Shaky Structures: The Wobbly World of Causal Graphs in Software Analytics

TL;DR

Shaky Structures investigates the instability of causal graphs in software analytics by applying four causal graph generators (PC, FCI, GES, LiNGAM) to 23 SE data sets across defect prediction, software configuration, and project management. The study employs Jaccard overlap to quantify edge stability and reveals pervasive instability: minor data or parameter changes can reverse edges, and inter-project graphs diverge significantly. The authors argue that generalizable causal conclusions require validation across many generated graphs, not a single graph, and advocate using causal graphs as verification tools. The work offers a transparent, repeatable methodology with public scripts for replication and extension, highlighting a critical caveat for causal analysis in software engineering.

Abstract

Causal graphs are widely used in software engineering to document and explore causal relationships. Though widely used, they may also be wildly misleading. Causal structures generated from SE data can be highly variable. This instability is so significant that conclusions drawn from one graph may be totally reversed in another, even when both graphs are learned from the same or very similar project data. To document this problem, this paper examines causal graphs found by four causal graph generators (PC, FCI, GES, and LiNGAM) when applied to 23 data sets, relating to three different SE tasks: (a) learning how configuration options are selected for different properties; (b) understanding how management choices affect software projects; and (c) defect prediction. Graphs were compared between (a) different projects exploring the same task; (b) version i and i + 1 of a system; (c) different 90% samples of the data; and (d) small variations in the causal graph generator. Measured in terms of the Jaccard index of the number of edges shared by two different graphs, over half the edges were changed by these treatments. Hence, we conclude two things. Firstly, specific conclusions found by causal graph generators about how two specific variables affect each other may not generalize since those conclusions could be reversed by minor changes in how those graphs are generated. Secondly, before researchers can report supposedly general conclusions from causal graphs (e.g., "long functions cause more defects"), they should test that such conclusions hold over the numerous causal graphs that might be generated from the same data.

Figures (5)

• Figure 1: Two (very different) causal graphs generated by the R pcalg package (using $\alpha = .01$). For an explanation of the terms used in this figure, see Table \ref{['words']}.
• Figure 2: Jaccard index results seen in 100 experiments where two causal graph generators are applied to the same data set.
• Figure 3: RQ1 results: Graph of Jaccard Values across release, comparing release i to i + 1
• Figure 4: RQ3 results: Graph of Jaccard values when shifting alpha $\Delta = 0.001$. All graphs compared to the graph from $\alpha=0.001$.
• Figure 5: RQ4 results: Graph of 90% subsample Jaccard values from datasets