Causal Programming
Files
Publication or External Link
Date
Authors
Advisor
Citation
DRUM DOI
Abstract
Causality is central to scientific inquiry. There is broad agreement on the
meaning of causal statements, such as “Smoking causes cancer”, or, “Applying
pesticides affects crop yields”. However, formalizing the intuition underlying such
statements and conducting rigorous inference is difficult in practice. Accordingly,
the overall goal of this dissertation is to reduce the difficulty of, and ambiguity
in, causal modeling and inference. In other words, the goal is to make it easy for
researchers to state precise causal assumptions, understand what they represent,
understand why they are necessary, and to yield precise causal conclusions with
minimal difficulty.
Using the framework of structural causal models, I introduce a causation coeffi-
cient as an analogue of the correlation coefficient, analyze its properties, and create
a taxonomy of correlation/causation relationships. Analyzing these relationships
provides insight into why correlation and causation are often conflated in practice,
as well as a principled argument as to why formal causal analysis is necessary. Next,
I introduce a theory of causal programming that unifies a large number of previ-
ously separate problems in causal modeling and inference. I describe the use and
implementation of a causal programming language as an embedded, domain-specific
language called ‘Whittemore’. Whittemore permits rigorously identifying and esti-
mating interventional queries without requiring the user to understand the details
of the underlying inference algorithms. Finally, I analyze the computational com-
plexity in determining the equilibrium distribution of cyclic causal models. I show
this is uncomputable in the general case, under mild assumptions about the distri-
butions of the model’s variables, suggesting that the structural causal model focus
on acyclic causal models is a ‘natural’ limitation. Further extensions of the concept
will have to give up either completeness or require the user to make additional —
likely parametric — model assumptions.
Together, this work supports the thesis that rigorous causal modeling and
inference can be effectively abstracted over, giving a researcher access to all of
the relevant details of causal modeling while encapsulating and automating the
irrelevant details of inference.