Epidemiology: Selection Bias

systematic error (bias)

a *flaw* in -study design -data collection -data analysis leads to *wrong conclusions*

Selection Bias in Case-Control Studies

due to -Selecting study participants -Individuals consenting to participate

Selection Bias in Cohort Studies

due to selecting the cohort and/or loss to follow-up

what are the 3 types of bias?

selection bias information bias confounding

when does selection bias occur?

when the measure of association from the available study population is meaningfully different from that which would have been obtained if all subjects had been included in the analysis underlying reason for selection bias usually differs in different study designs

Selection bias due to loss to follow-up only occurs when you lose what?

A different proportion of people in cells A and C *AND/OR* A different proportion of people in cells B and D

selection bias

A distortion in the measure of association, resulting from the manner in which subjects are selected for the study population and/or selective losses from the study population prior to analysis

addressing selection bias during data analysis

Correction using selection probabilities

Opportunities to address Bias (in general)

In the design stage During data analysis After the results are obtained

addressing selection bias in the design stage

Randomization High participation rates Complete follow-up Appropriate controls

random error

can be reduced by incr. sample size

Selection bias in cross-sectional studies

occurs if -exposed cases are likely to survive longer than non-exposed cases OR -non-exposed cases are likely to survive longer than exposed cases THEN Results of a cross-sectional study may suffer from survival bias

what are the 2 types of error?

random error systematic error

In order for the risk ratio to be biased, what must occur?

the loss to follow-up must be associated with *BOTH* exposure and disease If 10% of the cohort is lost, but it doesn't change the relative sizes of *each of your 4 groups* (E+D+, E+D-, E-D+, E-D-), then there is no bias