Collinearity Diagnostics in gretl

Collinearity is blamed for all sorts of trouble in empirical work: inconclusive or weak results, unexpected signs on coefficients, and general computational mayhem in nonlinear estimators. Collinearity is a matter of degree since perfect col linearity has a perfectly easy solution. Near perfect col linearity can be vexing however since it makes precise measurement of model parameters particularly difficult in some cases. A number of methods for detecting col-linearity have been proposed. Some of these are useful, others not. Hill and Adkins (2001) summarize the good and bad based on much of the relevant literature up to 2001. They also make some recommendations for the detection and amelioration of inadequate variation in the data. The purpose of our paper is twofold: 1) update any significant cant findings on col-linearity since the Hill and Adkins (2001) survey and 2) to write and document gretl functions that perform several regression diagnostic procedures not already present in the software. These include the diagnostics suggested in Hill and Adkins (2001). In particular, we introduce hansl routines to perform the variance decomposition of Belsely, Kuh, and Welch (1980) for both linear and nonlinear models and provide a function to compute critical values for the Belsley (1982) signal-to-noise ratio test. The use of these is explored in several examples.