Upload
barnard-martin
View
212
Download
0
Embed Size (px)
Citation preview
Case-Control
Studies
November 18 2004
Epidemiology 511W. A. Kukull
History
• Lane-Claypon (1926) first case control study: reproductive experience and breast ca
• Sociology used case control methods in 1920’s and 30’s
• Wynder& Graham (1950) and others linked cigarette smoking to lung ca
• Cornfield (1951) direct standardization to control confounding
• Mantel and Haenszel (JNCI, 1959) stratified analysis
Case Control Studies
Cases(with Disease)
Controls(no Disease)
Hx ofExposure
Hx ofExposure
Not Exposed
NotExposed
Study Base orPopulation
All New Cases of Disease “X”that meet studycriteria
“Sample” ofpersons withoutDisease “X”(Controls)
Population-based, Incident, Case-Control Study
“Controls” must be able to become diseased and must besampled without regard to “exposure”.
“Study base” concepts
Eligible Subjects
Cohort Studyenrollees
Follow-up Time
Loss, death, refusalsbefore disease develops
Disease cases
Non- diseased
Enroll all?
Sample?
Bias?
Case - Control
Nested Case Control
Cohort study population:Exposed and Unexposed
Developeddisease
Did Not develop Disease
Cases“Controls”
Sampled or Selectedfrom the remaining unaffectedwhen each case is diagnosed
Use data or biomarkers collected at cohort entry to determine exposurestatus for cases andcontrols
Selecting Cases
• Must cases be “representative” of all persons who have the disease?– What about female cases?– Severe cases?– early onset cases?– Cases from Omaha?
• They must be selected independent of exposure !! (and define a study base)
Principles of Comparability(after Wacholder et al, AJE;1992;135:1019-28)
• Case-control comparisons should be made within subjects from the same study base; (selection bias)
• Effects of other factors on the disease-exposure association being studied should be minimized; (confounding)
• Errors in exposure measurement should be non-differential; (information bias)
Case – Control comparability(after Koepsell & Weiss, 2003)
• Comparability and “representativeness”– If each of the study cases had not developed the
disease, would they have been included in the study base/population?
– If each of the non-cases in the study base/population had developed the disease, would they have been included as a case?
• Can we characterize the study base/population?
Representativeness?Ambiguous interpretation
• Cases may be restricted to any type of case– The case definition will define the “study base”
or source population for controls
• Cases do not need to be a “random sample” of the entire diseased universe to be valid– Case selection and inclusion criteria will affect
the research questions that can be answered
Selecting Cases
• Disease criteria: clinical or histopathological evidence?
• Hospitalized cases or cases from a registry?– may need to include more than one hospital
• Incident or Prevalent cases ?– survival bias? Prevalent case sample may miss
persons dying early in the course of disease– Health services factors plus risk factors
Cohort design similarities
• Suppose we could find all new cases of ALS in a particular town as they occur
• Could we conduct a cohort study ?– We know the population of the town– How would we measure exposures for
everyone?
• Could we take a “sample” of the persons without ALS and compare them to cases
Sources of Controls:Are they part of the same study base?
• Community or Population-based– RDD
• Friend or spouse • Neighborhood• Hospitalized patients
– unrelated to exposure; multiple diagnoses• Medicare and government lists• HMO enrollees
Timing
• Specify a reference time– e.g. diagnosis for cases, similar time for control
• Determine exposure before reference time– later ones don’t count
• What if enrolled controls later get the disease of interest?– Controls at enrollment are compared to cases at
enrollment
Exposure and Onset: how we think of onset influences potential relevant exposures
Biologic onset ofdisease
Disease Detectable by Screen
Sx/Dx
Outcome
Potentially effective Exposures (Critical period)
IrrelevantExposures
???
Comparable exposure periods for controls? Setting a “reference age/time”
Case age at biologic onset ofdisease
Disease first Detectable by Screen
Case age at
Sx/Dx
Outcome
Potentially effective Exposures (Critical Period)
IrrelevantExposures
???
Age/time
Proxy RespondentsWhen the subject can’t respond
• Spouse, sibling, child, friend
• Use for both case and control – when proxies systematically over or
underestimate exposure– when control responses and their proxies are
poorly correlated– when there is no information relating proxy and
subject responses
Choosing controls (1)
• We want to study computer use as a risk factor for carpal tunnel syndrome among 16 yo women
• We find cases through the hospital neurology clinic
• We enroll the best friend of each case (same age and sex), who has no symptoms, as a control—for a paired case-control design
Choosing controls (2)
• We want to study the effect of smoking on carotid artery stenosis/occlusion
• Cases selected from UWMC vascular clinic – carotid doppler > 70% stenosis
• Controls are selected from UWMC pulmonary clinic – carotid doppler : <20% stenosis
• Determine smoking Hx for each– What would we expect to find
Choosing controls (3)
• Selection Bias– controls representative of the study base?– selection related to exposure hx?
