Missing Data, Bayesian Imputation and People Analytics with PyMC [PyCon DE & PyData Berlin 2024]

Missing data in surveys can be classified into three main types:

• Missing Completely at Random (MCAR)
• Missing at Random (MAR)
• Missing Not at Random (MNAR)

Bayesian imputation methods are recommended for theory-informed missing data analysis because they:

• Allow flexible model specification
• Handle different types of distributions
• Provide built-in sensitivity analysis
• Enable workflow for model adequacy assessment

Hierarchical modeling is valuable for handling missing data because:

• It can account for team and management structures
• Helps isolate estimates of different impacts
• Can transform MNAR situations into MAR situations
• Allows for team-specific parameter estimates

In people analytics context:

• Decisions about careers need justifiable models
• Power relationships and hierarchies influence data collection
• Survey non-response patterns may reveal organizational inefficiencies
• Team-management mismatches can be identified

Technical implementation considerations:

• Variables should be ordered by degree of missingness
• Multiple distribution types can be handled in the same model
• Priors should be carefully selected based on domain knowledge
• Cross-validation and model adequacy checks are essential

Practical recommendations:

• Run pilot experiments to gather information for priors
• Consult subject matter experts for model construction
• Validate models and repeat as necessary
• Document your data-generating process assumptions