NU.Learning (version 1.5)
Nonparametric and Unsupervised Learning from Cross-Sectional
Observational Data
Description
Especially when cross-sectional data are observational, effects of treatment selection
bias and confounding are best revealed by using Nonparametric and Unsupervised methods to
"Design" the analysis of the given data ...rather than the collection of "designed data".
Specifically, the "effect-size distribution" that best quantifies a potentially causal
relationship between a numeric y-Outcome variable and either a binary t-Treatment or
continuous e-Exposure variable needs to consist of BLOCKS of relatively well-matched
experimental units (e.g. patients) that have the most similar X-confounder characteristics.
Since our NU Learning approach will form BLOCKS by "clustering" experimental units in
confounder X-space, the implicit statistical model for learning is One-Way ANOVA. Within
Block measures of effect-size are then either [a] LOCAL Treatment Differences (LTDs) between
Within-Cluster y-Outcome Means ("new" minus "control") when treatment choice is
Binary or else [b] LOCAL Rank Correlations (LRCs) when the e-Exposure variable is numeric
with (hopefully many) more than two levels. An Instrumental Variable (IV) method is also
provided so that Local Average y-Outcomes (LAOs) within BLOCKS may also contribute
information for effect-size inferences when X-Covariates are assumed to influence Treatment
choice or Exposure level but otherwise have no direct effects on y-Outcomes. Finally, a
"Most-Like-Me" function provides histograms of effect-size distributions to aid
Doctor-Patient (or Researcher-Society) communications about Heterogeneous Outcomes.
Obenchain and Young (2013) ; Obenchain, Young and Krstic
(2019) .