Near singular matrix in unit root test

[startz]

Run a unit root test on lowFreq in the attached work file. You get an error message. You can make the error message go away by reducing the maximum number of lags. Nonetheless, I think this is a bug.

[EViews Glenn]

After a(n admittedly) cursory check, I don't think this is a bug.

At first, I thought we might not be checking the required number of observations appropriately accounting for the lags. But I found that our check is appropriate. In this case, while there are enough observations to estimate the unit root regression at the maximum lags, it does appear that there is a singularity in obtaining the estimates for that regression. Since we can't compute the results for all of the lags, the lag selection procedure fails due to singularity.

For what it's worth, it appears that 13 lags is the knife edge here. If you specify a user-specified 14, you get the same failure with a slightly (only slightly) better error message. If you specify a user-specify 13 lags, you'll obtain results.

Hope this helps.

And remember "Non-normality and collinearity are NOT problems!"