To learn more about EpiX Analytics' work, please visit our modeling applications, white papers, and training schedule.
Page tree
Skip to end of metadata
Go to start of metadata

Bias occurs when problems during the data collection introduce systematic error in the data, so they are often caused by an inappropriate study design. Biases can be categorized as selection and measurement. Selection biases occur when the population sampled is not representative of the population targeted for the study, whereas measurement biases are caused by the data collection method or instrument. Both differ from cognitive bias, some types of which are discussed as sources of error in subjective estimation, as cognitive bias can influence thought and decision processes and the information collected from individuals, but are not created by the study design. 

Measurement bias introduces a unidirectional error through the method of assessment. For example, a survey that always overestimates the risk aversion of managers introduces a positive bias in that data, or a scale that is not properly calibrated may always underestimated the weight of an environmental sample.

Types of Measurement bias include: 

  1. Recall bias: Groups in the study have differing abilities to describe past experiences relevant to the study outcome. For example, patients with chronic gastrointestinal disorders may be able to describe their diet over the last six months in greater detail than patients with atopic dermatitis. 
  2. Attention bias (the Hawthorne effect): Observation of a person's behavior with their knowledge may change that behavior, especially in cases where the trait of interest is socially undesirable. 
  3. Reporting bias: Similar to attention bias, a person may be less likely to report past actions or behavior that are perceived as socially undesirable. 

Selection biases introduce error through examining subset of a population that is not representative of the whole population we are trying to estimate. For example, measuring vehicle traffic on a busy intersection on Sundays would be a poor estimator of the daily traffic any day of the week, even if the time of the day and the Sundays to be sampled are chosen at random. 

Types of Selection bias include: 

  1. Volunteer and non-response bias: These biases are two sides of the same problem, where the population who volunteers for a study or who fails to respond to study recruitment differs from the rest of the population in a way that is relevant to the study. For example, if a phone survey is being carried out at mid-day, the population choosing to respond might be less likely to work during the day, and therefore different in age, occupation, and other factors. 
  2. Attrition bias: This may occur if study participants drop out at different rates according to a relevant factor, for example, age, ethnicity, or assigned dose group in a clinical trial. 
  3. Healthy worker bias: This bias can be introduced if members of one occupation are compared with another, or with the general population, as those who are able to work, and who work in specific jobs may be healthier than those who cannot work for many reasons.

Confounding may also be considered a type of statistical bias. Confounding exists when a third variable is linked to both the outcome of interest and the variable under examination and distorts the true relationship between them. For example, if the outcome of interest is decision speed, and the variable of interest is sleep quality, caffeine intake could confound the association between the variables decision speed and sleep quality as people who sleep poorly might be more likely to drink coffee.

Related topics:

Modeling expert opinion

Sources of error in a subjective estimation

Check the quality of your data

  • No labels