Assessing the quality of regression estimates through a test for signal-to-noise and its application to detecting harmful collinearity