Errors of Design Studies

1) Potential confounders
We might fail to realise the confounders during our studies. We need to figure out whether our study design is biased against our research.

For example:
The NFIP Study.

A test to check if a newly developed vaccine is helpful against polio.
The test was split into 3 groups
- Children with consent (treatment)
- Control (consent and no consent)
- No consent

However, this design is flawed as children with consent are more at risk to contract polio in comparison to children without consent. This is due to the fact that children with consent are more likely to be rich families and thus being brought up in a clean environment, is not exposed to the virus early. Unlike children without consent that are poor children, are exposed to early age and gained immunity from the virus.

Thus children with consent group are more likely to contract the diseased compared to control group and no consent group. The design is flawed

Solution:

Ensure that only children with consent can be allowed to be in the control group or the treatment group. Exclude those who refuse treatment.

2) Randomisation
Based on the example above, it is important to make everything be randomised as human judgement is prone to bias. By randomising the subjects to be in the treatment group, the different groups tend to be similar in terms of fairness. This minimises confounding.

Examples of Randomisation:
- Placebo
- Blinding Subjects
- Blinding doctors

 Double binding
To further prevent biases, doctors who made diagnosis are blinded

3) Non-randomised controls
Its tempting to compare a treatment group to a historical control group however, the historical control group may differ from a current treatment group so confounding is a real problem

4) Controlled Experiment
Assignment by Investigators
An experiment conducted to test a short term effect
Examples can be like a new vaccine.

5) Observational Studies
Long term effects of something can only be investigated by observational studies.
Assignment by subjects
Subjects are already assigned to groups
Prone to confounding

5) Rates and association
Rate(A|B) means that the rate of A among B
=> (A and B) / B *100%
If rate(A|B) > rate(A| not B) or
rate(B | A) > rate(B | not A)
then
"A and B are positively associated"
However, positiely associated doesnt mean causation

But if
If rate(A | B) < rate(A | not B), or rate(B | A) < rate(B | not A)
then
"A and B are negatively associated"

In summary,
as long as r(A | B) ≠ r(A | not B), or r(B | A) ≠ r(B | not A)
then  A and B is associated

6) Confounding
Controlling it:
Slicing:
To deal with particular confounders, we use a method call slicing.
Separate the data into two different data group by its confounder.
For example:
Originally we have a table of Smokers and non-smokers
To deal with sex as a confounder, we can split the data into tow separate tables group according to sex. We can there recalculate the rate.

We can also make use of ranges to deal with variables that have a wide range of numbers like age



Next Association >