Observational studies are required when subject’s can’t be randomly allocated to treatment groups. This may be due to ethical reasons, e.g allocating who must smoke (treatment group) and who can’t smoke (control group) is not a feasible experimental design. Due to the nature of this study a causal link can never be drawn in observational studies since there may be confounding variables, only links can be made. For example an observational study into smoking and liver cancer has the confounding variable problem e.g smoking is linked with greater alcohol consumption which is linked with liver cancer. Making it difficult to say that A definitely causes C, however, we will go into techniques to control the confounding variable.
Last lesson we covered observational bias and selection bias. A new type of bias, which is more common in observational studies is survivor bias. Survivor bias can result in an “improvement” in the data despite the reality being the opposite. This could be that smokers move to non-smoking group. They then die in the non-smoking group from a smoking related death, however, they had dropped out of the smoking group so the fatality is reported for non smokers. In this case an improvement for smokers is caused by the worst subjects dropping out of the smoking category.
Adheres are subjects who continually take the treatment whereas non-adheres are subjects who only follow the treatment or program less then 80% of the time. Usually the adheres respond better to the treatment in both the control and the placebo groups.
During an observational experiment, confounding variables are common in observational studies since different lifestyle choices are more commonly associated with other variables. This makes it difficult to determine which one is causing the observed correlation. As previously stated, drinking alcohol being correlated with smoking which correlates with increased risk of liver cancer.
To control confounding variables, groups can be created in the data set to separate smokers and non-smokers who drink a lot, moderately, little and nothing. This controls this confounding variable, however, it is never known if all confounding variables were controlled. This means the results are always correlating to things never implying causation.
However, it is important to be aware of Simpson’s paradox which is when the observed trends in sub groups is reversed when the subgroups are combined.
|Treatment A||Treatment B|
|Day 1||60/90 = 66.7%||8/10 = 80%|
|Day 2||4/10 = 40%||45/90 = 50%|
|Avg of 2 days||64/100 = 64%||53/100 = 53%|
Observational studies are required when subject’s can’t be randomly allocated to treatment groups.
Explain what an observational study is with an example
Explain the effect of adheres and non adheres in a study.
Explain survivor bias and how it may mislead you in the results.
Explain Simpson’s paradox using an example