> The simple workaround for this was adding a fallback to GEE when f[1] orf[2] is 0. GEE is another estimator described in the paper, behaving muchbetter in those cases.
For completeness, what's the downside in just always using GEE?