To identify what is not there: A definition of missingness patterns and evaluation of missing value visualization

Abstract

While missing data is a commonly occurring issue in many domains, it is a topic that has been greatly overlooked by visualization scientists. Missing data values reduce the reliability of analysis results. A range of methods exist to replace the missing values with estimated values, but their appropriateness often depend on the patterns of missingness. Increased understanding of the missingness patterns and the distribution of missing values in data may greatly improve reliability, as well as provide valuable insight into potential problems in data gathering and analyses processes, and better understanding of the data as a whole. Visualization methods have a unique possibility to support investigation and understanding of missingness patterns by making the missing values and their relationship to recorded values visible. This article provides an overview of visualization of missing data values and defines a set of three missingness patterns of relevance for understanding missingness in data. It also contributes a usability evaluation which compares visualization methods representing missing values and how well they help users identify missingness patterns. The results indicate differences in performance depending on the visualization method as well as missingness pattern. Recommendations for future design of missing data visualization are provided based on the outcome of the study.

Missing Data Patterns: From Theory to an Application in the Steel Industry

“…Independently of the author (cf. [8,9,22]), at least an agreement that the MD patterns should be checked and considered, holds for all. In the following, we list MD patterns that commonly appear in related work as well as three patterns that are of specific interest for industrial data.…”

“…Several researchers (cf. [9,23]) stated without empirical proof that knowledge about MD patterns enhances imputation quality. Other researchers [17,22] suggested improved imputation procedures for specific MD patterns.…”

Where did my Lines go? Visualizing Missing Data in Parallel Coordinates

“…In terms of missing value techniques for parallel coordinates, Sjöbergh et al [ST17] introduced a horizontal axis below parallel coordinates and connected the missing value lines to this axis. This technique was also adopted by Muller et [Fer19]. This Missing Glyph is intended to be used either as a standalone visualization or as an enhancement for multivariate visualizations such as parallel coordinates.…”

Graphical Tools for Visualization of Missing Data in Large Longitudinal Phenomena

“…We propose three base different summary variables per case, which are related to the categories proposed by Fernstadt et al. [Fer19]: 1.The count of missing data ($$s_{L}$$) on a specified continuous segment of length $$L$$.…”

“…However, there is evidence that some plots are better suited to identifying missingness patterns in the data, as Fernstad et al. [Fer19] showed with an experimental design that matrix plot perform best when compared to profile plots with missing data, using the missingness categories previously mentioned to identify patterns. Twiddy et al.…”