Term
| what is recall bias? When it most important? |
|
Definition
> those with disease are more likely to recall an exposure -- people searching for any explination
> It is most signficant in retrospective studies such as case control studies. |
|
|
Term
|
Definition
> Innocent exposure leads to greater likelihood of detecting disease of detecting disease
- Estrogen use →bleeding -> endometrial
cancer detection
- Aspirin use → bleeding → colon cancer detection
>>the probability of detecting disease is related to exposure status |
|
|
Term
| diagnositic suspician bias |
|
Definition
>because of diagnositic suspician, some cases will be detected earlier than others due to more aggressive surveillance
>>the probability of detecting disease is related to exposure status |
|
|
Term
|
Definition
> Surveillance system case ascertainment may not always be complete
> If completeness of case ascertainment differs according to a causal exposure, estimates of the association between the exposure and the disease will be biased
Example:
– EXPOSURE = Socioeconomic status (SES) – DISEASE = Arthritis (ascertained by physician diagnosis)
– SES → access to medical care → biased arthritis diagnosis |
|
|
Term
| Diagnostic (Classification) Bias |
|
Definition
-> The probability of being classified as having the disease is related to EXPOSURE status
Example:
- EXPOSURE = Hypertension
– DISEASE = Stroke
– Case review: if hypertensive -> then reviewer more inclined to classify a suspected stroke as a definite stroke |
|
|
Term
| properties of a confounder |
|
Definition
1. Must be a cause of the DISEASE or, at least a marker (surrogate) of an actual cause of the DISEASE
2. Must be distributed differently in the EXPOSED and UNEXPOSED (dataset-related phenomenon)
3. Cannot be an intermediate step in the causal pathway between exposure and disease
EXPOSURE → Factor A → DISEASE
--> a nuisance association; artifact caused by a an incidental correlation between two variables (think maternal age and birth order) |
|
|
Term
| When is the odds ratio a good estimate of the cumilative incidence ratio? |
|
Definition
1) cases are representative of all people with the disease in the population from which the cases are drawn
2) control are representative of all people without the disease in the population
3) the disease being studied is rare (less than 10%) |
|
|
Term
|
Definition
| Pet theory being promated by the participant |
|
|
Term
| random (non-differential misclassification) |
|
Definition
mis-classification of exposure status NOT based up on disease status (same magnitude of mis-classification of exposure in diseased and non-diseased individuals)
causes bias towards the null |
|
|
Term
|
Definition
aka unnecessary matching
can cause loss of precision due to making the cases and controls artificially close |
|
|
Term
|
Definition
- most familiar type of case-control trial
- best when the source cohort is ill-defined or dynamic
- controls can be thought of a sample of the survivors at the end of a follow-up period
(a/c)/(b/d) --> OR |
|
|
Term
|
Definition
- best for a well defined source cohort where there is little variation in follow-up time between subjects
- controls can be thought of as a sample of the source cohort at the beginning of the follow-up period
(a/t1)/(c/t0) = (a/c)/(t1/t0) -> CIR
cases may also be controls in this design, especially in a prevalent disease
Since the cases can be controls, you can use the control group repeated for different diseases, getting around the usual limitation of case-control designs --> only one disease |
|
|
Term
|
Definition
- used in case-control studiew with a well defined source cohort and variable follow-up time
- controls can be thought as random sample of the person-time of the source cohort.
y1 = person-time in exposed
y0 = person-time in non-exposed
(a/c)/(y1/y0) -> IRR
controls are select from the instantaneous time period in which each case occurs
- controls may be select multiple times
- a control may be case in a later set
|
|
|
Term
|
Definition
- good for studying the influence of brief exposures on a risk of onset of acute incidents
- cases serve as their own controls
--> Were you doing anything unusual just before this happened?
- used for food poisoning investigations |
|
|
Term
| what is selection bias? what are the types? |
|
Definition
two types:
1. selection into: those who are selected (diseased and controls) are different than those not selected.
