BIAS AND CONFOUNDING - LEARNING OUTCOMES

Question 1

What are the possible explanations for an association?

Answer 1

1. Truth
2. Chance (sampling error)
3. Reverse Causation
4. Confounding
5. Bias

Difficult to know the truth.

Question 2

What does sampling error mean in terms of how true our findings are?

Answer 2

Sampling error means that any finding could be a chance finding.

Question 3

How can we assess the role of chance in our findings?

Answer 3

To help assess the role of chance in our findings we can look at the confidence interval around the measure of effect and also the p-value.

Question 4

What effect can sample size have on minimising errors cause by chance?

Answer 4

Larger sample sizes will minimise chance causing our sample to be unrepresentative.

Question 5

What is reverse causation?

Answer 5

Reverse causation is when, for example, the onset of disease actually caused the exposure. For example, the onset of lung cancer actually causes weight loss to occur rather than weight loss causing lung cancer. Timing of onset of disease in relation to exposure is often difficult to establish.

Question 6

In what kinds of studies is reverse causation likely to be a problem?

Answer 6

Any studies that take a snapshot of the population are those that are likely to be susceptible to problems of reverse causation as it may be hard to determine whether or not the exposure really did occur before the outcome.

Case reports, cross-sectional studies and case-control studies may all be subject to this kind of issue.

Question 7

What is confounding?

Answer 7

Confounding is when we see an association between exposure and disease, but it is not a true association. It is actually being caused because the estimate of exposure-outcome association is being distorted by some other exposure (confounder) which is associated with the outcome and with the exposure of interest.

Question 8

What is required for a feature to be a confounder?

Answer 8

For a feature to be a confounder, the factor must be associated with the exposure being investigated and the the factor must be independently associated with the disease being investigated.

Question 9

What are common confounding variables when dealing with individuals?

Answer 9

Need to consider confounders when designing study as we need to collect data about them:

• Age
• Sex
• Socio-economic status
• Occupation
• Ethnicity

There are bound to be many more depending on the specific research study that we want to carry out.

Question 10

What are the main to elements of a study that we can adjust to deal with confounding?

Answer 10

1. Design - we can look at the processes of randomisation, restriction and matching
2. Analysis - where we an do stratification or multivariate regression

Question 11

How does randomisation work at dealing with confounding during the design stage of a study?

Answer 11

Randomisation works because individuals in your study are assigned to the exposed or unexposed group in an entirely random way (we don’t just pick people who are already exposed or unexposed). This means that all confounding factors, both known and unknown are therefore distributed equally across both conditions.

This process essentially breaks the relationship between the exposure and the confounder so even though the confounder remains associated with the outcome, it won’t distort the relationship between the exposure and the outcome that you see.

This obviously won’t work for all study designs. We have to look at the exposure to see how practical it is - for example if our exposure is smoking habit we can’t randomise people into a smoking group, we have to use existing smokers. This process however does work very well for randomised control trials, for example if you are comparing a placebo (or existing drug) to a new drug then people would be randomised to the group which dictates which drug they receive. So hopefully confounding factors will also be distributed randomly between the 2 groups and that relationship between the exposure and the confounder will be broken.

Question 12

How can restriction at the study design stage reduce confounding in a study?

Answer 12

We can restrict our study to a possible level of a confounding variable - i.e. restrict our study to one level of a confounding variable only. For example we might decide to only look at non-smokers for a particular study, thus removing smoking as a confounding factor.

The problem with this however is that it makes the results less generalised / relevant to the actual population.

Question 13

What is matching? How can matching at the study design stage reduce confounding in a study?

Answer 13

Matching is often used in case-control studies. A control is recruited for each particular case based on potential confounders. For example it is quite common to choose a control of the same age and gender as the case that has already been recruited to the study. If we match then then we remove the possibility that the characteristics that match re the cause of any observed difference - i.e. if our case and control are both the same age then the observed difference cannot be due to age acting as a confounding factor.

For matching you must use special statistics to analyse data.

Question 14

How can stratification at the analysis stage of a study reduce confounding?

Answer 14

To deal with confounding at the analysis stage then we have to have collected all the information on our confounding variables for each of our individuals.

With stratification instead of restricting who we collect information from we still collect information from all of our participants, but then analyse each level separately.

Instead of doing a simple univariate analysis that doesn’t take into account any confounding factors we would essentially just split our participants into strata based on each confounding factor. We can then analyse each of these strata separately and see if the effect holds true for each strata, or the suspected confounding factor was causing the effect.

Question 15

Which measure of effect is easier to use in multivariate analysis? How can multivariate analysis at the analysis stage of a study reduce confounding?

