Mitigating bias in data collection

Conquering Data Bias

Konstantinos Kattidis

Data Analytics Lead

Identifying bias in data collection

• Selection bias, historical bias and measurement bias
• Understanding these biases creates awareness, enabling data experts to proactively identify them and take action

• Sensitivity analysis involves exploring how different assumptions, alternative subgroups, or weighting strategies affect the analysis results
• External validation compares data against independent sources to check for consistency and accuracy

Random and stratified sampling

• Selecting an appropriate sampling technique is important
• Random sampling involves selecting individuals or data points from a population randomly
• Stratified sampling divides the population into subgroups and then selects samples from each subgroup

Balancing subgroup representation

• Oversampling involves deliberately increasing the representation of certain groups or classes in a dataset to balance the distribution
• Undersampling involves reducing the representation of overrepresented groups to achieve a more balanced dataset
• Weighting involves assigning different weights to observations based on their importance, compensating for any imbalances in the sample distribution

Data augmentation

• To address historical bias, this technique enriches the dataset with additional data points
• The aim is to cover underrepresented periods or events
• It includes:
  • Filling data gaps
  • Diversifying perspectives
  • Updating and correcting errors

Data measurement practices

• Standardization of measurement tools and protocols
• Training and calibration of data collectors
• Pilot testing can be used to assess the accuracy and consistency of data collection procedures
• Regular quality assurance checks and automation of processes can further enhance data quality

Continuous monitoring and adjustment

• Continuous monitoring and adjustment are essential to address emerging biases
• Regular reviews of data quality metrics
• Bias assessments
• These enable immediate identification of biases

Let's practice!

Conquering Data Bias