cohort and case-control studies
DESCRIPTION
Cohort and Case-Control Studies. September 2013 Alexander M. Walker MD, DrPH With Sonia Hernández-Díaz MD, DrPH. Cohort Study. If you have People Observed for a health event For some amount of elapsed time in every one You have a cohort study. Graphical Representation. Randomized Trial. - PowerPoint PPT PresentationTRANSCRIPT
Cohort and Case-Control Studies
September 2013
Alexander M. Walker MD, DrPHWith Sonia Hernández-Díaz MD, DrPH
Cohort Study
If you have People Observed for a health event For some amount of elapsed time in every oneYou have a cohort study
2
Graphical Representation
Randomized Trial Cohort Study
Person-Time Person-Time
People People
Risks
Yes No
Yes a b n1
no c d n0
Outcome
Exp
osu
re
Risk Ratio = RR = [a/n1] / [c/n0]
Risk(exposed) r1 = a/n1
Risk(not exposed) r0 = c/n0People
Incident cases
Counts of people in different exposure/outcome categories
›
Rashkind Occluder
5
Outcome N=185
Successful closure 160 (86.5%)
Embolization of occluder 18 (9.7%)
Transfusion 26 (14.1%)
Average cost per patient $11,466
Gray DT, Fyler DC, Walker AM, Weinstein MC, Chalmers TC. Clinical outcomes and cost of transcatheter vs. surgical closure of patent ductus arteriosus. N Engl J Med 1993;329:1517-1523
Rates
Yes PT
Yes a t1
No c t0
Outcome
Exp
osu
re
Rate Ratio = RR = [a/t1] / [c/t0]
Person-Time
Incident cases
Rate(exposed) r1 = a/t1
Rate(not exposed) r0 = c/t0
Sum up all the individual times of observation
›
7
BMI and Seizures
8
Seizure Incidence and BMI
9
BMI
Seiz
ure
s p
er
100,0
00 p
ers
on
-years
95% Confidence
Interval
Time-Varying Exposures
Crossover Trial Cohort Study
Person-Time Person-Time
People
Rates with Time-Varying Exposures
Yes PT
Yes a t1
No c t0
Outcome
Exp
osu
re
Incident cases
Sum up all the individual times of observation
Person-Time
Rate Ratio = RR = [a/t1] / [c/t0]
Rate(exposed) r1 = a/t1
Rate(not exposed) r0 = c/t0
›
12
Walker et al. J Am Soc Nephrol 17: 2293–2298, 2006
13
Case-Control Studies
Cohort studies with sampling Instead of the full population denominator, we take
a sample, called the “controls” At random from the cohort members with no event At random from the cohort members At random times from random members Or optionally matched to cases
Personal characteristics Time of follow-up
Captures covariates that would be too expensive to ascertain for the full cohort Time-varying Resource intensive data collection
Sampling Person-Time
Yes PT
Yes a st1
No c st0
Outcome
Exp
osu
re
Person-Time
Take a person at random and take a day at random. If the person is under observation on the selected date, that person-day is a control.
The sampling rate of person-days (controls sampled divided by the number of person-days eligible) is the rate of control generation. It is a rate in the same sense that disease incidence is a rate, except that the control-sampling rate is by construction unrelated to exposure.
Call this sampling rate s.
The Odds Ratio with Cases and Sampled Person-Time
CasesControl
s
Yes λ1t1 st1
No λ0t0 st0
Outcome
Exp
osu
re
Person-Time
Consider the table of expected values, below. The odds ratio of this case-control table of expected values is
OR
1t1st1
0t0st0
10
From which it follows that an estimate of the OR is an estimate of ρ . Note that there is no “rare disease” assumption.
A Case-Control Study with Sampled Person-Time
17
The question at hand was whether inhaled corticosteroids had the effect of weakening bone to the point of making women more susceptible to fracture. Nonvertebral fracture was chosen because vertebral fracture is so often asymptomatic.
Random Sampling of Person-Time
18
Cases, Controls, OR
19
Risk Sets
Persons over TimeRisk Set
Each case is compared to all the people who were at risk to become cases at the time the case occurred.
