glm.fit: fitted probabilities numerically 0 or 1 occurred

[lorenzoFabbri]

I am using the latest version of lmtp as installed from GitHub. When using the TMLE estimator I get the following warning message: glm.fit: fitted probabilities numerically 0 or 1 occurred. Since this does not happen with the g-computation estimator and when I do not apply a shift function in TMLE, I believe it has to do with the logistic regression for estimating the density ratios. I am using continuous exposures and outcomes, and I noticed that if I reduce the shift parameter (e.g., multiply the natural value of the exposure by 10 rather than 100), the number of these warnings reduces. Some other perhaps relevant information:

• By visualizing the density ratios, they tend to be all very close to 0.
• I have a large fraction of missing outcomes due to loss to follow up.

I cannot reproduce this error with other, sharable, datasets, so I cannot provide a reproducible example. I was just wondering whether it happened to others.