Causal modeling, discovery & inference for software engineering.

Much empirical research in software engineering has focused on studies of \"naturalistic\" phenomena (e.g., [3]). But these studies collect only observational data, and so traditional analysis techniques yield only correlations between project practices and characteristics (on the one hand) and measurable outcomes (on the other hand). Without knowing the causal effects, it is difficult for a manager to act upon correlational evidence. For example, as source files in a software project increase in size, they tend to have more bugs and be touched by more developers. That is, file size, bugs, and number of developers are all strongly positively correlated. If one mistakes correlation for causation, then one might be tempted to conclude that bug rates could be lowered just by reducing the number of developers who are working on a file!