This collection of persons at risk is called the “Risk Set.”
Risk Set Sampling in Case-Control Studies
Proportional Hazards analysis
Each case is compared to a sample of the people who were at risk to become cases at the time the case occurred.
This sampled persons called matched controls. They have been matched on time and possibly other factors.
Matched case-control analysis
Graham et al. 2005
For every case, we randomly selected four controls from individuals under observation in the study cohort on the date of the case event (index date), and matched them for age (year of birth), sex, and health plan region (north or south). A given cohort member selected as a control for a case on one date could become a control for another case occurring on a later index date, as long as he or she remained in the study cohort and was therefore also at risk of becoming a case.
Graham DJ, Campen D, Hui R, Spence M, Cheetham C, Levy G, Shoor S, Ray WA. Risk of acute myocardial infarction and sudden cardiac death in patients treated with cyclo-oxygenase 2 selective and non-selective non-steroidal anti-inflammatory drugs: nested case-control study. Lancet 2005;365475-81
22
Graham et al. 2005
For every case, we randomly selected four controls from individuals under observation in the study cohort on the date of the case event (index date), and matched them for age (year of birth), sex, and health plan region (north or south). A given cohort member selected as a control for a case on one date could become a control for another case occurring on a later index date, as long as he or she remained in the study cohort and was therefore also at risk of becoming a case.
Graham DJ, Campen D, Hui R, Spence M, Cheetham C, Levy G, Shoor S, Ray WA. Risk of acute myocardial infarction and sudden cardiac death in patients treated with cyclo-oxygenase 2 selective and non-selective non-steroidal anti-inflammatory drugs: nested case-control study. Lancet 2005;365475-81
23
Risk set sampling
Graham et al. 2005
For every case, we randomly selected four controls from individuals under observation in the study cohort on the date of the case event (index date), and matched them for age (year of birth), sex, and health plan region (north or south). A given cohort member selected as a control for a case on one date could become a control for another case occurring on a later index date, as long as he or she remained in the study cohort and was therefore also at risk of becoming a case.
Graham DJ, Campen D, Hui R, Spence M, Cheetham C, Levy G, Shoor S, Ray WA. Risk of acute myocardial infarction and sudden cardiac death in patients treated with cyclo-oxygenase 2 selective and non-selective non-steroidal anti-inflammatory drugs: nested case-control study. Lancet 2005;365475-81
24
Matching in risk-set sampling is equivalent to sampling from strata, where the strata definitions are the levels of the matching factors.
Graham et al. 2005
For every case, we randomly selected four controls from individuals under observation in the study cohort on the date of the case event (index date), and matched them for age (year of birth), sex, and health plan region (north or south). A given cohort member selected as a control for a case on one date could become a control for another case occurring on a later index date, as long as he or she remained in the study cohort and was therefore also at risk of becoming a case.
Graham DJ, Campen D, Hui R, Spence M, Cheetham C, Levy G, Shoor S, Ray WA. Risk of acute myocardial infarction and sudden cardiac death in patients treated with cyclo-oxygenase 2 selective and non-selective non-steroidal anti-inflammatory drugs: nested case-control study. Lancet 2005;365475-81
25
Take Home Lessons: Cohort Sampling
All case-control studies can be seen as being samples drawn from specifiable cohorts.
Different modes of cohort analysis give rise to corresponding case-control designs. Closed cohort, fixed folow-up Sample noncases Variable follow-up time, time-varying exposure,
stable baseline incidence Sample person-time Variable baseline incidence Sample risk sets
Case-control analysis are directly tied to the corresponding cohort analysis, with further allowance for sampling
“Case-control” studies that do not use cohort sampling are valid only to the extent that the procedures used approximate cohort sampling.
26
Case-Control Advantages
Relatively inexpensive Relatively quick Particularly useful for rare outcomes Closely logical connection to cohort designs
However, you have to bear in mind that Focus on a single outcome can be very misleading
when overall cost and benefit is the real questions Poorly conceptualized studies can taint the field Negative attitudes that are a holdover from old
hierarchies of evidence that do not reflect modern understanding
Thank You!