Introduction to bias, cofounding and effect modification

Asking questions about…

• How studies are conducted

• How data has been collected

• How results have been interpreted

• What results really mean

• Overall implications of the study

Did the estimate occur due to:

What is Bias?

• Bias is a systematic error in data

• Bias consistently pulls the risk estimate away from its true value

Where could bias occur in a RCT?

• No blinding - can be focus on new treatment

• narrow no generalised population

Where could bias occur in cohort study?

• particular people choose

Where could bias occur in a case-control study?

What example of bias can you think of?

Think about

• Data collection methods

• Participants

Older people, cohort study, electronic data, case control study, interviews, all patients, questionnaires, randomised controlled trials

Length of follow up, interviewer bias, misclassification of exposure, data source limitations, who volunteers?

Measurement bias, non-compliance, recall bias, selecting controls who aren’t controls, selecting cases who aren’t cases

Volunteer bias

• Studies only include those who choose to participate

• Are people who volunteer to take part different from those who do not?

What are the different types of information bias?

• Interviewer bias

• Misclassification of exposure

• Recall bias

• Measurement bias

What are the different types of selection bias?

• Who volunteers?

• Selecting controls who aren’t controls

• Selecting cases who aren’t cases

Summarise the key points about Bias.

• The introduction of a systematic error in the study

  • At the point of study design

  • Cannot be eliminated at point of analysis

• Can be in terms of

  • Selection of study subjects

  • Exposure classification

  • Outcome classification

• Can result in over- or underestimated risks

Non-differential Misclassification

Misclassification to same degree for all groups

Differential Misclassification

Misclassification different between groups

What happens to our results if 10% of exposures are misclassified (non-differential)?

‘Estimated’ OR = (26x82)/(74x18) = 1.60

Non-differential misclassification will

underestimate effect (for two exposure categories)

would push all results closer to 1

Differential misclassification can bias result in

either direction

What is Cofounding?

The distortion of a risk estimate due to the mixture of the people in the study population

A confounder is a risk factor for the…

disease and is correlated with the exposure independent of disease

How can we determine cofounding?

If in the causal pathway then not a…

cofounder

Give examples of cofounder.

Age, gender, smoker

The indication for prescription is one of the most important factors to consider when…

evaluating medication exposures

A factor’s ability to be a confounder is entirely dependent on whether it is…

unevenly distributed between the study groups

What is Channelling?

• Where a drug is prescribed to a group of patients because of

  • a characteristic of the drug

  • a characteristic of that group of patients

  • usually is a new drug

• Often the drug

  • Has a claim of a better side-effect profile

  • Is heavily marketed

How can you control for cofounding using randomisation?

Evenly distribute confounders between study groups

How can you control for cofounding using restriction?

Gives much more control but cannot study variation between levels of that factor

How can you control for cofounding using matching?

• Each case is paired with a control subject(s) for specified constraints e.g. confounding factor(s)

• Can be difficult to find a match

• Risk of over-matching

• Can’t determine the association with a factor that is used for matching

• Only controls for the criteria that have been  matched on

How can you control for cofounding using analysis of data?

• Stratified analysis

  • Separating out factors so any mixture of their effect is removed

  • Can be used for one or two factors

• Multivariate analysis

  • Takes into account a number of factors simultaneously

What is Effect modifier?

• Effect modification occurs when the effect of the exposure is different in different groups of the population

• A factor that modifies the effect of a putative causal factor under study

• There is no average ‘true value’

Give an example of effect modifier.

Confounding distorts…

data

Unmeasured confounding is the most serious limitation in observational studies

Equal ‘estimated’ OR indicates…

cofounder

Effect of confounding: