Some Machine Learning Algorithms Find Relationships Between Variables When None Exist — CTA Doesn’t

Ariel Linden & Paul R. Yarnold

Linden Consulting Group, LLC & Optimal Data Analysis, LLC

Automated machine learning algorithms are widely promoted as the best approach for estimating propensity scores, because these methods detect patterns in the data which manual efforts fail to identify. If classification algorithms are indeed ideal for identifying relationships between treatment group participation and covariates which predict participation, then it stands to reason that these algorithms would also be unable to find relationships when none exist (i.e., covariates do not predict treatment group assignment). Accordingly, we compare the predictive accuracy of maximum-accuracy classification tree analysis (CTA) vs. classification algorithms most commonly used to obtain the propensity score (logistic regression, random forests, boosted regression, and support vector machines). However, here we use an artificial dataset in which ten continuous covariates are randomly generated and by design have no correlation with the binary dependent variable (i.e., treatment assignment). Among all of the algorithms tested, only CTA correctly failed to discriminate between treatment and control groups based on the covariates. These results lend further support to the use of CTA for generating propensity scores as an alternative to other common approaches which are currently in favor.

View journal article