Bias and Confounding
types of error,confounding factors and sample size calculations, and the factors that influence them
Bias is a systematic deviation from truth, and causes a study to lack internal validity.
In a research study, an observed difference between groups may be due to:
- A true difference between groups
- An error
Error can be due to:
- Normal random variation, i.e. chance
- A systematic difference, i.e. bias
Unlike error due to chance, the effect of bias cannot be reduced by increasing the sample size.
Types of Bias
|Type of bias||Description||Prevention|
|Selection||Where subject allocation results in treatment groups that are systematically different, apart from in the intervention being studied||Randomisation|
|Detection||Where measurements are taken differently between treatment groups||Blinding|
|Observer||Where the data collector is able to be subjective about the outcome||Blinding, Hard outcomes|
|Publication||When negative studies are less likely to be submitted or published than positive ones||Clinical trial registries|
|Recall||Altered reporting of symptoms by patients depending on which group they have been allocated to||Blinding|
|Response||When patients who enroll for a trial differ from the population, limiting generalisability||Random sampling|
|Hawthorne effect||When the process of actually doing the study improves the outcome||Control group, masking study intent from patients and observers|
A confounder is "a variable that, if removed, results in a change in the outcome variable by a clinically significant amount." It is a type of bias which will result in a distortion of the measured effect.
A confounding factor must be:
- Associated with the exposure but not a consequence of it
- A confounding factor cannot be on the causal pathway between exposure and disease
- It must be present unevenly between groups to cause distortion of the measured effect
- An independent predictor of outcome
The confounding factor must also be a risk factor for the disease, but independently from exposure.
Controlling for confounding
All confounders (known and unknown) are distributed evenly between groups.
Restricts participants to remove confounders.
- Results in reduced generalisablility and does not control all factors
Pairing of similar subjects between groups.
- May introduce additional confounding, and matching by multiple characteristics is difficult
Adjust for differences by transforming data.
Analyse the data in subgroups for each potential confounding factor.
- Sackett, D. L. (1979). Bias in analytic research. Journal of Chronic Diseases 32 (1–2): 51–63.
- PS Myles, T Gin. Statistical methods for anaesthesia and intensive care. 1st ed. Oxford: Butterworth-Heinemann, 2001.
- Stats notes from my MPh (University of Sydney). Probably a Timothy Schlub lecture, circa 2014.