Answer 15

The odds ratio is easier to use in multivariate analysis.

A multivariate analysis takes into account multiple potential confounding factors and gives you an odds ratio that is adjusted to take the confounding factors into account.

Confounding variables can also mask an association so when you do the multivariate analysis you find there actually is an association when the confounding factors are adjusted for - i.e. the odds ratio can increase as a result of multivariate analysis as well as decrease.

Question 16

What is bias?

Answer 16

Bias is systematic error in an estimate arising from problems in the study design / execution.

Bias can result in a measure of effect above or below the true value.

Unlike confounding we cannot correct bias in the analysis.

Question 17

Can we correct for bias in the analysis of a study?

Answer 17

Unlike confounding we cannot correct for bias in the analysis of the study.

Question 18

What are the two main types of bias?

Answer 18

1. Selection bias

2. Information bias

Question 19

How can selection bias be introduced as a result of the inclusion criteria for cases and controls?

Answer 19

Selection bias may result in an unrepresentative sample.

Bias is introduced into the study if the criteria for inclusion of cases and controls is also related to the exposure of interest - for example choosing controls for a study investigating lung cancer and smoking from a population of individuals with non-malignant respiratory disease (also related to smoking).

This is therefore of particular concern in case-control studies.

Question 20

Give an example of selection bias that has arisen as a result of the inclusion criteria for cases and controls?

Answer 20

In a study of smoking and lung cancer:

• controls not selected from general population but from patients with non-malignant respiratory disease
• smoking causes non-malignant respiratory disease - i.e. smoking is associated with both exposure and outcome
• therefore the effect of smoking on lung cancer will be underestimated because our smoking prevalence in our controls (without lung cancer) will be higher than if it was selected from the general population
• where we choose to recruit our controls from can have dramatic effects on the odds ratio

Question 21

Why is bias of particular concern for case-control studies?

Answer 21

Bias is introduced if the selection criteria for cases and controls is also related to the exposure of interest.

If our controls are recruited from a population using criteria related to the exposure of interest then our odds ratio will not give us a representative answer.

Question 22

Describe selection bias introduced as a result of non-response.

Answer 22

Non-response bias is often a problem in studies, particularly those using postal questionnaires. Bias will be introduced if response rate is related to exposure and outcome.

e.g. social class and breast cancer: those of low social class without the disease are less likely to respond.

Question 23

What can lead to selection bias?

Answer 23

• Sampling bias i.e. the inclusion criteria for cases and controls being linked to the exposure of interest
• Non-response
• Loss to follow up

Question 24

How might loss to follow-up result in selection bias?

Answer 24

Loss to follow-up is particularly a problem in cohort studies. It can lead to bias if drop-out is related to exposure / outcome.

Question 25

Define selection bias.

Answer 25

Selection bias is the selection of individuals, groups or data for analysis in such a way that proper randomisation is not achieved, thereby ensuring that the sample obtained is not representative of the population intended to be analysed. It is sometimes referred to as the selection effect.

Selection bias may result from sampling bias, non-response, loss to follow-up.

Question 26

What is ‘the healthy worker effect’ in relation to selection bias?

Answer 26

The healthy worker effect refers to a type of sampling bias that may occur when considering occupational exposures. If we are looking at occupational exposures it is a fact that workers are generally healthier than the population as a whole. If the comparison group is the general population then it may appear that the occupation ‘protects’ against the disease of interest.

Question 27

When does information bias occur?

Answer 27

Information bias occurs when the study subjects are misclassified either according to their disease status, their exposure status, or both.

They may be misclassifies due to error in instrument, subjects being given incorrect information, or a researcher recording information incorrectly.

Question 28

List types of information bias.

Answer 28

1. Recall bias
2. Reporting bias
3. Observer / interviewer bias

Question 29

How do we deal with bias?

Answer 29

Essentially the only way to deal with bias is to get the study design correct from the start. We need to:

1. Select an appropriate study population (to reduce selection bias)
2. Maximise response rates (to reduce selection bias)
3. Minimise loss to follow-up (to reduce selection bias)
4. Use objective rather than self-reported measures of exposure/outcome (to reduce information bias)
5. Follow standardised procedures in data collection (to reduce information bias)
6. Blind our interviewers (to reduce information bias)

Question 30

How can we deal with confounding?

Answer 30

Study design and / or analysis

Question 31

How can we deal with bias?

Answer 31

Study design

Question 32

How can we deal with chance?

Answer 32

Adequate sample size

Question 33

How can we deal with reverse causation?

Answer 33

Study design