• Information bias (recall and other)– is the exposure measured with the same
accuracy in cases as in controls
• Residual confounding: unmeasured factors• Statistical power
Example: study base(after MacMahon and Trichopoulos)
• Case control study of induced abortion and subsequent ectopic pregnancy– Cases: 26 women with EP and one previous
pregnancy – 3 controls for each case from same maternity
hospital (matched on age, education and pregnancy order)
• Result : odds ratio of 10.0
Example: IA an EP (2)
• Were controls representative of the case “study base”?
• Controls were generally completing pregnancy (in general, women completing were less likely to have had an IA)
• Cases of EP is diagnosed early in gestation
• Later study, with only “new” pregnancy controls showed OR= 1.9, not 10.0
Measuring Exposure
• Recall Bias– Are cases more/less likely to recall exposure
than controls?
• Limitations of recall– Is the person’s recall valid?
• Comparability in cases and controls
• Validity (accuracy) of measures
Comparability of Exposure Information
• Does incompleteness or inaccuracy occur to a different degree in cases vs controls?– Same degree?
• What would happen to the Odds Ratio in either case?– attenuation (reduction toward null value)?– exaggeration?– spurious association?
Comparability Considerations: Exposure Measurement
• Provisions for “sensitive” questions (illicit drug use, sex, income)
• Biologic specimens: lab methods, storage degradation
• Excessively long interviews– psychological testing– food frequency– occupational and extended family history
Obtaining ComparabilityExposure measurement
• Keep staff unaware of hypothesis and case/control status
• Place and circumstances of interview– equal proportions of cases and controls
interviewed by each staff member
• Use information recorded prior to time of diagnosis (hospital or pharmacy records)
• Direct Measurement (EMF, radiation)
Evidence of Comparability
• Is there similar proportions of “missing” data in cases and controls?– Does the time duration of interviews differ?
• Ask about etiologically irrelevant characteristics and assess response– e.g. if studying pancreatic ca and coffee ask
also about tonsillectomy and hemorrhoids
• More than one source
ExampleTrue classification
135
95
50
180
MI No MI
Illegaldrug
No Illegaldrug
OR= 5.1230 230
Example:Reporting accuracycases .90; controls .20Differential misclassification
112
118
10
220
MI(case)
No MI(control)
Illegaldrug
No Illegaldrug
OR = 20.8 230 230
Example:Reporting accuracycases .20; controls .20
Non-Differential misclassification
27
203
10
220
MI(case)
No MI(control)
Illegaldrug
No Illegaldrug
OR = 2.9 230 230
Example summary
• Differential Differential misclassification may bias Odds Ratio in either direction
• Non DifferentialNon Differential misclassification usually biases toward the NULL (1.0)– under some circumstances it may be biased
away from the null– Don’t always trust it to underestimate true
effect
Example: Recall Bias Comparability of information
• We are studying prenatal maternal infection and congenital malformations.
• True incidence of infection is 15% for both cases and controls
• IF Case mothers recall 60% or their true infections; controls recall 10% of theirs
• Results will show 9% infection rate in cases and 1.5% among controls (OR ~ 6.0)
Matchingto reduce potential confounding by design
• Individual matching requires a “matched” analysis– pairs are the unit of analysis OR= b/c
• Frequency matching uses a standard, stratified or unmatched analysis– May require that cases be enrolled before
controls
Indications for Matching
• If the unmatched groups have little overlap on the factor (and the factor is associated with disease)
• Small studies of rare diseases with several confounding variables
• To account for unmeasured confounders, through a surrogate measure, e.g., neighborhood
• Especially, when the matched factor is a STRONG confounder
MatchingPotential problems
• Finding a match for the 93 y/o man from Ballard, who was a fisherman, is married, has five children, and drinks socially– Age, neighborhood, occupation, marital status,
offspring, alcohol use—too much matching?
• Once you have “matched” on a factor you cannot study that factor– Why? because we have artificially established equal
proportions of that factor in the cases and the controls
OverOvermatching
• Matching on a variable intermediate in the causal pathway– Suppose smoking alters cholesterol, and
cholesterol is associated with CVD– What if we matched on cholesterol?
• Don’t match on factors related to exposure of interest but not to disease– Contraceptives, religion, -> embolism
Case Control Studies:Limitations
• Inefficient when exposure is rare
• Cannot compute incidence rates directly
• Sometimes difficult to establish temporal relationship between exposure and disease
• Prone to biases:– Selection bias– Information bias, specifically RECALL bias
Case Control Studies:Strengths
• Relatively quick and inexpensive
• Good for diseases with long latent periods
• Optimal for RARE diseases
• Can examine multiple etiologic factors for a single disease
Conclusion
• Define a study base• Select diseased and non diseased persons• Measure history of “exposure” • Compare exposure hx in cases and controls• Assess possibility of bias
– misclassification and non-comparability of exposure data
– inappropriate study base sampling; timing