A visual analytic method to investigate patterns and structures of missing data.

Datasets often have missing values. Understanding the locations and underlying reasons (missing data patterns and structures, respectively) is important to avoid analysis bias or building inappropriate models. However, investigating missing data patterns and structures presents significant challenges. To address them we developed a novel visual analytic method, which combines new purity metrics for three core types of pattern (block, monotone and disjoint) with seven different visualizations in an iterative workflow. We also conducted two case studies, one with a UCI Machine Learning Repository dataset and the other with a 21 million record/389 variable hospital dataset that had many interwoven missingness patterns. In the UCI case study, the method's purity metrics and heatmaps revealed a rare data quality issue that affected 0.08% of records, but was not apparent when an existing correlation-based approach was used. Our method also implements perceptual discontinuity to ensure that small values are visible in bar charts, and that showed that two variables were missing values in 0.02% and 0.3% of records, respectively, rather than being complete as reported in the original UCI paper. The hospital dataset contained 595,171 unique combinations of missing values, but that overwhelming number was reduced to 57 comprehensible patterns by iteratively using purity metrics and heatmaps with the three core pattern types in our method. That revealed insights that ranged from uncovering rare and common data quality issues, to finding data dictionary errors, and making explicit relationships between groups of variables that may otherwise be hidden. The method also provides an explanation graph, which aids reproducibility by documenting the order in which patterns were found and acts as a tangible artifact that should help communicate the findings to stakeholders.