Quantitative Health Sciences Calendar

Sensitive big data is often summarized by frequencies, means, and standard deviations of outcomes by univariate categories of interest. Privacy is maintained by requiring N>10 subjects for each statistic. This univariate approach precludes the use of data for linear multivariate regression models for hypothesis testing, prediction or verification. We demonstrate how information can instead be summarized in Z‘Z matrices of cross products of dependent and independent variables, with the privacy protected trimming the Z’Z matrix diagonal and off-diagonal cell statistics to zero when N≤10. Sharing Z’Z matrices permits a broad set of multivariate linear models to be estimated, enables external validation, enables pooling or comparison across distinct samples without requiring any sharing at the individual level, speeds up linear model estimation, and can potentially be used within Machine Learning (ML) iterations to speed up estimation. Trimming small cells causes only minor degradation of results in the N>60 million Risk Adjustment models estimated.  
Click here to join or call 1 301 7158592, Meeting ID: 96956824446#password:  840948