Bias is a form of nonrandom error in epidemiological studies. It can lead to incorrect or misleading results about the association between health factors and outcomes. 

While there are many types of bias that can occur in epidemiological studies, they can be categorized into three main types: Selection bias, information bias, and confounding. 

Below, we provide information about each of these three types of bias and provide some specific examples. 

• Selection bias: Occurs when the people in the study (the sample) does not represent the population that the researchers are trying to draw conclusions about. At the same time, the sampled group differs from the full population in terms of the factors being studied.

  Some specific examples of Selection Bias include:

  • Sampling bias: We want to gather opinions on a topic from the general population. If we conduct a study that uses only an internet survey, people without internet access will not be included. So, the study sample will not represent the whole population. People without internet access may differ from those with access regarding the issues being measured and studied. If this is the case, then the findings may be biased and not reflect the “truth” in the whole population. This type of selection bias can often result from convenience or “judgmental” sampling, where samples are selected for study in a nonrandom way.
  • Self-selection bias: We want to understand if college students are getting enough exercise. If the only people who choose to enroll in our study are student athletes, the study sample will not represent the entire college student population. 

  • Healthy worker effect: If researchers compare the working population to the general population, a bias will result. This is because the general population includes all people regardless of illness or disability. The working population includes only those capable of working.

     

• Information bias: Occurs when there are problems with how study information is collected. This can include issues with how people   collect and interpret information, the tools used to measure information, and how information is reported by people in the study.

  Examples:

  • Recall bias: If we are asking study participants about things that they did many years ago, it may be hard for them to provide accurate answers. They may not remember. A nutrition study asking people to list all the foods they ate last year, last month, or even last week would not collect accurate information.
  • Confirmation bias: Researchers may, intentionally or unintentionally, favor information that confirms their beliefs and ignore information that does not. This can influence what they choose to study, how they collect and record information, and the way they interpret and share study results.  
  • Hawthorne effect: Participants in a study modify their behaviors or responses due to their awareness of being studied or observed. A study looking to see if workers show up to work on time would be impacted by this bias if researchers told the workers in advance. Workers may modify their routine to ensure they are on time or early for work since they know they are being watched.
• Confounding: Occurs when the association between an exposure and outcome is actually the result of another factor. The other factor is known as a confounder. To be considered a confounder, the factor must be linked to both the exposure and the outcome. 

Example:

• There is an association between drinking alcohol and lung cancer. People who drink alcohol are more likely to have lung cancer. However, it should not be concluded that drinking alcohol causes lung cancer. People who drink alcohol are more likely to smoke cigarettes than people who do not drink alcohol. Cigarette smoking is the confounder in the association between alcohol and lung cancer and is the true cause of lung cancer. When we take cigarette smoking into account, the association between drinking alcohol and lung cancer goes away.

All studies have potential confounders. Epidemiologists use several tools to reduce the effects of confounding during different phases of research. 

While Setting Up the Study (Study Design):

• Randomization: A study with two groups randomly assigns participants to either group. The potential confounders are more likely to be evenly spread across both study groups. This balancing of the confounding factors can remove their ability to influence the study findings.

  Example: Randomly assigning people to a group that receives a new drug or a group that receives a placebo. This assignment helps ensure that both groups are fairly equal in potential confounding factors like age or health status of the people in the groups.

• Restriction: An epidemiologist can remove the impact of a confounder by restricting the group of people they are studying based on that confounder. Restriction can be completed before a study begins or when data are being analyzed. 

  Example: If biological sex is a confounder, we can remove the impact of the sex by restricting the group of people being studied to only one sex. This can be done when selecting participants or when analyzing results. However, this also limits the ability to apply the study results to a wider group of people. 

• Matching: In a case-control study, an epidemiologist can reduce the impact of a confounder by creating a similar distribution of confounders in both groups being studied. 

  Example: If biological sex and age are confounders, we can remove their impact by matching participants by these factors. For each 45-year-old male with the health outcome (case), the epidemiologist would select a matching 45-year-old male without the health outcome (control). 

While Evaluating the Data (Data Analysis):

• Stratification: We can analyze data separately for each group (or “strata”) of the confounder.

  Example: Analyze results separately for men and women, if sex might confound the association between the exposure and the outcome.

• Standardization: An epidemiologist can apply statistical methods to adjust for differences in the demographic composition between populations. This is done so they can be compared fairly, in an apples-to-apples way.     

  Example: A researcher may want to compare the rate of a disease between two counties, but one county has a much older population. We can compare the rate of disease between the two counties by adjusting to a “standard” age distribution.