People who drop out or die of other causes are ‘censored’. At each point numerator is all who have developed disease, while denominator is all without outcome in the interval just before
Cox regression analysis – multivariate analysis with same basic principles

47. Kaplan Meier survival analysis - example
Time
Number at start
During interval
Surviving
at end
Proportion surviving
Drop-outs
Deaths
Interval
Cumulative
100
1.0
3 months 10 90
6 months 70
0.88
10 months 60
0.86
0.75
12 months 40
0.8
0.6
18 months 30
Notes: Intervals are variable – defined by when subjects die
Proportion surviving interval – excludes drop-outs during the interval (censored)

48. Kaplan Meier survival analysis - example

49. Example of Kaplan-Meier analysis: General Hospital Ventilation and time to TST conversion

50. Selection Bias – Berkson’s
This is described in case control studies in hospitalized patients
First described on mathmatical basis.
Probability Hospitalization if Factor Z = 0.1 Probability Hospitalization if Factor Y = 0.05 Probability Hospitalization if both = higher
These two independent conditions will appear to be associated – but may not be.
In practice it is common that patients with 2 or more conditions ARE more likely hospitalized (eg CHF and pneumonia) so in hospital based Case-control study they appear to be strongly associated.
Fundamental problem is the same. P1 does not equal P2 does not equal P3 does not equal P4