Bias, confounding, and chance

Question 1

Validity

Answer 1

Measures the degree to which our study is free from bias, confounding, and chance (random error)

Question 2

Good protocol

Answer 2

Considers bias, confounding and random error and describes steps taken to tackle these in design and analysis.

Question 3

Internal validity

Answer 3

The results are correct for the particular group of subjects who are being studied. Happens by minimising the role of bias, confounding, and chance.

Question 4

External validity

Answer 4

Generalisable to the target population (internal validity does not mean external validity)

Question 5

Trade off between internal and external validity

Answer 5

The more you control for internal validity (bias, confounding, chance) the less generalisable to the entire population (external validity)

Question 6

Bias

Answer 6

Systematic variation from the truth

Leads to a mistaken (under/over)estimation of the true effect of the exposure and the outcome

Question 7

Random error/imprecision

Answer 7

Random (unpredictable) deviations from the truth

Question 8

Largest sources of bias

Answer 8

• selecting target population

- measuring exposures and outcomes in our study

Question 9

Selection bias

Answer 9

• the study population does not represent the target population
• the result is different to that obtained if you had enrolled the entire target population
• may result from procedures used to select subjects, and from factors that influences participation or likelihood of remaining in the study
• may occur while recruiting participants and/or while retaining them in the study

Question 10

How to minimise selection bias

Answer 10

1. Select participants independent to their affiliation to a medical centre
2. Use incentives to increase participation
3. Use data to compare those who did and those who didn’t volunteer to take part in your study (compare age and sex distribution of individuals in your study to those in your target population)
4. If goal is to extrapolate findings of the study to the entire UK -> External validity -> could use multiple UK regions

Question 11

Selection bias

Answer 11

• even if recruitment of patients was unbiased, might still still use participants to the study
• e.g. in RCT, cohort studies
• if loss to follow-up occurs randomly, that is similarly between control and exposure group, then it won’t lead to bias
• but if loss to follow-up is differential between groups and also associated to the outcome -> bias
• most applicable bias to longitudinal studies, often leads to loss of significance, might not able to detect the true effect of an intervention on an outcome

Question 12

How to minimise loss to follow up?

Answer 12

• Collect info to facilitate tracking of participants
• Recruit subjects that are easier to track (doctors, nurses, or living in places where people don’t tend to emigrate, (consider implications on external validity)
• Maintain regular contact, use tracking resources
• Send newsletters, multiple requests, etc. to non-responders

Question 13

Information (measurement) bias

Answer 13

If gather information (e.g. about exposure or outcome) differently in one group to another (e.g. among cases compared to controls), then bias may result

-occurs during data collection

Question 14

Misclassification bias

Answer 14

• commonest type of information bias
• individuals are wrongly classified into a category they do not belong to
• e.g. exposed is wrongly categorised as unexposed or disease

Question 15

Non-differential misclassification

Answer 15

Misclassified systematically differently to the true value, but the error rate or probability of being misclassified is the same in each study group
Estimates biased towards the null, which means that whatever association we observe we could actually argue that the true association is more extreme

Question 16

Differential Misclassification

Answer 16

Error rate or probability of misclassification differs across the study groups

Question 17

Recall bias

Answer 17

• type of information/ measurement bias

- all subjects might recall past exposures inaccurately, especially those that happened a long time ago

Question 18

When might recall bias be introduced?

Answer 18

Differential recall between groups is possible

May introduce bias, if recall depends on outcome or exposure of interest

Question 19

How to reduce recall bias?

