Confounding is the distortion of association between an exposure and health outcome by an extraneous third variable called a confounder.
This occurs when a variable is associated with the exposure of interest in the population and is a risk factor for the effect (outcome) among non-exposed persons without being affected by exposure or disease and is not an intermediate step in the causal pathway between exposure and effect.
Example: An investigator is examining the relationship between alcohol consumption (exposure) and oesophageal cancer (effect). Smoking is associated with alcohol consumption and is a risk factor for oesophageal cancer regardless of alcohol consumption status. Therefore, smoking is a confounder in this instance.
Confounding can be neutralized at the design stage (by matching or randomisation) and/or at the analysis stage if the confounders have been measured properly. However, statistical risk adjustment has some important limitations:
- Only recognized confounders can be addressed in a regression model
- Every potential confounding variable added to a statistical model decreases the model’s statistical power, increasing the chance of a type II error (false negative result).
- Regression models are not very reliable when there are very few outcome events. As a rule of thumb, logistic regression must have at least 10 outcome events for every variable adjusted in the model, whereas linear regression requires 10-15 outcomes per variable included in the model to prevent overfitting.
Confounding bias may be of the following types:
- Confounding by group: This is produced in an ecological study and may occur when the rate of disease in unexposed varies across groups due to differential distribution of external risk factors across groups.
- Confounding by indication: This is produced when an intervention (treatment) is indicated by some symptoms, poor prognosis, or a perceived high risk. Here, the confounder is the indication since it is related to the intervention and is a risk indicator for the disease. This kind of bias occurs mainly in retrospective observational studies. Confounding bias by indication may be mistaken for protopathic bias.
Example: In a study of association between cimetidine and gastric cancer, the indication peptic ulcer is considered the potential confounder.
B. Specific Biases in Trials
The following biases specifically occur in trials:
- Allocation of intervention bias: More common in non-randomized trials, this occurs when intervention is differentially assigned to the population. In randomized trials this may be avoided by concealing the allocation sequence of intervention. Trials without allocation concealment report larger estimates of treatment effects than trial with adequate concealment.
- Compliance bias: Seen in trials requiring adherence/compliance to intervention, the degree of adherence/compliance influences efficacy assessment of the intervention.
Example: Patients with high risk quit diet or exercise programmes.
- Contamination bias: More common in community intervention trials, this occurs when activities like the intervention find their way into the control group. It biases the estimate of the intervention effect towards the null hypothesis (there is no difference).
- Differential maturing: Seen in group randomized trials, this reflects uneven secular (long-term) trends among the groups in the trial favouring one condition or another.
- Lack of intention to treat analysis: In randomized trials the analysis should be performed keeping participants in the group they were assigned to. If non-compliant participants or those receiving a wrong intervention are excluded from the analysis, the arms of a randomized trial may not be comparable.
Links to relevant articles: