Description
Estimates model parameters and plausible values for incomplete session data by alternating between expected values of missing parts and maximum likelihood updates.
Session logs are routinely incomplete: dropped events, SDK failures, consent opt-outs, cross-device gaps. Deleting incomplete records biases analyses unless data are missing completely at random. The EM algorithm treats missing values as latent: the E-step computes their expected sufficient statistics given current parameters and observed data, and the M-step updates parameters to maximise likelihood. Iterating converges to maximum likelihood estimates under a missing-at-random assumption. EM also underlies Gaussian mixtures and many latent variable models. For inference, multiple imputation (e.g. MICE) often complements EM, so uncertainty from missingness is carried into standard errors. Rigour requires describing missingness patterns, reasoning about the mechanism and running sensitivity analysis for missing-not-at-random scenarios.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Goal
- Measure behaviour
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Quantify missingness per variable and per session and visualise patterns (e.g. missing by platform or app version).
- Assess the plausible mechanism (MCAR, MAR, MNAR) with domain knowledge and tests such as Little's MCAR test.
- Specify a model for the data (e.g. multivariate normal or mixture) and run EM from several starting values until log-likelihood changes fall below a tolerance (e.g. 1e-6).
- Use EM estimates for parameters, or create 5-20 multiple imputations to propagate uncertainty into analyses.
- Compare results with complete-case analysis and run a sensitivity analysis under plausible MNAR departures.
- Fix the instrumentation causes of missingness where possible.
Sources
- Deep Generative Modeling (Jakub Tomczak, Springer)
- Deep Learning (Ian Goodfellow, Yoshua Bengio & Aaron Courville)
- Methods Atlas (Vivek Gopinath)
- Dempster, Laird & Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm
- Statistical Analysis with Missing Data (Roderick Little & Donald Rubin)
- Deep Generative Modelling (Springer)
- Deep Learning (Goodfellow)
- Methods Atlas (Gopinath)
- The Inquiry Series, Ch 53 — Generative Models & Synthetic User Testing
Also known as
- Expectation-maximisation
- EM imputation
- Maximum likelihood with missing data
- EM algorithm
- Expectation-Maximisation for Incomplete Behavioural Data
Categories
- Analytics & Behaviour
- Statistics & Modelling