The DM Algorithm: A Causal Search Algorithm for the Discovery of MIMIC Models, with an Attempt to Recover a Protein Signalling Network from a High-Dimensional Ovarian Cancer Dataset
Alexander Murray-Watters · 2014
Latent variables have long confounded attempts to determine causal structure when experiments cannot be conducted. While some methods exist for dealing with exogenous latent variables, endogenous latents remain neglected. This thesis presents a new algorithm (the DM algorithm) designed to discover causal structure for a restricted class of models when endogenous latents are present. The algorithm is non-parametric, and in simulations outperformed one of the most popular methods for handling endogenous latents (namely, factor analysis). As the DM algorithm is also capable of handling a surprising number of variables, the algorithm was run on a high-dimensional genomic dataset. Popular methods in genomics lack the ability to address large numbers of variables and provide less information about the latent structure than the DM algorithm, so this represents an improvement on the state-of-the-art.