In the case of linear regression, sometimes multicollinearity cannot be detected even when it is present. Recall that in most cases, LIMDEP is not using X’X to compute the regression , so the presence of multicollinearity may not be obvious. (Even when the moment matrix is being used, the assessment of multicollinearity is only to within some tolerance .) Because the presence of unternal rounding error may leave some variation in the representation of the raw data, data may be multicollinear in theory, but only approximately so and, therefore, not, internally. The QR method can then become unstable, and report the condition number for X’X with a warning that the data are highly collinear.