2. loss to followup: those who drop out or die are different than those who remain in the study |
|
|
Term
|
Definition
| compute measures of association separately for each strata of the confounder |
|
|
Term
|
Definition
Compute summary estimate that is mathematically manipulated to
account for differential
distributions of the confounder
across 2 populations
– creates ‘comparable’ rates/risks
used primarily for age distributions in populations |
|
|
Term
| pooling (Mantel - Haenzsel) |
|
Definition
compute measures of association separately for each strata of the confounder then pool into a summary estimate
--assumes same effect for each strata
--harder to compare across studies
[image] |
|
|
Term
| effect measure modification |
|
Definition
aka interaction
--heterogeneity of effects: the effect of a risk factor in strata formed by another variable (an effect modifier)
--comparing observed and expected joint effects of the risk factor on the outcome
-implies a different or joint relation at the biological level. (think smoking and asbestos exposure and lung cancer) |
|
|
Term
|
Definition
conclude that a risk factor is associated with outcome when we shouldn't
-- related to p-value |
|
|
Term
| four types of measurement error |
|
Definition
--subject-related
-person
--environment-related
-white-coat hypertension
--observer-related
-person using the instrument or making the assesment
--instrument-related |
|
|
Term
|
Definition
(Io - Ie)/(1 - Ie)
-- proportion above and beyond expected by chance |
|
|
Term
|
Definition
| odds of being diseased among the exposed divided by the the odds of being diseased in unexposed. |
|
|
Term
|
Definition
| odds of being exposed among the diseased divided by the odds of being exposed among the non-diseased |
|
|
Term
| definition of effect modification |
|
Definition
incidence of a disease in the presence of 2 or more risk factors differs from the incidence expected to result from their individual effects
--the may be additive (use absolute measures of association) or multiplicative (use relative measures of association) |
|
|
Term
| assessing effect measure modification |
|
Definition
|
|
Term
|
Definition
| Making a generalization about a larger group of individuals (population) on the basis of a subset or sample |
|
|
Term
| quadrants of hypothesis testing |
|
Definition
|
|
Term
| When is hypothesis testing the most useful? |
|
Definition
| When a decision should be made between two course of action. It becomes much harder for nuanced decisions between more than two alternatives. |
|
|
Term
| name two criticisms of hypothesis testing |
|
Definition
1. inferences are more complex than yes/no dichotomous decisions
2. hypothesis testing provides no sense of the magnitude of the assocation
3. provides no proof of the clinical or biological significance of the association
4. rejection of the null hypothesis does not imply that you can embrace the point estimate
5. failure to reject the null hypothesis does not imply that one can reject the point estimate
|
|
|
Term
| why is the p-value better (and worse) than hypothesis testing? |
|
Definition
| the p-value encorporates the strength of association and role of chance in one number
-- encorporates more data but hides it.
[image] |
|
|
Term
| 95% Confidence Limit Ratio |
|
Definition
ratio of the 95% confidence limits
--if the confidence limits are (0.1 - 2) the ratio would be (2/0.1 = 20)
-- it is a measure of the precision of the estimate and lower the better |
|
|
Term
| decision tree to decide if a variable is a confounder or an effect measure modification |
|
Definition
|
|
Term
| what biologically does an effect measure modification imply |
|
Definition
| It mean that the two factors noted are interacting biologically to cause the disease |
|
|
Term
| statistical implication of effect measure modification |
|
Definition
| the mathomatical model a causal association between exposure and disease must be changed to include both factors |
|
|
Term
| public health implication of effect measure modification |
|
Definition
| it allows a better targeting of resources for disease prevention even if there is no biological explination of the assocation |
|
|
Term
| individual implication of effect measure modification |
|
Definition
| allows an individual to assess their individual risk better: if you have asbestos exposure don't smoke. |
|
|
Term
| study design implications of effect measure modification |
|
Definition
| Since most studies are only powered enough for the main assocation, they need to add subjects or only expect to see evidence of strong effect measure modification |
|
|
Term
| Attributable proportion in the total population |
|
Definition
That fraction of all cases in the total population that are a result of the
exposure.
(CI - CI0)/CI
where CI = crude cumulative incidence and CI0 = cumulative incidence in the unexposed |
|
|
Term
| attributable risk (aka cumulative incidence difference) |
|
Definition
amount of risk attributable to an exposure
CI1 - CI0
where CI1 = cumulative incidence in the exposed and CI0 = cumulative incidence in the unexposed |
|
|
Term
|
Definition
| (# of new cases)/(population at risk) |
|
|
Term
|
Definition
| (# of deaths)/(population at risk) |
|
|
Term
|
Definition
| (# of deaths)/(# of cases) |
|
|
Term
|
Definition
Null hypothesis is not true but you do not reject the null hypothesis.
resolved by adding statistical power to the study |
|
|
Term
| calculating the observed vs expected result on the additive scale |
|
Definition
|
|
Term
| Describe how you can use relative measures of association in the additive model |
|
Definition
- calculate relative measures of assocation and normalize to the group uneffected by either the exporsure or effect modifer
- create a 2 x 2 table filled with these relative measures
- use observed vs expected using the following formulas:
- observed OR = ORxy
- expected OR = ORx + ORy - 1 (to prevent double counting the background risk
|
|
|
