Epidemiology (VIII) Second Half of Class Flashcards
Attributable Risk, specific study usage, bias and confounding, etc.
Why do we prefer to use incident cases in case-control studies? What are the cons?
Pros:
- More representative of all persons w/disease
- Exposure is more likely to have preceded the disease
- Recall may be better
Cons:
-Gotta wait for incident cases to occur
- May still miss people who die before diagnoses
True or False: In a case-control study, the prevalence of the disease reflects the prevalence of the disease in the population.
False.
We select as many cases as possible, then an arbitrary # of controls to match. However, the prevalence of EXPOSURE in the D- and D+reflects the actual levels in the reference pop
Two rules about selecting controls:
1) They must come from the same reference population as the cases
2) Their exposure prevalence should be representative of the source population –> increase sample size!
Group matching vs Individual matching in case-control studies
Group matching: frq matching or stratification; select controls so that it matches the distribution of the factor in case
Individual matching: ‘matched pairs’
What study should you use if you want to…
- Study common diseases or common exposures
- Study many exposures & outcomes
General Cohort Study
What study should you use if you want to…
- Study rare exposures
- Study many outcomes
Specific Cohort Study
What study should you use if you want to…
- Study rare diseases
- Study many exposures
Case-control study
What study should you use if you want to…
- Study disease or exposure burden
- Don’t have time to do follow-up
Cross-Sectional
What study should you use if you want to…
- Identify correlations and generate hypotheses
- Don’t need individual-level exposure measurement
Ecologic Study
What measure of disease frequency is a direct measure of risk?
Cumulative Incidence is a direct measure; whereas, incidence rate is only an estimate.
When does OR provide a good approximation of relative risk?
When the cases’ and controls’ exposure prevalence is representative of the source populations’ exposure prevalence.
When the disease studied is relatively rare.
Ways to interpret an OR
“The odds of cases having been exposed is _x the odds of controls having been exposed”
“The odds of developing disease among people who were exposed is _x the odds of developing disease among those who were not exposed”
“The risk of disease is _x greater among the exposed compared to the unexposed.”
AR Total population takes into consideration both
Excess incidence associated with the exposure (additional risk above background)
Prevalence of exposure-How common is the exposure in the population
Case control vs Cohort - what’s the difference?
Cohort studies look at people’s exposures and then follows them over time (exposure -> outcome).
Case control studies look back at diseased and nondiseased people’s previous exposures to identify risk factors for that disease. (outcome -> exposure)
Selection Bias- What is it and what is the result o fit?
Systematic error in the selection/follow up of subjects. Thus, subjects are not representative of the source population with respect to exposure or outcomes. RR/OR among the study participants differs from that in the source population.
Ex) Nonexposed controls are more likely to participate than exposed controls.
Ex) Sampling from an area with higher lvls of exposure
Information Bias (Misclassification)
Systematic error in the measurement of exposures and/or outcomes
Ex) Subjects tend to underreport their drug use
Improper selection of subjects or nonparticipation are causes of selection bias in what types of studies?
Case control or cohort.
Selection bias is more likely to occur in case-control and retrospective cohort studies because both exposure & outcome have already occurred by the time of subject selection.
Losses to follow up (must be differential) causes selection bias in what types of studies?
Cohort or RCT
3 ways to prevent selection bias
- Representative sampling
- Maximize participation
- Minimize loss to follow up
Healthy worker effect
A form of selection bias.
People who work, esp in labor-intensive occupations, are healthier than the general population –> results in underestimates of harmful exposures.
Information bias- What is it and what are the types?
Collection of inaccurate information on study factors (exposure, disease, confounders).
Can be introduced by interviewers, participants, and instruments.
Recall bias, interviewer bias, differential and nondifferential misclassification.
Nondifferential misclassification
Errors in exposure or outcome status occur with approximately equal frequency in groups being compared.
Includes…
- Equally inaccurate memory of exposures in cases/controls.
