error in epidemiological studies II: bias, confounding and interaction

error in epidemiological studies

Note 1

Random error has no preferred direction, so net result approach zero when

averaging over a large number of observations. The estimate may be imprecise,

but not inaccurate

precision vs accuracy

Precision refers to how well repeated observations agree with one another

Accuracy refers to how well observed values agree with the true value

bias

a systematic deviation of results or inferences from the truth or processes

leading to such systematic deviation; any systematic tendency in the collection,

analysis, interpretation, publication, or review of data that can lead to

conclusions that are systematically different from the truth. In epidemiology,

does not imply intentional deviation

selection bias

systematic difference in the enrolment of participants in a study that leads to an incorrect result (e.g., risk ratio or odds ratio) or inference.

• The animal in the study are not representative of the population of

interest

• In case-control studies, controls are not drawn from the same

population as the cases

• In studies of occupational exposures when the general population is

used as the comparison group, as those who are not healthy are less

likely to be employed (Healthy Worker Effect)

information bias

systematic difference in the collection of data regarding the

participants in a study (e.g., about exposures in a case-control study, or about

health outcomes in a cohort study) that leads to an incorrect result (e.g., risk

ratio or odds ratio) or inference. In particular, measurements errors are referred

to as misclassification (e.g. imperfect testing).

• Incomplete medical records

• Misinterpretation of records

• Patients completing questionnaires incorrectly (perhaps because they

don’t remember or misunderstand the question)

• Differential recall of information by cases and controls: recall bias

• Bias in the information collected by the investigator (e.g. because

prior knowledge of the hypothesis under investigation): observer bias

examples of bias

Selection bias

Postal survey on the extent and impact on Schmallenberg virus: those who suffered the disease might be more/less likely to participate

Observer bias

Welfare assessment via visual inspection on farms. Minor non-compliances in good-welfare farms are more likely to be overseen than in poor-welfare ones

Recall bias

Case-control study of Johne’s disease (Paratuberculosis) in cattle. Farmers are

asked about movements of animals months to several years ago. Those affected

might make more effort in remembering animal movement

Selection bias

Age or weight of foetuses submitted to diagnostic laboratories younger/ older

heavier/ lighter than foetuses that aborted in the general population.

confounding bias

the distortion of the association between an exposure and a health outcome by a third variable that is related to both

Relationship between coffee drinking (exposure), heart disease (outcome), and a third variable (tobacco use). Advantages and disadvantages of different observational study designs

confounding

A confounding variable is an “extra” variable that you didn’t account for.

To be a confounder:

• The factor must be independently associated with the disease

• The factor must be also associated with the exposure being

investigated

• It should not lie on the causal pathway between exposure and disease

Effects of confounding:

• Confounding factors can lead to bias in the estimate of the impact of the exposure being studied.

how to deal with confounding?

At design stage (preferably!):

• Restriction: only select units similar in relation to the confounder

• Randomisation: clinical trials. Similar distribution of confounders in

both groups

• Pair matching - selecting for each case one or more controls with

similar characteristics (e.g. same age and habits)

• Frequency matching - ensuring that as a group the cases have similar

characteristics to the controls

At analytical stage:

• Stratification: estimate measures of association separately within

different levels of the confounding factor

• Statistical modelling: multivariate analyses can control for confounding

Detecting the presence of confounding

• Explore the degree of discrepancy between the crude and stratum-

specific estimates. example in image

cont

dont need to memorise formula

As rule of thumb, if the stratified estimates of association differ from the unadjusted estimate by 10% or more, then there is evidence of confounding.

The true odds ratio, accounting for the effect of smoking, is 1.0 (Maentel Hanzel OR , which is a weighted average of the stratum-specific ORs).

When confounding is present, as in this example, the adjusted odds ratio should be reported

Note

Smoking was a confounding factor and there appears (with this over simplified analysis) to be no association (odds ratio= 1.0) between alcohol and MI

interaction or effect modification

Two independent factors interact if the effect on disease of one of the variables differs depending on the level of the other variable.

Provides clues on biological mechanisms or pathways of action

Note I

The combined effect of Drink and Pill is not simply additive

Note II

A finding to be reported! It can provides clues on biological mechanisms or pathways of action.

• If there is only confounding: the stratum-specific measures of association

will be similar to one another, but they will be different from the overall crude

estimate by 10% or more

confounding vs interaction

If there is neither confounding nor effect modification: The crude estimate of association and the stratum-specific estimates will be similar. They don't have to be identical, just similar.

• If there is only effect modification: The stratum-specific estimates will differ

from one another significantly

Confounding vs Interaction

Confounding:

• Remove or account for the effect of the confounder in order to get

nearer to the truth (actual measurement of effect).

• Effect of the exposure is the same for all categories of the

confounding factor

Interaction

• Interaction is an important property of the association between two

factors (detect and describe)

• The factor modifies the effect of other: effect varies across category of

the other factor

summary-pathways of action

Bias:

• Systematic error.

• Consider always when you design your study

• Can not be evaluated / corrected at analytical stage, leading to incorrect/ spurious results

• Different types of bias

• Selection, information, misclassification

Confounding:

• Alternative explanations.

• Consider potential confounders, collect data for them

• Adjust our measures of effect by confounders

Interaction

• The effect of a risk factor varies according to the different categories of another factor.

• Analytical stage consider interaction

• If necessary report measures of effect separately