Final Exam Flashcards
Selection Bias
“..distortions that result from procedures used to select subjects and from factors that influence study participation”
Results in a measure of association between exposure and outcome that is meaningfully different from the association had all subjects been included in the analysis
Confounding
A mixing of effects/ distortion of the exposure-disease relationship due to the effect of a third variable
Is it a confounder?
Associated with outcome
Associated with exposure
Not a causal intermediate between the two
What kind of bias is misclassification bias?
Misclassification is a form of information bias. Self-reported data frequently raises information bias concerns
Non-differential misclassification
Classification errors do not depend on the subjects’ values for any other variables.
E.g. misclassification of exposure is not dependent on outcome status and misclassification of outcome is not dependent on exposure status
Example: you use a bad measurement tool, but use it on everyone. Likely there is measurement error and it will be non-differential.
If there is a large RR in the paper and you say this is due to non-differential misclassification due to use of a bad measurement tool for example, this would be WRONG. Because we would expect the RR to be larger than what is reported.
What direction does non differential misclassification bias go?
Results in bias toward the null
differential misclassification bias
misclassification of one variable depends on the values of other variables
Example: Recall bias – mothers of children with birth defects may be more likely to report/remember all exposures during pregnancy while mothers of children without birth defects are more likely to under-report exposures due to not remembering them
What direction does differential misclassification bias go?
any direction
What are the causes of selection bias and give an example of such in reproductive epi
Participation bias, Loss to follow-up
Issue those lost are systematically different than those kept
Repro epi example: “Unhealthy” worker effect
5 Methodologic Issues to Always Consider
Confounding
Detection Bias/Diagnostic Bias
Outcome assessment differential biased on exposure status
Ex. Study will catch more miscarriages among women whose pregnancies are diagnosed earlier Goldhaber
Participation Bias (type of selection bias)
Individuals who participate are different than those who do not
Impacts internal validity of the study
If not representative of the population to which you want to generalize results, participation bias can lead to external validity issues
Recall Bias (type of information bias)
Sample Size/Power
Not a type of bias, but something to keep in mind.
In experimental studies, randomization makes confounding less likely. Why is that?
Randomization makes confounding less likely because when you randomize each individual has an equal probability of being in either group so you are more likely to get two groups that are similar except for exposure. However still opportunity for confounding that by chance one group has more obese groups that would confound results bc they have higher risk for high blood pressure, impacting results … which is more likely in the case of a small sample size. The bigger sample size will help reduce this likelihood of confounding. But it can always occur by chance.
Why do observational studies have increased risk of confounding?
Risk of observational: no control and people exposed for different reasons which increases risk for confounding.
Intention to Treat (ITT) biases in what direction?
using this type of analysis biases toward the null (see a smaller effect) because it includes participants who are not adherent as well
**THIS IS THE BEST TYPE because although not completely unbias (bc adherenece is not 100%) but if there is any bias, it is a bias toward the null and generally in science we want to bias ourselves in a conservative way bc we do not want to see an association when there is not one.
When do we do ITT and why?
At allocation, because this is where we randomize
Flaw of the Goldberg calcium RCT in regards to internal validity
Did not really do any ITT or analysis of the full enrollment. And there are more in the treatment than in the placebo group. They were not treated for the full pregnancy length so a decision to exclude them based on adherence so that creates potentially greater bias. It would be better to include the data than just exclude them.
Not taking BP pre-pregnancy – randomization should reduce imbalance of women with preexisting hypertension
Exclusion of misclassified gestational age – definitely could create bias esp if they had risk factors for BP could bias toward the null because they took them out.
Example of selection bias in a study on anaesthetic gas & miscarriage
send survey to nurses asking for reproductive history and if they work in OR or were exposed to AG
Finding: rate was 2X as high as women unexposed and statistically significant
Issue: women who were exposed with normal pregnancies responded at much lower rate
Why is recall bias more common in studies of miscarriage and birth defects?
It is not that women exaggerate but rather they recall more completely bc they are more likely to recall than women who have healthy birth and not likely to recall something they do not think is important
Direct correlation of reporting spontaneous abortion and when it occurred. The earlier it occurs, the less likely to report. Also link between how far back in time you are asking to recall – further back you go the less accurate the memory. (2 separate effects: gestational age and recall time)
*Don’t use the word “overreport” for recall because the case is actually that control participants UNDERREPORT bc their recall may be different and more difficult bc of the normal outcome
Why was there diagnostic bias in the goldhaber study?
Time of diagnosis!
The VDT test group as a whole got their pregnancies diagnosed earlier but they were able to adjust for this in the analysis which is good because if not it would have biased away from the null
Diagnostic bias occurs because early pregnancy diagnosis has a higher rates so if the cases have earlier pregnancy diagnosis than the control, they will already have a higher rate of miscarriage and this is present in the Goldhaber study
Participation bias affects internal and external validity. How can you avoid this?
You can avoid this through recruitment
- Particularly a problem with retrospective studies bc participation influenced by having experienced the outcome
- To avoid participation bias, could do a prospective study; don’t communicate outcome or exposure the study is looking for
How can sample size affect internal validity
Sample size may affect if you are able to detect a difference
Which bias has the greatest bias?
Differential misclassification has the greatest bias. Affects only one group (control or case) and biases toward or away from the null depending on the group misclassified
What is the healthy worker effect?
A reproductive study of workersfinds weak effects but compared to outside the workforce, its higher. Because those outside the workforce are more fertile by definition because on pregnancy leave, stay at home mom, etc.)
Prevalence of miscarriage
20 to 30% of all pregnancies
