Bias and Confounding

Question 1

What is Bias?

Answer 1

Inclination or prejudice for or against one person or group, especially in a way considered to be unfair

Question 2

What is Error?

Answer 2

Error is the difference between the measurement and the ‘true value’

Question 3

What are the 2 Types of Error?

Answer 3

Systematic and Random Error

Question 4

When does Systematic Error Occur?

Answer 4

Systematic error arises when we do not include all groups of people, meaning certain groups are left out. There is a difference between the people in the study and those wuho are not (e.g. only studying rich and not poor, or men and not women). Therefore, the cohort is not representative of the whole population

Question 5

Systematic Error Occurs where there are limitations in what?

Answer 5

Limitations in the study design, in how patients were recruited, in how the data was collected/analysed

Question 6

What is a problem with Systematic Error?

Answer 6

Bias will persist even if we increase the sample size. We cannot correct for this type of error using statistical methods or mathematics, making it hard to determine the real effect of this error

Question 7

Systematic Error is ____ but not _____

Answer 7

Systematic Error is Precise but not Accurate

Question 8

What is Random Error?

Answer 8

Occurs due to random chance. (e.g. measuring your weight three times, and getting three different values); these differences are due to random error.

Random error changes the variability and distribution of your data, but not the average (i.e. the values will be more all over the place, but the average will come out to be the same).

Question 9

Random Error is ____ but not _____

Answer 9

Random Error is accurate but not precise

Question 10

What is an Advantage of Random Error?

Answer 10

The good thing about random error is that we can use statistics to see the effect that it has on our results. We use p-values, hypothesis testing and confidence intervals to determine whether or not any differences in the results are purely due to random error, or are legitimate.

Question 11

How do we Minimise the Effect of Random Error?

Answer 11

We can evaluate this using statistical methods, and reduce by increasing data/sample size.

Question 12

What causes Bias?

What does it result in?

How can we minimise this Bias?

Answer 12

Systematic error causes bias towards a particular group.

It results in a lack of accuracy, as all data will be either greater than or less than the true value.

We can try minimise this bias by following the CONSORT principles.

Question 13

What are the 3 Types of Bias?

Answer 13

Selection Bias

Measurement Bias (or misclassification or information bias)

Confounding Bias

Question 14

What is Selection Bias?

What is an Example?

Answer 14

Selection bias occurs when there is a different selection of people in the control group vs the test group, hence causing different results in each group

An example would be if in the control group, there was 70% males, but in the test group, there was 30% males. Selection bias has occurred; we would rather there be the same proportion in each. It will arise if you do not properly randomise the cohort, or properly exclude or include certain people

Question 15

What are Subtypes of Selection Bias?

Answer 15

Attrition bias (e.g. when people drop out due to death or other reasons)

Healthy worker bias (only selecting healthy people)

Nonresponse bias (people not completing a survey)

Question 16

Discuss an Example of the Effect of Attrition Bias

We took a study on the effect of smoking on lung cancer

RR = 9

People died along the way

Effect of this - RR decreased to 4.7

Answer 16

By just looking at the raw data, we see that the RR is 9, meaning smokers were 9x more likely to get lung cancer. However, this data included people that died along the way. When these people were removed from the data, there were less smokers that had lung cancer, since some had died. As such the RR decreased to 4.7, meaning we are now underestimating the risk of smoking (i.e. makes smoking look less bad). Since those people died and dropped out, attrition bias was created.

Question 17

What is Measurement Bias?

Answer 17

Measurement bias arises due to inaccuracies in any measurement collected from a participant

Question 18

What are the Types of Measurement Bias?

Answer 18

Digital bias (e.g. scales, pipettes)

Interviewer bias (e.g. interviewer better at interviewing women vs men)

Recall bias (ever get that feeling of de ja vu? Occurs when the person cannot remember details)

Reporting bias (e.g. food diary)

Question 19

What is an Example of Measurement Bias?

Answer 19

An example would be if a patient was asked if they were 182cm tall, before a doctor measured them. If the patient said yes, but the doctor said no, or vice versa, misclassification bias occurred. The patient should not be in the group they were put in

Question 20

What is the Classification of Measurement Bias?

Answer 20

Differential and Non-Differential Measurement Bias

Question 21

Measurement Bias: What is Differential Measurement Bias?

Answer 21

The probability for misclassification is different for each group. In other words, there are different ways of classifying for each group

Question 22

Measurement Bias: What is Non-Differential Measurement Bias?

Answer 22

The probability for misclassification is the same/similar between groups. In other words, both groups have the same ways of classifying (bias spread evenly across groups)

Question 23

How do we reduce Measurement/Misclassification Bias?

Answer 23

We can use blinding to help reduce measurement/misclassification bias

Question 24

What is Confounding?

Answer 24

Confounding mixes up the effects and results, due to the presence of another exposure that is not the primary exposure

Question 25

What can Confounding lead to?

Answer 25

Confounding can lead to over or under-estimation of the effect, as is the case with any type of bias

Question 26

To be classified as a confounder, the variable must meet what criteria?

Answer 26

Must be associated with the disease

Must be associated with the exposure

Must not lie on the causal pathway between the exposure and the disease.

Question 27

When is a confounder, a confounder?

Answer 27

The rule of thumb is that if the primary outcome changes by more than 10% when you adjust for the confounder, then it can be called a confounder

Question 28

What is the Causal Pathway and how does this relate to Confounders?

What is an Example with Smoking, Lung Cancer and Low socioeconomic status?

Answer 28

For a variable to be classed as a confounder, it must not lie on the causal pathway between the primary exposure and the disease. It must still be associated with both the exposure and disease, however it must not ‘occur in between’ the two.

An example would be Smoking -> Lung Cancer. Low socioeconomic status is a potential confounder and is associated with both smoking and lung cancer. However, the pathway does not go Smoking -> Low SES -> Lung Cancer, so therefore the variable does not lie on the causal pathway

Question 29

If there are potential confounders, what can we do?

Answer 29

In studies where there are potential confounders (i.e. other exposers confounding the result), we can stratify the data to assess the effect

Question 30

What does Stratifying the Data mean?

Answer 30

Stratifying the data involves splitting the data into 2 groups, one exposed to the potential confounder and one not, and analysing the results of each. You would then compare the results with the stratification to those without it (e.g. compare the RRs), and see how big of an effect the stratification had. If the effect is greater than 10%, the exposure must then be a confounder (assuming it satisfies all other criteria)

Question 31

Can we control for confounding?

Answer 31

There really isn’t a set way to control for a confounder. The main thing is to ensure good study design, with randomisation, stratification (to an appropriate level), restriction (look at one confounder at a time) and matching. We can also perform some extra statistical analysis (multivariable modelling) to control for the effect