Clinical Epidemiology Flashcards
Internal validity:
The degree to which the results obtained from a study sample reflect the truth.
Truth:
Defined as the results that would be obtained if everyone in the source population participated in the study, and all of the measures were completely accurate.
Source Population:
The population from which the sample was drawn
External validity:
(sometimes called generalizability) is the degree to which the results of the study are applicable to other (external) populations that are not even being sampled. This is an issue of judgment, and it depends on knowledge of the subject matter.
If the characteristics that predict the outcome (disease) or response to treatment are dramatically different between the sample and some external population, then the study has _____ external validity for that specific external population.
If the characteristics that predict the outcome (disease) or response to treatment are dramatically different between the sample and some external population, then the study has LOW external validity for that specific external population.
External validity in a clinical trial (an intervention study) is also based on the degree to which the intervention itself can be ________
External validity in a clinical trial (an intervention study) is also based on the degree to which the intervention itself can be REPLICATED
e.g. PTSD treatment with CBT by trained psychiatrists vs. untrained
Bias is:
Systematic error; the systematic difference between the observed findings and the truth
The greater the bias, the _______ the internal validity
The greater the bias, the LOWER the internal validity
Magnitude of Bias:
The magnitude of bias is the size of difference between the observed effect and the truth.
Direction of Bias:
Based on the relationship among: a) the truth; b) the observed findings; c) the null value. If the observed findings are farther from the null value than the truth is from the null value, then the bias is away from the null.
Confounding Bias:
A distortion in the results due to a mixing of the effects of exposure with the effects of some third, extraneous, factor. Confounding occurs when some third factor is associated with both the exposure and the disease.
Confounding is _____ a threat to the validity of observational studies
ALWAYS; because exposures which cause disease do not occur randomly
What are examples of observational studies?
- Cohort studies,
- case control studies,
- cross-sectional studies
3 ways that confounding can be prevented in the design of a study:
- Randomization
- Matching
- Restriction
Randomization:
If an exposure is randomly allocated to study participants, then any potential confounding factor, such as age, ethnicity, etc, should, on average, be distributed equally across the intervention groups. If a factor is equally distributed across intervention groups, it cannot be a confounder.
What is the best way of preventing confounding in human research?
Randomization
Matching:
If participants are matched on a potential confounding factor, then the distribution will be equal across groups and that factor cannot confound the association.
Restriction:
If sex is a potential confounder, and a study is restricted is women, then sex cannot be a confounder.
Stratified analysis:
In a stratified analysis, the data are analyzed separately in each level of the confounding variable, and then combined. For instance, in the study of gray hair and heart disease, the data would be analyzed separately in different age groups, and then mathematically combined to produce an overall result which is not confounded.
Regression analysis:
A statistical model; produces results in which the potential confounding factors are statistically controlled. One of the major advantages of modeling techniques is that MANY confounders can be controlled at the same time.
Selection bias:
A distortion in the results due to a difference between subjects who participated in a study and those who are theoretically eligible but did not participate (i.e., the source population).
Selection bias can develop in 3 ways:
- The way that subjects are selected for inclusion in a study (e.g. coffee related to pancreatic cancer, but all controls were also GI patients who reduced coffee),
- Refusal of eligibles to participate (e.g. people completing surveys are either thrilled or angry),
- Withdrawal of loss to follow-up of research subjects in a cohort study or randomized trial
What is the difference between confounding and selection bias?
-In confounding, the association between the extraneous factor and the exposure and between the extraneous factor and the outcome exist in the source population. Selection bias is the distortion in the association between an exposure and a disease when the procedures used to select or maintain the sample create differences between the exposed and unexposed groups on factors that are also associated with the disease.
Problem with selection bias:
it is frequently difficult to identify the imbalances in risk factors that are created by the processes used to create and maintain the sample, so it is difficult to control for.
How is selection bias different from external validity?
In selection bias, the issue is whether the results are distorted due to differences between the groups being compared that arose from the processes used to create the sample. For external validity, the issue is whether there are differences between the results in the sample and the results that we would expect in some external population.
Measurement bias:
A distortion in the results (e.g., the relative risk for the association between an exposure and a disease) due to error in measurement of the exposure, or measurement of the disease, or both.
Nondifferential measurement bias:
Occurs when errors in the measurement of exposure (second hand smoke) are independent of the disease (sinusitis), or when the errors in the measurement of disease (sinusitis) are independent of exposure. e.g. if people tend to under-report their exposure to second hand smoke, but that people who don’t have sinusitis under-report second hand smoke to the same extent as people who do have sinusitis.
In general, nondifferential measurement bias produces bias _____ the null (towards/away from)
In general, nondifferential measurement bias produces bias TOWARDS the null
Differential measurement bias:
Occurs when errors in measurement of exposure depend on the disease, or when errors in measurement of disease depend on the exposure. e.g. If persons who had sinusitis were less (or more) likely to under-report exposure to second hand smoke than persons without sinusitis, then there would be differential measurement bias of exposure.
Differential bias in measurement of exposure is more likely to occur when:
the disease status of the subject is known prior to measurement of exposure, as in a case control study.
Differential bias tends to produce bias _____ the null (towards/away from)
Both towards or away from - the direction of the bias depends on the particulars of the situation
Randomized trials are less likely to be plagued by ________ than ______ or _____ because the randomization process creates groups which should be similar for all factors which can affect the outcome.
randomized trials are less likely to be plagued by CONFOUNDING than COHORT STUDIES or CASE CONTROL STUDIES because the randomization process creates groups which should be similar for all factors which can affect the outcome.
(***less likely, not impossible)
Cohort Studies:
track people forward in time from exposure to outcome.
Case-control studies:
Work in reverse, tracing back from outcome to exposure.
Cross-sectional studies:
Like a snapshot, which measures both exposure and outcome at one time point.
Ratio:
a value obtained by dividing one number by another. (can either contain the numerator in the denominator or not, depending on study)
Rate:
Measures the frequency of an event in a population. Contains numerator in the denominator, and uses time and a multiplier. E.g. 11 cases of TB per 100,000 persons per year
Proportion:
Like a rate, must have the numerator contained in the denominator. But is unit-less, and does not have a time component.
E.g. prevalence - 27 of 100 at risk have hay fever.
