[Gretl-users] Re: OLS Results mis-match

Am 06.04.2026 um 11:25 schrieb Ramki S: Thanks for another interesting test case. Here's my summary of the situation in (lib)gretl as I understand it. The default numerical approach for ols estimation is the Cholesky decomposition. If the variables turn out to be too collinear, I believe gretl automatically falls back to the QR decomposition which is numerically more robust in that respect. Whether or not exact collinearity is "announced" to the user and the corresponding variable is dropped automatically depends on the threshold for the rank-determination code, and that threshold appears to be defined in lib/src/gretl_matrix.h, let me quote lines 30-33 here: /* minimum value of diagonal element of R (as in X = QR) that counts    as non-zero for the purpose of determining the rank of X */ #define R_DIAG_MIN 1.0e-8 So it seems to me that this is currently hard-coded as 1e-8. This is a very reasonable number I think, but apparently the other programs you tried are more tolerant here. So for comparison with other software one may wish to adjust it. I'm not aware that this is currently possible short of editing this piece of C source and re-compiling gretl yourself, which is not for the average user. Two possible exceptions: - For the matrix-oriented mols() function you can force the SVD instead of QR decomp with the script command "set svd on". Not sure whether this also affects the regular ols estimation command. - gretl also has the multiple-precision mpols() function (also matrix-oriented) which may have different thresholds. Thanks again for pointing to the relevant data, so more testing is possible. (Can't do that right now, though.) cheers sven