Catch error w/ stratify() — system is computationally singular

[tylermaule]

Tyler TODO

See screenshots attached.

[tylermaule]

Note -- suggested use of tolerance variable tol for use w/ the Mahalanobis distance func, unsure of needed threshold

https://stackoverflow.com/questions/22134398/mahalonobis-distance-in-r-error-system-is-computationally-singular

[tylermaule]

Still need to explore how to catch this if it will be an error in the future, despite the fix.

[tylermaule]

@TimRuel if the covariance matrix of the stratifying variables is singular, mahalanobis() will appropriately throw an error (system is computationally singular).

Seems like setting the tol parameter to a low enough value (1e-18 in the case of my example data) solved the issue, since the matrix was computationally singular but not actually singular.

With a different dataset, I did find that a duplicate stratifying variable will throw the singular matrix error (as it should). I guess we could eventually use dimension-reduction techniques or find a psuedoinverse, but that seems like a future enhancement rather than something to tackle now.

For now, can I set a low tolerance threshold and just write a more descriptive error message for truly singular covariance matrices?

[tylermaule]

^ OK, pushing try/catch with tol=1e-20 to my branch -- current error msg, which needs work:

ERROR: It appears that the covariance matrix associated with your selected variables is linearly dependent. That is, some combination(s) of selected variables are perfect predictors of one another. Please make sure you haven't included duplicate information across variables (columns), choosing fewer variables if necessary, and re-run stratify with a different set of variables.

Ideally this would be caught earlier, as soon as variables are entered and rows are dropped, though.