NU.Learning Analytical Methods:

Adjustment for Selection Bias and Confounding in Observational Studies

using methods that are Nonparametric and Unsupervised.

NEWS: My LocalControlStrategy R-package has been updated to become: NU.Learning

NU.Learning References and Software Downloads

NU.Learning provides tools to "design" a statistical analysis. After all, when one's data are observational, it's usually too late to design their collection!

Traditional Covariate Adjustment methods use Multivariable MODELS that make quite STRONG assumptions in order to simultaneously Estimate effects "parametrically" as well as Predict patient Y-outcomes.

In stark contrast,  NU Learning methods deliberately separate Estimation from Prediction. This allows NU treatment effect estimation to not only be non-parametric (distribution-free) and highly visual (graphical) but also to deliberately stress "fair" comparisons between [1] relatively well-matched patients who made different (binary) treatment choices or [2] relatively well-matched experimental units that receive various levels of exposure to a potential harm.

NU methods [a] estimate either Local Treatment Differences (LTDs) or Local (Spearman) Rank Correlations (LRCs) only within X-space CLUSTERS of most similar patients or observational units and [b] examine the joint DISTRIBUTION of these LTDs or LRCs that characterize the full spectrum of differential response to treatment or exposure.  Although somewhat computationally intensive, this NU Learning approach is nevertheless ideal when databases are very large and encompass patients or units from numerous, diverse sub-populations.  NU methods can be viewed as being based upon the Propensity Scoring theory of Rosenbaum and Rubin, Biometrika (1983). Specifically, cluster membership then becomes a guaranteed BALANCING Score, "finer" than the unknown true propensity score, in the limit as clusters become small, compact and numerous!

NU Learning consists of FOUR PHASES of activities which (when repeatedly applied, checked and redone) ultimately assure that robust and objective (statistically valid) results are being generated.  The four links below display pages that introduce and illustrate the specific sorts of statistical methods and graphical displays typically examined in each of the four phases of a NU Analysis...

Phase One: Aggregate

Phase Two: Confirm

Phase Three: Explore

Phase Four: Reveal


Want to learn more about methods for Observational Data Analysis?

Follow this Link to SAS Press BOOKS


KISS: Keep It Sophisticatedly Simple

Arnold Zellner, ASA Presidential Address, 1991