- Recording and coding errors in records and databases, but equally in cases & controls.
- Using surrogate measures of
- Recall bias is equal in both cases and controls
Differential misclassification
Errors in exposure and outcome occur with UNEQUAL frequency in the groups being compared (cases/ctrls; exp/nonexp)
Includes differential recall (e.g. rumination or social desirability bias); case records being collected differently from ctrl records; etc.
How to avoid information bias
- Masking interviewers and subjects to the hypothesis
- Using a control group composed of diseased people
- Careful designing questionnaire
- Not using an interview
- Using multiple measurements, the best info sources, and sensitive & specific criteria to define exposure and disease (prevents misclassification)
Nondifferential selection bias in a case control study
Participation is related to disease, but not exposure.
Unbiased.
Ex) Way more cases participate than controls, or vice versa.
Differential selection bias in a case-control study
Participation is related to both disease and exposure.
Biased.
Ex) Way more cases and nonexposed people participate.
Nondifferential selection bias in a cohort study
Losses to follow-up are related to disease, but not exposure. –> RR is unbiased.
Ex) Only cases, but no controls get lost to follow up.
Differential selection bias in a cohort study
Losses to follow up is related to both disease and exposure. –> RR is biased.
Ex) Exposed and cases are more likely to be lost to follow up.
Information bias in a case control or cohort study has what effect on the OR? (hint: depends on differential or nondifferential)
Nondifferential exposure misclassification: the error is related to disease, but not exposure –> OR/RR biased toward the null.
Differential exposure misclassification: error is related to disease and exposure –> OR/RR biased toward or away.
Bias towards the null means
The observed value is closer to H0: RR=1 than the true value.
True RR is above 1 & observed RR is below it.
or
True RR is less than 1 & observed RR above it.
Bias away from the null means
The observed value is farther from H0: RR=1 than the true value.
Requirements for an exposure to act as a confounder
1) The exposure (e.g. age) must be related to the exposure (e.g. country of residence) under study.
Ex) People from this country are older; Drinkers are more likely to smoke
2) The exposure must also be associated with the outcome.
Ex) Smoking is a risk factor for lung cancer. PTSD is a risk factor for suicide.
3) The confounder can’t be on the causal pathway between the exposure & disease.
Ex) Drinking doesn’t CAUSE smoking; whereas, hand-to-hand combat CAUSES PTSD.
Detecting Effect Modification (Interaction)
Homogeneity of effect: Are the stratum-specific measures association similar to each other?
Combined vs independent effects: Is the combined effect similar to what would be expected based on the independent contributions of each exposure alone?
If either of these are ‘no’, then there may be interaction.
Additive vs Multiplicative interaciton
Additive interaction: the difference in ABSOLUTE measures of risk between exposed & non exposed differs across strata formed by a 3rd variable.
- The combined incidence is higher/lower than the independent effects of each of the two exposures.
- compare risk differences
Multiplicative interaction: the RELATIVE risk between exposed & nonexposed differs across strata formed by a third variable.
- The combined relative risk is higher/lower than the independent effects of the two exposures.
- compare relative risk (risk ratio, relative ratio, odds ratio)
How to deal with confounding
During the design phase:
- Randomization
- Matching (removes possibility of confounding by that variable, but it can be inefficient, youcan’t study that variable anymore, you may accidentally match on other variables too)
- Restriction (exclude any confounding variables, but also limits generalizability)
During the analysis phase:
- Stratify the OR or RRs
- Direct adjustment: remove the effect of the confounding variable (Regression analysis)
Confounding variables are true associations- they’re just not causal. They are risk factors, so they can be used to
Identify an at-risk population
Effect modification
The operation of one or more factors to produce an effect; the incidence rate of disease in the presence of multiple risk factors differs from the incidence rate expected based on their individual effects.
Interaction
The magnitude of the association between the exposure and disease depends on a third factor; the AR or RR/OR is different across strata of a third variable!