Causal Discovery is a relatively recent form of statistical analysis for determining causa and effect relationships among a set of data. The output of such an analysis is a causal structure graph, where each variable in the data is a node, and discovered causes are edges between 2 nodes. There is an open-source tool called Tetrad that applies Causal Discovery to datasets to produce causal structure graphs.
Our research goal was to apply Causal Discovery to software and systems project data to determine the causes of software and systems effort and schedule. We ran "out of the box" Tetrad analyses on the COCOMO(R) II and COSYSMO 3.0 calibration datasets, with 161 and 68 projects, respectively. In each case, Tetrad found only a few variables that caused effort, perhaps because our datasets had few projects.
Tetrad can also run multiple searches on variants of the original dataset (bootstrapping) reporting the frequency with which a causal edge occurs between a pair of nodes. We have employed an innovation: inserting noise variables into the dataset and examining their causal structure in order to determine what cutoff for edge frequency would eliminate randomly-arising edges, and then applying that cutoff to produce the resulting causal graph). We explored 2 variants of this approach.
We show an example causal graph for each of our different approaches. We thereby show the specific causes of effort that we determined, and discuss the results of our investigation of frequency-based analysis.