Answer 19

• look at records instead of interviews
• if for instance comparing steroid use in cleft lip development, look at other disease instead of no cleft lip develpoment

Question 20

Reverse causality (protopathic) bias

Answer 20

• information bias
• x and y are associated, but not in the way you would expect
• Instead of x causing y it is really the other way around, Y is causing x
• in epidemiology, it is when the exposure-disease process is reversed, in other words the exposure causes the risk factor

Question 21

Observer bias

Answer 21

-type of misclassification bias
-knowledge of exposure status might lead to a search for the outcome
e.g. consider
Exposure: IV drug use
Outcome: HIV infection
-knowledge of IV drug use may lead for more intense search for HIV infection

Question 22

Control of bias - methods of data collection - When designing your study, consider and discuss the following (1/2)

Answer 22

1. What other factors may explain your findings
2. How can you capture information on these factors
3. What, how, when and who will collect it
4. How frequently will it be collected
5. How these methods will prevent or minimise potential bias

Question 23

Control of bias (2/2)

Answer 23

Objective questions (interviews) and objective measurements and biological samples

Standardised techniques

Use validated tools or validate yours

Blinding of investigator to treatment/exposure/outcome status

Blinding participants and data analysts

Use of existing records

Uses of multiple sources of data for exposure and outcome and compare

Memory aids (photographs)

Self-administered questionnaire for socially sensitive topics

Question 24

Confounding

Answer 24

• Occurs when the effect of an exposure on an outcome is mixed together with the effect of a third variable
• Is associated with the exposure and the disease but is not a consequence of either of them
• the distribution of the confounder differs between groups being compared or is unbalanced

Question 25

How to deal with confounding? (1/2) At the design stage

Answer 25

Randomisation
Ensures that the distribution of both known and unknown confounding factors will be similar in the groups to be compared

Restriction
Limit participation to subjects who are similar in relation to the confounding factor, e.g. only include smokers or non-smokers (limitations of external validity)

Matching
Subjects without the outcome of interest are selected so that the distribution of potential confounders is similar to those with the outcome

Question 26

How to deal with confounding? (2/2)

At the analysis stage

Answer 26

Stratification

• One or two categorical confounders
• Combines or pools to get an adjusted measure of effect

Statistical modelling
- continuous confounders or more than 2 categorical confounders -> multivariate models is the best result

Question 27

Confounding in observational studies

Answer 27

Residual confounding by known and unknown factors is always possible

Question 28

Covariate

Answer 28

A patient variable (e.g. age, sec, etc.) that may or may not be related to the outcome.

If the covariate is related to both the exposure/risk factor and the outcome, then covariate becomes as confounder.

Question 29

Confounder

Answer 29

A confounder is a covariate that is related to both the outcome and the exposure/risk factor. A confounder may either increase or decrease the likelihood of the outcome.

Question 30

Covariates

Answer 30

Not all covariates are confounders, one also needs to consider:

1) Effect modifiers
2) Prognostic variables or alternative predictors of outcome
3) Intermediate variables in the pathway

Question 31

Effect modifiers

Answer 31

Exposure has different effects among subgroups

Effect modification is associated with outcomes but not exposure

Question 32

Prognostic variables (alternative predictors) of outcome

Answer 32

Prognostic variables or prognostic markers are baseline characteristics known to be associated with the outcome of a health condition in the absence of treatment.

Question 33

Intermediate variables in the pathway

Answer 33

An intermediate variable serves as a causal link between other variables. It is acted on by the independent variable and then acts itself on the dependent variable to create change.

For example, suppose you wanted to look at the relationship between higher income and longer lifespans. A high income does not act directly on someone’s lifespan in a positive way, but it may allow access to better nutrition and health care. This better nutrition and health care, in turn, affect the life span of the person, and so these are intermediate variables.

The intermediate variable is sometimes also called the intervening variable, the mediating variable, or the intermediary variable.

Question 34

Chance

Answer 34

Use of sample to make inference about the underlying population

If we sampled again and again, purely by Play of chance may always affect results observed simply because of random variation from sample to sample

Major determinant of the degree to which chance affects findings in any particular study is sample size

Question 35

Chance (2)

Answer 35

larger sample -> reduce imprecision ->more valid inference

Use statistics to quantify the degree to which chance variability may account for the results observed in any individual study

p value
confidence interval